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::FREM , MVT::f32 , Expand);
76 setOperationAction(ISD::FREM , 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 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
86 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
88 // We don't support sin/cos/sqrt
89 setOperationAction(ISD::FSIN , MVT::f64, Expand);
90 setOperationAction(ISD::FCOS , MVT::f64, Expand);
91 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
92 setOperationAction(ISD::FSIN , MVT::f32, Expand);
93 setOperationAction(ISD::FCOS , MVT::f32, Expand);
94 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
96 //IA64 has these, but they are not implemented
97 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
98 setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
99 setOperationAction(ISD::ROTL , MVT::i64 , Expand);
100 setOperationAction(ISD::ROTR , MVT::i64 , Expand);
101 // FIXME: implement mulhs (xma.h) and mulhu (xma.hu)
102 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
103 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
105 // We don't have line number support yet.
106 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
107 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
108 setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
110 // Not implemented yet.
111 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
112 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
114 computeRegisterProperties();
116 addLegalFPImmediate(+0.0);
117 addLegalFPImmediate(+1.0);
118 addLegalFPImmediate(-0.0);
119 addLegalFPImmediate(-1.0);
122 /// LowerArguments - This hook must be implemented to indicate how we should
123 /// lower the arguments for the specified function, into the specified DAG.
124 virtual std::vector<SDOperand>
125 LowerArguments(Function &F, SelectionDAG &DAG);
127 /// LowerCallTo - This hook lowers an abstract call to a function into an
129 virtual std::pair<SDOperand, SDOperand>
130 LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
131 bool isTailCall, SDOperand Callee, ArgListTy &Args,
134 virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
135 Value *VAListV, SelectionDAG &DAG);
136 virtual std::pair<SDOperand,SDOperand>
137 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
138 const Type *ArgTy, SelectionDAG &DAG);
140 void restoreGP_SP_RP(MachineBasicBlock* BB)
142 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
143 BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
144 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
147 void restoreSP_RP(MachineBasicBlock* BB)
149 BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
150 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
153 void restoreRP(MachineBasicBlock* BB)
155 BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
158 void restoreGP(MachineBasicBlock* BB)
160 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
167 std::vector<SDOperand>
168 IA64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
169 std::vector<SDOperand> ArgValues;
172 // add beautiful description of IA64 stack frame format
173 // here (from intel 24535803.pdf most likely)
175 MachineFunction &MF = DAG.getMachineFunction();
176 MachineFrameInfo *MFI = MF.getFrameInfo();
178 GP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
179 SP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
180 RP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
182 MachineBasicBlock& BB = MF.front();
184 unsigned args_int[] = {IA64::r32, IA64::r33, IA64::r34, IA64::r35,
185 IA64::r36, IA64::r37, IA64::r38, IA64::r39};
187 unsigned args_FP[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
188 IA64::F12,IA64::F13,IA64::F14, IA64::F15};
194 unsigned used_FPArgs = 0; // how many FP args have been used so far?
196 unsigned ArgOffset = 0;
199 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
201 SDOperand newroot, argt;
202 if(count < 8) { // need to fix this logic? maybe.
204 switch (getValueType(I->getType())) {
206 std::cerr << "ERROR in LowerArgs: unknown type "
207 << getValueType(I->getType()) << "\n";
210 // fixme? (well, will need to for weird FP structy stuff,
211 // see intel ABI docs)
213 //XXX BuildMI(&BB, IA64::IDEF, 0, args_FP[used_FPArgs]);
214 MF.addLiveIn(args_FP[used_FPArgs]); // mark this reg as liveIn
215 // floating point args go into f8..f15 as-needed, the increment
216 argVreg[count] = // is below..:
217 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::f64));
218 // FP args go into f8..f15 as needed: (hence the ++)
219 argPreg[count] = args_FP[used_FPArgs++];
220 argOpc[count] = IA64::FMOV;
221 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), argVreg[count],
223 if (I->getType() == Type::FloatTy)
224 argt = DAG.getNode(ISD::FP_ROUND, MVT::f32, argt);
226 case MVT::i1: // NOTE: as far as C abi stuff goes,
227 // bools are just boring old ints
232 //XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
233 MF.addLiveIn(args_int[count]); // mark this register as liveIn
235 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
236 argPreg[count] = args_int[count];
237 argOpc[count] = IA64::MOV;
239 DAG.getCopyFromReg(DAG.getRoot(), argVreg[count], MVT::i64);
240 if ( getValueType(I->getType()) != MVT::i64)
241 argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
245 } else { // more than 8 args go into the frame
246 // Create the frame index object for this incoming parameter...
247 ArgOffset = 16 + 8 * (count - 8);
248 int FI = MFI->CreateFixedObject(8, ArgOffset);
250 // Create the SelectionDAG nodes corresponding to a load
251 //from this parameter
252 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
253 argt = newroot = DAG.getLoad(getValueType(I->getType()),
254 DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
257 DAG.setRoot(newroot.getValue(1));
258 ArgValues.push_back(argt);
262 // Create a vreg to hold the output of (what will become)
263 // the "alloc" instruction
264 VirtGPR = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
265 BuildMI(&BB, IA64::PSEUDO_ALLOC, 0, VirtGPR);
266 // we create a PSEUDO_ALLOC (pseudo)instruction for now
268 BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
271 BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
272 BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
275 BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
278 BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
279 BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
282 unsigned tempOffset=0;
284 // if this is a varargs function, we simply lower llvm.va_start by
285 // pointing to the first entry
288 VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
291 // here we actually do the moving of args, and store them to the stack
292 // too if this is a varargs function:
293 for (int i = 0; i < count && i < 8; ++i) {
294 BuildMI(&BB, argOpc[i], 1, argVreg[i]).addReg(argPreg[i]);
296 // if this is a varargs function, we copy the input registers to the stack
297 int FI = MFI->CreateFixedObject(8, tempOffset);
298 tempOffset+=8; //XXX: is it safe to use r22 like this?
299 BuildMI(&BB, IA64::MOV, 1, IA64::r22).addFrameIndex(FI);
300 // FIXME: we should use st8.spill here, one day
301 BuildMI(&BB, IA64::ST8, 1, IA64::r22).addReg(argPreg[i]);
305 // Finally, inform the code generator which regs we return values in.
306 // (see the ISD::RET: case down below)
307 switch (getValueType(F.getReturnType())) {
308 default: assert(0 && "i have no idea where to return this type!");
309 case MVT::isVoid: break;
315 MF.addLiveOut(IA64::r8);
319 MF.addLiveOut(IA64::F8);
326 std::pair<SDOperand, SDOperand>
327 IA64TargetLowering::LowerCallTo(SDOperand Chain,
328 const Type *RetTy, bool isVarArg,
329 unsigned CallingConv, bool isTailCall,
330 SDOperand Callee, ArgListTy &Args,
333 MachineFunction &MF = DAG.getMachineFunction();
335 unsigned NumBytes = 16;
336 unsigned outRegsUsed = 0;
338 if (Args.size() > 8) {
339 NumBytes += (Args.size() - 8) * 8;
342 outRegsUsed = Args.size();
345 // FIXME? this WILL fail if we ever try to pass around an arg that
346 // consumes more than a single output slot (a 'real' double, int128
347 // some sort of aggregate etc.), as we'll underestimate how many 'outX'
348 // registers we use. Hopefully, the assembler will notice.
349 MF.getInfo<IA64FunctionInfo>()->outRegsUsed=
350 std::max(outRegsUsed, MF.getInfo<IA64FunctionInfo>()->outRegsUsed);
352 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
353 DAG.getConstant(NumBytes, getPointerTy()));
355 std::vector<SDOperand> args_to_use;
356 for (unsigned i = 0, e = Args.size(); i != e; ++i)
358 switch (getValueType(Args[i].second)) {
359 default: assert(0 && "unexpected argument type!");
364 //promote to 64-bits, sign/zero extending based on type
366 if(Args[i].second->isSigned())
367 Args[i].first = DAG.getNode(ISD::SIGN_EXTEND, MVT::i64,
370 Args[i].first = DAG.getNode(ISD::ZERO_EXTEND, MVT::i64,
375 Args[i].first = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Args[i].first);
380 args_to_use.push_back(Args[i].first);
383 std::vector<MVT::ValueType> RetVals;
384 MVT::ValueType RetTyVT = getValueType(RetTy);
385 if (RetTyVT != MVT::isVoid)
386 RetVals.push_back(RetTyVT);
387 RetVals.push_back(MVT::Other);
389 SDOperand TheCall = SDOperand(DAG.getCall(RetVals, Chain,
390 Callee, args_to_use), 0);
391 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
392 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
393 DAG.getConstant(NumBytes, getPointerTy()));
394 return std::make_pair(TheCall, Chain);
398 IA64TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
399 Value *VAListV, SelectionDAG &DAG) {
400 // vastart just stores the address of the VarArgsFrameIndex slot.
401 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
402 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR,
403 VAListP, DAG.getSrcValue(VAListV));
406 std::pair<SDOperand,SDOperand> IA64TargetLowering::
407 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
408 const Type *ArgTy, SelectionDAG &DAG) {
410 MVT::ValueType ArgVT = getValueType(ArgTy);
411 SDOperand Val = DAG.getLoad(MVT::i64, Chain,
412 VAListP, DAG.getSrcValue(VAListV));
413 SDOperand Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), Val,
414 DAG.getSrcValue(NULL));
416 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
419 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
420 "Other types should have been promoted for varargs!");
423 Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
424 DAG.getConstant(Amt, Val.getValueType()));
425 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
426 Val, VAListP, DAG.getSrcValue(VAListV));
427 return std::make_pair(Result, Chain);
432 //===--------------------------------------------------------------------===//
433 /// ISel - IA64 specific code to select IA64 machine instructions for
434 /// SelectionDAG operations.
