1 //===-- IA64ISelPattern.cpp - A pattern matching inst selector for IA64 ---===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Duraid Madina and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a pattern matching instruction selector for IA64.
12 //===----------------------------------------------------------------------===//
15 #include "IA64InstrBuilder.h"
16 #include "IA64RegisterInfo.h"
17 #include "IA64MachineFunctionInfo.h"
18 #include "llvm/Constants.h" // FIXME: REMOVE
19 #include "llvm/Function.h"
20 #include "llvm/CodeGen/MachineConstantPool.h" // FIXME: REMOVE
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/CodeGen/SelectionDAGISel.h"
25 #include "llvm/CodeGen/SSARegMap.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetLowering.h"
28 #include "llvm/Support/MathExtras.h"
29 #include "llvm/ADT/Statistic.h"
35 //===----------------------------------------------------------------------===//
36 // IA64TargetLowering - IA64 Implementation of the TargetLowering interface
38 class IA64TargetLowering : public TargetLowering {
39 int VarArgsFrameIndex; // FrameIndex for start of varargs area.
41 //int ReturnAddrIndex; // FrameIndex for return slot.
42 unsigned GP, SP, RP; // FIXME - clean this mess up
45 unsigned VirtGPR; // this is public so it can be accessed in the selector
46 // for ISD::RET down below. add an accessor instead? FIXME
48 IA64TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
50 // register class for general registers
51 addRegisterClass(MVT::i64, IA64::GRRegisterClass);
53 // register class for FP registers
54 addRegisterClass(MVT::f64, IA64::FPRegisterClass);
56 // register class for predicate registers
57 addRegisterClass(MVT::i1, IA64::PRRegisterClass);
59 setOperationAction(ISD::BRCONDTWOWAY , MVT::Other, Expand);
60 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
62 setSetCCResultType(MVT::i1);
63 setShiftAmountType(MVT::i64);
65 setOperationAction(ISD::EXTLOAD , MVT::i1 , Promote);
67 setOperationAction(ISD::ZEXTLOAD , MVT::i1 , Expand);
69 setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
70 setOperationAction(ISD::SEXTLOAD , MVT::i8 , Expand);
71 setOperationAction(ISD::SEXTLOAD , MVT::i16 , Expand);
72 setOperationAction(ISD::SEXTLOAD , MVT::i32 , Expand);
74 setOperationAction(ISD::SREM , MVT::f32 , Expand);
75 setOperationAction(ISD::SREM , MVT::f64 , Expand);
77 setOperationAction(ISD::UREM , MVT::f32 , Expand);
78 setOperationAction(ISD::UREM , MVT::f64 , Expand);
80 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
81 setOperationAction(ISD::MEMSET , MVT::Other, Expand);
82 setOperationAction(ISD::MEMCPY , MVT::Other, Expand);
84 // We don't support sin/cos/sqrt
85 setOperationAction(ISD::FSIN , MVT::f64, Expand);
86 setOperationAction(ISD::FCOS , MVT::f64, Expand);
87 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
88 setOperationAction(ISD::FSIN , MVT::f32, Expand);
89 setOperationAction(ISD::FCOS , MVT::f32, Expand);
90 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
92 //IA64 has these, but they are not implemented
93 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
94 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
96 computeRegisterProperties();
98 addLegalFPImmediate(+0.0);
99 addLegalFPImmediate(+1.0);
100 addLegalFPImmediate(-0.0);
101 addLegalFPImmediate(-1.0);
104 /// LowerArguments - This hook must be implemented to indicate how we should
105 /// lower the arguments for the specified function, into the specified DAG.
106 virtual std::vector<SDOperand>
107 LowerArguments(Function &F, SelectionDAG &DAG);
109 /// LowerCallTo - This hook lowers an abstract call to a function into an
111 virtual std::pair<SDOperand, SDOperand>
112 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
113 bool isTailCall, SDOperand Callee, ArgListTy &Args,
116 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
117 Value *VAListV, SelectionDAG &DAG);
118 virtual std::pair<SDOperand,SDOperand>
119 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
120 const Type *ArgTy, SelectionDAG &DAG);
122 void restoreGP_SP_RP(MachineBasicBlock* BB)
124 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
125 BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
126 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
129 void restoreSP_RP(MachineBasicBlock* BB)
131 BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
132 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
135 void restoreRP(MachineBasicBlock* BB)
137 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
140 void restoreGP(MachineBasicBlock* BB)
142 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
149 std::vector<SDOperand>
150 IA64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
151 std::vector<SDOperand> ArgValues;
154 // add beautiful description of IA64 stack frame format
155 // here (from intel 24535803.pdf most likely)
157 MachineFunction &MF = DAG.getMachineFunction();
158 MachineFrameInfo *MFI = MF.getFrameInfo();
160 GP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
161 SP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
162 RP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
164 MachineBasicBlock& BB = MF.front();
166 unsigned args_int[] = {IA64::r32, IA64::r33, IA64::r34, IA64::r35,
167 IA64::r36, IA64::r37, IA64::r38, IA64::r39};
169 unsigned args_FP[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
170 IA64::F12,IA64::F13,IA64::F14, IA64::F15};
176 unsigned used_FPArgs = 0; // how many FP args have been used so far?
178 unsigned ArgOffset = 0;
181 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
183 SDOperand newroot, argt;
184 if(count < 8) { // need to fix this logic? maybe.
186 switch (getValueType(I->getType())) {
188 std::cerr << "ERROR in LowerArgs: unknown type "
189 << getValueType(I->getType()) << "\n";
192 // fixme? (well, will need to for weird FP structy stuff,
193 // see intel ABI docs)
195 //XXX BuildMI(&BB, IA64::IDEF, 0, args_FP[used_FPArgs]);
196 MF.addLiveIn(args_FP[used_FPArgs]); // mark this reg as liveIn
197 // floating point args go into f8..f15 as-needed, the increment
198 argVreg[count] = // is below..:
199 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::f64));
200 // FP args go into f8..f15 as needed: (hence the ++)
201 argPreg[count] = args_FP[used_FPArgs++];
202 argOpc[count] = IA64::FMOV;
203 argt = newroot = DAG.getCopyFromReg(argVreg[count],
204 getValueType(I->getType()), DAG.getRoot());
206 case MVT::i1: // NOTE: as far as C abi stuff goes,
207 // bools are just boring old ints
212 //XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
213 MF.addLiveIn(args_int[count]); // mark this register as liveIn
215 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
216 argPreg[count] = args_int[count];
217 argOpc[count] = IA64::MOV;
219 DAG.getCopyFromReg(argVreg[count], MVT::i64, DAG.getRoot());
220 if ( getValueType(I->getType()) != MVT::i64)
221 argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
225 } else { // more than 8 args go into the frame
226 // Create the frame index object for this incoming parameter...
227 ArgOffset = 16 + 8 * (count - 8);
228 int FI = MFI->CreateFixedObject(8, ArgOffset);
230 // Create the SelectionDAG nodes corresponding to a load
231 //from this parameter
232 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
233 argt = newroot = DAG.getLoad(getValueType(I->getType()),
234 DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
237 DAG.setRoot(newroot.getValue(1));
238 ArgValues.push_back(argt);
242 // Create a vreg to hold the output of (what will become)
243 // the "alloc" instruction
244 VirtGPR = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
245 BuildMI(&BB, IA64::PSEUDO_ALLOC, 0, VirtGPR);
246 // we create a PSEUDO_ALLOC (pseudo)instruction for now
248 BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
251 BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
252 BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
255 BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
258 BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
259 BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
262 unsigned tempOffset=0;
264 // if this is a varargs function, we simply lower llvm.va_start by
265 // pointing to the first entry
268 VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
271 // here we actually do the moving of args, and store them to the stack
272 // too if this is a varargs function:
273 for (int i = 0; i < count && i < 8; ++i) {
274 BuildMI(&BB, argOpc[i], 1, argVreg[i]).addReg(argPreg[i]);
276 // if this is a varargs function, we copy the input registers to the stack
277 int FI = MFI->CreateFixedObject(8, tempOffset);
278 tempOffset+=8; //XXX: is it safe to use r22 like this?
279 BuildMI(&BB, IA64::MOV, 1, IA64::r22).addFrameIndex(FI);
280 // FIXME: we should use st8.spill here, one day
281 BuildMI(&BB, IA64::ST8, 1, IA64::r22).addReg(argPreg[i]);
285 // Finally, inform the code generator which regs we return values in.
286 // (see the ISD::RET: case down below)
287 switch (getValueType(F.getReturnType())) {
288 default: assert(0 && "i have no idea where to return this type!");
289 case MVT::isVoid: break;
295 MF.addLiveOut(IA64::r8);
299 MF.addLiveOut(IA64::F8);
306 std::pair<SDOperand, SDOperand>
307 IA64TargetLowering::LowerCallTo(SDOperand Chain,
308 const Type *RetTy, bool isVarArg,
309 unsigned CallingConv, bool isTailCall,
310 SDOperand Callee, ArgListTy &Args,
313 MachineFunction &MF = DAG.getMachineFunction();
315 unsigned NumBytes = 16;
316 unsigned outRegsUsed = 0;
318 if (Args.size() > 8) {
319 NumBytes += (Args.size() - 8) * 8;
322 outRegsUsed = Args.size();
325 // FIXME? this WILL fail if we ever try to pass around an arg that
326 // consumes more than a single output slot (a 'real' double, int128
327 // some sort of aggregate etc.), as we'll underestimate how many 'outX'
328 // registers we use. Hopefully, the assembler will notice.
