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(argVreg[count],
205 getValueType(I->getType()), DAG.getRoot());
207 case MVT::i1: // NOTE: as far as C abi stuff goes,
208 // bools are just boring old ints
213 //XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
214 MF.addLiveIn(args_int[count]); // mark this register as liveIn
216 MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
217 argPreg[count] = args_int[count];
218 argOpc[count] = IA64::MOV;
220 DAG.getCopyFromReg(argVreg[count], MVT::i64, DAG.getRoot());
221 if ( getValueType(I->getType()) != MVT::i64)
222 argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
226 } else { // more than 8 args go into the frame
227 // Create the frame index object for this incoming parameter...
228 ArgOffset = 16 + 8 * (count - 8);
229 int FI = MFI->CreateFixedObject(8, ArgOffset);
231 // Create the SelectionDAG nodes corresponding to a load
232 //from this parameter
233 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
234 argt = newroot = DAG.getLoad(getValueType(I->getType()),
235 DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
238 DAG.setRoot(newroot.getValue(1));
239 ArgValues.push_back(argt);
243 // Create a vreg to hold the output of (what will become)
244 // the "alloc" instruction
245 VirtGPR = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
246 BuildMI(&BB, IA64::PSEUDO_ALLOC, 0, VirtGPR);
247 // we create a PSEUDO_ALLOC (pseudo)instruction for now
249 BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
252 BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
253 BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
256 BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
259 BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
260 BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
263 unsigned tempOffset=0;
265 // if this is a varargs function, we simply lower llvm.va_start by
266 // pointing to the first entry
269 VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
272 // here we actually do the moving of args, and store them to the stack
273 // too if this is a varargs function:
274 for (int i = 0; i < count && i < 8; ++i) {
275 BuildMI(&BB, argOpc[i], 1, argVreg[i]).addReg(argPreg[i]);
277 // if this is a varargs function, we copy the input registers to the stack
278 int FI = MFI->CreateFixedObject(8, tempOffset);
279 tempOffset+=8; //XXX: is it safe to use r22 like this?
280 BuildMI(&BB, IA64::MOV, 1, IA64::r22).addFrameIndex(FI);
281 // FIXME: we should use st8.spill here, one day
282 BuildMI(&BB, IA64::ST8, 1, IA64::r22).addReg(argPreg[i]);
286 // Finally, inform the code generator which regs we return values in.
287 // (see the ISD::RET: case down below)
288 switch (getValueType(F.getReturnType())) {
289 default: assert(0 && "i have no idea where to return this type!");
290 case MVT::isVoid: break;
296 MF.addLiveOut(IA64::r8);
300 MF.addLiveOut(IA64::F8);
307 std::pair<SDOperand, SDOperand>
308 IA64TargetLowering::LowerCallTo(SDOperand Chain,
309 const Type *RetTy, bool isVarArg,
310 unsigned CallingConv, bool isTailCall,
311 SDOperand Callee, ArgListTy &Args,
314 MachineFunction &MF = DAG.getMachineFunction();
316 unsigned NumBytes = 16;
317 unsigned outRegsUsed = 0;
319 if (Args.size() > 8) {
320 NumBytes += (Args.size() - 8) * 8;
323 outRegsUsed = Args.size();
326 // FIXME? this WILL fail if we ever try to pass around an arg that
327 // consumes more than a single output slot (a 'real' double, int128
328 // some sort of aggregate etc.), as we'll underestimate how many 'outX'
329 // registers we use. Hopefully, the assembler will notice.
330 MF.getInfo<IA64FunctionInfo>()->outRegsUsed=
331 std::max(outRegsUsed, MF.getInfo<IA64FunctionInfo>()->outRegsUsed);
333 Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
334 DAG.getConstant(NumBytes, getPointerTy()));
336 std::vector<SDOperand> args_to_use;
337 for (unsigned i = 0, e = Args.size(); i != e; ++i)
339 switch (getValueType(Args[i].second)) {
340 default: assert(0 && "unexpected argument type!");
345 //promote to 64-bits, sign/zero extending based on type
347 if(Args[i].second->isSigned())
348 Args[i].first = DAG.getNode(ISD::SIGN_EXTEND, MVT::i64,
351 Args[i].first = DAG.getNode(ISD::ZERO_EXTEND, MVT::i64,
356 Args[i].first = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Args[i].first);
361 args_to_use.push_back(Args[i].first);
364 std::vector<MVT::ValueType> RetVals;
365 MVT::ValueType RetTyVT = getValueType(RetTy);
366 if (RetTyVT != MVT::isVoid)
367 RetVals.push_back(RetTyVT);
368 RetVals.push_back(MVT::Other);
370 SDOperand TheCall = SDOperand(DAG.getCall(RetVals, Chain,
371 Callee, args_to_use), 0);
372 Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
373 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
374 DAG.getConstant(NumBytes, getPointerTy()));
375 return std::make_pair(TheCall, Chain);
379 IA64TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
380 Value *VAListV, SelectionDAG &DAG) {
381 // vastart just stores the address of the VarArgsFrameIndex slot.
