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::BRTWOWAY_CC , MVT::Other, Expand);
61 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
63 setSetCCResultType(MVT::i1);
64 setShiftAmountType(MVT::i64);
66 setOperationAction(ISD::EXTLOAD , MVT::i1 , Promote);
68 setOperationAction(ISD::ZEXTLOAD , MVT::i1 , Expand);
70 setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
71 setOperationAction(ISD::SEXTLOAD , MVT::i8 , Expand);
72 setOperationAction(ISD::SEXTLOAD , MVT::i16 , Expand);
73 setOperationAction(ISD::SEXTLOAD , MVT::i32 , Expand);
75 setOperationAction(ISD::SREM , MVT::f32 , Expand);
76 setOperationAction(ISD::SREM , MVT::f64 , Expand);
78 setOperationAction(ISD::UREM , MVT::f32 , Expand);
79 setOperationAction(ISD::UREM , MVT::f64 , Expand);
81 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
82 setOperationAction(ISD::MEMSET , MVT::Other, Expand);
83 setOperationAction(ISD::MEMCPY , MVT::Other, Expand);
85 // We don't support sin/cos/sqrt
86 setOperationAction(ISD::FSIN , MVT::f64, Expand);
87 setOperationAction(ISD::FCOS , MVT::f64, Expand);
88 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
89 setOperationAction(ISD::FSIN , MVT::f32, Expand);
90 setOperationAction(ISD::FCOS , MVT::f32, Expand);
91 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
93 //IA64 has these, but they are not implemented
94 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
95 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
97 computeRegisterProperties();
99 addLegalFPImmediate(+0.0);
100 addLegalFPImmediate(+1.0);
101 addLegalFPImmediate(-0.0);
102 addLegalFPImmediate(-1.0);
105 /// LowerArguments - This hook must be implemented to indicate how we should
106 /// lower the arguments for the specified function, into the specified DAG.
107 virtual std::vector<SDOperand>
108 LowerArguments(Function &F, SelectionDAG &DAG);
110 /// LowerCallTo - This hook lowers an abstract call to a function into an
112 virtual std::pair<SDOperand, SDOperand>
113 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
114 bool isTailCall, SDOperand Callee, ArgListTy &Args,
117 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
118 Value *VAListV, SelectionDAG &DAG);
119 virtual std::pair<SDOperand,SDOperand>
120 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
121 const Type *ArgTy, SelectionDAG &DAG);
123 void restoreGP_SP_RP(MachineBasicBlock* BB)
125 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
126 BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
127 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
130 void restoreSP_RP(MachineBasicBlock* BB)
132 BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
133 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
136 void restoreRP(MachineBasicBlock* BB)
138 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
141 void restoreGP(MachineBasicBlock* BB)
143 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
150 std::vector<SDOperand>
151 IA64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
152 std::vector<SDOperand> ArgValues;
155 // add beautiful description of IA64 stack frame format
156 // here (from intel 24535803.pdf most likely)
158 MachineFunction &MF = DAG.getMachineFunction();
159 MachineFrameInfo *MFI = MF.getFrameInfo();
161 GP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
162 SP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
163 RP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
165 MachineBasicBlock& BB = MF.front();
167 unsigned args_int[] = {IA64::r32, IA64::r33, IA64::r34, IA64::r35,
168 IA64::r36, IA64::r37, IA64::r38, IA64::r39};
170 unsigned args_FP[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
171 IA64::F12,IA64::F13,IA64::F14, IA64::F15};
177 unsigned used_FPArgs = 0; // how many FP args have been used so far?
179 unsigned ArgOffset = 0;
182 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
184 SDOperand newroot, argt;
185 if(count < 8) { // need to fix this logic? maybe.
187 switch (getValueType(I->getType())) {
189 std::cerr << "ERROR in LowerArgs: unknown type "
190 << getValueType(I->getType()) << "\n";
193 // fixme? (well, will need to for weird FP structy stuff,
194 // see intel ABI docs)
196 //XXX BuildMI(&BB, IA64::IDEF, 0, args_FP[used_FPArgs]);
197 MF.addLiveIn(args_FP[used_FPArgs]); // mark this reg as liveIn
198 // floating point args go into f8..f15 as-needed, the increment
199 argVreg[count] = // is below..:
200 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::f64));
201 // FP args go into f8..f15 as needed: (hence the ++)
202 argPreg[count] = args_FP[used_FPArgs++];
203 argOpc[count] = IA64::FMOV;
204 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), argVreg[count],
206 if (I->getType() == Type::FloatTy)
207 argt = DAG.getNode(ISD::FP_ROUND, MVT::f32, argt);
209 case MVT::i1: // NOTE: as far as C abi stuff goes,
210 // bools are just boring old ints
215 //XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
216 MF.addLiveIn(args_int[count]); // mark this register as liveIn
218 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
219 argPreg[count] = args_int[count];
220 argOpc[count] = IA64::MOV;
222 DAG.getCopyFromReg(DAG.getRoot(), argVreg[count], MVT::i64);
223 if ( getValueType(I->getType()) != MVT::i64)
224 argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
228 } else { // more than 8 args go into the frame
229 // Create the frame index object for this incoming parameter...
230 ArgOffset = 16 + 8 * (count - 8);
231 int FI = MFI->CreateFixedObject(8, ArgOffset);
233 // Create the SelectionDAG nodes corresponding to a load
234 //from this parameter
235 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
236 argt = newroot = DAG.getLoad(getValueType(I->getType()),
237 DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
240 DAG.setRoot(newroot.getValue(1));
241 ArgValues.push_back(argt);
245 // Create a vreg to hold the output of (what will become)
246 // the "alloc" instruction
247 VirtGPR = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
248 BuildMI(&BB, IA64::PSEUDO_ALLOC, 0, VirtGPR);
249 // we create a PSEUDO_ALLOC (pseudo)instruction for now
251 BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
254 BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
255 BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
258 BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
261 BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
262 BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
265 unsigned tempOffset=0;
267 // if this is a varargs function, we simply lower llvm.va_start by
268 // pointing to the first entry
271 VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
274 // here we actually do the moving of args, and store them to the stack
275 // too if this is a varargs function:
276 for (int i = 0; i < count && i < 8; ++i) {
277 BuildMI(&BB, argOpc[i], 1, argVreg[i]).addReg(argPreg[i]);
279 // if this is a varargs function, we copy the input registers to the stack
280 int FI = MFI->CreateFixedObject(8, tempOffset);
281 tempOffset+=8; //XXX: is it safe to use r22 like this?
282 BuildMI(&BB, IA64::MOV, 1, IA64::r22).addFrameIndex(FI);
283 // FIXME: we should use st8.spill here, one day
284 BuildMI(&BB, IA64::ST8, 1, IA64::r22).addReg(argPreg[i]);
288 // Finally, inform the code generator which regs we return values in.
289 // (see the ISD::RET: case down below)
290 switch (getValueType(F.getReturnType())) {
291 default: assert(0 && "i have no idea where to return this type!");
292 case MVT::isVoid: break;
298 MF.addLiveOut(IA64::r8);
302 MF.addLiveOut(IA64::F8);
309 std::pair<SDOperand, SDOperand>
310 IA64TargetLowering::LowerCallTo(SDOperand Chain,
311 const Type *RetTy, bool isVarArg,
312 unsigned CallingConv, bool isTailCall,
313 SDOperand Callee, ArgListTy &Args,
316 MachineFunction &MF = DAG.getMachineFunction();
318 unsigned NumBytes = 16;
319 unsigned outRegsUsed = 0;
321 if (Args.size() > 8) {
322 NumBytes += (Args.size() - 8) * 8;
325 outRegsUsed = Args.size();
328 // FIXME? this WILL fail if we ever try to pass around an arg that
329 // consumes more than a single output slot (a 'real' double, int128
330 // some sort of aggregate etc.), as we'll underestimate how many 'outX'
331 // registers we use. Hopefully, the assembler will notice.