329 MF.getInfo<IA64FunctionInfo>()->outRegsUsed=
330 std::max(outRegsUsed, MF.getInfo<IA64FunctionInfo>()->outRegsUsed);
332 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
333 DAG.getConstant(NumBytes, getPointerTy()));
335 std::vector<SDOperand> args_to_use;
336 for (unsigned i = 0, e = Args.size(); i != e; ++i)
338 switch (getValueType(Args[i].second)) {
339 default: assert(0 && "unexpected argument type!");
344 //promote to 64-bits, sign/zero extending based on type
346 if(Args[i].second->isSigned())
347 Args[i].first = DAG.getNode(ISD::SIGN_EXTEND, MVT::i64,
350 Args[i].first = DAG.getNode(ISD::ZERO_EXTEND, MVT::i64,
355 Args[i].first = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Args[i].first);
360 args_to_use.push_back(Args[i].first);
363 std::vector<MVT::ValueType> RetVals;
364 MVT::ValueType RetTyVT = getValueType(RetTy);
365 if (RetTyVT != MVT::isVoid)
366 RetVals.push_back(RetTyVT);
367 RetVals.push_back(MVT::Other);
369 SDOperand TheCall = SDOperand(DAG.getCall(RetVals, Chain,
370 Callee, args_to_use), 0);
371 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
372 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
373 DAG.getConstant(NumBytes, getPointerTy()));
374 return std::make_pair(TheCall, Chain);
378 IA64TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
379 Value *VAListV, SelectionDAG &DAG) {
380 // vastart just stores the address of the VarArgsFrameIndex slot.
381 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
382 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR,
383 VAListP, DAG.getSrcValue(VAListV));
386 std::pair<SDOperand,SDOperand> IA64TargetLowering::
387 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
388 const Type *ArgTy, SelectionDAG &DAG) {
390 MVT::ValueType ArgVT = getValueType(ArgTy);
391 SDOperand Val = DAG.getLoad(MVT::i64, Chain,
392 VAListP, DAG.getSrcValue(VAListV));
393 SDOperand Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), Val,
394 DAG.getSrcValue(NULL));
396 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
399 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
400 "Other types should have been promoted for varargs!");
403 Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
404 DAG.getConstant(Amt, Val.getValueType()));
405 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
406 Val, VAListP, DAG.getSrcValue(VAListV));
407 return std::make_pair(Result, Chain);
412 //===--------------------------------------------------------------------===//
413 /// ISel - IA64 specific code to select IA64 machine instructions for
414 /// SelectionDAG operations.
416 class ISel : public SelectionDAGISel {
417 /// IA64Lowering - This object fully describes how to lower LLVM code to an
418 /// IA64-specific SelectionDAG.
419 IA64TargetLowering IA64Lowering;
420 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
421 // for sdiv and udiv until it is put into the future
424 /// ExprMap - As shared expressions are codegen'd, we keep track of which
425 /// vreg the value is produced in, so we only emit one copy of each compiled
427 std::map<SDOperand, unsigned> ExprMap;
428 std::set<SDOperand> LoweredTokens;
431 ISel(TargetMachine &TM) : SelectionDAGISel(IA64Lowering), IA64Lowering(TM),
434 /// InstructionSelectBasicBlock - This callback is invoked by
435 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
436 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
438 unsigned SelectExpr(SDOperand N);
439 void Select(SDOperand N);
440 // a dag->dag to transform mul-by-constant-int to shifts+adds/subs
441 SDOperand BuildConstmulSequence(SDOperand N);
446 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
447 /// when it has created a SelectionDAG for us to codegen.
448 void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
450 // Codegen the basic block.
452 Select(DAG.getRoot());
454 // Clear state used for selection.
456 LoweredTokens.clear();
460 // strip leading '0' characters from a string
461 void munchLeadingZeros(std::string& inString) {
462 while(inString.c_str()[0]=='0') {
463 inString.erase(0, 1);
467 // strip trailing '0' characters from a string
468 void munchTrailingZeros(std::string& inString) {
469 int curPos=inString.length()-1;
471 while(inString.c_str()[curPos]=='0') {
472 inString.erase(curPos, 1);
477 // return how many consecutive '0' characters are at the end of a string
478 unsigned int countTrailingZeros(std::string& inString) {
479 int curPos=inString.length()-1;
480 unsigned int zeroCount=0;
482 while(inString.c_str()[curPos--]=='0') {
488 // booth encode a string of '1' and '0' characters (returns string of 'P' (+1)
489 // '0' and 'N' (-1) characters)
490 void boothEncode(std::string inString, std::string& boothEncodedString) {
494 int lim=inString.size();
497 if(inString[curpos]=='1') { // if we see a '1', look for a run of them
499 std::string replaceString="N";
501 // find the run length
502 for(;inString[curpos+runlength]=='1';runlength++) ;
504 for(int i=0; i<runlength-1; i++)
509 inString.replace(curpos, runlength+1, replaceString);
513 } else { // a zero, we just keep chugging along
518 // clean up (trim the string, reverse it and turn '1's into 'P's)
519 munchTrailingZeros(inString);
520 boothEncodedString="";
522 for(int i=inString.size()-1;i>=0;i--)
524 boothEncodedString+="P";
526 boothEncodedString+=inString[i];
530 struct shiftaddblob { // this encodes stuff like (x=) "A << B [+-] C << D"
531 unsigned firstVal; // A
532 unsigned firstShift; // B
533 unsigned secondVal; // C
534 unsigned secondShift; // D
538 /* this implements Lefevre's "pattern-based" constant multiplication,
539 * see "Multiplication by an Integer Constant", INRIA report 1999-06
541 * TODO: implement a method to try rewriting P0N<->0PP / N0P<->0NN
542 * to get better booth encodings - this does help in practice
543 * TODO: weight shifts appropriately (most architectures can't
544 * fuse a shift and an add for arbitrary shift amounts) */
545 unsigned lefevre(const std::string inString,
546 std::vector<struct shiftaddblob> &ops) {
547 std::string retstring;
548 std::string s = inString;
549 munchTrailingZeros(s);
551 int length=s.length()-1;
557 std::vector<int> p,n;
559 for(int i=0; i<=length; i++) {
560 if (s.c_str()[length-i]=='P') {
562 } else if (s.c_str()[length-i]=='N') {
570 std::map<const int, int> w;
572 for(unsigned i=0; i<p.size(); i++) {
573 for(unsigned j=0; j<i; j++) {
578 for(unsigned i=1; i<n.size(); i++) {
579 for(unsigned j=0; j<i; j++) {
584 for(unsigned i=0; i<p.size(); i++) {
585 for(unsigned j=0; j<n.size(); j++) {
586 w[-abs(p[i]-n[j])]++;
590 std::map<const int, int>::const_iterator ii;
592 std::multimap<int, int> sorted_by_value;
594 for(ii = w.begin(); ii!=w.end(); ii++)
595 sorted_by_value.insert(std::pair<int, int>((*ii).second,(*ii).first));
597 for (std::multimap<int, int>::iterator it = sorted_by_value.begin();
598 it != sorted_by_value.end(); ++it) {
599 d.push_back((*it).second);
605 while(d.size()>0 && (w[int_d=d.back()] > int_W)) {
613 for(unsigned base=0; base<retstring.size(); base++) {
614 if( ((base+z+1) < retstring.size()) &&
615 retstring.c_str()[base]=='P' &&
616 retstring.c_str()[base+z+1]=='P')
620 retstring.replace(base, 1, "0");
621 retstring.replace(base+z+1, 1, "p");
625 for(unsigned base=0; base<retstring.size(); base++) {
626 if( ((base+z+1) < retstring.size()) &&
627 retstring.c_str()[base]=='N' &&
628 retstring.c_str()[base+z+1]=='N')
632 retstring.replace(base, 1, "0");
633 retstring.replace(base+z+1, 1, "n");
638 for(unsigned base=0; base<retstring.size(); base++) {
639 if( ((base+z+1) < retstring.size()) &&
640 ((retstring.c_str()[base]=='P' &&
641 retstring.c_str()[base+z+1]=='N') ||
642 (retstring.c_str()[base]=='N' &&
643 retstring.c_str()[base+z+1]=='P')) ) {
647 if(retstring.c_str()[base]=='P') {
648 retstring.replace(base, 1, "0");
649 retstring.replace(base+z+1, 1, "p");
650 } else { // retstring[base]=='N'
651 retstring.replace(base, 1, "0");
652 retstring.replace(base+z+1, 1, "n");
664 } d.pop_back(); // hmm
668 for(unsigned i=0; i<t.length(); i++) {
669 if(t.c_str()[i]=='p' || t.c_str()[i]=='n')
670 t.replace(i, 1, "0");
673 for(unsigned i=0; i<u.length(); i++) {
674 if(u[i]=='P' || u[i]=='N')
675 u.replace(i, 1, "0");
677 u.replace(i, 1, "P");
679 u.replace(i, 1, "N");
692 bool hit=(u[pos]=='N');
697 for(unsigned p=0; p<u.length(); p++) {
698 bool isP=(u[p]=='P');
699 bool isN=(u[p]=='N');
702 u.replace(p, 1, "N");
704 u.replace(p, 1, "P");
708 munchLeadingZeros(u);
710 int i = lefevre(u, ops);
714 blob.firstVal=i; blob.firstShift=c;
716 blob.secondVal=i; blob.secondShift=0;
722 munchLeadingZeros(t);
727 if(t.c_str()[0]!='P') {
729 for(unsigned p=0; p<t.length(); p++) {
730 bool isP=(t.c_str()[p]=='P');
731 bool isN=(t.c_str()[p]=='N');
734 t.replace(p, 1, "N");
736 t.replace(p, 1, "P");
740 int j = lefevre(t, ops);
742 int trail=countTrailingZeros(u);
743 blob.secondVal=i; blob.secondShift=trail;
745 trail=countTrailingZeros(t);
746 blob.firstVal=j; blob.firstShift=trail;
750 blob.isSub=false; // first + second
753 blob.isSub=true; // first - second
756 blob.isSub=true; // second - first
757 int tmpval, tmpshift;
758 tmpval=blob.firstVal;
759 tmpshift=blob.firstShift;
760 blob.firstVal=blob.secondVal;
761 blob.firstShift=blob.secondShift;
762 blob.secondVal=tmpval;
763 blob.secondShift=tmpshift;
772 SDOperand ISel::BuildConstmulSequence(SDOperand N) {
773 //FIXME: we should shortcut this stuff for multiplies by 2^n+1
774 // in particular, *3 is nicer as *2+1, not *4-1
775 int64_t constant=cast<ConstantSDNode>(N.getOperand(1))->getValue();
778 unsigned preliminaryShift=0;
780 assert(constant != 0 && "erk, you're trying to multiply by constant zero\n");
782 // first, we make the constant to multiply by positive
790 // next, we make it odd.