382 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
383 return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR,
384 VAListP, DAG.getSrcValue(VAListV));
387 std::pair<SDOperand,SDOperand> IA64TargetLowering::
388 LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
389 const Type *ArgTy, SelectionDAG &DAG) {
391 MVT::ValueType ArgVT = getValueType(ArgTy);
392 SDOperand Val = DAG.getLoad(MVT::i64, Chain,
393 VAListP, DAG.getSrcValue(VAListV));
394 SDOperand Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), Val,
395 DAG.getSrcValue(NULL));
397 if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
400 assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
401 "Other types should have been promoted for varargs!");
404 Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
405 DAG.getConstant(Amt, Val.getValueType()));
406 Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
407 Val, VAListP, DAG.getSrcValue(VAListV));
408 return std::make_pair(Result, Chain);
413 //===--------------------------------------------------------------------===//
414 /// ISel - IA64 specific code to select IA64 machine instructions for
415 /// SelectionDAG operations.
417 class ISel : public SelectionDAGISel {
418 /// IA64Lowering - This object fully describes how to lower LLVM code to an
419 /// IA64-specific SelectionDAG.
420 IA64TargetLowering IA64Lowering;
421 SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
422 // for sdiv and udiv until it is put into the future
425 /// ExprMap - As shared expressions are codegen'd, we keep track of which
426 /// vreg the value is produced in, so we only emit one copy of each compiled
428 std::map<SDOperand, unsigned> ExprMap;
429 std::set<SDOperand> LoweredTokens;
432 ISel(TargetMachine &TM) : SelectionDAGISel(IA64Lowering), IA64Lowering(TM),
435 /// InstructionSelectBasicBlock - This callback is invoked by
436 /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
437 virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
439 unsigned SelectExpr(SDOperand N);
440 void Select(SDOperand N);
441 // a dag->dag to transform mul-by-constant-int to shifts+adds/subs
442 SDOperand BuildConstmulSequence(SDOperand N);
447 /// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
448 /// when it has created a SelectionDAG for us to codegen.
449 void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
451 // Codegen the basic block.
453 Select(DAG.getRoot());
455 // Clear state used for selection.
457 LoweredTokens.clear();
461 // strip leading '0' characters from a string
462 void munchLeadingZeros(std::string& inString) {
463 while(inString.c_str()[0]=='0') {
464 inString.erase(0, 1);
468 // strip trailing '0' characters from a string
469 void munchTrailingZeros(std::string& inString) {
470 int curPos=inString.length()-1;
472 while(inString.c_str()[curPos]=='0') {
473 inString.erase(curPos, 1);
478 // return how many consecutive '0' characters are at the end of a string
479 unsigned int countTrailingZeros(std::string& inString) {
480 int curPos=inString.length()-1;
481 unsigned int zeroCount=0;
483 while(inString.c_str()[curPos--]=='0') {
489 // booth encode a string of '1' and '0' characters (returns string of 'P' (+1)
490 // '0' and 'N' (-1) characters)
491 void boothEncode(std::string inString, std::string& boothEncodedString) {
495 int lim=inString.size();
498 if(inString[curpos]=='1') { // if we see a '1', look for a run of them
500 std::string replaceString="N";
502 // find the run length
503 for(;inString[curpos+runlength]=='1';runlength++) ;
505 for(int i=0; i<runlength-1; i++)
510 inString.replace(curpos, runlength+1, replaceString);
514 } else { // a zero, we just keep chugging along
519 // clean up (trim the string, reverse it and turn '1's into 'P's)
520 munchTrailingZeros(inString);
521 boothEncodedString="";
523 for(int i=inString.size()-1;i>=0;i--)
525 boothEncodedString+="P";
527 boothEncodedString+=inString[i];
531 struct shiftaddblob { // this encodes stuff like (x=) "A << B [+-] C << D"
532 unsigned firstVal; // A
533 unsigned firstShift; // B
534 unsigned secondVal; // C
535 unsigned secondShift; // D
539 /* this implements Lefevre's "pattern-based" constant multiplication,