436 class ISel : public SelectionDAGISel {
437 /// IA64Lowering - This object fully describes how to lower LLVM code to an
438 /// IA64-specific SelectionDAG.
439 IA64TargetLowering IA64Lowering;
440 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
441 // for sdiv and udiv until it is put into the future
444 /// ExprMap - As shared expressions are codegen'd, we keep track of which
445 /// vreg the value is produced in, so we only emit one copy of each compiled
447 std::map<SDOperand, unsigned> ExprMap;
448 std::set<SDOperand> LoweredTokens;
451 ISel(TargetMachine &TM) : SelectionDAGISel(IA64Lowering), IA64Lowering(TM),
454 /// InstructionSelectBasicBlock - This callback is invoked by
455 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
456 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
458 unsigned SelectExpr(SDOperand N);
459 void Select(SDOperand N);
460 // a dag->dag to transform mul-by-constant-int to shifts+adds/subs
461 SDOperand BuildConstmulSequence(SDOperand N);
463 const char *getPassName() const { return "IA64 Instruction Selector"; }
467 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
468 /// when it has created a SelectionDAG for us to codegen.
469 void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
471 // Codegen the basic block.
473 Select(DAG.getRoot());
475 // Clear state used for selection.
477 LoweredTokens.clear();
481 // strip leading '0' characters from a string
482 void munchLeadingZeros(std::string& inString) {
483 while(inString.c_str()[0]=='0') {
484 inString.erase(0, 1);
488 // strip trailing '0' characters from a string
489 void munchTrailingZeros(std::string& inString) {
490 int curPos=inString.length()-1;
492 while(inString.c_str()[curPos]=='0') {
493 inString.erase(curPos, 1);
498 // return how many consecutive '0' characters are at the end of a string
499 unsigned int countTrailingZeros(std::string& inString) {
500 int curPos=inString.length()-1;
501 unsigned int zeroCount=0;
503 while(inString.c_str()[curPos--]=='0') {
509 // booth encode a string of '1' and '0' characters (returns string of 'P' (+1)
510 // '0' and 'N' (-1) characters)
511 void boothEncode(std::string inString, std::string& boothEncodedString) {
515 int lim=inString.size();
518 if(inString[curpos]=='1') { // if we see a '1', look for a run of them
520 std::string replaceString="N";
522 // find the run length
523 for(;inString[curpos+runlength]=='1';runlength++) ;
525 for(int i=0; i<runlength-1; i++)
530 inString.replace(curpos, runlength+1, replaceString);
534 } else { // a zero, we just keep chugging along
539 // clean up (trim the string, reverse it and turn '1's into 'P's)
540 munchTrailingZeros(inString);
541 boothEncodedString="";
543 for(int i=inString.size()-1;i>=0;i--)
545 boothEncodedString+="P";
547 boothEncodedString+=inString[i];
551 struct shiftaddblob { // this encodes stuff like (x=) "A << B [+-] C << D"
552 unsigned firstVal; // A
553 unsigned firstShift; // B
554 unsigned secondVal; // C
555 unsigned secondShift; // D
559 /* this implements Lefevre's "pattern-based" constant multiplication,
560 * see "Multiplication by an Integer Constant", INRIA report 1999-06
562 * TODO: implement a method to try rewriting P0N<->0PP / N0P<->0NN
563 * to get better booth encodings - this does help in practice
564 * TODO: weight shifts appropriately (most architectures can't
565 * fuse a shift and an add for arbitrary shift amounts) */
566 unsigned lefevre(const std::string inString,
567 std::vector<struct shiftaddblob> &ops) {
568 std::string retstring;
569 std::string s = inString;
570 munchTrailingZeros(s);
572 int length=s.length()-1;
578 std::vector<int> p,n;
580 for(int i=0; i<=length; i++) {
581 if (s.c_str()[length-i]=='P') {
583 } else if (s.c_str()[length-i]=='N') {
591 std::map<const int, int> w;
593 for(unsigned i=0; i<p.size(); i++) {
594 for(unsigned j=0; j<i; j++) {
599 for(unsigned i=1; i<n.size(); i++) {
600 for(unsigned j=0; j<i; j++) {
605 for(unsigned i=0; i<p.size(); i++) {
606 for(unsigned j=0; j<n.size(); j++) {
607 w[-abs(p[i]-n[j])]++;
611 std::map<const int, int>::const_iterator ii;
613 std::multimap<int, int> sorted_by_value;
615 for(ii = w.begin(); ii!=w.end(); ii++)
616 sorted_by_value.insert(std::pair<int, int>((*ii).second,(*ii).first));
618 for (std::multimap<int, int>::iterator it = sorted_by_value.begin();
619 it != sorted_by_value.end(); ++it) {
620 d.push_back((*it).second);
626 while(d.size()>0 && (w[int_d=d.back()] > int_W)) {
634 for(unsigned base=0; base<retstring.size(); base++) {
635 if( ((base+z+1) < retstring.size()) &&
636 retstring.c_str()[base]=='P' &&
637 retstring.c_str()[base+z+1]=='P')
641 retstring.replace(base, 1, "0");
642 retstring.replace(base+z+1, 1, "p");
646 for(unsigned base=0; base<retstring.size(); base++) {
647 if( ((base+z+1) < retstring.size()) &&
648 retstring.c_str()[base]=='N' &&
649 retstring.c_str()[base+z+1]=='N')
653 retstring.replace(base, 1, "0");
654 retstring.replace(base+z+1, 1, "n");
659 for(unsigned base=0; base<retstring.size(); base++) {
660 if( ((base+z+1) < retstring.size()) &&
661 ((retstring.c_str()[base]=='P' &&
662 retstring.c_str()[base+z+1]=='N') ||
663 (retstring.c_str()[base]=='N' &&
664 retstring.c_str()[base+z+1]=='P')) ) {
668 if(retstring.c_str()[base]=='P') {
669 retstring.replace(base, 1, "0");
670 retstring.replace(base+z+1, 1, "p");
671 } else { // retstring[base]=='N'
672 retstring.replace(base, 1, "0");
673 retstring.replace(base+z+1, 1, "n");
685 } d.pop_back(); // hmm
689 for(unsigned i=0; i<t.length(); i++) {
690 if(t.c_str()[i]=='p' || t.c_str()[i]=='n')
691 t.replace(i, 1, "0");
694 for(unsigned i=0; i<u.length(); i++) {
695 if(u[i]=='P' || u[i]=='N')
696 u.replace(i, 1, "0");
698 u.replace(i, 1, "P");
700 u.replace(i, 1, "N");
713 bool hit=(u[pos]=='N');
718 for(unsigned p=0; p<u.length(); p++) {
719 bool isP=(u[p]=='P');
720 bool isN=(u[p]=='N');
723 u.replace(p, 1, "N");
725 u.replace(p, 1, "P");
729 munchLeadingZeros(u);
731 int i = lefevre(u, ops);
735 blob.firstVal=i; blob.firstShift=c;
737 blob.secondVal=i; blob.secondShift=0;
743 munchLeadingZeros(t);
748 if(t.c_str()[0]!='P') {
750 for(unsigned p=0; p<t.length(); p++) {
751 bool isP=(t.c_str()[p]=='P');
752 bool isN=(t.c_str()[p]=='N');
755 t.replace(p, 1, "N");
757 t.replace(p, 1, "P");
761 int j = lefevre(t, ops);
763 int trail=countTrailingZeros(u);
764 blob.secondVal=i; blob.secondShift=trail;
766 trail=countTrailingZeros(t);
767 blob.firstVal=j; blob.firstShift=trail;
771 blob.isSub=false; // first + second
774 blob.isSub=true; // first - second
777 blob.isSub=true; // second - first
778 int tmpval, tmpshift;
779 tmpval=blob.firstVal;
780 tmpshift=blob.firstShift;
781 blob.firstVal=blob.secondVal;
782 blob.firstShift=blob.secondShift;
783 blob.secondVal=tmpval;
784 blob.secondShift=tmpshift;
793 SDOperand ISel::BuildConstmulSequence(SDOperand N) {
794 //FIXME: we should shortcut this stuff for multiplies by 2^n+1
795 // in particular, *3 is nicer as *2+1, not *4-1
796 int64_t constant=cast<ConstantSDNode>(N.getOperand(1))->getValue();
799 unsigned preliminaryShift=0;
801 assert(constant != 0 && "erk, you're trying to multiply by constant zero\n");
803 // first, we make the constant to multiply by positive
811 // next, we make it odd.
812 for(; (constant%2==0); preliminaryShift++)
815 //OK, we have a positive, odd number of 64 bits or less. Convert it
816 //to a binary string, constantString[0] is the LSB
817 char constantString[65];
818 for(int i=0; i<64; i++)
819 constantString[i]='0'+((constant>>i)&0x1);
820 constantString[64]=0;
822 // now, Booth encode it
823 std::string boothEncodedString;
824 boothEncode(constantString, boothEncodedString);
826 std::vector<struct shiftaddblob> ops;
827 // do the transformation, filling out 'ops'
828 lefevre(boothEncodedString, ops);
830 assert(ops.size() < 80 && "constmul code has gone haywire\n");
831 SDOperand results[80]; // temporary results (of adds/subs of shifts)
833 // now turn 'ops' into DAG bits
834 for(unsigned i=0; i<ops.size(); i++) {
835 SDOperand amt = ISelDAG->getConstant(ops[i].firstShift, MVT::i64);
836 SDOperand val = (ops[i].firstVal == 0) ? N.getOperand(0) :
837 results[ops[i].firstVal-1];
838 SDOperand left = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
839 amt = ISelDAG->getConstant(ops[i].secondShift, MVT::i64);
840 val = (ops[i].secondVal == 0) ? N.getOperand(0) :
841 results[ops[i].secondVal-1];
842 SDOperand right = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
844 results[i] = ISelDAG->getNode(ISD::SUB, MVT::i64, left, right);
846 results[i] = ISelDAG->getNode(ISD::ADD, MVT::i64, left, right);
849 // don't forget flippedSign and preliminaryShift!