791 for(; (constant%2==0); preliminaryShift++)
794 //OK, we have a positive, odd number of 64 bits or less. Convert it
795 //to a binary string, constantString[0] is the LSB
796 char constantString[65];
797 for(int i=0; i<64; i++)
798 constantString[i]='0'+((constant>>i)&0x1);
799 constantString[64]=0;
801 // now, Booth encode it
802 std::string boothEncodedString;
803 boothEncode(constantString, boothEncodedString);
805 std::vector<struct shiftaddblob> ops;
806 // do the transformation, filling out 'ops'
807 lefevre(boothEncodedString, ops);
809 assert(ops.size() < 80 && "constmul code has gone haywire\n");
810 SDOperand results[80]; // temporary results (of adds/subs of shifts)
812 // now turn 'ops' into DAG bits
813 for(unsigned i=0; i<ops.size(); i++) {
814 SDOperand amt = ISelDAG->getConstant(ops[i].firstShift, MVT::i64);
815 SDOperand val = (ops[i].firstVal == 0) ? N.getOperand(0) :
816 results[ops[i].firstVal-1];
817 SDOperand left = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
818 amt = ISelDAG->getConstant(ops[i].secondShift, MVT::i64);
819 val = (ops[i].secondVal == 0) ? N.getOperand(0) :
820 results[ops[i].secondVal-1];
821 SDOperand right = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
823 results[i] = ISelDAG->getNode(ISD::SUB, MVT::i64, left, right);
825 results[i] = ISelDAG->getNode(ISD::ADD, MVT::i64, left, right);
828 // don't forget flippedSign and preliminaryShift!
829 SDOperand shiftedresult;
830 if(preliminaryShift) {
831 SDOperand finalshift = ISelDAG->getConstant(preliminaryShift, MVT::i64);
832 shiftedresult = ISelDAG->getNode(ISD::SHL, MVT::i64,
833 results[ops.size()-1], finalshift);
834 } else { // there was no preliminary divide-by-power-of-2 required
835 shiftedresult = results[ops.size()-1];
838 SDOperand finalresult;
839 if(flippedSign) { // if we were multiplying by a negative constant:
840 SDOperand zero = ISelDAG->getConstant(0, MVT::i64);
841 // subtract the result from 0 to flip its sign
842 finalresult = ISelDAG->getNode(ISD::SUB, MVT::i64, zero, shiftedresult);
843 } else { // there was no preliminary multiply by -1 required
844 finalresult = shiftedresult;
850 /// ponderIntegerDivisionBy - When handling integer divides, if the divide
851 /// is by a constant such that we can efficiently codegen it, this
852 /// function says what to do. Currently, it returns 0 if the division must
853 /// become a genuine divide, and 1 if the division can be turned into a
855 static unsigned ponderIntegerDivisionBy(SDOperand N, bool isSigned,
857 if (N.getOpcode() != ISD::Constant) return 0; // if not a divide by
858 // a constant, give up.
860 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
862 if (isPowerOf2_64(v)) { // if a division by a power of two, say so
867 return 0; // fallthrough
870 static unsigned ponderIntegerAndWith(SDOperand N, unsigned& Imm) {
871 if (N.getOpcode() != ISD::Constant) return 0; // if not ANDing with
872 // a constant, give up.
874 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
876 if (isMask_64(v)) { // if ANDing with ((2^n)-1) for some n
881 return 0; // fallthrough
884 static unsigned ponderIntegerAdditionWith(SDOperand N, unsigned& Imm) {
885 if (N.getOpcode() != ISD::Constant) return 0; // if not adding a
886 // constant, give up.
887 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
889 if (v <= 8191 && v >= -8192) { // if this constants fits in 14 bits, say so
890 Imm = v & 0x3FFF; // 14 bits
893 return 0; // fallthrough
896 static unsigned ponderIntegerSubtractionFrom(SDOperand N, unsigned& Imm) {
897 if (N.getOpcode() != ISD::Constant) return 0; // if not subtracting a
898 // constant, give up.
899 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
901 if (v <= 127 && v >= -128) { // if this constants fits in 8 bits, say so
902 Imm = v & 0xFF; // 8 bits
905 return 0; // fallthrough
908 unsigned ISel::SelectExpr(SDOperand N) {
910 unsigned Tmp1, Tmp2, Tmp3;
912 MVT::ValueType DestType = N.getValueType();
914 unsigned opcode = N.getOpcode();
916 SDNode *Node = N.Val;
919 if (Node->getOpcode() == ISD::CopyFromReg)
920 // Just use the specified register as our input.
921 return dyn_cast<RegSDNode>(Node)->getReg();
923 unsigned &Reg = ExprMap[N];
926 if (N.getOpcode() != ISD::CALL && N.getOpcode() != ISD::TAILCALL)
927 Reg = Result = (N.getValueType() != MVT::Other) ?
928 MakeReg(N.getValueType()) : 1;
930 // If this is a call instruction, make sure to prepare ALL of the result
931 // values as well as the chain.
932 if (Node->getNumValues() == 1)
933 Reg = Result = 1; // Void call, just a chain.
935 Result = MakeReg(Node->getValueType(0));
936 ExprMap[N.getValue(0)] = Result;
937 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
938 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
939 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
943 switch (N.getOpcode()) {
946 assert(0 && "Node not handled!\n");
948 case ISD::FrameIndex: {
949 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
950 BuildMI(BB, IA64::MOV, 1, Result).addFrameIndex(Tmp1);
954 case ISD::ConstantPool: {
955 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
956 IA64Lowering.restoreGP(BB); // FIXME: do i really need this?
957 BuildMI(BB, IA64::ADD, 2, Result).addConstantPoolIndex(Tmp1)
962 case ISD::ConstantFP: {
963 Tmp1 = Result; // Intermediate Register
964 if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
965 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
966 Tmp1 = MakeReg(MVT::f64);
968 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
969 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
970 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F0); // load 0.0
971 else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
972 cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
973 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F1); // load 1.0
975 assert(0 && "Unexpected FP constant!");
977 // we multiply by +1.0, negate (this is FNMA), and then add 0.0
978 BuildMI(BB, IA64::FNMA, 3, Result).addReg(Tmp1).addReg(IA64::F1)
983 case ISD::DYNAMIC_STACKALLOC: {
984 // Generate both result values.
986 ExprMap[N.getValue(1)] = 1; // Generate the token
988 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
990 // FIXME: We are currently ignoring the requested alignment for handling
991 // greater than the stack alignment. This will need to be revisited at some
992 // point. Align = N.getOperand(2);
994 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
995 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
996 std::cerr << "Cannot allocate stack object with greater alignment than"
997 << " the stack alignment yet!";
1002 Select(N.getOperand(0));
1003 if (ConstantSDNode* CN = dyn_cast<ConstantSDNode>(N.getOperand(1)))
1005 if (CN->getValue() < 32000)
1007 BuildMI(BB, IA64::ADDIMM22, 2, IA64::r12).addReg(IA64::r12)
1008 .addImm(-CN->getValue());
1010 Tmp1 = SelectExpr(N.getOperand(1));
1011 // Subtract size from stack pointer, thereby allocating some space.
1012 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1015 Tmp1 = SelectExpr(N.getOperand(1));
1016 // Subtract size from stack pointer, thereby allocating some space.
1017 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1020 Select(N.getOperand(0));
1021 Tmp1 = SelectExpr(N.getOperand(1));
1022 // Subtract size from stack pointer, thereby allocating some space.
1023 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1024 // Put a pointer to the space into the result register, by copying the
1026 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r12);
1031 Tmp1 = SelectExpr(N.getOperand(0)); //Cond
1032 Tmp2 = SelectExpr(N.getOperand(1)); //Use if TRUE
1033 Tmp3 = SelectExpr(N.getOperand(2)); //Use if FALSE
1035 unsigned bogoResult;
1037 switch (N.getOperand(1).getValueType()) {
1038 default: assert(0 &&
1039 "ISD::SELECT: 'select'ing something other than i1, i64 or f64!\n");
1040 // for i1, we load the condition into an integer register, then
1041 // conditionally copy Tmp2 and Tmp3 to Tmp1 in parallel (only one
1042 // of them will go through, since the integer register will hold
1045 bogoResult=MakeReg(MVT::i1);
1047 // load the condition into an integer register
1048 unsigned condReg=MakeReg(MVT::i64);
1049 unsigned dummy=MakeReg(MVT::i64);
1050 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1051 BuildMI(BB, IA64::TPCADDIMM22, 2, condReg).addReg(dummy)
1052 .addImm(1).addReg(Tmp1);
1054 // initialize Result (bool) to false (hence UNC) and if
1055 // the select condition (condReg) is false (0), copy Tmp3
1056 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoResult)
1057 .addReg(condReg).addReg(IA64::r0).addReg(Tmp3);
1059 // now, if the selection condition is true, write 1 to the
1060 // result if Tmp2 is 1
1061 BuildMI(BB, IA64::TPCMPNE, 3, Result).addReg(bogoResult)
1062 .addReg(condReg).addReg(IA64::r0).addReg(Tmp2);
1065 // for i64/f64, we just copy Tmp3 and then conditionally overwrite it
1066 // with Tmp2 if Tmp1 is true
1068 bogoResult=MakeReg(MVT::i64);
1069 BuildMI(BB, IA64::MOV, 1, bogoResult).addReg(Tmp3);
1070 BuildMI(BB, IA64::CMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1074 bogoResult=MakeReg(MVT::f64);
1075 BuildMI(BB, IA64::FMOV, 1, bogoResult).addReg(Tmp3);
1076 BuildMI(BB, IA64::CFMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1084 case ISD::Constant: {
1085 unsigned depositPos=0;
1086 unsigned depositLen=0;
1087 switch (N.getValueType()) {
1088 default: assert(0 && "Cannot use constants of this type!");
1089 case MVT::i1: { // if a bool, we don't 'load' so much as generate
1091 if(cast<ConstantSDNode>(N)->getValue()) // true:
1092 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1094 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1095 return Result; // early exit
1097 case MVT::i64: break;
1100 int64_t immediate = cast<ConstantSDNode>(N)->getValue();
1102 if(immediate==0) { // if the constant is just zero,
1103 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r0); // just copy r0
1104 return Result; // early exit
1107 if (immediate <= 8191 && immediate >= -8192) {
1108 // if this constants fits in 14 bits, we use a mov the assembler will
1109 // turn into: "adds rDest=imm,r0" (and _not_ "andl"...)