540 * see "Multiplication by an Integer Constant", INRIA report 1999-06
542 * TODO: implement a method to try rewriting P0N<->0PP / N0P<->0NN
543 * to get better booth encodings - this does help in practice
544 * TODO: weight shifts appropriately (most architectures can't
545 * fuse a shift and an add for arbitrary shift amounts) */
546 unsigned lefevre(const std::string inString,
547 std::vector<struct shiftaddblob> &ops) {
548 std::string retstring;
549 std::string s = inString;
550 munchTrailingZeros(s);
552 int length=s.length()-1;
558 std::vector<int> p,n;
560 for(int i=0; i<=length; i++) {
561 if (s.c_str()[length-i]=='P') {
563 } else if (s.c_str()[length-i]=='N') {
571 std::map<const int, int> w;
573 for(unsigned i=0; i<p.size(); i++) {
574 for(unsigned j=0; j<i; j++) {
579 for(unsigned i=1; i<n.size(); i++) {
580 for(unsigned j=0; j<i; j++) {
585 for(unsigned i=0; i<p.size(); i++) {
586 for(unsigned j=0; j<n.size(); j++) {
587 w[-abs(p[i]-n[j])]++;
591 std::map<const int, int>::const_iterator ii;
593 std::multimap<int, int> sorted_by_value;
595 for(ii = w.begin(); ii!=w.end(); ii++)
596 sorted_by_value.insert(std::pair<int, int>((*ii).second,(*ii).first));
598 for (std::multimap<int, int>::iterator it = sorted_by_value.begin();
599 it != sorted_by_value.end(); ++it) {
600 d.push_back((*it).second);
606 while(d.size()>0 && (w[int_d=d.back()] > int_W)) {
614 for(unsigned base=0; base<retstring.size(); base++) {
615 if( ((base+z+1) < retstring.size()) &&
616 retstring.c_str()[base]=='P' &&
617 retstring.c_str()[base+z+1]=='P')
621 retstring.replace(base, 1, "0");
622 retstring.replace(base+z+1, 1, "p");
626 for(unsigned base=0; base<retstring.size(); base++) {
627 if( ((base+z+1) < retstring.size()) &&
628 retstring.c_str()[base]=='N' &&
629 retstring.c_str()[base+z+1]=='N')
633 retstring.replace(base, 1, "0");
634 retstring.replace(base+z+1, 1, "n");
639 for(unsigned base=0; base<retstring.size(); base++) {
640 if( ((base+z+1) < retstring.size()) &&
641 ((retstring.c_str()[base]=='P' &&
642 retstring.c_str()[base+z+1]=='N') ||
643 (retstring.c_str()[base]=='N' &&
644 retstring.c_str()[base+z+1]=='P')) ) {
648 if(retstring.c_str()[base]=='P') {
649 retstring.replace(base, 1, "0");
650 retstring.replace(base+z+1, 1, "p");
651 } else { // retstring[base]=='N'
652 retstring.replace(base, 1, "0");
653 retstring.replace(base+z+1, 1, "n");
665 } d.pop_back(); // hmm
669 for(unsigned i=0; i<t.length(); i++) {
670 if(t.c_str()[i]=='p' || t.c_str()[i]=='n')
671 t.replace(i, 1, "0");
674 for(unsigned i=0; i<u.length(); i++) {
675 if(u[i]=='P' || u[i]=='N')
676 u.replace(i, 1, "0");
678 u.replace(i, 1, "P");
680 u.replace(i, 1, "N");
693 bool hit=(u[pos]=='N');
698 for(unsigned p=0; p<u.length(); p++) {
699 bool isP=(u[p]=='P');
700 bool isN=(u[p]=='N');
703 u.replace(p, 1, "N");
705 u.replace(p, 1, "P");
709 munchLeadingZeros(u);
711 int i = lefevre(u, ops);
715 blob.firstVal=i; blob.firstShift=c;
717 blob.secondVal=i; blob.secondShift=0;
723 munchLeadingZeros(t);
728 if(t.c_str()[0]!='P') {
730 for(unsigned p=0; p<t.length(); p++) {
731 bool isP=(t.c_str()[p]=='P');
732 bool isN=(t.c_str()[p]=='N');
735 t.replace(p, 1, "N");
737 t.replace(p, 1, "P");
741 int j = lefevre(t, ops);
743 int trail=countTrailingZeros(u);
744 blob.secondVal=i; blob.secondShift=trail;
746 trail=countTrailingZeros(t);
747 blob.firstVal=j; blob.firstShift=trail;
751 blob.isSub=false; // first + second
754 blob.isSub=true; // first - second
757 blob.isSub=true; // second - first
758 int tmpval, tmpshift;
759 tmpval=blob.firstVal;
760 tmpshift=blob.firstShift;
761 blob.firstVal=blob.secondVal;
762 blob.firstShift=blob.secondShift;
763 blob.secondVal=tmpval;
764 blob.secondShift=tmpshift;
773 SDOperand ISel::BuildConstmulSequence(SDOperand N) {
774 //FIXME: we should shortcut this stuff for multiplies by 2^n+1
775 // in particular, *3 is nicer as *2+1, not *4-1
776 int64_t constant=cast<ConstantSDNode>(N.getOperand(1))->getValue();
779 unsigned preliminaryShift=0;
781 assert(constant != 0 && "erk, you're trying to multiply by constant zero\n");
783 // first, we make the constant to multiply by positive
791 // next, we make it odd.