332 MF.getInfo<IA64FunctionInfo>()->outRegsUsed=
333 std::max(outRegsUsed, MF.getInfo<IA64FunctionInfo>()->outRegsUsed);
335 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
336 DAG.getConstant(NumBytes, getPointerTy()));
338 std::vector<SDOperand> args_to_use;
339 for (unsigned i = 0, e = Args.size(); i != e; ++i)
341 switch (getValueType(Args[i].second)) {
342 default: assert(0 && "unexpected argument type!");
347 //promote to 64-bits, sign/zero extending based on type
349 if(Args[i].second->isSigned())
350 Args[i].first = DAG.getNode(ISD::SIGN_EXTEND, MVT::i64,
353 Args[i].first = DAG.getNode(ISD::ZERO_EXTEND, MVT::i64,
358 Args[i].first = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Args[i].first);
363 args_to_use.push_back(Args[i].first);
366 std::vector<MVT::ValueType> RetVals;
367 MVT::ValueType RetTyVT = getValueType(RetTy);
368 if (RetTyVT != MVT::isVoid)
369 RetVals.push_back(RetTyVT);
370 RetVals.push_back(MVT::Other);
372 SDOperand TheCall = SDOperand(DAG.getCall(RetVals, Chain,
373 Callee, args_to_use), 0);
374 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
375 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
376 DAG.getConstant(NumBytes, getPointerTy()));
377 return std::make_pair(TheCall, Chain);
381 IA64TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
382 Value *VAListV, SelectionDAG &DAG) {
383 // vastart just stores the address of the VarArgsFrameIndex slot.
384 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
385 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR,
386 VAListP, DAG.getSrcValue(VAListV));
389 std::pair<SDOperand,SDOperand> IA64TargetLowering::
390 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
391 const Type *ArgTy, SelectionDAG &DAG) {
393 MVT::ValueType ArgVT = getValueType(ArgTy);
394 SDOperand Val = DAG.getLoad(MVT::i64, Chain,
395 VAListP, DAG.getSrcValue(VAListV));
396 SDOperand Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), Val,
397 DAG.getSrcValue(NULL));
399 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
402 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
403 "Other types should have been promoted for varargs!");
406 Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
407 DAG.getConstant(Amt, Val.getValueType()));
408 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
409 Val, VAListP, DAG.getSrcValue(VAListV));
410 return std::make_pair(Result, Chain);
415 //===--------------------------------------------------------------------===//
416 /// ISel - IA64 specific code to select IA64 machine instructions for
417 /// SelectionDAG operations.
419 class ISel : public SelectionDAGISel {
420 /// IA64Lowering - This object fully describes how to lower LLVM code to an
421 /// IA64-specific SelectionDAG.
422 IA64TargetLowering IA64Lowering;
423 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
424 // for sdiv and udiv until it is put into the future
427 /// ExprMap - As shared expressions are codegen'd, we keep track of which
428 /// vreg the value is produced in, so we only emit one copy of each compiled
430 std::map<SDOperand, unsigned> ExprMap;
431 std::set<SDOperand> LoweredTokens;
434 ISel(TargetMachine &TM) : SelectionDAGISel(IA64Lowering), IA64Lowering(TM),
437 /// InstructionSelectBasicBlock - This callback is invoked by
438 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
439 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
441 unsigned SelectExpr(SDOperand N);
442 void Select(SDOperand N);
443 // a dag->dag to transform mul-by-constant-int to shifts+adds/subs
444 SDOperand BuildConstmulSequence(SDOperand N);
446 const char *getPassName() const { return "IA64 Instruction Selector"; }
450 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
451 /// when it has created a SelectionDAG for us to codegen.
452 void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
454 // Codegen the basic block.
456 Select(DAG.getRoot());
458 // Clear state used for selection.
460 LoweredTokens.clear();
464 // strip leading '0' characters from a string
465 void munchLeadingZeros(std::string& inString) {
466 while(inString.c_str()[0]=='0') {
467 inString.erase(0, 1);
471 // strip trailing '0' characters from a string
472 void munchTrailingZeros(std::string& inString) {
473 int curPos=inString.length()-1;
475 while(inString.c_str()[curPos]=='0') {
476 inString.erase(curPos, 1);
481 // return how many consecutive '0' characters are at the end of a string
482 unsigned int countTrailingZeros(std::string& inString) {
483 int curPos=inString.length()-1;
484 unsigned int zeroCount=0;
486 while(inString.c_str()[curPos--]=='0') {
492 // booth encode a string of '1' and '0' characters (returns string of 'P' (+1)
493 // '0' and 'N' (-1) characters)
494 void boothEncode(std::string inString, std::string& boothEncodedString) {
498 int lim=inString.size();
501 if(inString[curpos]=='1') { // if we see a '1', look for a run of them
503 std::string replaceString="N";
505 // find the run length
506 for(;inString[curpos+runlength]=='1';runlength++) ;
508 for(int i=0; i<runlength-1; i++)
513 inString.replace(curpos, runlength+1, replaceString);
517 } else { // a zero, we just keep chugging along
522 // clean up (trim the string, reverse it and turn '1's into 'P's)
523 munchTrailingZeros(inString);
524 boothEncodedString="";
526 for(int i=inString.size()-1;i>=0;i--)
528 boothEncodedString+="P";
530 boothEncodedString+=inString[i];
534 struct shiftaddblob { // this encodes stuff like (x=) "A << B [+-] C << D"
535 unsigned firstVal; // A
536 unsigned firstShift; // B
537 unsigned secondVal; // C
538 unsigned secondShift; // D
542 /* this implements Lefevre's "pattern-based" constant multiplication,
543 * see "Multiplication by an Integer Constant", INRIA report 1999-06
545 * TODO: implement a method to try rewriting P0N<->0PP / N0P<->0NN
546 * to get better booth encodings - this does help in practice
547 * TODO: weight shifts appropriately (most architectures can't
548 * fuse a shift and an add for arbitrary shift amounts) */
549 unsigned lefevre(const std::string inString,
550 std::vector<struct shiftaddblob> &ops) {
551 std::string retstring;
552 std::string s = inString;
553 munchTrailingZeros(s);
555 int length=s.length()-1;
561 std::vector<int> p,n;
563 for(int i=0; i<=length; i++) {
564 if (s.c_str()[length-i]=='P') {
566 } else if (s.c_str()[length-i]=='N') {
574 std::map<const int, int> w;
576 for(unsigned i=0; i<p.size(); i++) {
577 for(unsigned j=0; j<i; j++) {
582 for(unsigned i=1; i<n.size(); i++) {
583 for(unsigned j=0; j<i; j++) {
588 for(unsigned i=0; i<p.size(); i++) {
589 for(unsigned j=0; j<n.size(); j++) {
590 w[-abs(p[i]-n[j])]++;
594 std::map<const int, int>::const_iterator ii;
596 std::multimap<int, int> sorted_by_value;
598 for(ii = w.begin(); ii!=w.end(); ii++)
599 sorted_by_value.insert(std::pair<int, int>((*ii).second,(*ii).first));
601 for (std::multimap<int, int>::iterator it = sorted_by_value.begin();
602 it != sorted_by_value.end(); ++it) {
603 d.push_back((*it).second);
609 while(d.size()>0 && (w[int_d=d.back()] > int_W)) {
617 for(unsigned base=0; base<retstring.size(); base++) {
618 if( ((base+z+1) < retstring.size()) &&
619 retstring.c_str()[base]=='P' &&
620 retstring.c_str()[base+z+1]=='P')
624 retstring.replace(base, 1, "0");
625 retstring.replace(base+z+1, 1, "p");
629 for(unsigned base=0; base<retstring.size(); base++) {
630 if( ((base+z+1) < retstring.size()) &&
631 retstring.c_str()[base]=='N' &&
632 retstring.c_str()[base+z+1]=='N')
636 retstring.replace(base, 1, "0");
637 retstring.replace(base+z+1, 1, "n");
642 for(unsigned base=0; base<retstring.size(); base++) {
643 if( ((base+z+1) < retstring.size()) &&
644 ((retstring.c_str()[base]=='P' &&
645 retstring.c_str()[base+z+1]=='N') ||
646 (retstring.c_str()[base]=='N' &&
647 retstring.c_str()[base+z+1]=='P')) ) {
651 if(retstring.c_str()[base]=='P') {
652 retstring.replace(base, 1, "0");
653 retstring.replace(base+z+1, 1, "p");
654 } else { // retstring[base]=='N'
655 retstring.replace(base, 1, "0");
656 retstring.replace(base+z+1, 1, "n");
668 } d.pop_back(); // hmm
672 for(unsigned i=0; i<t.length(); i++) {
673 if(t.c_str()[i]=='p' || t.c_str()[i]=='n')
674 t.replace(i, 1, "0");
677 for(unsigned i=0; i<u.length(); i++) {
678 if(u[i]=='P' || u[i]=='N')
679 u.replace(i, 1, "0");
681 u.replace(i, 1, "P");
683 u.replace(i, 1, "N");
696 bool hit=(u[pos]=='N');
701 for(unsigned p=0; p<u.length(); p++) {
702 bool isP=(u[p]=='P');
703 bool isN=(u[p]=='N');
706 u.replace(p, 1, "N");
708 u.replace(p, 1, "P");
712 munchLeadingZeros(u);
714 int i = lefevre(u, ops);
718 blob.firstVal=i; blob.firstShift=c;
720 blob.secondVal=i; blob.secondShift=0;
726 munchLeadingZeros(t);
731 if(t.c_str()[0]!='P') {
733 for(unsigned p=0; p<t.length(); p++) {
734 bool isP=(t.c_str()[p]=='P');
735 bool isN=(t.c_str()[p]=='N');
738 t.replace(p, 1, "N");
740 t.replace(p, 1, "P");
744 int j = lefevre(t, ops);
746 int trail=countTrailingZeros(u);
747 blob.secondVal=i; blob.secondShift=trail;
749 trail=countTrailingZeros(t);
750 blob.firstVal=j; blob.firstShift=trail;
754 blob.isSub=false; // first + second
757 blob.isSub=true; // first - second
760 blob.isSub=true; // second - first
761 int tmpval, tmpshift;
762 tmpval=blob.firstVal;
763 tmpshift=blob.firstShift;
764 blob.firstVal=blob.secondVal;
765 blob.firstShift=blob.secondShift;
766 blob.secondVal=tmpval;
767 blob.secondShift=tmpshift;
776 SDOperand ISel::BuildConstmulSequence(SDOperand N) {
777 //FIXME: we should shortcut this stuff for multiplies by 2^n+1
778 // in particular, *3 is nicer as *2+1, not *4-1
779 int64_t constant=cast<ConstantSDNode>(N.getOperand(1))->getValue();
782 unsigned preliminaryShift=0;
784 assert(constant != 0 && "erk, you're trying to multiply by constant zero\n");
786 // first, we make the constant to multiply by positive
794 // next, we make it odd.