1110 BuildMI(BB, IA64::MOVSIMM14, 1, Result).addSImm(immediate);
1111 return Result; // early exit
1114 if (immediate <= 2097151 && immediate >= -2097152) {
1115 // if this constants fits in 22 bits, we use a mov the assembler will
1116 // turn into: "addl rDest=imm,r0"
1117 BuildMI(BB, IA64::MOVSIMM22, 1, Result).addSImm(immediate);
1118 return Result; // early exit
1121 /* otherwise, our immediate is big, so we use movl */
1122 uint64_t Imm = immediate;
1123 BuildMI(BB, IA64::MOVLIMM64, 1, Result).addImm64(Imm);
1128 BuildMI(BB, IA64::IDEF, 0, Result);
1132 case ISD::GlobalAddress: {
1133 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1134 unsigned Tmp1 = MakeReg(MVT::i64);
1136 BuildMI(BB, IA64::ADD, 2, Tmp1).addGlobalAddress(GV).addReg(IA64::r1);
1137 BuildMI(BB, IA64::LD8, 1, Result).addReg(Tmp1);
1142 case ISD::ExternalSymbol: {
1143 const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
1144 // assert(0 && "sorry, but what did you want an ExternalSymbol for again?");
1145 BuildMI(BB, IA64::MOV, 1, Result).addExternalSymbol(Sym); // XXX
1149 case ISD::FP_EXTEND: {
1150 Tmp1 = SelectExpr(N.getOperand(0));
1151 BuildMI(BB, IA64::FMOV, 1, Result).addReg(Tmp1);
1155 case ISD::ZERO_EXTEND: {
1156 Tmp1 = SelectExpr(N.getOperand(0)); // value
1158 switch (N.getOperand(0).getValueType()) {
1159 default: assert(0 && "Cannot zero-extend this type!");
1160 case MVT::i8: Opc = IA64::ZXT1; break;
1161 case MVT::i16: Opc = IA64::ZXT2; break;
1162 case MVT::i32: Opc = IA64::ZXT4; break;
1164 // we handle bools differently! :
1165 case MVT::i1: { // if the predicate reg has 1, we want a '1' in our GR.
1166 unsigned dummy = MakeReg(MVT::i64);
1168 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1169 // ...then conditionally (PR:Tmp1) add 1:
1170 BuildMI(BB, IA64::TPCADDIMM22, 2, Result).addReg(dummy)
1171 .addImm(1).addReg(Tmp1);
1172 return Result; // XXX early exit!
1176 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1180 case ISD::SIGN_EXTEND: { // we should only have to handle i1 -> i64 here!!!
1182 assert(0 && "hmm, ISD::SIGN_EXTEND: shouldn't ever be reached. bad luck!\n");
1184 Tmp1 = SelectExpr(N.getOperand(0)); // value
1186 switch (N.getOperand(0).getValueType()) {
1187 default: assert(0 && "Cannot sign-extend this type!");
1188 case MVT::i1: assert(0 && "trying to sign extend a bool? ow.\n");
1189 Opc = IA64::SXT1; break;
1190 // FIXME: for now, we treat bools the same as i8s
1191 case MVT::i8: Opc = IA64::SXT1; break;
1192 case MVT::i16: Opc = IA64::SXT2; break;
1193 case MVT::i32: Opc = IA64::SXT4; break;
1196 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1200 case ISD::TRUNCATE: {
1201 // we use the funky dep.z (deposit (zero)) instruction to deposit bits
1202 // of R0 appropriately.
1203 switch (N.getOperand(0).getValueType()) {
1204 default: assert(0 && "Unknown truncate!");
1205 case MVT::i64: break;
1207 Tmp1 = SelectExpr(N.getOperand(0));
1208 unsigned depositPos, depositLen;
1210 switch (N.getValueType()) {
1211 default: assert(0 && "Unknown truncate!");
1213 // if input (normal reg) is 0, 0!=0 -> false (0), if 1, 1!=0 ->true (1):
1214 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1)
1216 return Result; // XXX early exit!
1218 case MVT::i8: depositPos=0; depositLen=8; break;
1219 case MVT::i16: depositPos=0; depositLen=16; break;
1220 case MVT::i32: depositPos=0; depositLen=32; break;
1222 BuildMI(BB, IA64::DEPZ, 1, Result).addReg(Tmp1)
1223 .addImm(depositPos).addImm(depositLen);
1228 case ISD::FP_ROUND: {
1229 assert (DestType == MVT::f32 && N.getOperand(0).getValueType() == MVT::f64 &&
1230 "error: trying to FP_ROUND something other than f64 -> f32!\n");
1231 Tmp1 = SelectExpr(N.getOperand(0));
1232 BuildMI(BB, IA64::FADDS, 2, Result).addReg(Tmp1).addReg(IA64::F0);
1233 // we add 0.0 using a single precision add to do rounding
1238 // FIXME: the following 4 cases need cleaning
1239 case ISD::SINT_TO_FP: {
1240 Tmp1 = SelectExpr(N.getOperand(0));
1241 Tmp2 = MakeReg(MVT::f64);
1242 unsigned dummy = MakeReg(MVT::f64);
1243 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1244 BuildMI(BB, IA64::FCVTXF, 1, dummy).addReg(Tmp2);
1245 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1249 case ISD::UINT_TO_FP: {
1250 Tmp1 = SelectExpr(N.getOperand(0));
1251 Tmp2 = MakeReg(MVT::f64);
1252 unsigned dummy = MakeReg(MVT::f64);
1253 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1254 BuildMI(BB, IA64::FCVTXUF, 1, dummy).addReg(Tmp2);
1255 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1259 case ISD::FP_TO_SINT: {
1260 Tmp1 = SelectExpr(N.getOperand(0));
1261 Tmp2 = MakeReg(MVT::f64);
1262 BuildMI(BB, IA64::FCVTFXTRUNC, 1, Tmp2).addReg(Tmp1);
1263 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1267 case ISD::FP_TO_UINT: {
1268 Tmp1 = SelectExpr(N.getOperand(0));
1269 Tmp2 = MakeReg(MVT::f64);
1270 BuildMI(BB, IA64::FCVTFXUTRUNC, 1, Tmp2).addReg(Tmp1);
1271 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1276 if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
1277 N.getOperand(0).Val->hasOneUse()) { // if we can fold this add
1278 // into an fma, do so:
1279 // ++FusedFP; // Statistic
1280 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1281 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1282 Tmp3 = SelectExpr(N.getOperand(1));
1283 BuildMI(BB, IA64::FMA, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1284 return Result; // early exit
1287 if(DestType != MVT::f64 && N.getOperand(0).getOpcode() == ISD::SHL &&
1288 N.getOperand(0).Val->hasOneUse()) { // if we might be able to fold
1289 // this add into a shladd, try:
1290 ConstantSDNode *CSD = NULL;
1291 if((CSD = dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) &&
1292 (CSD->getValue() >= 1) && (CSD->getValue() <= 4) ) { // we can:
1294 // ++FusedSHLADD; // Statistic
1295 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1296 int shl_amt = CSD->getValue();
1297 Tmp3 = SelectExpr(N.getOperand(1));
1299 BuildMI(BB, IA64::SHLADD, 3, Result)
1300 .addReg(Tmp1).addImm(shl_amt).addReg(Tmp3);
1301 return Result; // early exit
1305 //else, fallthrough:
1306 Tmp1 = SelectExpr(N.getOperand(0));
1307 if(DestType != MVT::f64) { // integer addition:
1308 switch (ponderIntegerAdditionWith(N.getOperand(1), Tmp3)) {
1309 case 1: // adding a constant that's 14 bits
1310 BuildMI(BB, IA64::ADDIMM14, 2, Result).addReg(Tmp1).addSImm(Tmp3);
1311 return Result; // early exit
1312 } // fallthrough and emit a reg+reg ADD:
1313 Tmp2 = SelectExpr(N.getOperand(1));
1314 BuildMI(BB, IA64::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1315 } else { // this is a floating point addition
1316 Tmp2 = SelectExpr(N.getOperand(1));
1317 BuildMI(BB, IA64::FADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1324 if(DestType != MVT::f64) { // TODO: speed!
1325 if(N.getOperand(1).getOpcode() != ISD::Constant) { // if not a const mul
1326 // boring old integer multiply with xma
1327 Tmp1 = SelectExpr(N.getOperand(0));
1328 Tmp2 = SelectExpr(N.getOperand(1));
1330 unsigned TempFR1=MakeReg(MVT::f64);
1331 unsigned TempFR2=MakeReg(MVT::f64);
1332 unsigned TempFR3=MakeReg(MVT::f64);
1333 BuildMI(BB, IA64::SETFSIG, 1, TempFR1).addReg(Tmp1);
1334 BuildMI(BB, IA64::SETFSIG, 1, TempFR2).addReg(Tmp2);
1335 BuildMI(BB, IA64::XMAL, 1, TempFR3).addReg(TempFR1).addReg(TempFR2)
1337 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TempFR3);
1338 return Result; // early exit
1339 } else { // we are multiplying by an integer constant! yay
1340 return Reg = SelectExpr(BuildConstmulSequence(N)); // avert your eyes!