792 for(; (constant%2==0); preliminaryShift++)
795 //OK, we have a positive, odd number of 64 bits or less. Convert it
796 //to a binary string, constantString[0] is the LSB
797 char constantString[65];
798 for(int i=0; i<64; i++)
799 constantString[i]='0'+((constant>>i)&0x1);
800 constantString[64]=0;
802 // now, Booth encode it
803 std::string boothEncodedString;
804 boothEncode(constantString, boothEncodedString);
806 std::vector<struct shiftaddblob> ops;
807 // do the transformation, filling out 'ops'
808 lefevre(boothEncodedString, ops);
810 assert(ops.size() < 80 && "constmul code has gone haywire\n");
811 SDOperand results[80]; // temporary results (of adds/subs of shifts)
813 // now turn 'ops' into DAG bits
814 for(unsigned i=0; i<ops.size(); i++) {
815 SDOperand amt = ISelDAG->getConstant(ops[i].firstShift, MVT::i64);
816 SDOperand val = (ops[i].firstVal == 0) ? N.getOperand(0) :
817 results[ops[i].firstVal-1];
818 SDOperand left = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
819 amt = ISelDAG->getConstant(ops[i].secondShift, MVT::i64);
820 val = (ops[i].secondVal == 0) ? N.getOperand(0) :
821 results[ops[i].secondVal-1];
822 SDOperand right = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
824 results[i] = ISelDAG->getNode(ISD::SUB, MVT::i64, left, right);
826 results[i] = ISelDAG->getNode(ISD::ADD, MVT::i64, left, right);
829 // don't forget flippedSign and preliminaryShift!
830 SDOperand shiftedresult;
831 if(preliminaryShift) {
832 SDOperand finalshift = ISelDAG->getConstant(preliminaryShift, MVT::i64);
833 shiftedresult = ISelDAG->getNode(ISD::SHL, MVT::i64,
834 results[ops.size()-1], finalshift);
835 } else { // there was no preliminary divide-by-power-of-2 required
836 shiftedresult = results[ops.size()-1];
839 SDOperand finalresult;
840 if(flippedSign) { // if we were multiplying by a negative constant:
841 SDOperand zero = ISelDAG->getConstant(0, MVT::i64);
842 // subtract the result from 0 to flip its sign
843 finalresult = ISelDAG->getNode(ISD::SUB, MVT::i64, zero, shiftedresult);
844 } else { // there was no preliminary multiply by -1 required
845 finalresult = shiftedresult;
851 /// ponderIntegerDivisionBy - When handling integer divides, if the divide
852 /// is by a constant such that we can efficiently codegen it, this
853 /// function says what to do. Currently, it returns 0 if the division must
854 /// become a genuine divide, and 1 if the division can be turned into a
856 static unsigned ponderIntegerDivisionBy(SDOperand N, bool isSigned,
858 if (N.getOpcode() != ISD::Constant) return 0; // if not a divide by
859 // a constant, give up.
861 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
863 if (isPowerOf2_64(v)) { // if a division by a power of two, say so
868 return 0; // fallthrough
871 static unsigned ponderIntegerAndWith(SDOperand N, unsigned& Imm) {
872 if (N.getOpcode() != ISD::Constant) return 0; // if not ANDing with
873 // a constant, give up.
875 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
877 if (isMask_64(v)) { // if ANDing with ((2^n)-1) for some n
882 return 0; // fallthrough
885 static unsigned ponderIntegerAdditionWith(SDOperand N, unsigned& Imm) {
886 if (N.getOpcode() != ISD::Constant) return 0; // if not adding a
887 // constant, give up.