795 for(; (constant%2==0); preliminaryShift++)
798 //OK, we have a positive, odd number of 64 bits or less. Convert it
799 //to a binary string, constantString[0] is the LSB
800 char constantString[65];
801 for(int i=0; i<64; i++)
802 constantString[i]='0'+((constant>>i)&0x1);
803 constantString[64]=0;
805 // now, Booth encode it
806 std::string boothEncodedString;
807 boothEncode(constantString, boothEncodedString);
809 std::vector<struct shiftaddblob> ops;
810 // do the transformation, filling out 'ops'
811 lefevre(boothEncodedString, ops);
813 assert(ops.size() < 80 && "constmul code has gone haywire\n");
814 SDOperand results[80]; // temporary results (of adds/subs of shifts)
816 // now turn 'ops' into DAG bits
817 for(unsigned i=0; i<ops.size(); i++) {
818 SDOperand amt = ISelDAG->getConstant(ops[i].firstShift, MVT::i64);
819 SDOperand val = (ops[i].firstVal == 0) ? N.getOperand(0) :
820 results[ops[i].firstVal-1];
821 SDOperand left = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
822 amt = ISelDAG->getConstant(ops[i].secondShift, MVT::i64);
823 val = (ops[i].secondVal == 0) ? N.getOperand(0) :
824 results[ops[i].secondVal-1];
825 SDOperand right = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
827 results[i] = ISelDAG->getNode(ISD::SUB, MVT::i64, left, right);
829 results[i] = ISelDAG->getNode(ISD::ADD, MVT::i64, left, right);
832 // don't forget flippedSign and preliminaryShift!
833 SDOperand shiftedresult;
834 if(preliminaryShift) {
835 SDOperand finalshift = ISelDAG->getConstant(preliminaryShift, MVT::i64);
836 shiftedresult = ISelDAG->getNode(ISD::SHL, MVT::i64,
837 results[ops.size()-1], finalshift);
838 } else { // there was no preliminary divide-by-power-of-2 required
839 shiftedresult = results[ops.size()-1];
842 SDOperand finalresult;
843 if(flippedSign) { // if we were multiplying by a negative constant:
844 SDOperand zero = ISelDAG->getConstant(0, MVT::i64);
845 // subtract the result from 0 to flip its sign
846 finalresult = ISelDAG->getNode(ISD::SUB, MVT::i64, zero, shiftedresult);
847 } else { // there was no preliminary multiply by -1 required
848 finalresult = shiftedresult;
854 /// ponderIntegerDivisionBy - When handling integer divides, if the divide
855 /// is by a constant such that we can efficiently codegen it, this
856 /// function says what to do. Currently, it returns 0 if the division must
857 /// become a genuine divide, and 1 if the division can be turned into a
859 static unsigned ponderIntegerDivisionBy(SDOperand N, bool isSigned,
861 if (N.getOpcode() != ISD::Constant) return 0; // if not a divide by
862 // a constant, give up.
864 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
866 if (isPowerOf2_64(v)) { // if a division by a power of two, say so
871 return 0; // fallthrough
874 static unsigned ponderIntegerAndWith(SDOperand N, unsigned& Imm) {
875 if (N.getOpcode() != ISD::Constant) return 0; // if not ANDing with
876 // a constant, give up.
878 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
880 if (isMask_64(v)) { // if ANDing with ((2^n)-1) for some n
881 Imm = Log2_64(v) + 1;
885 return 0; // fallthrough
888 static unsigned ponderIntegerAdditionWith(SDOperand N, unsigned& Imm) {
889 if (N.getOpcode() != ISD::Constant) return 0; // if not adding a
890 // constant, give up.
891 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
893 if (v <= 8191 && v >= -8192) { // if this constants fits in 14 bits, say so
894 Imm = v & 0x3FFF; // 14 bits
897 return 0; // fallthrough
900 static unsigned ponderIntegerSubtractionFrom(SDOperand N, unsigned& Imm) {
901 if (N.getOpcode() != ISD::Constant) return 0; // if not subtracting a
902 // constant, give up.
903 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
905 if (v <= 127 && v >= -128) { // if this constants fits in 8 bits, say so
906 Imm = v & 0xFF; // 8 bits
909 return 0; // fallthrough
912 unsigned ISel::SelectExpr(SDOperand N) {
914 unsigned Tmp1, Tmp2, Tmp3;
916 MVT::ValueType DestType = N.getValueType();
918 unsigned opcode = N.getOpcode();
920 SDNode *Node = N.Val;
923 if (Node->getOpcode() == ISD::CopyFromReg)
924 // Just use the specified register as our input.
925 return cast<RegisterSDNode>(Node->getOperand(1))->getReg();
927 unsigned &Reg = ExprMap[N];
930 if (N.getOpcode() != ISD::CALL && N.getOpcode() != ISD::TAILCALL)
931 Reg = Result = (N.getValueType() != MVT::Other) ?
932 MakeReg(N.getValueType()) : 1;
934 // If this is a call instruction, make sure to prepare ALL of the result
935 // values as well as the chain.
936 if (Node->getNumValues() == 1)
937 Reg = Result = 1; // Void call, just a chain.
939 Result = MakeReg(Node->getValueType(0));
940 ExprMap[N.getValue(0)] = Result;
941 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
942 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
943 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
947 switch (N.getOpcode()) {
950 assert(0 && "Node not handled!\n");
952 case ISD::FrameIndex: {
953 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
954 BuildMI(BB, IA64::MOV, 1, Result).addFrameIndex(Tmp1);
958 case ISD::ConstantPool: {
959 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
960 IA64Lowering.restoreGP(BB); // FIXME: do i really need this?
961 BuildMI(BB, IA64::ADD, 2, Result).addConstantPoolIndex(Tmp1)
966 case ISD::ConstantFP: {
967 Tmp1 = Result; // Intermediate Register
968 if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
969 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
970 Tmp1 = MakeReg(MVT::f64);
972 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
973 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
974 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F0); // load 0.0
975 else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
976 cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
977 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F1); // load 1.0
979 assert(0 && "Unexpected FP constant!");
981 // we multiply by +1.0, negate (this is FNMA), and then add 0.0
982 BuildMI(BB, IA64::FNMA, 3, Result).addReg(Tmp1).addReg(IA64::F1)
987 case ISD::DYNAMIC_STACKALLOC: {
988 // Generate both result values.