1343 else { // floating point multiply
1344 Tmp1 = SelectExpr(N.getOperand(0));
1345 Tmp2 = SelectExpr(N.getOperand(1));
1346 BuildMI(BB, IA64::FMPY, 2, Result).addReg(Tmp1).addReg(Tmp2);
1352 if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
1353 N.getOperand(0).Val->hasOneUse()) { // if we can fold this sub
1354 // into an fms, do so:
1355 // ++FusedFP; // Statistic
1356 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1357 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1358 Tmp3 = SelectExpr(N.getOperand(1));
1359 BuildMI(BB, IA64::FMS, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1360 return Result; // early exit
1362 Tmp2 = SelectExpr(N.getOperand(1));
1363 if(DestType != MVT::f64) { // integer subtraction:
1364 switch (ponderIntegerSubtractionFrom(N.getOperand(0), Tmp3)) {
1365 case 1: // subtracting *from* an 8 bit constant:
1366 BuildMI(BB, IA64::SUBIMM8, 2, Result).addSImm(Tmp3).addReg(Tmp2);
1367 return Result; // early exit
1368 } // fallthrough and emit a reg+reg SUB:
1369 Tmp1 = SelectExpr(N.getOperand(0));
1370 BuildMI(BB, IA64::SUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1371 } else { // this is a floating point subtraction
1372 Tmp1 = SelectExpr(N.getOperand(0));
1373 BuildMI(BB, IA64::FSUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1379 Tmp1 = SelectExpr(N.getOperand(0));
1380 assert(DestType == MVT::f64 && "trying to fabs something other than f64?");
1381 BuildMI(BB, IA64::FABS, 1, Result).addReg(Tmp1);
1386 assert(DestType == MVT::f64 && "trying to fneg something other than f64?");
1388 if (ISD::FABS == N.getOperand(0).getOpcode()) { // && hasOneUse()?
1389 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1390 BuildMI(BB, IA64::FNEGABS, 1, Result).addReg(Tmp1); // fold in abs
1392 Tmp1 = SelectExpr(N.getOperand(0));
1393 BuildMI(BB, IA64::FNEG, 1, Result).addReg(Tmp1); // plain old fneg
1400 switch (N.getValueType()) {
1401 default: assert(0 && "Cannot AND this type!");
1402 case MVT::i1: { // if a bool, we emit a pseudocode AND
1403 unsigned pA = SelectExpr(N.getOperand(0));
1404 unsigned pB = SelectExpr(N.getOperand(1));
1406 /* our pseudocode for AND is:
1408 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1409 cmp.eq pTemp,p0 = r0,r0 // pTemp = NOT pB
1411 (pB) cmp.ne pTemp,p0 = r0,r0
1413 (pTemp)cmp.ne pC,p0 = r0,r0 // if (NOT pB) pC = 0
1416 unsigned pTemp = MakeReg(MVT::i1);
1418 unsigned bogusTemp1 = MakeReg(MVT::i1);
1419 unsigned bogusTemp2 = MakeReg(MVT::i1);
1420 unsigned bogusTemp3 = MakeReg(MVT::i1);
1421 unsigned bogusTemp4 = MakeReg(MVT::i1);
1423 BuildMI(BB, IA64::PCMPEQUNC, 3, bogusTemp1)
1424 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1425 BuildMI(BB, IA64::CMPEQ, 2, bogusTemp2)
1426 .addReg(IA64::r0).addReg(IA64::r0);
1427 BuildMI(BB, IA64::TPCMPNE, 3, pTemp)
1428 .addReg(bogusTemp2).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1429 BuildMI(BB, IA64::TPCMPNE, 3, Result)
1430 .addReg(bogusTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pTemp);
1434 // if not a bool, we just AND away:
1439 Tmp1 = SelectExpr(N.getOperand(0));
1440 switch (ponderIntegerAndWith(N.getOperand(1), Tmp3)) {
1441 case 1: // ANDing a constant that is 2^n-1 for some n
1443 case 8: // if AND 0x00000000000000FF, be quaint and use zxt1
1444 BuildMI(BB, IA64::ZXT1, 1, Result).addReg(Tmp1);
1446 case 16: // if AND 0x000000000000FFFF, be quaint and use zxt2
1447 BuildMI(BB, IA64::ZXT2, 1, Result).addReg(Tmp1);
1449 case 32: // if AND 0x00000000FFFFFFFF, be quaint and use zxt4
1450 BuildMI(BB, IA64::ZXT4, 1, Result).addReg(Tmp1);
1452 default: // otherwise, use dep.z to paste zeros
1453 BuildMI(BB, IA64::DEPZ, 3, Result).addReg(Tmp1)
1454 .addImm(0).addImm(Tmp3);
1457 return Result; // early exit
1458 } // fallthrough and emit a simple AND:
1459 Tmp2 = SelectExpr(N.getOperand(1));
1460 BuildMI(BB, IA64::AND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1467 switch (N.getValueType()) {
1468 default: assert(0 && "Cannot OR this type!");
1469 case MVT::i1: { // if a bool, we emit a pseudocode OR
1470 unsigned pA = SelectExpr(N.getOperand(0));
1471 unsigned pB = SelectExpr(N.getOperand(1));
1473 unsigned pTemp1 = MakeReg(MVT::i1);
1475 /* our pseudocode for OR is:
1481 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1483 (pB) cmp.eq pC,p0 = r0,r0 // if (pB) pC = 1
1486 BuildMI(BB, IA64::PCMPEQUNC, 3, pTemp1)
1487 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1488 BuildMI(BB, IA64::TPCMPEQ, 3, Result)
1489 .addReg(pTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1492 // if not a bool, we just OR away:
1497 Tmp1 = SelectExpr(N.getOperand(0));
1498 Tmp2 = SelectExpr(N.getOperand(1));
1499 BuildMI(BB, IA64::OR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1507 switch (N.getValueType()) {
1508 default: assert(0 && "Cannot XOR this type!");
1509 case MVT::i1: { // if a bool, we emit a pseudocode XOR
1510 unsigned pY = SelectExpr(N.getOperand(0));
1511 unsigned pZ = SelectExpr(N.getOperand(1));
1513 /* one possible routine for XOR is:
1515 // Compute px = py ^ pz
1516 // using sum of products: px = (py & !pz) | (pz & !py)
1517 // Uses 5 instructions in 3 cycles.
1519 (pz) cmp.eq.unc px = r0, r0 // px = pz
1520 (py) cmp.eq.unc pt = r0, r0 // pt = py
1523 (pt) cmp.ne.and px = r0, r0 // px = px & !pt (px = pz & !pt)
1524 (pz) cmp.ne.and pt = r0, r0 // pt = pt & !pz
1528 (pt) cmp.eq.or px = r0, r0 // px = px | pt
1530 *** Another, which we use here, requires one scratch GR. it is:
1532 mov rt = 0 // initialize rt off critical path
1536 (pz) cmp.eq.unc px = r0, r0 // px = pz
1537 (pz) mov rt = 1 // rt = pz
1540 (py) cmp.ne px = 1, rt // if (py) px = !pz
1542 .. these routines kindly provided by Jim Hull
1544 unsigned rt = MakeReg(MVT::i64);
1546 // these two temporaries will never actually appear,
1547 // due to the two-address form of some of the instructions below
1548 unsigned bogoPR = MakeReg(MVT::i1); // becomes Result
1549 unsigned bogoGR = MakeReg(MVT::i64); // becomes rt
1551 BuildMI(BB, IA64::MOV, 1, bogoGR).addReg(IA64::r0);
1552 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoPR)
1553 .addReg(IA64::r0).addReg(IA64::r0).addReg(pZ);
1554 BuildMI(BB, IA64::TPCADDIMM22, 2, rt)
1555 .addReg(bogoGR).addImm(1).addReg(pZ);
1556 BuildMI(BB, IA64::TPCMPIMM8NE, 3, Result)
1557 .addReg(bogoPR).addImm(1).addReg(rt).addReg(pY);
1560 // if not a bool, we just XOR away:
1565 Tmp1 = SelectExpr(N.getOperand(0));
1566 Tmp2 = SelectExpr(N.getOperand(1));
1567 BuildMI(BB, IA64::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1575 Tmp1 = SelectExpr(N.getOperand(0));
1576 BuildMI(BB, IA64::POPCNT, 1, Result).addReg(Tmp1);
1581 Tmp1 = SelectExpr(N.getOperand(0));
1582 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1583 Tmp2 = CN->getValue();
1584 BuildMI(BB, IA64::SHLI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1586 Tmp2 = SelectExpr(N.getOperand(1));
1587 BuildMI(BB, IA64::SHL, 2, Result).addReg(Tmp1).addReg(Tmp2);
1593 Tmp1 = SelectExpr(N.getOperand(0));
1594 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1595 Tmp2 = CN->getValue();
1596 BuildMI(BB, IA64::SHRUI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1598 Tmp2 = SelectExpr(N.getOperand(1));
1599 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1605 Tmp1 = SelectExpr(N.getOperand(0));
1606 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1607 Tmp2 = CN->getValue();
1608 BuildMI(BB, IA64::SHRSI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1610 Tmp2 = SelectExpr(N.getOperand(1));
1611 BuildMI(BB, IA64::SHRS, 2, Result).addReg(Tmp1).addReg(Tmp2);
1621 Tmp1 = SelectExpr(N.getOperand(0));
1622 Tmp2 = SelectExpr(N.getOperand(1));
1626 if(DestType == MVT::f64) // XXX: we're not gonna be fed MVT::f32, are we?
1629 bool isModulus=false; // is it a division or a modulus?