850 SDOperand shiftedresult;
851 if(preliminaryShift) {
852 SDOperand finalshift = ISelDAG->getConstant(preliminaryShift, MVT::i64);
853 shiftedresult = ISelDAG->getNode(ISD::SHL, MVT::i64,
854 results[ops.size()-1], finalshift);
855 } else { // there was no preliminary divide-by-power-of-2 required
856 shiftedresult = results[ops.size()-1];
859 SDOperand finalresult;
860 if(flippedSign) { // if we were multiplying by a negative constant:
861 SDOperand zero = ISelDAG->getConstant(0, MVT::i64);
862 // subtract the result from 0 to flip its sign
863 finalresult = ISelDAG->getNode(ISD::SUB, MVT::i64, zero, shiftedresult);
864 } else { // there was no preliminary multiply by -1 required
865 finalresult = shiftedresult;
871 /// ponderIntegerDivisionBy - When handling integer divides, if the divide
872 /// is by a constant such that we can efficiently codegen it, this
873 /// function says what to do. Currently, it returns 0 if the division must
874 /// become a genuine divide, and 1 if the division can be turned into a
876 static unsigned ponderIntegerDivisionBy(SDOperand N, bool isSigned,
878 if (N.getOpcode() != ISD::Constant) return 0; // if not a divide by
879 // a constant, give up.
881 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
883 if (isPowerOf2_64(v)) { // if a division by a power of two, say so
888 return 0; // fallthrough
891 static unsigned ponderIntegerAndWith(SDOperand N, unsigned& Imm) {
892 if (N.getOpcode() != ISD::Constant) return 0; // if not ANDing with
893 // a constant, give up.
895 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
897 if (isMask_64(v)) { // if ANDing with ((2^n)-1) for some n
898 Imm = Log2_64(v) + 1;
902 return 0; // fallthrough
905 static unsigned ponderIntegerAdditionWith(SDOperand N, unsigned& Imm) {
906 if (N.getOpcode() != ISD::Constant) return 0; // if not adding a
907 // constant, give up.
908 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
910 if (v <= 8191 && v >= -8192) { // if this constants fits in 14 bits, say so
911 Imm = v & 0x3FFF; // 14 bits
914 return 0; // fallthrough
917 static unsigned ponderIntegerSubtractionFrom(SDOperand N, unsigned& Imm) {
918 if (N.getOpcode() != ISD::Constant) return 0; // if not subtracting a
919 // constant, give up.
920 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
922 if (v <= 127 && v >= -128) { // if this constants fits in 8 bits, say so
923 Imm = v & 0xFF; // 8 bits
926 return 0; // fallthrough
929 unsigned ISel::SelectExpr(SDOperand N) {
931 unsigned Tmp1, Tmp2, Tmp3;
933 MVT::ValueType DestType = N.getValueType();
935 unsigned opcode = N.getOpcode();
937 SDNode *Node = N.Val;
940 if (Node->getOpcode() == ISD::CopyFromReg)
941 // Just use the specified register as our input.
942 return cast<RegisterSDNode>(Node->getOperand(1))->getReg();
944 unsigned &Reg = ExprMap[N];
947 if (N.getOpcode() != ISD::CALL && N.getOpcode() != ISD::TAILCALL)
948 Reg = Result = (N.getValueType() != MVT::Other) ?
949 MakeReg(N.getValueType()) : 1;
951 // If this is a call instruction, make sure to prepare ALL of the result
952 // values as well as the chain.
953 if (Node->getNumValues() == 1)
954 Reg = Result = 1; // Void call, just a chain.
956 Result = MakeReg(Node->getValueType(0));
957 ExprMap[N.getValue(0)] = Result;
958 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
959 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
960 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
964 switch (N.getOpcode()) {
967 assert(0 && "Node not handled!\n");
969 case ISD::FrameIndex: {
970 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
971 BuildMI(BB, IA64::MOV, 1, Result).addFrameIndex(Tmp1);
975 case ISD::ConstantPool: {
976 Tmp1 = BB->getParent()->getConstantPool()->
977 getConstantPoolIndex(cast<ConstantPoolSDNode>(N)->get());
978 IA64Lowering.restoreGP(BB); // FIXME: do i really need this?
979 BuildMI(BB, IA64::ADD, 2, Result).addConstantPoolIndex(Tmp1)
984 case ISD::ConstantFP: {
985 Tmp1 = Result; // Intermediate Register
986 if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
987 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
988 Tmp1 = MakeReg(MVT::f64);
990 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
991 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
992 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F0); // load 0.0
993 else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
994 cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
995 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F1); // load 1.0
997 assert(0 && "Unexpected FP constant!");
999 // we multiply by +1.0, negate (this is FNMA), and then add 0.0
1000 BuildMI(BB, IA64::FNMA, 3, Result).addReg(Tmp1).addReg(IA64::F1)
1005 case ISD::DYNAMIC_STACKALLOC: {
1006 // Generate both result values.
1008 ExprMap[N.getValue(1)] = 1; // Generate the token
1010 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1012 // FIXME: We are currently ignoring the requested alignment for handling
1013 // greater than the stack alignment. This will need to be revisited at some
1014 // point. Align = N.getOperand(2);
1016 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
1017 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
1018 std::cerr << "Cannot allocate stack object with greater alignment than"
1019 << " the stack alignment yet!";
1024 Select(N.getOperand(0));
1025 if (ConstantSDNode* CN = dyn_cast<ConstantSDNode>(N.getOperand(1)))
1027 if (CN->getValue() < 32000)
1029 BuildMI(BB, IA64::ADDIMM22, 2, IA64::r12).addReg(IA64::r12)
1030 .addImm(-CN->getValue());
1032 Tmp1 = SelectExpr(N.getOperand(1));
1033 // Subtract size from stack pointer, thereby allocating some space.
1034 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1037 Tmp1 = SelectExpr(N.getOperand(1));
1038 // Subtract size from stack pointer, thereby allocating some space.
1039 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1042 Select(N.getOperand(0));
1043 Tmp1 = SelectExpr(N.getOperand(1));
1044 // Subtract size from stack pointer, thereby allocating some space.
1045 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1046 // Put a pointer to the space into the result register, by copying the
1048 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r12);
1053 Tmp1 = SelectExpr(N.getOperand(0)); //Cond
1054 Tmp2 = SelectExpr(N.getOperand(1)); //Use if TRUE
1055 Tmp3 = SelectExpr(N.getOperand(2)); //Use if FALSE
1057 unsigned bogoResult;
1059 switch (N.getOperand(1).getValueType()) {
1060 default: assert(0 &&
1061 "ISD::SELECT: 'select'ing something other than i1, i64 or f64!\n");
1062 // for i1, we load the condition into an integer register, then
1063 // conditionally copy Tmp2 and Tmp3 to Tmp1 in parallel (only one
1064 // of them will go through, since the integer register will hold
1067 bogoResult=MakeReg(MVT::i1);
1069 // load the condition into an integer register
1070 unsigned condReg=MakeReg(MVT::i64);
1071 unsigned dummy=MakeReg(MVT::i64);
1072 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1073 BuildMI(BB, IA64::TPCADDIMM22, 2, condReg).addReg(dummy)
1074 .addImm(1).addReg(Tmp1);
1076 // initialize Result (bool) to false (hence UNC) and if
1077 // the select condition (condReg) is false (0), copy Tmp3
1078 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoResult)
1079 .addReg(condReg).addReg(IA64::r0).addReg(Tmp3);
1081 // now, if the selection condition is true, write 1 to the
1082 // result if Tmp2 is 1
1083 BuildMI(BB, IA64::TPCMPNE, 3, Result).addReg(bogoResult)
1084 .addReg(condReg).addReg(IA64::r0).addReg(Tmp2);
1087 // for i64/f64, we just copy Tmp3 and then conditionally overwrite it
1088 // with Tmp2 if Tmp1 is true
1090 bogoResult=MakeReg(MVT::i64);
1091 BuildMI(BB, IA64::MOV, 1, bogoResult).addReg(Tmp3);
1092 BuildMI(BB, IA64::CMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1096 bogoResult=MakeReg(MVT::f64);
1097 BuildMI(BB, IA64::FMOV, 1, bogoResult).addReg(Tmp3);
1098 BuildMI(BB, IA64::CFMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1106 case ISD::Constant: {
1107 unsigned depositPos=0;
1108 unsigned depositLen=0;
1109 switch (N.getValueType()) {
1110 default: assert(0 && "Cannot use constants of this type!");
1111 case MVT::i1: { // if a bool, we don't 'load' so much as generate
1113 if(cast<ConstantSDNode>(N)->getValue()) // true:
1114 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1116 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1117 return Result; // early exit
1119 case MVT::i64: break;
1122 int64_t immediate = cast<ConstantSDNode>(N)->getValue();
1124 if(immediate==0) { // if the constant is just zero,
1125 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r0); // just copy r0
1126 return Result; // early exit
1129 if (immediate <= 8191 && immediate >= -8192) {
1130 // if this constants fits in 14 bits, we use a mov the assembler will
1131 // turn into: "adds rDest=imm,r0" (and _not_ "andl"...)