990 ExprMap[N.getValue(1)] = 1; // Generate the token
992 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
994 // FIXME: We are currently ignoring the requested alignment for handling
995 // greater than the stack alignment. This will need to be revisited at some
996 // point. Align = N.getOperand(2);
998 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
999 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
1000 std::cerr << "Cannot allocate stack object with greater alignment than"
1001 << " the stack alignment yet!";
1006 Select(N.getOperand(0));
1007 if (ConstantSDNode* CN = dyn_cast<ConstantSDNode>(N.getOperand(1)))
1009 if (CN->getValue() < 32000)
1011 BuildMI(BB, IA64::ADDIMM22, 2, IA64::r12).addReg(IA64::r12)
1012 .addImm(-CN->getValue());
1014 Tmp1 = SelectExpr(N.getOperand(1));
1015 // Subtract size from stack pointer, thereby allocating some space.
1016 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1019 Tmp1 = SelectExpr(N.getOperand(1));
1020 // Subtract size from stack pointer, thereby allocating some space.
1021 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1024 Select(N.getOperand(0));
1025 Tmp1 = SelectExpr(N.getOperand(1));
1026 // Subtract size from stack pointer, thereby allocating some space.
1027 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1028 // Put a pointer to the space into the result register, by copying the
1030 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r12);
1035 Tmp1 = SelectExpr(N.getOperand(0)); //Cond
1036 Tmp2 = SelectExpr(N.getOperand(1)); //Use if TRUE
1037 Tmp3 = SelectExpr(N.getOperand(2)); //Use if FALSE
1039 unsigned bogoResult;
1041 switch (N.getOperand(1).getValueType()) {
1042 default: assert(0 &&
1043 "ISD::SELECT: 'select'ing something other than i1, i64 or f64!\n");
1044 // for i1, we load the condition into an integer register, then
1045 // conditionally copy Tmp2 and Tmp3 to Tmp1 in parallel (only one
1046 // of them will go through, since the integer register will hold
1049 bogoResult=MakeReg(MVT::i1);
1051 // load the condition into an integer register
1052 unsigned condReg=MakeReg(MVT::i64);
1053 unsigned dummy=MakeReg(MVT::i64);
1054 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1055 BuildMI(BB, IA64::TPCADDIMM22, 2, condReg).addReg(dummy)
1056 .addImm(1).addReg(Tmp1);
1058 // initialize Result (bool) to false (hence UNC) and if
1059 // the select condition (condReg) is false (0), copy Tmp3
1060 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoResult)
1061 .addReg(condReg).addReg(IA64::r0).addReg(Tmp3);
1063 // now, if the selection condition is true, write 1 to the
1064 // result if Tmp2 is 1
1065 BuildMI(BB, IA64::TPCMPNE, 3, Result).addReg(bogoResult)
1066 .addReg(condReg).addReg(IA64::r0).addReg(Tmp2);
1069 // for i64/f64, we just copy Tmp3 and then conditionally overwrite it
1070 // with Tmp2 if Tmp1 is true
1072 bogoResult=MakeReg(MVT::i64);
1073 BuildMI(BB, IA64::MOV, 1, bogoResult).addReg(Tmp3);
1074 BuildMI(BB, IA64::CMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1078 bogoResult=MakeReg(MVT::f64);
1079 BuildMI(BB, IA64::FMOV, 1, bogoResult).addReg(Tmp3);
1080 BuildMI(BB, IA64::CFMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1088 case ISD::Constant: {
1089 unsigned depositPos=0;
1090 unsigned depositLen=0;
1091 switch (N.getValueType()) {
1092 default: assert(0 && "Cannot use constants of this type!");
1093 case MVT::i1: { // if a bool, we don't 'load' so much as generate
1095 if(cast<ConstantSDNode>(N)->getValue()) // true:
1096 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1098 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1099 return Result; // early exit
1101 case MVT::i64: break;
1104 int64_t immediate = cast<ConstantSDNode>(N)->getValue();
1106 if(immediate==0) { // if the constant is just zero,
1107 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r0); // just copy r0
1108 return Result; // early exit
1111 if (immediate <= 8191 && immediate >= -8192) {
1112 // if this constants fits in 14 bits, we use a mov the assembler will
1113 // turn into: "adds rDest=imm,r0" (and _not_ "andl"...)
1114 BuildMI(BB, IA64::MOVSIMM14, 1, Result).addSImm(immediate);
1115 return Result; // early exit
1118 if (immediate <= 2097151 && immediate >= -2097152) {
1119 // if this constants fits in 22 bits, we use a mov the assembler will
1120 // turn into: "addl rDest=imm,r0"
1121 BuildMI(BB, IA64::MOVSIMM22, 1, Result).addSImm(immediate);
1122 return Result; // early exit
1125 /* otherwise, our immediate is big, so we use movl */
1126 uint64_t Imm = immediate;
1127 BuildMI(BB, IA64::MOVLIMM64, 1, Result).addImm64(Imm);
1132 BuildMI(BB, IA64::IDEF, 0, Result);
1136 case ISD::GlobalAddress: {
1137 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1138 unsigned Tmp1 = MakeReg(MVT::i64);
1140 BuildMI(BB, IA64::ADD, 2, Tmp1).addGlobalAddress(GV).addReg(IA64::r1);
1141 BuildMI(BB, IA64::LD8, 1, Result).addReg(Tmp1);
1146 case ISD::ExternalSymbol: {
1147 const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
1148 // assert(0 && "sorry, but what did you want an ExternalSymbol for again?");
1149 BuildMI(BB, IA64::MOV, 1, Result).addExternalSymbol(Sym); // XXX
1153 case ISD::FP_EXTEND: {
1154 Tmp1 = SelectExpr(N.getOperand(0));
1155 BuildMI(BB, IA64::FMOV, 1, Result).addReg(Tmp1);
1159 case ISD::ZERO_EXTEND: {
1160 Tmp1 = SelectExpr(N.getOperand(0)); // value
1162 switch (N.getOperand(0).getValueType()) {
1163 default: assert(0 && "Cannot zero-extend this type!");
1164 case MVT::i8: Opc = IA64::ZXT1; break;
1165 case MVT::i16: Opc = IA64::ZXT2; break;
1166 case MVT::i32: Opc = IA64::ZXT4; break;
1168 // we handle bools differently! :
1169 case MVT::i1: { // if the predicate reg has 1, we want a '1' in our GR.
1170 unsigned dummy = MakeReg(MVT::i64);
1172 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1173 // ...then conditionally (PR:Tmp1) add 1:
1174 BuildMI(BB, IA64::TPCADDIMM22, 2, Result).addReg(dummy)
1175 .addImm(1).addReg(Tmp1);
1176 return Result; // XXX early exit!
1180 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1184 case ISD::SIGN_EXTEND: { // we should only have to handle i1 -> i64 here!!!
1186 assert(0 && "hmm, ISD::SIGN_EXTEND: shouldn't ever be reached. bad luck!\n");
1188 Tmp1 = SelectExpr(N.getOperand(0)); // value
1190 switch (N.getOperand(0).getValueType()) {
1191 default: assert(0 && "Cannot sign-extend this type!");
1192 case MVT::i1: assert(0 && "trying to sign extend a bool? ow.\n");
1193 Opc = IA64::SXT1; break;
1194 // FIXME: for now, we treat bools the same as i8s
1195 case MVT::i8: Opc = IA64::SXT1; break;
1196 case MVT::i16: Opc = IA64::SXT2; break;
1197 case MVT::i32: Opc = IA64::SXT4; break;
1200 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1204 case ISD::TRUNCATE: {
1205 // we use the funky dep.z (deposit (zero)) instruction to deposit bits
1206 // of R0 appropriately.
1207 switch (N.getOperand(0).getValueType()) {
1208 default: assert(0 && "Unknown truncate!");
1209 case MVT::i64: break;
1211 Tmp1 = SelectExpr(N.getOperand(0));
1212 unsigned depositPos, depositLen;
1214 switch (N.getValueType()) {
1215 default: assert(0 && "Unknown truncate!");
1217 // if input (normal reg) is 0, 0!=0 -> false (0), if 1, 1!=0 ->true (1):
1218 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1)
1220 return Result; // XXX early exit!