888 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
890 if (v <= 8191 && v >= -8192) { // if this constants fits in 14 bits, say so
891 Imm = v & 0x3FFF; // 14 bits
894 return 0; // fallthrough
897 static unsigned ponderIntegerSubtractionFrom(SDOperand N, unsigned& Imm) {
898 if (N.getOpcode() != ISD::Constant) return 0; // if not subtracting a
899 // constant, give up.
900 int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
902 if (v <= 127 && v >= -128) { // if this constants fits in 8 bits, say so
903 Imm = v & 0xFF; // 8 bits
906 return 0; // fallthrough
909 unsigned ISel::SelectExpr(SDOperand N) {
911 unsigned Tmp1, Tmp2, Tmp3;
913 MVT::ValueType DestType = N.getValueType();
915 unsigned opcode = N.getOpcode();
917 SDNode *Node = N.Val;
920 if (Node->getOpcode() == ISD::CopyFromReg)
921 // Just use the specified register as our input.
922 return dyn_cast<RegSDNode>(Node)->getReg();
924 unsigned &Reg = ExprMap[N];
927 if (N.getOpcode() != ISD::CALL && N.getOpcode() != ISD::TAILCALL)
928 Reg = Result = (N.getValueType() != MVT::Other) ?
929 MakeReg(N.getValueType()) : 1;
931 // If this is a call instruction, make sure to prepare ALL of the result
932 // values as well as the chain.
933 if (Node->getNumValues() == 1)
934 Reg = Result = 1; // Void call, just a chain.
936 Result = MakeReg(Node->getValueType(0));
937 ExprMap[N.getValue(0)] = Result;
938 for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
939 ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
940 ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
944 switch (N.getOpcode()) {
947 assert(0 && "Node not handled!\n");
949 case ISD::FrameIndex: {
950 Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
951 BuildMI(BB, IA64::MOV, 1, Result).addFrameIndex(Tmp1);
955 case ISD::ConstantPool: {
956 Tmp1 = cast<ConstantPoolSDNode>(N)->getIndex();
957 IA64Lowering.restoreGP(BB); // FIXME: do i really need this?
958 BuildMI(BB, IA64::ADD, 2, Result).addConstantPoolIndex(Tmp1)
963 case ISD::ConstantFP: {
964 Tmp1 = Result; // Intermediate Register
965 if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
966 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
967 Tmp1 = MakeReg(MVT::f64);
969 if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
970 cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
971 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F0); // load 0.0
972 else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
973 cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
974 BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F1); // load 1.0
976 assert(0 && "Unexpected FP constant!");
978 // we multiply by +1.0, negate (this is FNMA), and then add 0.0
979 BuildMI(BB, IA64::FNMA, 3, Result).addReg(Tmp1).addReg(IA64::F1)
984 case ISD::DYNAMIC_STACKALLOC: {
985 // Generate both result values.
987 ExprMap[N.getValue(1)] = 1; // Generate the token
989 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
991 // FIXME: We are currently ignoring the requested alignment for handling
992 // greater than the stack alignment. This will need to be revisited at some
993 // point. Align = N.getOperand(2);
995 if (!isa<ConstantSDNode>(N.getOperand(2)) ||
996 cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
997 std::cerr << "Cannot allocate stack object with greater alignment than"
998 << " the stack alignment yet!";
1003 Select(N.getOperand(0));
1004 if (ConstantSDNode* CN = dyn_cast<ConstantSDNode>(N.getOperand(1)))
1006 if (CN->getValue() < 32000)
1008 BuildMI(BB, IA64::ADDIMM22, 2, IA64::r12).addReg(IA64::r12)
1009 .addImm(-CN->getValue());
1011 Tmp1 = SelectExpr(N.getOperand(1));
1012 // Subtract size from stack pointer, thereby allocating some space.
1013 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1016 Tmp1 = SelectExpr(N.getOperand(1));
1017 // Subtract size from stack pointer, thereby allocating some space.
1018 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1021 Select(N.getOperand(0));
1022 Tmp1 = SelectExpr(N.getOperand(1));
1023 // Subtract size from stack pointer, thereby allocating some space.