1132 BuildMI(BB, IA64::MOVSIMM14, 1, Result).addSImm(immediate);
1133 return Result; // early exit
1136 if (immediate <= 2097151 && immediate >= -2097152) {
1137 // if this constants fits in 22 bits, we use a mov the assembler will
1138 // turn into: "addl rDest=imm,r0"
1139 BuildMI(BB, IA64::MOVSIMM22, 1, Result).addSImm(immediate);
1140 return Result; // early exit
1143 /* otherwise, our immediate is big, so we use movl */
1144 uint64_t Imm = immediate;
1145 BuildMI(BB, IA64::MOVLIMM64, 1, Result).addImm64(Imm);
1150 BuildMI(BB, IA64::IDEF, 0, Result);
1154 case ISD::GlobalAddress: {
1155 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1156 unsigned Tmp1 = MakeReg(MVT::i64);
1158 BuildMI(BB, IA64::ADD, 2, Tmp1).addGlobalAddress(GV).addReg(IA64::r1);
1159 BuildMI(BB, IA64::LD8, 1, Result).addReg(Tmp1);
1164 case ISD::ExternalSymbol: {
1165 const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
1166 // assert(0 && "sorry, but what did you want an ExternalSymbol for again?");
1167 BuildMI(BB, IA64::MOV, 1, Result).addExternalSymbol(Sym); // XXX
1171 case ISD::FP_EXTEND: {
1172 Tmp1 = SelectExpr(N.getOperand(0));
1173 BuildMI(BB, IA64::FMOV, 1, Result).addReg(Tmp1);
1177 case ISD::ANY_EXTEND:
1178 case ISD::ZERO_EXTEND: {
1179 Tmp1 = SelectExpr(N.getOperand(0)); // value
1181 assert(N.getOperand(0).getValueType() == MVT::i1 &&
1182 "Cannot zero-extend this type!");
1184 // if the predicate reg has 1, we want a '1' in our GR.
1185 unsigned dummy = MakeReg(MVT::i64);
1187 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1188 // ...then conditionally (PR:Tmp1) add 1:
1189 BuildMI(BB, IA64::TPCADDIMM22, 2, Result).addReg(dummy)
1190 .addImm(1).addReg(Tmp1);
1191 return Result; // XXX early exit!
1194 case ISD::SIGN_EXTEND:
1195 assert(N.getOperand(0).getValueType() == MVT::i1 &&
1196 "Cannot zero-extend this type!");
1198 Tmp1 = SelectExpr(N.getOperand(0)); // value
1199 assert(0 && "don't know how to sign_extend from bool yet!");
1203 // we use the funky dep.z (deposit (zero)) instruction to deposit bits
1204 // of R0 appropriately.
1205 assert(N.getOperand(0).getValueType() == MVT::i64 &&
1206 N.getValueType() == MVT::i1 && "Unknown truncate!");
1207 Tmp1 = SelectExpr(N.getOperand(0));
1209 // if input (normal reg) is 0, 0!=0 -> false (0), if 1, 1!=0 ->true (1):
1210 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1).addReg(IA64::r0);
1211 return Result; // XXX early exit!
1214 case ISD::FP_ROUND: {
1215 assert (DestType == MVT::f32 && N.getOperand(0).getValueType() == MVT::f64 &&
1216 "error: trying to FP_ROUND something other than f64 -> f32!\n");
1217 Tmp1 = SelectExpr(N.getOperand(0));
1218 BuildMI(BB, IA64::FADDS, 2, Result).addReg(Tmp1).addReg(IA64::F0);
1219 // we add 0.0 using a single precision add to do rounding
1224 // FIXME: the following 4 cases need cleaning
1225 case ISD::SINT_TO_FP: {
1226 Tmp1 = SelectExpr(N.getOperand(0));
1227 Tmp2 = MakeReg(MVT::f64);
1228 unsigned dummy = MakeReg(MVT::f64);
1229 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1230 BuildMI(BB, IA64::FCVTXF, 1, dummy).addReg(Tmp2);
1231 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1235 case ISD::UINT_TO_FP: {
1236 Tmp1 = SelectExpr(N.getOperand(0));
1237 Tmp2 = MakeReg(MVT::f64);
1238 unsigned dummy = MakeReg(MVT::f64);
1239 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1240 BuildMI(BB, IA64::FCVTXUF, 1, dummy).addReg(Tmp2);
1241 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1245 case ISD::FP_TO_SINT: {
1246 Tmp1 = SelectExpr(N.getOperand(0));
1247 Tmp2 = MakeReg(MVT::f64);
1248 BuildMI(BB, IA64::FCVTFXTRUNC, 1, Tmp2).addReg(Tmp1);
1249 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1253 case ISD::FP_TO_UINT: {
1254 Tmp1 = SelectExpr(N.getOperand(0));
1255 Tmp2 = MakeReg(MVT::f64);
1256 BuildMI(BB, IA64::FCVTFXUTRUNC, 1, Tmp2).addReg(Tmp1);
1257 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1262 if (N.getOperand(0).getOpcode() == ISD::FMUL &&
1263 N.getOperand(0).Val->hasOneUse()) { // if we can fold this add
1264 // into an fma, do so:
1265 // ++FusedFP; // Statistic
1266 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1267 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1268 Tmp3 = SelectExpr(N.getOperand(1));
1269 BuildMI(BB, IA64::FMA, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1270 return Result; // early exit
1273 //else, fallthrough:
1274 Tmp1 = SelectExpr(N.getOperand(0));
1275 Tmp2 = SelectExpr(N.getOperand(1));
1276 BuildMI(BB, IA64::FADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1281 if (N.getOperand(0).getOpcode() == ISD::SHL &&
1282 N.getOperand(0).Val->hasOneUse()) { // if we might be able to fold
1283 // this add into a shladd, try:
1284 ConstantSDNode *CSD = NULL;
1285 if((CSD = dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) &&
1286 (CSD->getValue() >= 1) && (CSD->getValue() <= 4) ) { // we can:
1288 // ++FusedSHLADD; // Statistic
1289 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1290 int shl_amt = CSD->getValue();
1291 Tmp3 = SelectExpr(N.getOperand(1));
1293 BuildMI(BB, IA64::SHLADD, 3, Result)
1294 .addReg(Tmp1).addImm(shl_amt).addReg(Tmp3);
1295 return Result; // early exit
1299 //else, fallthrough:
1300 Tmp1 = SelectExpr(N.getOperand(0));
1301 switch (ponderIntegerAdditionWith(N.getOperand(1), Tmp3)) {
1302 case 1: // adding a constant that's 14 bits
1303 BuildMI(BB, IA64::ADDIMM14, 2, Result).addReg(Tmp1).addSImm(Tmp3);
1304 return Result; // early exit
1305 } // fallthrough and emit a reg+reg ADD:
1306 Tmp2 = SelectExpr(N.getOperand(1));
1307 BuildMI(BB, IA64::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1312 Tmp1 = SelectExpr(N.getOperand(0));
1313 Tmp2 = SelectExpr(N.getOperand(1));
1314 BuildMI(BB, IA64::FMPY, 2, Result).addReg(Tmp1).addReg(Tmp2);
1320 /* FIXME if(N.getOperand(1).getOpcode() != ISD::Constant) { // if not a const mul
1322 // boring old integer multiply with xma
1323 Tmp1 = SelectExpr(N.getOperand(0));
1324 Tmp2 = SelectExpr(N.getOperand(1));
1326 unsigned TempFR1=MakeReg(MVT::f64);
1327 unsigned TempFR2=MakeReg(MVT::f64);
1328 unsigned TempFR3=MakeReg(MVT::f64);
1329 BuildMI(BB, IA64::SETFSIG, 1, TempFR1).addReg(Tmp1);
1330 BuildMI(BB, IA64::SETFSIG, 1, TempFR2).addReg(Tmp2);
1331 BuildMI(BB, IA64::XMAL, 1, TempFR3).addReg(TempFR1).addReg(TempFR2)
1333 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TempFR3);
1334 return Result; // early exit
1335 /* FIXME } else { // we are multiplying by an integer constant! yay
1336 return Reg = SelectExpr(BuildConstmulSequence(N)); // avert your eyes!
1341 if(N.getOperand(0).getOpcode() == ISD::FMUL &&
1342 N.getOperand(0).Val->hasOneUse()) { // if we can fold this sub
1343 // into an fms, do so:
1344 // ++FusedFP; // Statistic
1345 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1346 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1347 Tmp3 = SelectExpr(N.getOperand(1));
1348 BuildMI(BB, IA64::FMS, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1349 return Result; // early exit
1352 Tmp2 = SelectExpr(N.getOperand(1));
1353 Tmp1 = SelectExpr(N.getOperand(0));
1354 BuildMI(BB, IA64::FSUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1358 Tmp2 = SelectExpr(N.getOperand(1));
1359 switch (ponderIntegerSubtractionFrom(N.getOperand(0), Tmp3)) {
1360 case 1: // subtracting *from* an 8 bit constant:
1361 BuildMI(BB, IA64::SUBIMM8, 2, Result).addSImm(Tmp3).addReg(Tmp2);
1362 return Result; // early exit
1363 } // fallthrough and emit a reg+reg SUB:
1364 Tmp1 = SelectExpr(N.getOperand(0));
1365 BuildMI(BB, IA64::SUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1370 Tmp1 = SelectExpr(N.getOperand(0));
1371 assert(DestType == MVT::f64 && "trying to fabs something other than f64?");
1372 BuildMI(BB, IA64::FABS, 1, Result).addReg(Tmp1);
1377 assert(DestType == MVT::f64 && "trying to fneg something other than f64?");
1379 if (ISD::FABS == N.getOperand(0).getOpcode()) { // && hasOneUse()?