1222 case MVT::i8: depositPos=0; depositLen=8; break;
1223 case MVT::i16: depositPos=0; depositLen=16; break;
1224 case MVT::i32: depositPos=0; depositLen=32; break;
1226 BuildMI(BB, IA64::DEPZ, 3, Result).addReg(Tmp1)
1227 .addImm(depositPos).addImm(depositLen);
1232 case ISD::FP_ROUND: {
1233 assert (DestType == MVT::f32 && N.getOperand(0).getValueType() == MVT::f64 &&
1234 "error: trying to FP_ROUND something other than f64 -> f32!\n");
1235 Tmp1 = SelectExpr(N.getOperand(0));
1236 BuildMI(BB, IA64::FADDS, 2, Result).addReg(Tmp1).addReg(IA64::F0);
1237 // we add 0.0 using a single precision add to do rounding
1242 // FIXME: the following 4 cases need cleaning
1243 case ISD::SINT_TO_FP: {
1244 Tmp1 = SelectExpr(N.getOperand(0));
1245 Tmp2 = MakeReg(MVT::f64);
1246 unsigned dummy = MakeReg(MVT::f64);
1247 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1248 BuildMI(BB, IA64::FCVTXF, 1, dummy).addReg(Tmp2);
1249 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1253 case ISD::UINT_TO_FP: {
1254 Tmp1 = SelectExpr(N.getOperand(0));
1255 Tmp2 = MakeReg(MVT::f64);
1256 unsigned dummy = MakeReg(MVT::f64);
1257 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1258 BuildMI(BB, IA64::FCVTXUF, 1, dummy).addReg(Tmp2);
1259 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1263 case ISD::FP_TO_SINT: {
1264 Tmp1 = SelectExpr(N.getOperand(0));
1265 Tmp2 = MakeReg(MVT::f64);
1266 BuildMI(BB, IA64::FCVTFXTRUNC, 1, Tmp2).addReg(Tmp1);
1267 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1271 case ISD::FP_TO_UINT: {
1272 Tmp1 = SelectExpr(N.getOperand(0));
1273 Tmp2 = MakeReg(MVT::f64);
1274 BuildMI(BB, IA64::FCVTFXUTRUNC, 1, Tmp2).addReg(Tmp1);
1275 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1280 if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
1281 N.getOperand(0).Val->hasOneUse()) { // if we can fold this add
1282 // into an fma, do so:
1283 // ++FusedFP; // Statistic
1284 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1285 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1286 Tmp3 = SelectExpr(N.getOperand(1));
1287 BuildMI(BB, IA64::FMA, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1288 return Result; // early exit
1291 if(DestType != MVT::f64 && N.getOperand(0).getOpcode() == ISD::SHL &&
1292 N.getOperand(0).Val->hasOneUse()) { // if we might be able to fold
1293 // this add into a shladd, try:
1294 ConstantSDNode *CSD = NULL;
1295 if((CSD = dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) &&
1296 (CSD->getValue() >= 1) && (CSD->getValue() <= 4) ) { // we can:
1298 // ++FusedSHLADD; // Statistic
1299 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1300 int shl_amt = CSD->getValue();
1301 Tmp3 = SelectExpr(N.getOperand(1));
1303 BuildMI(BB, IA64::SHLADD, 3, Result)
1304 .addReg(Tmp1).addImm(shl_amt).addReg(Tmp3);
1305 return Result; // early exit
1309 //else, fallthrough:
1310 Tmp1 = SelectExpr(N.getOperand(0));
1311 if(DestType != MVT::f64) { // integer addition:
1312 switch (ponderIntegerAdditionWith(N.getOperand(1), Tmp3)) {
1313 case 1: // adding a constant that's 14 bits
1314 BuildMI(BB, IA64::ADDIMM14, 2, Result).addReg(Tmp1).addSImm(Tmp3);
1315 return Result; // early exit
1316 } // fallthrough and emit a reg+reg ADD:
1317 Tmp2 = SelectExpr(N.getOperand(1));
1318 BuildMI(BB, IA64::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1319 } else { // this is a floating point addition
1320 Tmp2 = SelectExpr(N.getOperand(1));
1321 BuildMI(BB, IA64::FADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1328 if(DestType != MVT::f64) { // TODO: speed!
1329 /* FIXME if(N.getOperand(1).getOpcode() != ISD::Constant) { // if not a const mul
1331 // boring old integer multiply with xma
1332 Tmp1 = SelectExpr(N.getOperand(0));
1333 Tmp2 = SelectExpr(N.getOperand(1));
1335 unsigned TempFR1=MakeReg(MVT::f64);
1336 unsigned TempFR2=MakeReg(MVT::f64);
1337 unsigned TempFR3=MakeReg(MVT::f64);
1338 BuildMI(BB, IA64::SETFSIG, 1, TempFR1).addReg(Tmp1);
1339 BuildMI(BB, IA64::SETFSIG, 1, TempFR2).addReg(Tmp2);
1340 BuildMI(BB, IA64::XMAL, 1, TempFR3).addReg(TempFR1).addReg(TempFR2)
1342 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TempFR3);
1343 return Result; // early exit
1344 /* FIXME } else { // we are multiplying by an integer constant! yay
1345 return Reg = SelectExpr(BuildConstmulSequence(N)); // avert your eyes!
1348 else { // floating point multiply
1349 Tmp1 = SelectExpr(N.getOperand(0));
1350 Tmp2 = SelectExpr(N.getOperand(1));
1351 BuildMI(BB, IA64::FMPY, 2, Result).addReg(Tmp1).addReg(Tmp2);
1357 if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
1358 N.getOperand(0).Val->hasOneUse()) { // if we can fold this sub
1359 // into an fms, do so:
1360 // ++FusedFP; // Statistic
1361 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1362 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1363 Tmp3 = SelectExpr(N.getOperand(1));
1364 BuildMI(BB, IA64::FMS, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1365 return Result; // early exit
1367 Tmp2 = SelectExpr(N.getOperand(1));
1368 if(DestType != MVT::f64) { // integer subtraction:
1369 switch (ponderIntegerSubtractionFrom(N.getOperand(0), Tmp3)) {
1370 case 1: // subtracting *from* an 8 bit constant:
1371 BuildMI(BB, IA64::SUBIMM8, 2, Result).addSImm(Tmp3).addReg(Tmp2);
1372 return Result; // early exit
1373 } // fallthrough and emit a reg+reg SUB:
1374 Tmp1 = SelectExpr(N.getOperand(0));
1375 BuildMI(BB, IA64::SUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1376 } else { // this is a floating point subtraction
1377 Tmp1 = SelectExpr(N.getOperand(0));
1378 BuildMI(BB, IA64::FSUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1384 Tmp1 = SelectExpr(N.getOperand(0));
1385 assert(DestType == MVT::f64 && "trying to fabs something other than f64?");
1386 BuildMI(BB, IA64::FABS, 1, Result).addReg(Tmp1);
1391 assert(DestType == MVT::f64 && "trying to fneg something other than f64?");
1393 if (ISD::FABS == N.getOperand(0).getOpcode()) { // && hasOneUse()?