1630 bool isSigned=false;
1632 switch(N.getOpcode()) {
1633 case ISD::SDIV: isModulus=false; isSigned=true; break;
1634 case ISD::UDIV: isModulus=false; isSigned=false; break;
1635 case ISD::SREM: isModulus=true; isSigned=true; break;
1636 case ISD::UREM: isModulus=true; isSigned=false; break;
1639 if(!isModulus && !isFP) { // if this is an integer divide,
1640 switch (ponderIntegerDivisionBy(N.getOperand(1), isSigned, Tmp3)) {
1641 case 1: // division by a constant that's a power of 2
1642 Tmp1 = SelectExpr(N.getOperand(0));
1643 if(isSigned) { // argument could be negative, so emit some code:
1644 unsigned divAmt=Tmp3;
1645 unsigned tempGR1=MakeReg(MVT::i64);
1646 unsigned tempGR2=MakeReg(MVT::i64);
1647 unsigned tempGR3=MakeReg(MVT::i64);
1648 BuildMI(BB, IA64::SHRS, 2, tempGR1)
1649 .addReg(Tmp1).addImm(divAmt-1);
1650 BuildMI(BB, IA64::EXTRU, 3, tempGR2)
1651 .addReg(tempGR1).addImm(64-divAmt).addImm(divAmt);
1652 BuildMI(BB, IA64::ADD, 2, tempGR3)
1653 .addReg(Tmp1).addReg(tempGR2);
1654 BuildMI(BB, IA64::SHRS, 2, Result)
1655 .addReg(tempGR3).addImm(divAmt);
1657 else // unsigned div-by-power-of-2 becomes a simple shift right:
1658 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addImm(Tmp3);
1659 return Result; // early exit
1663 unsigned TmpPR=MakeReg(MVT::i1); // we need two scratch
1664 unsigned TmpPR2=MakeReg(MVT::i1); // predicate registers,
1665 unsigned TmpF1=MakeReg(MVT::f64); // and one metric truckload of FP regs.
1666 unsigned TmpF2=MakeReg(MVT::f64); // lucky we have IA64?
1667 unsigned TmpF3=MakeReg(MVT::f64); // well, the real FIXME is to have
1668 unsigned TmpF4=MakeReg(MVT::f64); // isTwoAddress forms of these
1669 unsigned TmpF5=MakeReg(MVT::f64); // FP instructions so we can end up with
1670 unsigned TmpF6=MakeReg(MVT::f64); // stuff like setf.sig f10=f10 etc.
1671 unsigned TmpF7=MakeReg(MVT::f64);
1672 unsigned TmpF8=MakeReg(MVT::f64);
1673 unsigned TmpF9=MakeReg(MVT::f64);
1674 unsigned TmpF10=MakeReg(MVT::f64);
1675 unsigned TmpF11=MakeReg(MVT::f64);
1676 unsigned TmpF12=MakeReg(MVT::f64);
1677 unsigned TmpF13=MakeReg(MVT::f64);
1678 unsigned TmpF14=MakeReg(MVT::f64);
1679 unsigned TmpF15=MakeReg(MVT::f64);
1681 // OK, emit some code:
1684 // first, load the inputs into FP regs.
1685 BuildMI(BB, IA64::SETFSIG, 1, TmpF1).addReg(Tmp1);
1686 BuildMI(BB, IA64::SETFSIG, 1, TmpF2).addReg(Tmp2);
1688 // next, convert the inputs to FP
1690 BuildMI(BB, IA64::FCVTXF, 1, TmpF3).addReg(TmpF1);
1691 BuildMI(BB, IA64::FCVTXF, 1, TmpF4).addReg(TmpF2);
1693 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF3).addReg(TmpF1);
1694 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF4).addReg(TmpF2);
1697 } else { // this is an FP divide/remainder, so we 'leak' some temp
1698 // regs and assign TmpF3=Tmp1, TmpF4=Tmp2
1703 // we start by computing an approximate reciprocal (good to 9 bits?)
1704 // note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
1705 BuildMI(BB, IA64::FRCPAS1, 4)
1706 .addReg(TmpF5, MachineOperand::Def)
1707 .addReg(TmpPR, MachineOperand::Def)
1708 .addReg(TmpF3).addReg(TmpF4);
1710 if(!isModulus) { // if this is a divide, we worry about div-by-zero
1711 unsigned bogusPR=MakeReg(MVT::i1); // won't appear, due to twoAddress
1713 BuildMI(BB, IA64::CMPEQ, 2, bogusPR).addReg(IA64::r0).addReg(IA64::r0);
1714 BuildMI(BB, IA64::TPCMPNE, 3, TmpPR2).addReg(bogusPR)
1715 .addReg(IA64::r0).addReg(IA64::r0).addReg(TmpPR);
1718 // now we apply newton's method, thrice! (FIXME: this is ~72 bits of
1719 // precision, don't need this much for f32/i32)
1720 BuildMI(BB, IA64::CFNMAS1, 4, TmpF6)
1721 .addReg(TmpF4).addReg(TmpF5).addReg(IA64::F1).addReg(TmpPR);
1722 BuildMI(BB, IA64::CFMAS1, 4, TmpF7)
1723 .addReg(TmpF3).addReg(TmpF5).addReg(IA64::F0).addReg(TmpPR);
1724 BuildMI(BB, IA64::CFMAS1, 4, TmpF8)
1725 .addReg(TmpF6).addReg(TmpF6).addReg(IA64::F0).addReg(TmpPR);
1726 BuildMI(BB, IA64::CFMAS1, 4, TmpF9)
1727 .addReg(TmpF6).addReg(TmpF7).addReg(TmpF7).addReg(TmpPR);
1728 BuildMI(BB, IA64::CFMAS1, 4,TmpF10)
1729 .addReg(TmpF6).addReg(TmpF5).addReg(TmpF5).addReg(TmpPR);
1730 BuildMI(BB, IA64::CFMAS1, 4,TmpF11)
1731 .addReg(TmpF8).addReg(TmpF9).addReg(TmpF9).addReg(TmpPR);
1732 BuildMI(BB, IA64::CFMAS1, 4,TmpF12)
1733 .addReg(TmpF8).addReg(TmpF10).addReg(TmpF10).addReg(TmpPR);
1734 BuildMI(BB, IA64::CFNMAS1, 4,TmpF13)
1735 .addReg(TmpF4).addReg(TmpF11).addReg(TmpF3).addReg(TmpPR);
1737 // FIXME: this is unfortunate :(
1738 // the story is that the dest reg of the fnma above and the fma below
1739 // (and therefore possibly the src of the fcvt.fx[u] as well) cannot
1740 // be the same register, or this code breaks if the first argument is
1741 // zero. (e.g. without this hack, 0%8 yields -64, not 0.)
1742 BuildMI(BB, IA64::CFMAS1, 4,TmpF14)
1743 .addReg(TmpF13).addReg(TmpF12).addReg(TmpF11).addReg(TmpPR);
1745 if(isModulus) { // XXX: fragile! fixes _only_ mod, *breaks* div! !
1746 BuildMI(BB, IA64::IUSE, 1).addReg(TmpF13); // hack :(
1750 // round to an integer
1752 BuildMI(BB, IA64::FCVTFXTRUNCS1, 1, TmpF15).addReg(TmpF14);
1754 BuildMI(BB, IA64::FCVTFXUTRUNCS1, 1, TmpF15).addReg(TmpF14);
1756 BuildMI(BB, IA64::FMOV, 1, TmpF15).addReg(TmpF14);
1757 // EXERCISE: can you see why TmpF15=TmpF14 does not work here, and
1758 // we really do need the above FMOV? ;)
1762 if(isFP) { // extra worrying about div-by-zero
1763 unsigned bogoResult=MakeReg(MVT::f64);
1765 // we do a 'conditional fmov' (of the correct result, depending
1766 // on how the frcpa predicate turned out)
1767 BuildMI(BB, IA64::PFMOV, 2, bogoResult)
1768 .addReg(TmpF12).addReg(TmpPR2);
1769 BuildMI(BB, IA64::CFMOV, 2, Result)
1770 .addReg(bogoResult).addReg(TmpF15).addReg(TmpPR);
1773 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TmpF15);
1775 } else { // this is a modulus
1777 // answer = q * (-b) + a
1778 unsigned ModulusResult = MakeReg(MVT::f64);
1779 unsigned TmpF = MakeReg(MVT::f64);
1780 unsigned TmpI = MakeReg(MVT::i64);
1782 BuildMI(BB, IA64::SUB, 2, TmpI).addReg(IA64::r0).addReg(Tmp2);
1783 BuildMI(BB, IA64::SETFSIG, 1, TmpF).addReg(TmpI);
1784 BuildMI(BB, IA64::XMAL, 3, ModulusResult)
1785 .addReg(TmpF15).addReg(TmpF).addReg(TmpF1);
1786 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(ModulusResult);
1787 } else { // FP modulus! The horror... the horror....
1788 assert(0 && "sorry, no FP modulus just yet!\n!\n");
1795 case ISD::SIGN_EXTEND_INREG: {
1796 Tmp1 = SelectExpr(N.getOperand(0));
1797 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1800 assert(0 && "don't know how to sign extend this type");
1802 case MVT::i8: Opc = IA64::SXT1; break;
1803 case MVT::i16: Opc = IA64::SXT2; break;
1804 case MVT::i32: Opc = IA64::SXT4; break;
1806 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1811 Tmp1 = SelectExpr(N.getOperand(0));
1813 if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Node)) {
1814 if (MVT::isInteger(SetCC->getOperand(0).getValueType())) {
1816 if(ConstantSDNode *CSDN =
1817 dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1818 // if we are comparing against a constant zero
1819 if(CSDN->getValue()==0)
1820 Tmp2 = IA64::r0; // then we can just compare against r0
1822 Tmp2 = SelectExpr(N.getOperand(1));
1823 } else // not comparing against a constant
1824 Tmp2 = SelectExpr(N.getOperand(1));
1826 switch (SetCC->getCondition()) {
1827 default: assert(0 && "Unknown integer comparison!");
1829 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1832 BuildMI(BB, IA64::CMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1835 BuildMI(BB, IA64::CMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1838 BuildMI(BB, IA64::CMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1841 BuildMI(BB, IA64::CMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1844 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1847 BuildMI(BB, IA64::CMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1850 BuildMI(BB, IA64::CMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1853 BuildMI(BB, IA64::CMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1856 BuildMI(BB, IA64::CMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1860 else { // if not integer, should be FP. FIXME: what about bools? ;)
1861 assert(SetCC->getOperand(0).getValueType() != MVT::f32 &&
1862 "error: SETCC should have had incoming f32 promoted to f64!\n");
1864 if(ConstantFPSDNode *CFPSDN =
1865 dyn_cast<ConstantFPSDNode>(N.getOperand(1))) {
1867 // if we are comparing against a constant +0.0 or +1.0
1868 if(CFPSDN->isExactlyValue(+0.0))
1869 Tmp2 = IA64::F0; // then we can just compare against f0
1870 else if(CFPSDN->isExactlyValue(+1.0))
1871 Tmp2 = IA64::F1; // or f1
1873 Tmp2 = SelectExpr(N.getOperand(1));
1874 } else // not comparing against a constant
1875 Tmp2 = SelectExpr(N.getOperand(1));
1877 switch (SetCC->getCondition()) {
1878 default: assert(0 && "Unknown FP comparison!");
1880 BuildMI(BB, IA64::FCMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1883 BuildMI(BB, IA64::FCMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1886 BuildMI(BB, IA64::FCMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1889 BuildMI(BB, IA64::FCMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1892 BuildMI(BB, IA64::FCMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1895 BuildMI(BB, IA64::FCMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1898 BuildMI(BB, IA64::FCMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1901 BuildMI(BB, IA64::FCMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1904 BuildMI(BB, IA64::FCMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1907 BuildMI(BB, IA64::FCMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1913 assert(0 && "this setcc not implemented yet");
1921 // Make sure we generate both values.