1380 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1381 BuildMI(BB, IA64::FNEGABS, 1, Result).addReg(Tmp1); // fold in abs
1383 Tmp1 = SelectExpr(N.getOperand(0));
1384 BuildMI(BB, IA64::FNEG, 1, Result).addReg(Tmp1); // plain old fneg
1391 switch (N.getValueType()) {
1392 default: assert(0 && "Cannot AND this type!");
1393 case MVT::i1: { // if a bool, we emit a pseudocode AND
1394 unsigned pA = SelectExpr(N.getOperand(0));
1395 unsigned pB = SelectExpr(N.getOperand(1));
1397 /* our pseudocode for AND is:
1399 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1400 cmp.eq pTemp,p0 = r0,r0 // pTemp = NOT pB
1402 (pB) cmp.ne pTemp,p0 = r0,r0
1404 (pTemp)cmp.ne pC,p0 = r0,r0 // if (NOT pB) pC = 0
1407 unsigned pTemp = MakeReg(MVT::i1);
1409 unsigned bogusTemp1 = MakeReg(MVT::i1);
1410 unsigned bogusTemp2 = MakeReg(MVT::i1);
1411 unsigned bogusTemp3 = MakeReg(MVT::i1);
1412 unsigned bogusTemp4 = MakeReg(MVT::i1);
1414 BuildMI(BB, IA64::PCMPEQUNC, 3, bogusTemp1)
1415 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1416 BuildMI(BB, IA64::CMPEQ, 2, bogusTemp2)
1417 .addReg(IA64::r0).addReg(IA64::r0);
1418 BuildMI(BB, IA64::TPCMPNE, 3, pTemp)
1419 .addReg(bogusTemp2).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1420 BuildMI(BB, IA64::TPCMPNE, 3, Result)
1421 .addReg(bogusTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pTemp);
1425 // if not a bool, we just AND away:
1430 Tmp1 = SelectExpr(N.getOperand(0));
1431 switch (ponderIntegerAndWith(N.getOperand(1), Tmp3)) {
1432 case 1: // ANDing a constant that is 2^n-1 for some n
1434 case 8: // if AND 0x00000000000000FF, be quaint and use zxt1
1435 BuildMI(BB, IA64::ZXT1, 1, Result).addReg(Tmp1);
1437 case 16: // if AND 0x000000000000FFFF, be quaint and use zxt2
1438 BuildMI(BB, IA64::ZXT2, 1, Result).addReg(Tmp1);
1440 case 32: // if AND 0x00000000FFFFFFFF, be quaint and use zxt4
1441 BuildMI(BB, IA64::ZXT4, 1, Result).addReg(Tmp1);
1443 default: // otherwise, use dep.z to paste zeros
1444 // FIXME: assert the dep.z is in bounds
1445 BuildMI(BB, IA64::DEPZ, 3, Result).addReg(Tmp1)
1446 .addImm(0).addImm(Tmp3);
1449 return Result; // early exit
1450 } // fallthrough and emit a simple AND:
1451 Tmp2 = SelectExpr(N.getOperand(1));
1452 BuildMI(BB, IA64::AND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1459 switch (N.getValueType()) {
1460 default: assert(0 && "Cannot OR this type!");
1461 case MVT::i1: { // if a bool, we emit a pseudocode OR
1462 unsigned pA = SelectExpr(N.getOperand(0));
1463 unsigned pB = SelectExpr(N.getOperand(1));
1465 unsigned pTemp1 = MakeReg(MVT::i1);
1467 /* our pseudocode for OR is:
1473 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1475 (pB) cmp.eq pC,p0 = r0,r0 // if (pB) pC = 1
1478 BuildMI(BB, IA64::PCMPEQUNC, 3, pTemp1)
1479 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1480 BuildMI(BB, IA64::TPCMPEQ, 4, Result)
1481 .addReg(pTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1484 // if not a bool, we just OR away:
1489 Tmp1 = SelectExpr(N.getOperand(0));
1490 Tmp2 = SelectExpr(N.getOperand(1));
1491 BuildMI(BB, IA64::OR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1499 switch (N.getValueType()) {
1500 default: assert(0 && "Cannot XOR this type!");
1501 case MVT::i1: { // if a bool, we emit a pseudocode XOR
1502 unsigned pY = SelectExpr(N.getOperand(0));
1503 unsigned pZ = SelectExpr(N.getOperand(1));
1505 /* one possible routine for XOR is:
1507 // Compute px = py ^ pz
1508 // using sum of products: px = (py & !pz) | (pz & !py)
1509 // Uses 5 instructions in 3 cycles.
1511 (pz) cmp.eq.unc px = r0, r0 // px = pz
1512 (py) cmp.eq.unc pt = r0, r0 // pt = py
1515 (pt) cmp.ne.and px = r0, r0 // px = px & !pt (px = pz & !pt)
1516 (pz) cmp.ne.and pt = r0, r0 // pt = pt & !pz
1520 (pt) cmp.eq.or px = r0, r0 // px = px | pt
1522 *** Another, which we use here, requires one scratch GR. it is:
1524 mov rt = 0 // initialize rt off critical path
1528 (pz) cmp.eq.unc px = r0, r0 // px = pz
1529 (pz) mov rt = 1 // rt = pz
1532 (py) cmp.ne px = 1, rt // if (py) px = !pz
1534 .. these routines kindly provided by Jim Hull
1536 unsigned rt = MakeReg(MVT::i64);
1538 // these two temporaries will never actually appear,
1539 // due to the two-address form of some of the instructions below
1540 unsigned bogoPR = MakeReg(MVT::i1); // becomes Result
1541 unsigned bogoGR = MakeReg(MVT::i64); // becomes rt
1543 BuildMI(BB, IA64::MOV, 1, bogoGR).addReg(IA64::r0);
1544 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoPR)
1545 .addReg(IA64::r0).addReg(IA64::r0).addReg(pZ);
1546 BuildMI(BB, IA64::TPCADDIMM22, 2, rt)
1547 .addReg(bogoGR).addImm(1).addReg(pZ);
1548 BuildMI(BB, IA64::TPCMPIMM8NE, 3, Result)
1549 .addReg(bogoPR).addImm(1).addReg(rt).addReg(pY);
1552 // if not a bool, we just XOR away:
1557 Tmp1 = SelectExpr(N.getOperand(0));
1558 Tmp2 = SelectExpr(N.getOperand(1));
1559 BuildMI(BB, IA64::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1567 Tmp1 = SelectExpr(N.getOperand(0));
1568 BuildMI(BB, IA64::POPCNT, 1, Result).addReg(Tmp1);
1573 Tmp1 = SelectExpr(N.getOperand(0));
1574 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1575 Tmp2 = CN->getValue();
1576 BuildMI(BB, IA64::SHLI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1578 Tmp2 = SelectExpr(N.getOperand(1));
1579 BuildMI(BB, IA64::SHL, 2, Result).addReg(Tmp1).addReg(Tmp2);
1585 Tmp1 = SelectExpr(N.getOperand(0));
1586 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1587 Tmp2 = CN->getValue();
1588 BuildMI(BB, IA64::SHRUI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1590 Tmp2 = SelectExpr(N.getOperand(1));
1591 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1597 Tmp1 = SelectExpr(N.getOperand(0));
1598 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1599 Tmp2 = CN->getValue();
1600 BuildMI(BB, IA64::SHRSI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1602 Tmp2 = SelectExpr(N.getOperand(1));
1603 BuildMI(BB, IA64::SHRS, 2, Result).addReg(Tmp1).addReg(Tmp2);
1614 Tmp1 = SelectExpr(N.getOperand(0));
1615 Tmp2 = SelectExpr(N.getOperand(1));
1619 if(DestType == MVT::f64) // XXX: we're not gonna be fed MVT::f32, are we?
1622 bool isModulus=false; // is it a division or a modulus?
1623 bool isSigned=false;
1625 switch(N.getOpcode()) {
1627 case ISD::SDIV: isModulus=false; isSigned=true; break;
1628 case ISD::UDIV: isModulus=false; isSigned=false; break;
1630 case ISD::SREM: isModulus=true; isSigned=true; break;
1631 case ISD::UREM: isModulus=true; isSigned=false; break;
1634 if(!isModulus && !isFP) { // if this is an integer divide,
1635 switch (ponderIntegerDivisionBy(N.getOperand(1), isSigned, Tmp3)) {
1636 case 1: // division by a constant that's a power of 2
1637 Tmp1 = SelectExpr(N.getOperand(0));
1638 if(isSigned) { // argument could be negative, so emit some code:
1639 unsigned divAmt=Tmp3;
1640 unsigned tempGR1=MakeReg(MVT::i64);
1641 unsigned tempGR2=MakeReg(MVT::i64);
1642 unsigned tempGR3=MakeReg(MVT::i64);
1643 BuildMI(BB, IA64::SHRS, 2, tempGR1)
1644 .addReg(Tmp1).addImm(divAmt-1);
1645 BuildMI(BB, IA64::EXTRU, 3, tempGR2)
1646 .addReg(tempGR1).addImm(64-divAmt).addImm(divAmt);
1647 BuildMI(BB, IA64::ADD, 2, tempGR3)
1648 .addReg(Tmp1).addReg(tempGR2);
1649 BuildMI(BB, IA64::SHRS, 2, Result)
1650 .addReg(tempGR3).addImm(divAmt);
1652 else // unsigned div-by-power-of-2 becomes a simple shift right:
1653 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addImm(Tmp3);
1654 return Result; // early exit
1658 unsigned TmpPR=MakeReg(MVT::i1); // we need two scratch
1659 unsigned TmpPR2=MakeReg(MVT::i1); // predicate registers,
1660 unsigned TmpF1=MakeReg(MVT::f64); // and one metric truckload of FP regs.