1394 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1395 BuildMI(BB, IA64::FNEGABS, 1, Result).addReg(Tmp1); // fold in abs
1397 Tmp1 = SelectExpr(N.getOperand(0));
1398 BuildMI(BB, IA64::FNEG, 1, Result).addReg(Tmp1); // plain old fneg
1405 switch (N.getValueType()) {
1406 default: assert(0 && "Cannot AND this type!");
1407 case MVT::i1: { // if a bool, we emit a pseudocode AND
1408 unsigned pA = SelectExpr(N.getOperand(0));
1409 unsigned pB = SelectExpr(N.getOperand(1));
1411 /* our pseudocode for AND is:
1413 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1414 cmp.eq pTemp,p0 = r0,r0 // pTemp = NOT pB
1416 (pB) cmp.ne pTemp,p0 = r0,r0
1418 (pTemp)cmp.ne pC,p0 = r0,r0 // if (NOT pB) pC = 0
1421 unsigned pTemp = MakeReg(MVT::i1);
1423 unsigned bogusTemp1 = MakeReg(MVT::i1);
1424 unsigned bogusTemp2 = MakeReg(MVT::i1);
1425 unsigned bogusTemp3 = MakeReg(MVT::i1);
1426 unsigned bogusTemp4 = MakeReg(MVT::i1);
1428 BuildMI(BB, IA64::PCMPEQUNC, 3, bogusTemp1)
1429 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1430 BuildMI(BB, IA64::CMPEQ, 2, bogusTemp2)
1431 .addReg(IA64::r0).addReg(IA64::r0);
1432 BuildMI(BB, IA64::TPCMPNE, 3, pTemp)
1433 .addReg(bogusTemp2).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1434 BuildMI(BB, IA64::TPCMPNE, 3, Result)
1435 .addReg(bogusTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pTemp);
1439 // if not a bool, we just AND away:
1444 Tmp1 = SelectExpr(N.getOperand(0));
1445 switch (ponderIntegerAndWith(N.getOperand(1), Tmp3)) {
1446 case 1: // ANDing a constant that is 2^n-1 for some n
1448 case 8: // if AND 0x00000000000000FF, be quaint and use zxt1
1449 BuildMI(BB, IA64::ZXT1, 1, Result).addReg(Tmp1);
1451 case 16: // if AND 0x000000000000FFFF, be quaint and use zxt2
1452 BuildMI(BB, IA64::ZXT2, 1, Result).addReg(Tmp1);
1454 case 32: // if AND 0x00000000FFFFFFFF, be quaint and use zxt4
1455 BuildMI(BB, IA64::ZXT4, 1, Result).addReg(Tmp1);
1457 default: // otherwise, use dep.z to paste zeros
1458 // FIXME: assert the dep.z is in bounds
1459 BuildMI(BB, IA64::DEPZ, 3, Result).addReg(Tmp1)
1460 .addImm(0).addImm(Tmp3);
1463 return Result; // early exit
1464 } // fallthrough and emit a simple AND:
1465 Tmp2 = SelectExpr(N.getOperand(1));
1466 BuildMI(BB, IA64::AND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1473 switch (N.getValueType()) {
1474 default: assert(0 && "Cannot OR this type!");
1475 case MVT::i1: { // if a bool, we emit a pseudocode OR
1476 unsigned pA = SelectExpr(N.getOperand(0));
1477 unsigned pB = SelectExpr(N.getOperand(1));
1479 unsigned pTemp1 = MakeReg(MVT::i1);
1481 /* our pseudocode for OR is:
1487 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1489 (pB) cmp.eq pC,p0 = r0,r0 // if (pB) pC = 1
1492 BuildMI(BB, IA64::PCMPEQUNC, 3, pTemp1)
1493 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1494 BuildMI(BB, IA64::TPCMPEQ, 3, Result)
1495 .addReg(pTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1498 // if not a bool, we just OR away:
1503 Tmp1 = SelectExpr(N.getOperand(0));
1504 Tmp2 = SelectExpr(N.getOperand(1));
1505 BuildMI(BB, IA64::OR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1513 switch (N.getValueType()) {
1514 default: assert(0 && "Cannot XOR this type!");
1515 case MVT::i1: { // if a bool, we emit a pseudocode XOR
1516 unsigned pY = SelectExpr(N.getOperand(0));
1517 unsigned pZ = SelectExpr(N.getOperand(1));
1519 /* one possible routine for XOR is:
1521 // Compute px = py ^ pz
1522 // using sum of products: px = (py & !pz) | (pz & !py)
1523 // Uses 5 instructions in 3 cycles.
1525 (pz) cmp.eq.unc px = r0, r0 // px = pz
1526 (py) cmp.eq.unc pt = r0, r0 // pt = py
1529 (pt) cmp.ne.and px = r0, r0 // px = px & !pt (px = pz & !pt)
1530 (pz) cmp.ne.and pt = r0, r0 // pt = pt & !pz
1534 (pt) cmp.eq.or px = r0, r0 // px = px | pt
1536 *** Another, which we use here, requires one scratch GR. it is:
1538 mov rt = 0 // initialize rt off critical path
1542 (pz) cmp.eq.unc px = r0, r0 // px = pz
1543 (pz) mov rt = 1 // rt = pz
1546 (py) cmp.ne px = 1, rt // if (py) px = !pz
1548 .. these routines kindly provided by Jim Hull
1550 unsigned rt = MakeReg(MVT::i64);
1552 // these two temporaries will never actually appear,
1553 // due to the two-address form of some of the instructions below
1554 unsigned bogoPR = MakeReg(MVT::i1); // becomes Result
1555 unsigned bogoGR = MakeReg(MVT::i64); // becomes rt
1557 BuildMI(BB, IA64::MOV, 1, bogoGR).addReg(IA64::r0);
1558 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoPR)
1559 .addReg(IA64::r0).addReg(IA64::r0).addReg(pZ);
1560 BuildMI(BB, IA64::TPCADDIMM22, 2, rt)
1561 .addReg(bogoGR).addImm(1).addReg(pZ);
1562 BuildMI(BB, IA64::TPCMPIMM8NE, 3, Result)
1563 .addReg(bogoPR).addImm(1).addReg(rt).addReg(pY);
1566 // if not a bool, we just XOR away:
1571 Tmp1 = SelectExpr(N.getOperand(0));
1572 Tmp2 = SelectExpr(N.getOperand(1));
1573 BuildMI(BB, IA64::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1581 Tmp1 = SelectExpr(N.getOperand(0));
1582 BuildMI(BB, IA64::POPCNT, 1, Result).addReg(Tmp1);
1587 Tmp1 = SelectExpr(N.getOperand(0));
1588 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1589 Tmp2 = CN->getValue();
1590 BuildMI(BB, IA64::SHLI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1592 Tmp2 = SelectExpr(N.getOperand(1));
1593 BuildMI(BB, IA64::SHL, 2, Result).addReg(Tmp1).addReg(Tmp2);
1599 Tmp1 = SelectExpr(N.getOperand(0));
1600 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1601 Tmp2 = CN->getValue();
1602 BuildMI(BB, IA64::SHRUI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1604 Tmp2 = SelectExpr(N.getOperand(1));
1605 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1611 Tmp1 = SelectExpr(N.getOperand(0));
1612 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1613 Tmp2 = CN->getValue();
1614 BuildMI(BB, IA64::SHRSI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1616 Tmp2 = SelectExpr(N.getOperand(1));
1617 BuildMI(BB, IA64::SHRS, 2, Result).addReg(Tmp1).addReg(Tmp2);
1627 Tmp1 = SelectExpr(N.getOperand(0));
1628 Tmp2 = SelectExpr(N.getOperand(1));
1632 if(DestType == MVT::f64) // XXX: we're not gonna be fed MVT::f32, are we?
1635 bool isModulus=false; // is it a division or a modulus?
1636 bool isSigned=false;
1638 switch(N.getOpcode()) {
1639 case ISD::SDIV: isModulus=false; isSigned=true; break;
1640 case ISD::UDIV: isModulus=false; isSigned=false; break;
1641 case ISD::SREM: isModulus=true; isSigned=true; break;
1642 case ISD::UREM: isModulus=true; isSigned=false; break;
1645 if(!isModulus && !isFP) { // if this is an integer divide,
1646 switch (ponderIntegerDivisionBy(N.getOperand(1), isSigned, Tmp3)) {
1647 case 1: // division by a constant that's a power of 2
1648 Tmp1 = SelectExpr(N.getOperand(0));
1649 if(isSigned) { // argument could be negative, so emit some code:
1650 unsigned divAmt=Tmp3;
1651 unsigned tempGR1=MakeReg(MVT::i64);
1652 unsigned tempGR2=MakeReg(MVT::i64);
1653 unsigned tempGR3=MakeReg(MVT::i64);
1654 BuildMI(BB, IA64::SHRS, 2, tempGR1)
1655 .addReg(Tmp1).addImm(divAmt-1);
1656 BuildMI(BB, IA64::EXTRU, 3, tempGR2)
1657 .addReg(tempGR1).addImm(64-divAmt).addImm(divAmt);
1658 BuildMI(BB, IA64::ADD, 2, tempGR3)
1659 .addReg(Tmp1).addReg(tempGR2);
1660 BuildMI(BB, IA64::SHRS, 2, Result)
1661 .addReg(tempGR3).addImm(divAmt);
1663 else // unsigned div-by-power-of-2 becomes a simple shift right:
1664 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addImm(Tmp3);
1665 return Result; // early exit
1669 unsigned TmpPR=MakeReg(MVT::i1); // we need two scratch
1670 unsigned TmpPR2=MakeReg(MVT::i1); // predicate registers,
1671 unsigned TmpF1=MakeReg(MVT::f64); // and one metric truckload of FP regs.
1672 unsigned TmpF2=MakeReg(MVT::f64); // lucky we have IA64?