1923 ExprMap[N.getValue(1)] = 1; // Generate the token
1925 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1929 if(opcode == ISD::LOAD) { // this is a LOAD
1930 switch (Node->getValueType(0)) {
1931 default: assert(0 && "Cannot load this type!");
1932 case MVT::i1: Opc = IA64::LD1; isBool=true; break;
1933 // FIXME: for now, we treat bool loads the same as i8 loads */
1934 case MVT::i8: Opc = IA64::LD1; break;
1935 case MVT::i16: Opc = IA64::LD2; break;
1936 case MVT::i32: Opc = IA64::LD4; break;
1937 case MVT::i64: Opc = IA64::LD8; break;
1939 case MVT::f32: Opc = IA64::LDF4; break;
1940 case MVT::f64: Opc = IA64::LDF8; break;
1942 } else { // this is an EXTLOAD or ZEXTLOAD
1943 MVT::ValueType TypeBeingLoaded =
1944 cast<VTSDNode>(Node->getOperand(3))->getVT();
1945 switch (TypeBeingLoaded) {
1946 default: assert(0 && "Cannot extload/zextload this type!");
1948 case MVT::i8: Opc = IA64::LD1; break;
1949 case MVT::i16: Opc = IA64::LD2; break;
1950 case MVT::i32: Opc = IA64::LD4; break;
1951 case MVT::f32: Opc = IA64::LDF4; break;
1955 SDOperand Chain = N.getOperand(0);
1956 SDOperand Address = N.getOperand(1);
1958 if(Address.getOpcode() == ISD::GlobalAddress) {
1960 unsigned dummy = MakeReg(MVT::i64);
1961 unsigned dummy2 = MakeReg(MVT::i64);
1962 BuildMI(BB, IA64::ADD, 2, dummy)
1963 .addGlobalAddress(cast<GlobalAddressSDNode>(Address)->getGlobal())
1965 BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
1967 BuildMI(BB, Opc, 1, Result).addReg(dummy2);
1968 else { // emit a little pseudocode to load a bool (stored in one byte)
1969 // into a predicate register
1970 assert(Opc==IA64::LD1 && "problem loading a bool");
1971 unsigned dummy3 = MakeReg(MVT::i64);
1972 BuildMI(BB, Opc, 1, dummy3).addReg(dummy2);
1973 // we compare to 0. true? 0. false? 1.
1974 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
1976 } else if(ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1978 IA64Lowering.restoreGP(BB);
1979 unsigned dummy = MakeReg(MVT::i64);
1980 BuildMI(BB, IA64::ADD, 2, dummy).addConstantPoolIndex(CP->getIndex())
1981 .addReg(IA64::r1); // CPI+GP
1983 BuildMI(BB, Opc, 1, Result).addReg(dummy);
1984 else { // emit a little pseudocode to load a bool (stored in one byte)
1985 // into a predicate register
1986 assert(Opc==IA64::LD1 && "problem loading a bool");
1987 unsigned dummy3 = MakeReg(MVT::i64);
1988 BuildMI(BB, Opc, 1, dummy3).addReg(dummy);
1989 // we compare to 0. true? 0. false? 1.
1990 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
1992 } else if(Address.getOpcode() == ISD::FrameIndex) {
1993 Select(Chain); // FIXME ? what about bools?
1994 unsigned dummy = MakeReg(MVT::i64);
1995 BuildMI(BB, IA64::MOV, 1, dummy)
1996 .addFrameIndex(cast<FrameIndexSDNode>(Address)->getIndex());
1998 BuildMI(BB, Opc, 1, Result).addReg(dummy);
1999 else { // emit a little pseudocode to load a bool (stored in one byte)
2000 // into a predicate register
2001 assert(Opc==IA64::LD1 && "problem loading a bool");
2002 unsigned dummy3 = MakeReg(MVT::i64);
2003 BuildMI(BB, Opc, 1, dummy3).addReg(dummy);
2004 // we compare to 0. true? 0. false? 1.
2005 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
2007 } else { // none of the above...
2009 Tmp2 = SelectExpr(Address);
2011 BuildMI(BB, Opc, 1, Result).addReg(Tmp2);
2012 else { // emit a little pseudocode to load a bool (stored in one byte)
2013 // into a predicate register
2014 assert(Opc==IA64::LD1 && "problem loading a bool");
2015 unsigned dummy = MakeReg(MVT::i64);
2016 BuildMI(BB, Opc, 1, dummy).addReg(Tmp2);
2017 // we compare to 0. true? 0. false? 1.
2018 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy).addReg(IA64::r0);
2025 case ISD::CopyFromReg: {
2027 Result = ExprMap[N.getValue(0)] =
2028 MakeReg(N.getValue(0).getValueType());
2030 SDOperand Chain = N.getOperand(0);
2033 unsigned r = dyn_cast<RegSDNode>(Node)->getReg();
2035 if(N.getValueType() == MVT::i1) // if a bool, we use pseudocode
2036 BuildMI(BB, IA64::PCMPEQUNC, 3, Result)
2037 .addReg(IA64::r0).addReg(IA64::r0).addReg(r);
2038 // (r) Result =cmp.eq.unc(r0,r0)
2040 BuildMI(BB, IA64::MOV, 1, Result).addReg(r); // otherwise MOV
2046 Select(N.getOperand(0));
2048 // The chain for this call is now lowered.
2049 ExprMap.insert(std::make_pair(N.getValue(Node->getNumValues()-1), 1));
2051 //grab the arguments
2052 std::vector<unsigned> argvregs;
2054 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
2055 argvregs.push_back(SelectExpr(N.getOperand(i)));
2057 // see section 8.5.8 of "Itanium Software Conventions and
2058 // Runtime Architecture Guide to see some examples of what's going
2059 // on here. (in short: int args get mapped 1:1 'slot-wise' to out0->out7,
2060 // while FP args get mapped to F8->F15 as needed)
2062 unsigned used_FPArgs=0; // how many FP Args have been used so far?
2065 for(int i = 0, e = std::min(8, (int)argvregs.size()); i < e; ++i)
2067 unsigned intArgs[] = {IA64::out0, IA64::out1, IA64::out2, IA64::out3,
2068 IA64::out4, IA64::out5, IA64::out6, IA64::out7 };
2069 unsigned FPArgs[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
2070 IA64::F12, IA64::F13, IA64::F14, IA64::F15 };
2072 switch(N.getOperand(i+2).getValueType())
2074 default: // XXX do we need to support MVT::i1 here?
2076 N.getOperand(i).Val->dump();
2077 std::cerr << "Type for " << i << " is: " <<
2078 N.getOperand(i+2).getValueType() << std::endl;
2079 assert(0 && "Unknown value type for call");
2081 BuildMI(BB, IA64::MOV, 1, intArgs[i]).addReg(argvregs[i]);
2084 BuildMI(BB, IA64::FMOV, 1, FPArgs[used_FPArgs++])
2085 .addReg(argvregs[i]);
2086 // FIXME: we don't need to do this _all_ the time:
2087 BuildMI(BB, IA64::GETFD, 1, intArgs[i]).addReg(argvregs[i]);
2093 for (int i = 8, e = argvregs.size(); i < e; ++i)
2095 unsigned tempAddr = MakeReg(MVT::i64);
2097 switch(N.getOperand(i+2).getValueType()) {
2100 N.getOperand(i).Val->dump();
2101 std::cerr << "Type for " << i << " is: " <<
2102 N.getOperand(i+2).getValueType() << "\n";
2103 assert(0 && "Unknown value type for call");
2104 case MVT::i1: // FIXME?