1661 unsigned TmpF2=MakeReg(MVT::f64); // lucky we have IA64?
1662 unsigned TmpF3=MakeReg(MVT::f64); // well, the real FIXME is to have
1663 unsigned TmpF4=MakeReg(MVT::f64); // isTwoAddress forms of these
1664 unsigned TmpF5=MakeReg(MVT::f64); // FP instructions so we can end up with
1665 unsigned TmpF6=MakeReg(MVT::f64); // stuff like setf.sig f10=f10 etc.
1666 unsigned TmpF7=MakeReg(MVT::f64);
1667 unsigned TmpF8=MakeReg(MVT::f64);
1668 unsigned TmpF9=MakeReg(MVT::f64);
1669 unsigned TmpF10=MakeReg(MVT::f64);
1670 unsigned TmpF11=MakeReg(MVT::f64);
1671 unsigned TmpF12=MakeReg(MVT::f64);
1672 unsigned TmpF13=MakeReg(MVT::f64);
1673 unsigned TmpF14=MakeReg(MVT::f64);
1674 unsigned TmpF15=MakeReg(MVT::f64);
1676 // OK, emit some code:
1679 // first, load the inputs into FP regs.
1680 BuildMI(BB, IA64::SETFSIG, 1, TmpF1).addReg(Tmp1);
1681 BuildMI(BB, IA64::SETFSIG, 1, TmpF2).addReg(Tmp2);
1683 // next, convert the inputs to FP
1685 BuildMI(BB, IA64::FCVTXF, 1, TmpF3).addReg(TmpF1);
1686 BuildMI(BB, IA64::FCVTXF, 1, TmpF4).addReg(TmpF2);
1688 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF3).addReg(TmpF1);
1689 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF4).addReg(TmpF2);
1692 } else { // this is an FP divide/remainder, so we 'leak' some temp
1693 // regs and assign TmpF3=Tmp1, TmpF4=Tmp2
1698 // we start by computing an approximate reciprocal (good to 9 bits?)
1699 // note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
1700 BuildMI(BB, IA64::FRCPAS1, 4)
1701 .addReg(TmpF5, MachineOperand::Def)
1702 .addReg(TmpPR, MachineOperand::Def)
1703 .addReg(TmpF3).addReg(TmpF4);
1705 if(!isModulus) { // if this is a divide, we worry about div-by-zero
1706 unsigned bogusPR=MakeReg(MVT::i1); // won't appear, due to twoAddress
1708 BuildMI(BB, IA64::CMPEQ, 2, bogusPR).addReg(IA64::r0).addReg(IA64::r0);
1709 BuildMI(BB, IA64::TPCMPNE, 3, TmpPR2).addReg(bogusPR)
1710 .addReg(IA64::r0).addReg(IA64::r0).addReg(TmpPR);
1713 // now we apply newton's method, thrice! (FIXME: this is ~72 bits of
1714 // precision, don't need this much for f32/i32)
1715 BuildMI(BB, IA64::CFNMAS1, 4, TmpF6)
1716 .addReg(TmpF4).addReg(TmpF5).addReg(IA64::F1).addReg(TmpPR);
1717 BuildMI(BB, IA64::CFMAS1, 4, TmpF7)
1718 .addReg(TmpF3).addReg(TmpF5).addReg(IA64::F0).addReg(TmpPR);
1719 BuildMI(BB, IA64::CFMAS1, 4, TmpF8)
1720 .addReg(TmpF6).addReg(TmpF6).addReg(IA64::F0).addReg(TmpPR);
1721 BuildMI(BB, IA64::CFMAS1, 4, TmpF9)
1722 .addReg(TmpF6).addReg(TmpF7).addReg(TmpF7).addReg(TmpPR);
1723 BuildMI(BB, IA64::CFMAS1, 4,TmpF10)
1724 .addReg(TmpF6).addReg(TmpF5).addReg(TmpF5).addReg(TmpPR);
1725 BuildMI(BB, IA64::CFMAS1, 4,TmpF11)
1726 .addReg(TmpF8).addReg(TmpF9).addReg(TmpF9).addReg(TmpPR);
1727 BuildMI(BB, IA64::CFMAS1, 4,TmpF12)
1728 .addReg(TmpF8).addReg(TmpF10).addReg(TmpF10).addReg(TmpPR);
1729 BuildMI(BB, IA64::CFNMAS1, 4,TmpF13)
1730 .addReg(TmpF4).addReg(TmpF11).addReg(TmpF3).addReg(TmpPR);
1732 // FIXME: this is unfortunate :(
1733 // the story is that the dest reg of the fnma above and the fma below
1734 // (and therefore possibly the src of the fcvt.fx[u] as well) cannot
1735 // be the same register, or this code breaks if the first argument is
1736 // zero. (e.g. without this hack, 0%8 yields -64, not 0.)
1737 BuildMI(BB, IA64::CFMAS1, 4,TmpF14)
1738 .addReg(TmpF13).addReg(TmpF12).addReg(TmpF11).addReg(TmpPR);
1740 if(isModulus) { // XXX: fragile! fixes _only_ mod, *breaks* div! !
1741 BuildMI(BB, IA64::IUSE, 1).addReg(TmpF13); // hack :(
1745 // round to an integer
1747 BuildMI(BB, IA64::FCVTFXTRUNCS1, 1, TmpF15).addReg(TmpF14);
1749 BuildMI(BB, IA64::FCVTFXUTRUNCS1, 1, TmpF15).addReg(TmpF14);
1751 BuildMI(BB, IA64::FMOV, 1, TmpF15).addReg(TmpF14);
1752 // EXERCISE: can you see why TmpF15=TmpF14 does not work here, and
1753 // we really do need the above FMOV? ;)
1757 if(isFP) { // extra worrying about div-by-zero
1758 unsigned bogoResult=MakeReg(MVT::f64);
1760 // we do a 'conditional fmov' (of the correct result, depending
1761 // on how the frcpa predicate turned out)
1762 BuildMI(BB, IA64::PFMOV, 2, bogoResult)
1763 .addReg(TmpF12).addReg(TmpPR2);
1764 BuildMI(BB, IA64::CFMOV, 2, Result)
1765 .addReg(bogoResult).addReg(TmpF15).addReg(TmpPR);
1768 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TmpF15);
1770 } else { // this is a modulus
1772 // answer = q * (-b) + a
1773 unsigned ModulusResult = MakeReg(MVT::f64);
1774 unsigned TmpF = MakeReg(MVT::f64);
1775 unsigned TmpI = MakeReg(MVT::i64);
1777 BuildMI(BB, IA64::SUB, 2, TmpI).addReg(IA64::r0).addReg(Tmp2);
1778 BuildMI(BB, IA64::SETFSIG, 1, TmpF).addReg(TmpI);
1779 BuildMI(BB, IA64::XMAL, 3, ModulusResult)
1780 .addReg(TmpF15).addReg(TmpF).addReg(TmpF1);
1781 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(ModulusResult);
1782 } else { // FP modulus! The horror... the horror....
1783 assert(0 && "sorry, no FP modulus just yet!\n!\n");
1790 case ISD::SIGN_EXTEND_INREG: {
1791 Tmp1 = SelectExpr(N.getOperand(0));
1792 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1795 assert(0 && "don't know how to sign extend this type");
1797 case MVT::i8: Opc = IA64::SXT1; break;
1798 case MVT::i16: Opc = IA64::SXT2; break;
1799 case MVT::i32: Opc = IA64::SXT4; break;
1801 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1806 Tmp1 = SelectExpr(N.getOperand(0));
1807 ISD::CondCode CC = cast<CondCodeSDNode>(Node->getOperand(2))->get();
1808 if (MVT::isInteger(N.getOperand(0).getValueType())) {
1810 if(ConstantSDNode *CSDN =
1811 dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1812 // if we are comparing against a constant zero
1813 if(CSDN->getValue()==0)
1814 Tmp2 = IA64::r0; // then we can just compare against r0
1816 Tmp2 = SelectExpr(N.getOperand(1));
1817 } else // not comparing against a constant
1818 Tmp2 = SelectExpr(N.getOperand(1));
1821 default: assert(0 && "Unknown integer comparison!");
1823 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1826 BuildMI(BB, IA64::CMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1829 BuildMI(BB, IA64::CMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1832 BuildMI(BB, IA64::CMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1835 BuildMI(BB, IA64::CMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1838 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1841 BuildMI(BB, IA64::CMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1844 BuildMI(BB, IA64::CMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1847 BuildMI(BB, IA64::CMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1850 BuildMI(BB, IA64::CMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1853 } else { // if not integer, should be FP.
1854 assert(N.getOperand(0).getValueType() != MVT::f32 &&
1855 "error: SETCC should have had incoming f32 promoted to f64!\n");
1857 if(ConstantFPSDNode *CFPSDN =
1858 dyn_cast<ConstantFPSDNode>(N.getOperand(1))) {
1860 // if we are comparing against a constant +0.0 or +1.0
1861 if(CFPSDN->isExactlyValue(+0.0))
1862 Tmp2 = IA64::F0; // then we can just compare against f0
1863 else if(CFPSDN->isExactlyValue(+1.0))
1864 Tmp2 = IA64::F1; // or f1
1866 Tmp2 = SelectExpr(N.getOperand(1));
1867 } else // not comparing against a constant
1868 Tmp2 = SelectExpr(N.getOperand(1));
1871 default: assert(0 && "Unknown FP comparison!");
1873 BuildMI(BB, IA64::FCMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1876 BuildMI(BB, IA64::FCMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1879 BuildMI(BB, IA64::FCMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1882 BuildMI(BB, IA64::FCMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1885 BuildMI(BB, IA64::FCMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1888 BuildMI(BB, IA64::FCMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1891 BuildMI(BB, IA64::FCMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1894 BuildMI(BB, IA64::FCMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1897 BuildMI(BB, IA64::FCMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1900 BuildMI(BB, IA64::FCMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1910 // Make sure we generate both values.