1673 unsigned TmpF3=MakeReg(MVT::f64); // well, the real FIXME is to have
1674 unsigned TmpF4=MakeReg(MVT::f64); // isTwoAddress forms of these
1675 unsigned TmpF5=MakeReg(MVT::f64); // FP instructions so we can end up with
1676 unsigned TmpF6=MakeReg(MVT::f64); // stuff like setf.sig f10=f10 etc.
1677 unsigned TmpF7=MakeReg(MVT::f64);
1678 unsigned TmpF8=MakeReg(MVT::f64);
1679 unsigned TmpF9=MakeReg(MVT::f64);
1680 unsigned TmpF10=MakeReg(MVT::f64);
1681 unsigned TmpF11=MakeReg(MVT::f64);
1682 unsigned TmpF12=MakeReg(MVT::f64);
1683 unsigned TmpF13=MakeReg(MVT::f64);
1684 unsigned TmpF14=MakeReg(MVT::f64);
1685 unsigned TmpF15=MakeReg(MVT::f64);
1687 // OK, emit some code:
1690 // first, load the inputs into FP regs.
1691 BuildMI(BB, IA64::SETFSIG, 1, TmpF1).addReg(Tmp1);
1692 BuildMI(BB, IA64::SETFSIG, 1, TmpF2).addReg(Tmp2);
1694 // next, convert the inputs to FP
1696 BuildMI(BB, IA64::FCVTXF, 1, TmpF3).addReg(TmpF1);
1697 BuildMI(BB, IA64::FCVTXF, 1, TmpF4).addReg(TmpF2);
1699 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF3).addReg(TmpF1);
1700 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF4).addReg(TmpF2);
1703 } else { // this is an FP divide/remainder, so we 'leak' some temp
1704 // regs and assign TmpF3=Tmp1, TmpF4=Tmp2
1709 // we start by computing an approximate reciprocal (good to 9 bits?)
1710 // note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
1711 BuildMI(BB, IA64::FRCPAS1, 4)
1712 .addReg(TmpF5, MachineOperand::Def)
1713 .addReg(TmpPR, MachineOperand::Def)
1714 .addReg(TmpF3).addReg(TmpF4);
1716 if(!isModulus) { // if this is a divide, we worry about div-by-zero
1717 unsigned bogusPR=MakeReg(MVT::i1); // won't appear, due to twoAddress
1719 BuildMI(BB, IA64::CMPEQ, 2, bogusPR).addReg(IA64::r0).addReg(IA64::r0);
1720 BuildMI(BB, IA64::TPCMPNE, 3, TmpPR2).addReg(bogusPR)
1721 .addReg(IA64::r0).addReg(IA64::r0).addReg(TmpPR);
1724 // now we apply newton's method, thrice! (FIXME: this is ~72 bits of
1725 // precision, don't need this much for f32/i32)
1726 BuildMI(BB, IA64::CFNMAS1, 4, TmpF6)
1727 .addReg(TmpF4).addReg(TmpF5).addReg(IA64::F1).addReg(TmpPR);
1728 BuildMI(BB, IA64::CFMAS1, 4, TmpF7)
1729 .addReg(TmpF3).addReg(TmpF5).addReg(IA64::F0).addReg(TmpPR);
1730 BuildMI(BB, IA64::CFMAS1, 4, TmpF8)
1731 .addReg(TmpF6).addReg(TmpF6).addReg(IA64::F0).addReg(TmpPR);
1732 BuildMI(BB, IA64::CFMAS1, 4, TmpF9)
1733 .addReg(TmpF6).addReg(TmpF7).addReg(TmpF7).addReg(TmpPR);
1734 BuildMI(BB, IA64::CFMAS1, 4,TmpF10)
1735 .addReg(TmpF6).addReg(TmpF5).addReg(TmpF5).addReg(TmpPR);
1736 BuildMI(BB, IA64::CFMAS1, 4,TmpF11)
1737 .addReg(TmpF8).addReg(TmpF9).addReg(TmpF9).addReg(TmpPR);
1738 BuildMI(BB, IA64::CFMAS1, 4,TmpF12)
1739 .addReg(TmpF8).addReg(TmpF10).addReg(TmpF10).addReg(TmpPR);
1740 BuildMI(BB, IA64::CFNMAS1, 4,TmpF13)
1741 .addReg(TmpF4).addReg(TmpF11).addReg(TmpF3).addReg(TmpPR);
1743 // FIXME: this is unfortunate :(
1744 // the story is that the dest reg of the fnma above and the fma below
1745 // (and therefore possibly the src of the fcvt.fx[u] as well) cannot
1746 // be the same register, or this code breaks if the first argument is
1747 // zero. (e.g. without this hack, 0%8 yields -64, not 0.)
1748 BuildMI(BB, IA64::CFMAS1, 4,TmpF14)
1749 .addReg(TmpF13).addReg(TmpF12).addReg(TmpF11).addReg(TmpPR);
1751 if(isModulus) { // XXX: fragile! fixes _only_ mod, *breaks* div! !
1752 BuildMI(BB, IA64::IUSE, 1).addReg(TmpF13); // hack :(
1756 // round to an integer
1758 BuildMI(BB, IA64::FCVTFXTRUNCS1, 1, TmpF15).addReg(TmpF14);
1760 BuildMI(BB, IA64::FCVTFXUTRUNCS1, 1, TmpF15).addReg(TmpF14);
1762 BuildMI(BB, IA64::FMOV, 1, TmpF15).addReg(TmpF14);
1763 // EXERCISE: can you see why TmpF15=TmpF14 does not work here, and
1764 // we really do need the above FMOV? ;)
1768 if(isFP) { // extra worrying about div-by-zero
1769 unsigned bogoResult=MakeReg(MVT::f64);
1771 // we do a 'conditional fmov' (of the correct result, depending
1772 // on how the frcpa predicate turned out)
1773 BuildMI(BB, IA64::PFMOV, 2, bogoResult)
1774 .addReg(TmpF12).addReg(TmpPR2);
1775 BuildMI(BB, IA64::CFMOV, 2, Result)
1776 .addReg(bogoResult).addReg(TmpF15).addReg(TmpPR);
1779 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TmpF15);
1781 } else { // this is a modulus
1783 // answer = q * (-b) + a
1784 unsigned ModulusResult = MakeReg(MVT::f64);
1785 unsigned TmpF = MakeReg(MVT::f64);
1786 unsigned TmpI = MakeReg(MVT::i64);
1788 BuildMI(BB, IA64::SUB, 2, TmpI).addReg(IA64::r0).addReg(Tmp2);
1789 BuildMI(BB, IA64::SETFSIG, 1, TmpF).addReg(TmpI);
1790 BuildMI(BB, IA64::XMAL, 3, ModulusResult)
1791 .addReg(TmpF15).addReg(TmpF).addReg(TmpF1);
1792 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(ModulusResult);
1793 } else { // FP modulus! The horror... the horror....
1794 assert(0 && "sorry, no FP modulus just yet!\n!\n");
1801 case ISD::SIGN_EXTEND_INREG: {
1802 Tmp1 = SelectExpr(N.getOperand(0));
1803 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1806 assert(0 && "don't know how to sign extend this type");
1808 case MVT::i8: Opc = IA64::SXT1; break;
1809 case MVT::i16: Opc = IA64::SXT2; break;
1810 case MVT::i32: Opc = IA64::SXT4; break;
1812 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1817 Tmp1 = SelectExpr(N.getOperand(0));
1818 ISD::CondCode CC = cast<CondCodeSDNode>(Node->getOperand(2))->get();
1819 if (MVT::isInteger(N.getOperand(0).getValueType())) {
1821 if(ConstantSDNode *CSDN =
1822 dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1823 // if we are comparing against a constant zero
1824 if(CSDN->getValue()==0)
1825 Tmp2 = IA64::r0; // then we can just compare against r0
1827 Tmp2 = SelectExpr(N.getOperand(1));
1828 } else // not comparing against a constant
1829 Tmp2 = SelectExpr(N.getOperand(1));
1832 default: assert(0 && "Unknown integer comparison!");
1834 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1837 BuildMI(BB, IA64::CMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1840 BuildMI(BB, IA64::CMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1843 BuildMI(BB, IA64::CMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1846 BuildMI(BB, IA64::CMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1849 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1852 BuildMI(BB, IA64::CMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1855 BuildMI(BB, IA64::CMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1858 BuildMI(BB, IA64::CMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1861 BuildMI(BB, IA64::CMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1864 } else { // if not integer, should be FP.