1024 BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
1025 // Put a pointer to the space into the result register, by copying the
1027 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r12);
1032 Tmp1 = SelectExpr(N.getOperand(0)); //Cond
1033 Tmp2 = SelectExpr(N.getOperand(1)); //Use if TRUE
1034 Tmp3 = SelectExpr(N.getOperand(2)); //Use if FALSE
1036 unsigned bogoResult;
1038 switch (N.getOperand(1).getValueType()) {
1039 default: assert(0 &&
1040 "ISD::SELECT: 'select'ing something other than i1, i64 or f64!\n");
1041 // for i1, we load the condition into an integer register, then
1042 // conditionally copy Tmp2 and Tmp3 to Tmp1 in parallel (only one
1043 // of them will go through, since the integer register will hold
1046 bogoResult=MakeReg(MVT::i1);
1048 // load the condition into an integer register
1049 unsigned condReg=MakeReg(MVT::i64);
1050 unsigned dummy=MakeReg(MVT::i64);
1051 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1052 BuildMI(BB, IA64::TPCADDIMM22, 2, condReg).addReg(dummy)
1053 .addImm(1).addReg(Tmp1);
1055 // initialize Result (bool) to false (hence UNC) and if
1056 // the select condition (condReg) is false (0), copy Tmp3
1057 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoResult)
1058 .addReg(condReg).addReg(IA64::r0).addReg(Tmp3);
1060 // now, if the selection condition is true, write 1 to the
1061 // result if Tmp2 is 1
1062 BuildMI(BB, IA64::TPCMPNE, 3, Result).addReg(bogoResult)
1063 .addReg(condReg).addReg(IA64::r0).addReg(Tmp2);
1066 // for i64/f64, we just copy Tmp3 and then conditionally overwrite it
1067 // with Tmp2 if Tmp1 is true
1069 bogoResult=MakeReg(MVT::i64);
1070 BuildMI(BB, IA64::MOV, 1, bogoResult).addReg(Tmp3);
1071 BuildMI(BB, IA64::CMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1075 bogoResult=MakeReg(MVT::f64);
1076 BuildMI(BB, IA64::FMOV, 1, bogoResult).addReg(Tmp3);
1077 BuildMI(BB, IA64::CFMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
1085 case ISD::Constant: {
1086 unsigned depositPos=0;
1087 unsigned depositLen=0;
1088 switch (N.getValueType()) {
1089 default: assert(0 && "Cannot use constants of this type!");
1090 case MVT::i1: { // if a bool, we don't 'load' so much as generate
1092 if(cast<ConstantSDNode>(N)->getValue()) // true:
1093 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1095 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
1096 return Result; // early exit
1098 case MVT::i64: break;
1101 int64_t immediate = cast<ConstantSDNode>(N)->getValue();
1103 if(immediate==0) { // if the constant is just zero,
1104 BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r0); // just copy r0
1105 return Result; // early exit
1108 if (immediate <= 8191 && immediate >= -8192) {
1109 // if this constants fits in 14 bits, we use a mov the assembler will
1110 // turn into: "adds rDest=imm,r0" (and _not_ "andl"...)
1111 BuildMI(BB, IA64::MOVSIMM14, 1, Result).addSImm(immediate);
1112 return Result; // early exit
1115 if (immediate <= 2097151 && immediate >= -2097152) {
1116 // if this constants fits in 22 bits, we use a mov the assembler will
1117 // turn into: "addl rDest=imm,r0"
1118 BuildMI(BB, IA64::MOVSIMM22, 1, Result).addSImm(immediate);
1119 return Result; // early exit
1122 /* otherwise, our immediate is big, so we use movl */
1123 uint64_t Imm = immediate;
1124 BuildMI(BB, IA64::MOVLIMM64, 1, Result).addImm64(Imm);
1129 BuildMI(BB, IA64::IDEF, 0, Result);
1133 case ISD::GlobalAddress: {
1134 GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
1135 unsigned Tmp1 = MakeReg(MVT::i64);
1137 BuildMI(BB, IA64::ADD, 2, Tmp1).addGlobalAddress(GV).addReg(IA64::r1);
1138 BuildMI(BB, IA64::LD8, 1, Result).addReg(Tmp1);
1143 case ISD::ExternalSymbol: {
1144 const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
1145 // assert(0 && "sorry, but what did you want an ExternalSymbol for again?");
1146 BuildMI(BB, IA64::MOV, 1, Result).addExternalSymbol(Sym); // XXX
1150 case ISD::FP_EXTEND: {
1151 Tmp1 = SelectExpr(N.getOperand(0));
1152 BuildMI(BB, IA64::FMOV, 1, Result).addReg(Tmp1);
1156 case ISD::ZERO_EXTEND: {
1157 Tmp1 = SelectExpr(N.getOperand(0)); // value
1159 switch (N.getOperand(0).getValueType()) {
1160 default: assert(0 && "Cannot zero-extend this type!");
1161 case MVT::i8: Opc = IA64::ZXT1; break;
1162 case MVT::i16: Opc = IA64::ZXT2; break;
1163 case MVT::i32: Opc = IA64::ZXT4; break;
1165 // we handle bools differently! :
1166 case MVT::i1: { // if the predicate reg has 1, we want a '1' in our GR.