1912 ExprMap[N.getValue(1)] = 1; // Generate the token
1914 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1918 if(opcode == ISD::LOAD) { // this is a LOAD
1919 switch (Node->getValueType(0)) {
1920 default: assert(0 && "Cannot load this type!");
1921 case MVT::i1: Opc = IA64::LD1; isBool=true; break;
1922 // FIXME: for now, we treat bool loads the same as i8 loads */
1923 case MVT::i8: Opc = IA64::LD1; break;
1924 case MVT::i16: Opc = IA64::LD2; break;
1925 case MVT::i32: Opc = IA64::LD4; break;
1926 case MVT::i64: Opc = IA64::LD8; break;
1928 case MVT::f32: Opc = IA64::LDF4; break;
1929 case MVT::f64: Opc = IA64::LDF8; break;
1931 } else { // this is an EXTLOAD or ZEXTLOAD
1932 MVT::ValueType TypeBeingLoaded =
1933 cast<VTSDNode>(Node->getOperand(3))->getVT();
1934 switch (TypeBeingLoaded) {
1935 default: assert(0 && "Cannot extload/zextload this type!");
1937 case MVT::i8: Opc = IA64::LD1; break;
1938 case MVT::i16: Opc = IA64::LD2; break;
1939 case MVT::i32: Opc = IA64::LD4; break;
1940 case MVT::f32: Opc = IA64::LDF4; break;
1944 SDOperand Chain = N.getOperand(0);
1945 SDOperand Address = N.getOperand(1);
1947 if(Address.getOpcode() == ISD::GlobalAddress) {
1949 unsigned dummy = MakeReg(MVT::i64);
1950 unsigned dummy2 = MakeReg(MVT::i64);
1951 BuildMI(BB, IA64::ADD, 2, dummy)
1952 .addGlobalAddress(cast<GlobalAddressSDNode>(Address)->getGlobal())
1954 BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
1956 BuildMI(BB, Opc, 1, Result).addReg(dummy2);
1957 else { // emit a little pseudocode to load a bool (stored in one byte)
1958 // into a predicate register
1959 assert(Opc==IA64::LD1 && "problem loading a bool");
1960 unsigned dummy3 = MakeReg(MVT::i64);
1961 BuildMI(BB, Opc, 1, dummy3).addReg(dummy2);
1962 // we compare to 0. true? 0. false? 1.
1963 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
1965 } else if(ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1966 unsigned CPIdx = BB->getParent()->getConstantPool()->
1967 getConstantPoolIndex(CP->get());
1969 IA64Lowering.restoreGP(BB);
1970 unsigned dummy = MakeReg(MVT::i64);
1971 unsigned dummy2 = MakeReg(MVT::i64);
1972 BuildMI(BB, IA64::MOVLIMM64, 1, dummy2).addConstantPoolIndex(CPIdx);
1973 BuildMI(BB, IA64::ADD, 2, dummy).addReg(dummy2).addReg(IA64::r1); //CPI+GP
1976 // OLD BuildMI(BB, IA64::ADD, 2, dummy).addConstantPoolIndex(CPIdx)
1977 // (FIXME!) .addReg(IA64::r1); // CPI+GP
1979 BuildMI(BB, Opc, 1, Result).addReg(dummy);
1980 else { // emit a little pseudocode to load a bool (stored in one byte)
1981 // into a predicate register
1982 assert(Opc==IA64::LD1 && "problem loading a bool");
1983 unsigned dummy3 = MakeReg(MVT::i64);
1984 BuildMI(BB, Opc, 1, dummy3).addReg(dummy);
1985 // we compare to 0. true? 0. false? 1.
1986 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
1988 } else if(Address.getOpcode() == ISD::FrameIndex) {
1989 Select(Chain); // FIXME ? what about bools?
1990 unsigned dummy = MakeReg(MVT::i64);
1991 BuildMI(BB, IA64::MOV, 1, dummy)
1992 .addFrameIndex(cast<FrameIndexSDNode>(Address)->getIndex());
1994 BuildMI(BB, Opc, 1, Result).addReg(dummy);
1995 else { // emit a little pseudocode to load a bool (stored in one byte)
1996 // into a predicate register
1997 assert(Opc==IA64::LD1 && "problem loading a bool");
1998 unsigned dummy3 = MakeReg(MVT::i64);
1999 BuildMI(BB, Opc, 1, dummy3).addReg(dummy);
2000 // we compare to 0. true? 0. false? 1.
2001 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
2003 } else { // none of the above...
2005 Tmp2 = SelectExpr(Address);
2007 BuildMI(BB, Opc, 1, Result).addReg(Tmp2);
2008 else { // emit a little pseudocode to load a bool (stored in one byte)
2009 // into a predicate register
2010 assert(Opc==IA64::LD1 && "problem loading a bool");
2011 unsigned dummy = MakeReg(MVT::i64);
2012 BuildMI(BB, Opc, 1, dummy).addReg(Tmp2);
2013 // we compare to 0. true? 0. false? 1.
2014 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy).addReg(IA64::r0);
2021 case ISD::CopyFromReg: {
2023 Result = ExprMap[N.getValue(0)] =
2024 MakeReg(N.getValue(0).getValueType());
2026 SDOperand Chain = N.getOperand(0);
2029 unsigned r = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
2031 if(N.getValueType() == MVT::i1) // if a bool, we use pseudocode
2032 BuildMI(BB, IA64::PCMPEQUNC, 3, Result)
2033 .addReg(IA64::r0).addReg(IA64::r0).addReg(r);
2034 // (r) Result =cmp.eq.unc(r0,r0)
2036 BuildMI(BB, IA64::MOV, 1, Result).addReg(r); // otherwise MOV
2042 Select(N.getOperand(0));
2044 // The chain for this call is now lowered.
2045 ExprMap.insert(std::make_pair(N.getValue(Node->getNumValues()-1), 1));
2047 //grab the arguments
2048 std::vector<unsigned> argvregs;
2050 for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
2051 argvregs.push_back(SelectExpr(N.getOperand(i)));
2053 // see section 8.5.8 of "Itanium Software Conventions and
2054 // Runtime Architecture Guide to see some examples of what's going
2055 // on here. (in short: int args get mapped 1:1 'slot-wise' to out0->out7,
2056 // while FP args get mapped to F8->F15 as needed)
2058 unsigned used_FPArgs=0; // how many FP Args have been used so far?
2061 for(int i = 0, e = std::min(8, (int)argvregs.size()); i < e; ++i)
2063 unsigned intArgs[] = {IA64::out0, IA64::out1, IA64::out2, IA64::out3,
2064 IA64::out4, IA64::out5, IA64::out6, IA64::out7 };
2065 unsigned FPArgs[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
2066 IA64::F12, IA64::F13, IA64::F14, IA64::F15 };
2068 switch(N.getOperand(i+2).getValueType())
2070 default: // XXX do we need to support MVT::i1 here?
2072 N.getOperand(i).Val->dump();
2073 std::cerr << "Type for " << i << " is: " <<
2074 N.getOperand(i+2).getValueType() << std::endl;
2075 assert(0 && "Unknown value type for call");
2077 BuildMI(BB, IA64::MOV, 1, intArgs[i]).addReg(argvregs[i]);
2080 BuildMI(BB, IA64::FMOV, 1, FPArgs[used_FPArgs++])
2081 .addReg(argvregs[i]);
2082 // FIXME: we don't need to do this _all_ the time:
2083 BuildMI(BB, IA64::GETFD, 1, intArgs[i]).addReg(argvregs[i]);
2089 for (int i = 8, e = argvregs.size(); i < e; ++i)
2091 unsigned tempAddr = MakeReg(MVT::i64);
2093 switch(N.getOperand(i+2).getValueType()) {
2096 N.getOperand(i).Val->dump();
2097 std::cerr << "Type for " << i << " is: " <<
2098 N.getOperand(i+2).getValueType() << "\n";
2099 assert(0 && "Unknown value type for call");
2100 case MVT::i1: // FIXME?