2109 BuildMI(BB, IA64::ADDIMM22, 2, tempAddr)
2110 .addReg(IA64::r12).addImm(16 + (i - 8) * 8); // r12 is SP
2111 BuildMI(BB, IA64::ST8, 2).addReg(tempAddr).addReg(argvregs[i]);
2115 BuildMI(BB, IA64::ADDIMM22, 2, tempAddr)
2116 .addReg(IA64::r12).addImm(16 + (i - 8) * 8); // r12 is SP
2117 BuildMI(BB, IA64::STF8, 2).addReg(tempAddr).addReg(argvregs[i]);
2122 // build the right kind of call. if we can branch directly, do so:
2123 if (GlobalAddressSDNode *GASD =
2124 dyn_cast<GlobalAddressSDNode>(N.getOperand(1)))
2126 BuildMI(BB, IA64::BRCALL, 1).addGlobalAddress(GASD->getGlobal(),true);
2127 IA64Lowering.restoreGP_SP_RP(BB);
2129 if (ExternalSymbolSDNode *ESSDN =
2130 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1)))
2131 { // FIXME : currently need this case for correctness, to avoid
2132 // "non-pic code with imm relocation against dynamic symbol" errors
2133 BuildMI(BB, IA64::BRCALL, 1)
2134 .addExternalSymbol(ESSDN->getSymbol(), true);
2135 IA64Lowering.restoreGP_SP_RP(BB);
2137 else { // otherwise we need to get the function descriptor
2138 // load the branch target (function)'s entry point and
2140 Tmp1 = SelectExpr(N.getOperand(1));
2142 unsigned targetEntryPoint=MakeReg(MVT::i64);
2143 unsigned targetGPAddr=MakeReg(MVT::i64);
2144 unsigned currentGP=MakeReg(MVT::i64);
2146 // b6 is a scratch branch register, we load the target entry point
2147 // from the base of the function descriptor
2148 BuildMI(BB, IA64::LD8, 1, targetEntryPoint).addReg(Tmp1);
2149 BuildMI(BB, IA64::MOV, 1, IA64::B6).addReg(targetEntryPoint);
2151 // save the current GP:
2152 BuildMI(BB, IA64::MOV, 1, currentGP).addReg(IA64::r1);
2154 /* TODO: we need to make sure doing this never, ever loads a
2155 * bogus value into r1 (GP). */
2156 // load the target GP (which is at mem[functiondescriptor+8])
2157 BuildMI(BB, IA64::ADDIMM22, 2, targetGPAddr)
2158 .addReg(Tmp1).addImm(8); // FIXME: addimm22? why not postincrement ld
2159 BuildMI(BB, IA64::LD8, 1, IA64::r1).addReg(targetGPAddr);
2161 // and then jump: (well, call)
2162 BuildMI(BB, IA64::BRCALL, 1).addReg(IA64::B6);
2163 // and finally restore the old GP
2164 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(currentGP);
2165 IA64Lowering.restoreSP_RP(BB);
2168 switch (Node->getValueType(0)) {
2169 default: assert(0 && "Unknown value type for call result!");
2170 case MVT::Other: return 1;
2172 BuildMI(BB, IA64::CMPNE, 2, Result)
2173 .addReg(IA64::r8).addReg(IA64::r0);
2179 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r8);
2182 BuildMI(BB, IA64::FMOV, 1, Result).addReg(IA64::F8);
2185 return Result+N.ResNo;
2192 void ISel::Select(SDOperand N) {
2193 unsigned Tmp1, Tmp2, Opc;
2194 unsigned opcode = N.getOpcode();
2196 if (!LoweredTokens.insert(N).second)
2197 return; // Already selected.
2199 SDNode *Node = N.Val;
2201 switch (Node->getOpcode()) {
2203 Node->dump(); std::cerr << "\n";
2204 assert(0 && "Node not handled yet!");
2206 case ISD::EntryToken: return; // Noop
2208 case ISD::TokenFactor: {
2209 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2210 Select(Node->getOperand(i));
2214 case ISD::CopyToReg: {
2215 Select(N.getOperand(0));
2216 Tmp1 = SelectExpr(N.getOperand(1));
2217 Tmp2 = cast<RegSDNode>(N)->getReg();
2220 if(N.getValueType() == MVT::i1) // if a bool, we use pseudocode
2221 BuildMI(BB, IA64::PCMPEQUNC, 3, Tmp2)
2222 .addReg(IA64::r0).addReg(IA64::r0).addReg(Tmp1);
2223 // (Tmp1) Tmp2 = cmp.eq.unc(r0,r0)
2225 BuildMI(BB, IA64::MOV, 1, Tmp2).addReg(Tmp1);
2226 // XXX is this the right way 'round? ;)
2233 /* what the heck is going on here:
2235 <_sabre_> ret with two operands is obvious: chain and value
2237 <_sabre_> ret with 3 values happens when 'expansion' occurs
2238 <_sabre_> e.g. i64 gets split into 2x i32
2240 <_sabre_> you don't have this case on ia64
2242 <_sabre_> so the two returned values go into EAX/EDX on ia32
2243 <camel_> ahhh *memories*
2245 <camel_> ok, thanks :)
2246 <_sabre_> so yeah, everything that has a side effect takes a 'token chain'
2247 <_sabre_> this is the first operand always
2248 <_sabre_> these operand often define chains, they are the last operand
2249 <_sabre_> they are printed as 'ch' if you do DAG.dump()
2252 switch (N.getNumOperands()) {
2254 assert(0 && "Unknown return instruction!");
2256 Select(N.getOperand(0));
2257 Tmp1 = SelectExpr(N.getOperand(1));
2258 switch (N.getOperand(1).getValueType()) {
2259 default: assert(0 && "All other types should have been promoted!!");
2260 // FIXME: do I need to add support for bools here?
2261 // (return '0' or '1' r8, basically...)
2263 // FIXME: need to round floats - 80 bits is bad, the tester
2266 // we mark r8 as live on exit up above in LowerArguments()
2267 BuildMI(BB, IA64::MOV, 1, IA64::r8).addReg(Tmp1);
2270 // we mark F8 as live on exit up above in LowerArguments()
2271 BuildMI(BB, IA64::FMOV, 1, IA64::F8).addReg(Tmp1);
2275 Select(N.getOperand(0));
2278 // before returning, restore the ar.pfs register (set by the 'alloc' up top)
2279 BuildMI(BB, IA64::MOV, 1).addReg(IA64::AR_PFS).addReg(IA64Lowering.VirtGPR);
2280 BuildMI(BB, IA64::RET, 0); // and then just emit a 'ret' instruction
2285 Select(N.getOperand(0));
2286 MachineBasicBlock *Dest =
2287 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2288 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(IA64::p0).addMBB(Dest);
2289 // XXX HACK! we do _not_ need long branches all the time
2293 case ISD::ImplicitDef: {
2294 Select(N.getOperand(0));
2295 BuildMI(BB, IA64::IDEF, 0, cast<RegSDNode>(N)->getReg());
2300 MachineBasicBlock *Dest =
2301 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
2303 Select(N.getOperand(0));
2304 Tmp1 = SelectExpr(N.getOperand(1));
2305 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(Tmp1).addMBB(Dest);
2306 // XXX HACK! we do _not_ need long branches all the time
2316 case ISD::CopyFromReg:
2317 case ISD::DYNAMIC_STACKALLOC:
2321 case ISD::TRUNCSTORE:
2323 Select(N.getOperand(0));
2324 Tmp1 = SelectExpr(N.getOperand(1)); // value
2328 if(opcode == ISD::STORE) {
2329 switch (N.getOperand(1).getValueType()) {
2330 default: assert(0 && "Cannot store this type!");
2331 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2332 // FIXME?: for now, we treat bool loads the same as i8 stores */
2333 case MVT::i8: Opc = IA64::ST1; break;
2334 case MVT::i16: Opc = IA64::ST2; break;
2335 case MVT::i32: Opc = IA64::ST4; break;
2336 case MVT::i64: Opc = IA64::ST8; break;
2338 case MVT::f32: Opc = IA64::STF4; break;
2339 case MVT::f64: Opc = IA64::STF8; break;
2341 } else { // truncstore
2342 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2343 default: assert(0 && "unknown type in truncstore");
2344 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2345 //FIXME: DAG does not promote this load?
2346 case MVT::i8: Opc = IA64::ST1; break;
2347 case MVT::i16: Opc = IA64::ST2; break;
2348 case MVT::i32: Opc = IA64::ST4; break;
2349 case MVT::f32: Opc = IA64::STF4; break;
2353 if(N.getOperand(2).getOpcode() == ISD::GlobalAddress) {
2354 unsigned dummy = MakeReg(MVT::i64);
2355 unsigned dummy2 = MakeReg(MVT::i64);
2356 BuildMI(BB, IA64::ADD, 2, dummy)
2357 .addGlobalAddress(cast<GlobalAddressSDNode>
2358 (N.getOperand(2))->getGlobal()).addReg(IA64::r1);
2359 BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
2362 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(Tmp1);
2363 else { // we are storing a bool, so emit a little pseudocode
2364 // to store a predicate register as one byte
2365 assert(Opc==IA64::ST1);
2366 unsigned dummy3 = MakeReg(MVT::i64);
2367 unsigned dummy4 = MakeReg(MVT::i64);
2368 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2369 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2370 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2371 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(dummy4);
2373 } else if(N.getOperand(2).getOpcode() == ISD::FrameIndex) {
2375 // FIXME? (what about bools?)
2377 unsigned dummy = MakeReg(MVT::i64);
2378 BuildMI(BB, IA64::MOV, 1, dummy)
2379 .addFrameIndex(cast<FrameIndexSDNode>(N.getOperand(2))->getIndex());
2380 BuildMI(BB, Opc, 2).addReg(dummy).addReg(Tmp1);
2381 } else { // otherwise
2382 Tmp2 = SelectExpr(N.getOperand(2)); //address
2384 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(Tmp1);
2385 else { // we are storing a bool, so emit a little pseudocode
2386 // to store a predicate register as one byte
2387 assert(Opc==IA64::ST1);
2388 unsigned dummy3 = MakeReg(MVT::i64);
2389 unsigned dummy4 = MakeReg(MVT::i64);
2390 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2391 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2392 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2393 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(dummy4);
2399 case ISD::CALLSEQ_START:
2400 case ISD::CALLSEQ_END: {
2401 Select(N.getOperand(0));
2402 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2404 Opc = N.getOpcode() == ISD::CALLSEQ_START ? IA64::ADJUSTCALLSTACKDOWN :
2405 IA64::ADJUSTCALLSTACKUP;
2406 BuildMI(BB, Opc, 1).addImm(Tmp1);
2412 assert(0 && "GAME OVER. INSERT COIN?");
2416 /// createIA64PatternInstructionSelector - This pass converts an LLVM function
2417 /// into a machine code representation using pattern matching and a machine
2418 /// description file.
2420 FunctionPass *llvm::createIA64PatternInstructionSelector(TargetMachine &TM) {
2421 return new ISel(TM);