1167 unsigned dummy = MakeReg(MVT::i64);
1169 BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
1170 // ...then conditionally (PR:Tmp1) add 1:
1171 BuildMI(BB, IA64::TPCADDIMM22, 2, Result).addReg(dummy)
1172 .addImm(1).addReg(Tmp1);
1173 return Result; // XXX early exit!
1177 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1181 case ISD::SIGN_EXTEND: { // we should only have to handle i1 -> i64 here!!!
1183 assert(0 && "hmm, ISD::SIGN_EXTEND: shouldn't ever be reached. bad luck!\n");
1185 Tmp1 = SelectExpr(N.getOperand(0)); // value
1187 switch (N.getOperand(0).getValueType()) {
1188 default: assert(0 && "Cannot sign-extend this type!");
1189 case MVT::i1: assert(0 && "trying to sign extend a bool? ow.\n");
1190 Opc = IA64::SXT1; break;
1191 // FIXME: for now, we treat bools the same as i8s
1192 case MVT::i8: Opc = IA64::SXT1; break;
1193 case MVT::i16: Opc = IA64::SXT2; break;
1194 case MVT::i32: Opc = IA64::SXT4; break;
1197 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1201 case ISD::TRUNCATE: {
1202 // we use the funky dep.z (deposit (zero)) instruction to deposit bits
1203 // of R0 appropriately.
1204 switch (N.getOperand(0).getValueType()) {
1205 default: assert(0 && "Unknown truncate!");
1206 case MVT::i64: break;
1208 Tmp1 = SelectExpr(N.getOperand(0));
1209 unsigned depositPos, depositLen;
1211 switch (N.getValueType()) {
1212 default: assert(0 && "Unknown truncate!");
1214 // if input (normal reg) is 0, 0!=0 -> false (0), if 1, 1!=0 ->true (1):
1215 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1)
1217 return Result; // XXX early exit!
1219 case MVT::i8: depositPos=0; depositLen=8; break;
1220 case MVT::i16: depositPos=0; depositLen=16; break;
1221 case MVT::i32: depositPos=0; depositLen=32; break;
1223 BuildMI(BB, IA64::DEPZ, 1, Result).addReg(Tmp1)
1224 .addImm(depositPos).addImm(depositLen);
1229 case ISD::FP_ROUND: {
1230 assert (DestType == MVT::f32 && N.getOperand(0).getValueType() == MVT::f64 &&
1231 "error: trying to FP_ROUND something other than f64 -> f32!\n");
1232 Tmp1 = SelectExpr(N.getOperand(0));
1233 BuildMI(BB, IA64::FADDS, 2, Result).addReg(Tmp1).addReg(IA64::F0);
1234 // we add 0.0 using a single precision add to do rounding
1239 // FIXME: the following 4 cases need cleaning
1240 case ISD::SINT_TO_FP: {
1241 Tmp1 = SelectExpr(N.getOperand(0));
1242 Tmp2 = MakeReg(MVT::f64);
1243 unsigned dummy = MakeReg(MVT::f64);
1244 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1245 BuildMI(BB, IA64::FCVTXF, 1, dummy).addReg(Tmp2);
1246 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1250 case ISD::UINT_TO_FP: {
1251 Tmp1 = SelectExpr(N.getOperand(0));
1252 Tmp2 = MakeReg(MVT::f64);
1253 unsigned dummy = MakeReg(MVT::f64);
1254 BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
1255 BuildMI(BB, IA64::FCVTXUF, 1, dummy).addReg(Tmp2);
1256 BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
1260 case ISD::FP_TO_SINT: {
1261 Tmp1 = SelectExpr(N.getOperand(0));
1262 Tmp2 = MakeReg(MVT::f64);
1263 BuildMI(BB, IA64::FCVTFXTRUNC, 1, Tmp2).addReg(Tmp1);
1264 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1268 case ISD::FP_TO_UINT: {
1269 Tmp1 = SelectExpr(N.getOperand(0));
1270 Tmp2 = MakeReg(MVT::f64);
1271 BuildMI(BB, IA64::FCVTFXUTRUNC, 1, Tmp2).addReg(Tmp1);
1272 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
1277 if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
1278 N.getOperand(0).Val->hasOneUse()) { // if we can fold this add
1279 // into an fma, do so:
1280 // ++FusedFP; // Statistic
1281 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1282 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1283 Tmp3 = SelectExpr(N.getOperand(1));
1284 BuildMI(BB, IA64::FMA, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1285 return Result; // early exit