2105 BuildMI(BB, IA64::ADDIMM22, 2, tempAddr)
2106 .addReg(IA64::r12).addImm(16 + (i - 8) * 8); // r12 is SP
2107 BuildMI(BB, IA64::ST8, 2).addReg(tempAddr).addReg(argvregs[i]);
2111 BuildMI(BB, IA64::ADDIMM22, 2, tempAddr)
2112 .addReg(IA64::r12).addImm(16 + (i - 8) * 8); // r12 is SP
2113 BuildMI(BB, IA64::STF8, 2).addReg(tempAddr).addReg(argvregs[i]);
2118 // build the right kind of call. if we can branch directly, do so:
2119 if (GlobalAddressSDNode *GASD =
2120 dyn_cast<GlobalAddressSDNode>(N.getOperand(1)))
2122 BuildMI(BB, IA64::BRCALL, 1).addGlobalAddress(GASD->getGlobal(),true);
2123 IA64Lowering.restoreGP_SP_RP(BB);
2125 if (ExternalSymbolSDNode *ESSDN =
2126 dyn_cast<ExternalSymbolSDNode>(N.getOperand(1)))
2127 { // FIXME : currently need this case for correctness, to avoid
2128 // "non-pic code with imm relocation against dynamic symbol" errors
2129 BuildMI(BB, IA64::BRCALL, 1)
2130 .addExternalSymbol(ESSDN->getSymbol(), true);
2131 IA64Lowering.restoreGP_SP_RP(BB);
2133 else { // otherwise we need to get the function descriptor
2134 // load the branch target (function)'s entry point and
2136 Tmp1 = SelectExpr(N.getOperand(1));
2138 unsigned targetEntryPoint=MakeReg(MVT::i64);
2139 unsigned targetGPAddr=MakeReg(MVT::i64);
2140 unsigned currentGP=MakeReg(MVT::i64);
2142 // b6 is a scratch branch register, we load the target entry point
2143 // from the base of the function descriptor
2144 BuildMI(BB, IA64::LD8, 1, targetEntryPoint).addReg(Tmp1);
2145 BuildMI(BB, IA64::MOV, 1, IA64::B6).addReg(targetEntryPoint);
2147 // save the current GP:
2148 BuildMI(BB, IA64::MOV, 1, currentGP).addReg(IA64::r1);
2150 /* TODO: we need to make sure doing this never, ever loads a
2151 * bogus value into r1 (GP). */
2152 // load the target GP (which is at mem[functiondescriptor+8])
2153 BuildMI(BB, IA64::ADDIMM22, 2, targetGPAddr)
2154 .addReg(Tmp1).addImm(8); // FIXME: addimm22? why not postincrement ld
2155 BuildMI(BB, IA64::LD8, 1, IA64::r1).addReg(targetGPAddr);
2157 // and then jump: (well, call)
2158 BuildMI(BB, IA64::BRCALL, 1).addReg(IA64::B6);
2159 // and finally restore the old GP
2160 BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(currentGP);
2161 IA64Lowering.restoreSP_RP(BB);
2164 switch (Node->getValueType(0)) {
2165 default: assert(0 && "Unknown value type for call result!");
2166 case MVT::Other: return 1;
2168 BuildMI(BB, IA64::CMPNE, 2, Result)
2169 .addReg(IA64::r8).addReg(IA64::r0);
2175 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r8);
2178 BuildMI(BB, IA64::FMOV, 1, Result).addReg(IA64::F8);
2181 return Result+N.ResNo;
2188 void ISel::Select(SDOperand N) {
2189 unsigned Tmp1, Tmp2, Opc;
2190 unsigned opcode = N.getOpcode();
2192 if (!LoweredTokens.insert(N).second)
2193 return; // Already selected.
2195 SDNode *Node = N.Val;
2197 switch (Node->getOpcode()) {
2199 Node->dump(); std::cerr << "\n";
2200 assert(0 && "Node not handled yet!");
2202 case ISD::EntryToken: return; // Noop
2204 case ISD::TokenFactor: {
2205 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
2206 Select(Node->getOperand(i));
2210 case ISD::CopyToReg: {
2211 Select(N.getOperand(0));
2212 Tmp1 = SelectExpr(N.getOperand(2));
2213 Tmp2 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
2216 // if a bool, we use pseudocode
2217 if (N.getOperand(2).getValueType() == MVT::i1)
2218 BuildMI(BB, IA64::PCMPEQUNC, 3, Tmp2)
2219 .addReg(IA64::r0).addReg(IA64::r0).addReg(Tmp1);
2220 // (Tmp1) Tmp2 = cmp.eq.unc(r0,r0)
2222 BuildMI(BB, IA64::MOV, 1, Tmp2).addReg(Tmp1);
2223 // XXX is this the right way 'round? ;)
2224 // FIXME: WHAT ABOUT FLOATING POINT?
2231 /* what the heck is going on here:
2233 <_sabre_> ret with two operands is obvious: chain and value
2235 <_sabre_> ret with 3 values happens when 'expansion' occurs
2236 <_sabre_> e.g. i64 gets split into 2x i32
2238 <_sabre_> you don't have this case on ia64
2240 <_sabre_> so the two returned values go into EAX/EDX on ia32
2241 <camel_> ahhh *memories*
2243 <camel_> ok, thanks :)
2244 <_sabre_> so yeah, everything that has a side effect takes a 'token chain'
2245 <_sabre_> this is the first operand always
2246 <_sabre_> these operand often define chains, they are the last operand
2247 <_sabre_> they are printed as 'ch' if you do DAG.dump()
2250 switch (N.getNumOperands()) {
2252 assert(0 && "Unknown return instruction!");
2254 Select(N.getOperand(0));
2255 Tmp1 = SelectExpr(N.getOperand(1));
2256 switch (N.getOperand(1).getValueType()) {
2257 default: assert(0 && "All other types should have been promoted!!");
2258 // FIXME: do I need to add support for bools here?
2259 // (return '0' or '1' r8, basically...)
2261 // FIXME: need to round floats - 80 bits is bad, the tester
2264 // we mark r8 as live on exit up above in LowerArguments()
2265 BuildMI(BB, IA64::MOV, 1, IA64::r8).addReg(Tmp1);
2268 // we mark F8 as live on exit up above in LowerArguments()
2269 BuildMI(BB, IA64::FMOV, 1, IA64::F8).addReg(Tmp1);
2273 Select(N.getOperand(0));
2276 // before returning, restore the ar.pfs register (set by the 'alloc' up top)
2277 BuildMI(BB, IA64::MOV, 1).addReg(IA64::AR_PFS).addReg(IA64Lowering.VirtGPR);
2278 BuildMI(BB, IA64::RET, 0); // and then just emit a 'ret' instruction
2283 Select(N.getOperand(0));
2284 MachineBasicBlock *Dest =
2285 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2286 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(IA64::p0).addMBB(Dest);
2287 // XXX HACK! we do _not_ need long branches all the time
2292 MachineBasicBlock *Dest =
2293 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
2295 Select(N.getOperand(0));
2296 Tmp1 = SelectExpr(N.getOperand(1));
2297 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(Tmp1).addMBB(Dest);
2298 // XXX HACK! we do _not_ need long branches all the time
2308 case ISD::CopyFromReg:
2309 case ISD::DYNAMIC_STACKALLOC:
2313 case ISD::TRUNCSTORE:
2315 Select(N.getOperand(0));
2316 Tmp1 = SelectExpr(N.getOperand(1)); // value
2320 if(opcode == ISD::STORE) {
2321 switch (N.getOperand(1).getValueType()) {
2322 default: assert(0 && "Cannot store this type!");
2323 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2324 // FIXME?: for now, we treat bool loads the same as i8 stores */
2325 case MVT::i8: Opc = IA64::ST1; break;
2326 case MVT::i16: Opc = IA64::ST2; break;
2327 case MVT::i32: Opc = IA64::ST4; break;
2328 case MVT::i64: Opc = IA64::ST8; break;
2330 case MVT::f32: Opc = IA64::STF4; break;
2331 case MVT::f64: Opc = IA64::STF8; break;
2333 } else { // truncstore
2334 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2335 default: assert(0 && "unknown type in truncstore");
2336 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2337 //FIXME: DAG does not promote this load?
2338 case MVT::i8: Opc = IA64::ST1; break;
2339 case MVT::i16: Opc = IA64::ST2; break;
2340 case MVT::i32: Opc = IA64::ST4; break;
2341 case MVT::f32: Opc = IA64::STF4; break;
2345 if(N.getOperand(2).getOpcode() == ISD::GlobalAddress) {
2346 unsigned dummy = MakeReg(MVT::i64);
2347 unsigned dummy2 = MakeReg(MVT::i64);
2348 BuildMI(BB, IA64::ADD, 2, dummy)
2349 .addGlobalAddress(cast<GlobalAddressSDNode>
2350 (N.getOperand(2))->getGlobal()).addReg(IA64::r1);
2351 BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
2354 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(Tmp1);
2355 else { // we are storing a bool, so emit a little pseudocode
2356 // to store a predicate register as one byte
2357 assert(Opc==IA64::ST1);
2358 unsigned dummy3 = MakeReg(MVT::i64);
2359 unsigned dummy4 = MakeReg(MVT::i64);
2360 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2361 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2362 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2363 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(dummy4);
2365 } else if(N.getOperand(2).getOpcode() == ISD::FrameIndex) {
2367 // FIXME? (what about bools?)
2369 unsigned dummy = MakeReg(MVT::i64);
2370 BuildMI(BB, IA64::MOV, 1, dummy)
2371 .addFrameIndex(cast<FrameIndexSDNode>(N.getOperand(2))->getIndex());
2372 BuildMI(BB, Opc, 2).addReg(dummy).addReg(Tmp1);
2373 } else { // otherwise
2374 Tmp2 = SelectExpr(N.getOperand(2)); //address
2376 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(Tmp1);
2377 else { // we are storing a bool, so emit a little pseudocode
2378 // to store a predicate register as one byte
2379 assert(Opc==IA64::ST1);
2380 unsigned dummy3 = MakeReg(MVT::i64);
2381 unsigned dummy4 = MakeReg(MVT::i64);
2382 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2383 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2384 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2385 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(dummy4);
2391 case ISD::CALLSEQ_START:
2392 case ISD::CALLSEQ_END: {
2393 Select(N.getOperand(0));
2394 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2396 Opc = N.getOpcode() == ISD::CALLSEQ_START ? IA64::ADJUSTCALLSTACKDOWN :
2397 IA64::ADJUSTCALLSTACKUP;
2398 BuildMI(BB, Opc, 1).addImm(Tmp1);
2404 assert(0 && "GAME OVER. INSERT COIN?");
2408 /// createIA64PatternInstructionSelector - This pass converts an LLVM function
2409 /// into a machine code representation using pattern matching and a machine
2410 /// description file.
2412 FunctionPass *llvm::createIA64PatternInstructionSelector(TargetMachine &TM) {
2413 return new ISel(TM);