1865 assert(N.getOperand(0).getValueType() != MVT::f32 &&
1866 "error: SETCC should have had incoming f32 promoted to f64!\n");
1868 if(ConstantFPSDNode *CFPSDN =
1869 dyn_cast<ConstantFPSDNode>(N.getOperand(1))) {
1871 // if we are comparing against a constant +0.0 or +1.0
1872 if(CFPSDN->isExactlyValue(+0.0))
1873 Tmp2 = IA64::F0; // then we can just compare against f0
1874 else if(CFPSDN->isExactlyValue(+1.0))
1875 Tmp2 = IA64::F1; // or f1
1877 Tmp2 = SelectExpr(N.getOperand(1));
1878 } else // not comparing against a constant
1879 Tmp2 = SelectExpr(N.getOperand(1));
1882 default: assert(0 && "Unknown FP comparison!");
1884 BuildMI(BB, IA64::FCMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1887 BuildMI(BB, IA64::FCMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1890 BuildMI(BB, IA64::FCMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1893 BuildMI(BB, IA64::FCMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1896 BuildMI(BB, IA64::FCMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1899 BuildMI(BB, IA64::FCMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1902 BuildMI(BB, IA64::FCMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1905 BuildMI(BB, IA64::FCMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1908 BuildMI(BB, IA64::FCMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1911 BuildMI(BB, IA64::FCMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
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 = cast<RegisterSDNode>(Node->getOperand(1))->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(2));
2217 Tmp2 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
2220 // if a bool, we use pseudocode
2221 if (N.getOperand(2).getValueType() == MVT::i1)
2222 BuildMI(BB, IA64::PCMPEQUNC, 3, Tmp2)
2223 .addReg(IA64::r0).addReg(IA64::r0).addReg(Tmp1);
2224 // (Tmp1) Tmp2 = cmp.eq.unc(r0,r0)
2226 BuildMI(BB, IA64::MOV, 1, Tmp2).addReg(Tmp1);
2227 // XXX is this the right way 'round? ;)
2228 // FIXME: WHAT ABOUT FLOATING POINT?
2235 /* what the heck is going on here:
2237 <_sabre_> ret with two operands is obvious: chain and value
2239 <_sabre_> ret with 3 values happens when 'expansion' occurs
2240 <_sabre_> e.g. i64 gets split into 2x i32
2242 <_sabre_> you don't have this case on ia64
2244 <_sabre_> so the two returned values go into EAX/EDX on ia32
2245 <camel_> ahhh *memories*
2247 <camel_> ok, thanks :)
2248 <_sabre_> so yeah, everything that has a side effect takes a 'token chain'
2249 <_sabre_> this is the first operand always
2250 <_sabre_> these operand often define chains, they are the last operand
2251 <_sabre_> they are printed as 'ch' if you do DAG.dump()
2254 switch (N.getNumOperands()) {
2256 assert(0 && "Unknown return instruction!");
2258 Select(N.getOperand(0));
2259 Tmp1 = SelectExpr(N.getOperand(1));
2260 switch (N.getOperand(1).getValueType()) {
2261 default: assert(0 && "All other types should have been promoted!!");
2262 // FIXME: do I need to add support for bools here?
2263 // (return '0' or '1' r8, basically...)
2265 // FIXME: need to round floats - 80 bits is bad, the tester
2268 // we mark r8 as live on exit up above in LowerArguments()
2269 BuildMI(BB, IA64::MOV, 1, IA64::r8).addReg(Tmp1);
2272 // we mark F8 as live on exit up above in LowerArguments()
2273 BuildMI(BB, IA64::FMOV, 1, IA64::F8).addReg(Tmp1);
2277 Select(N.getOperand(0));
2280 // before returning, restore the ar.pfs register (set by the 'alloc' up top)
2281 BuildMI(BB, IA64::MOV, 1).addReg(IA64::AR_PFS).addReg(IA64Lowering.VirtGPR);
2282 BuildMI(BB, IA64::RET, 0); // and then just emit a 'ret' instruction
2287 Select(N.getOperand(0));
2288 MachineBasicBlock *Dest =
2289 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2290 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(IA64::p0).addMBB(Dest);
2291 // XXX HACK! we do _not_ need long branches all the time
2295 case ISD::ImplicitDef: {
2296 Select(N.getOperand(0));
2297 BuildMI(BB, IA64::IDEF, 0,
2298 cast<RegisterSDNode>(N.getOperand(1))->getReg());
2303 MachineBasicBlock *Dest =
2304 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
2306 Select(N.getOperand(0));
2307 Tmp1 = SelectExpr(N.getOperand(1));
2308 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(Tmp1).addMBB(Dest);
2309 // XXX HACK! we do _not_ need long branches all the time
2319 case ISD::CopyFromReg:
2320 case ISD::DYNAMIC_STACKALLOC:
2324 case ISD::TRUNCSTORE:
2326 Select(N.getOperand(0));
2327 Tmp1 = SelectExpr(N.getOperand(1)); // value
2331 if(opcode == ISD::STORE) {
2332 switch (N.getOperand(1).getValueType()) {
2333 default: assert(0 && "Cannot store this type!");
2334 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2335 // FIXME?: for now, we treat bool loads the same as i8 stores */
2336 case MVT::i8: Opc = IA64::ST1; break;
2337 case MVT::i16: Opc = IA64::ST2; break;
2338 case MVT::i32: Opc = IA64::ST4; break;
2339 case MVT::i64: Opc = IA64::ST8; break;
2341 case MVT::f32: Opc = IA64::STF4; break;
2342 case MVT::f64: Opc = IA64::STF8; break;
2344 } else { // truncstore
2345 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2346 default: assert(0 && "unknown type in truncstore");
2347 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2348 //FIXME: DAG does not promote this load?
2349 case MVT::i8: Opc = IA64::ST1; break;
2350 case MVT::i16: Opc = IA64::ST2; break;
2351 case MVT::i32: Opc = IA64::ST4; break;
2352 case MVT::f32: Opc = IA64::STF4; break;
2356 if(N.getOperand(2).getOpcode() == ISD::GlobalAddress) {
2357 unsigned dummy = MakeReg(MVT::i64);
2358 unsigned dummy2 = MakeReg(MVT::i64);
2359 BuildMI(BB, IA64::ADD, 2, dummy)
2360 .addGlobalAddress(cast<GlobalAddressSDNode>
2361 (N.getOperand(2))->getGlobal()).addReg(IA64::r1);
2362 BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
2365 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(Tmp1);
2366 else { // we are storing a bool, so emit a little pseudocode
2367 // to store a predicate register as one byte
2368 assert(Opc==IA64::ST1);
2369 unsigned dummy3 = MakeReg(MVT::i64);
2370 unsigned dummy4 = MakeReg(MVT::i64);
2371 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2372 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2373 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2374 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(dummy4);
2376 } else if(N.getOperand(2).getOpcode() == ISD::FrameIndex) {
2378 // FIXME? (what about bools?)
2380 unsigned dummy = MakeReg(MVT::i64);
2381 BuildMI(BB, IA64::MOV, 1, dummy)
2382 .addFrameIndex(cast<FrameIndexSDNode>(N.getOperand(2))->getIndex());
2383 BuildMI(BB, Opc, 2).addReg(dummy).addReg(Tmp1);
2384 } else { // otherwise
2385 Tmp2 = SelectExpr(N.getOperand(2)); //address
2387 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(Tmp1);
2388 else { // we are storing a bool, so emit a little pseudocode
2389 // to store a predicate register as one byte
2390 assert(Opc==IA64::ST1);
2391 unsigned dummy3 = MakeReg(MVT::i64);
2392 unsigned dummy4 = MakeReg(MVT::i64);
2393 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2394 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2395 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2396 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(dummy4);
2402 case ISD::CALLSEQ_START:
2403 case ISD::CALLSEQ_END: {
2404 Select(N.getOperand(0));
2405 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2407 Opc = N.getOpcode() == ISD::CALLSEQ_START ? IA64::ADJUSTCALLSTACKDOWN :
2408 IA64::ADJUSTCALLSTACKUP;
2409 BuildMI(BB, Opc, 1).addImm(Tmp1);
2415 assert(0 && "GAME OVER. INSERT COIN?");
2419 /// createIA64PatternInstructionSelector - This pass converts an LLVM function
2420 /// into a machine code representation using pattern matching and a machine
2421 /// description file.
2423 FunctionPass *llvm::createIA64PatternInstructionSelector(TargetMachine &TM) {
2424 return new ISel(TM);