1288 if(DestType != MVT::f64 && N.getOperand(0).getOpcode() == ISD::SHL &&
1289 N.getOperand(0).Val->hasOneUse()) { // if we might be able to fold
1290 // this add into a shladd, try:
1291 ConstantSDNode *CSD = NULL;
1292 if((CSD = dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) &&
1293 (CSD->getValue() >= 1) && (CSD->getValue() <= 4) ) { // we can:
1295 // ++FusedSHLADD; // Statistic
1296 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1297 int shl_amt = CSD->getValue();
1298 Tmp3 = SelectExpr(N.getOperand(1));
1300 BuildMI(BB, IA64::SHLADD, 3, Result)
1301 .addReg(Tmp1).addImm(shl_amt).addReg(Tmp3);
1302 return Result; // early exit
1306 //else, fallthrough:
1307 Tmp1 = SelectExpr(N.getOperand(0));
1308 if(DestType != MVT::f64) { // integer addition:
1309 switch (ponderIntegerAdditionWith(N.getOperand(1), Tmp3)) {
1310 case 1: // adding a constant that's 14 bits
1311 BuildMI(BB, IA64::ADDIMM14, 2, Result).addReg(Tmp1).addSImm(Tmp3);
1312 return Result; // early exit
1313 } // fallthrough and emit a reg+reg ADD:
1314 Tmp2 = SelectExpr(N.getOperand(1));
1315 BuildMI(BB, IA64::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1316 } else { // this is a floating point addition
1317 Tmp2 = SelectExpr(N.getOperand(1));
1318 BuildMI(BB, IA64::FADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
1325 if(DestType != MVT::f64) { // TODO: speed!
1326 /* FIXME if(N.getOperand(1).getOpcode() != ISD::Constant) { // if not a const mul
1328 // boring old integer multiply with xma
1329 Tmp1 = SelectExpr(N.getOperand(0));
1330 Tmp2 = SelectExpr(N.getOperand(1));
1332 unsigned TempFR1=MakeReg(MVT::f64);
1333 unsigned TempFR2=MakeReg(MVT::f64);
1334 unsigned TempFR3=MakeReg(MVT::f64);
1335 BuildMI(BB, IA64::SETFSIG, 1, TempFR1).addReg(Tmp1);
1336 BuildMI(BB, IA64::SETFSIG, 1, TempFR2).addReg(Tmp2);
1337 BuildMI(BB, IA64::XMAL, 1, TempFR3).addReg(TempFR1).addReg(TempFR2)
1339 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TempFR3);
1340 return Result; // early exit
1341 /* FIXME } else { // we are multiplying by an integer constant! yay
1342 return Reg = SelectExpr(BuildConstmulSequence(N)); // avert your eyes!
1345 else { // floating point multiply
1346 Tmp1 = SelectExpr(N.getOperand(0));
1347 Tmp2 = SelectExpr(N.getOperand(1));
1348 BuildMI(BB, IA64::FMPY, 2, Result).addReg(Tmp1).addReg(Tmp2);
1354 if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
1355 N.getOperand(0).Val->hasOneUse()) { // if we can fold this sub
1356 // into an fms, do so:
1357 // ++FusedFP; // Statistic
1358 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1359 Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
1360 Tmp3 = SelectExpr(N.getOperand(1));
1361 BuildMI(BB, IA64::FMS, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
1362 return Result; // early exit
1364 Tmp2 = SelectExpr(N.getOperand(1));
1365 if(DestType != MVT::f64) { // integer subtraction:
1366 switch (ponderIntegerSubtractionFrom(N.getOperand(0), Tmp3)) {
1367 case 1: // subtracting *from* an 8 bit constant:
1368 BuildMI(BB, IA64::SUBIMM8, 2, Result).addSImm(Tmp3).addReg(Tmp2);
1369 return Result; // early exit
1370 } // fallthrough and emit a reg+reg SUB:
1371 Tmp1 = SelectExpr(N.getOperand(0));
1372 BuildMI(BB, IA64::SUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1373 } else { // this is a floating point subtraction
1374 Tmp1 = SelectExpr(N.getOperand(0));
1375 BuildMI(BB, IA64::FSUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
1381 Tmp1 = SelectExpr(N.getOperand(0));
1382 assert(DestType == MVT::f64 && "trying to fabs something other than f64?");
1383 BuildMI(BB, IA64::FABS, 1, Result).addReg(Tmp1);
1388 assert(DestType == MVT::f64 && "trying to fneg something other than f64?");
1390 if (ISD::FABS == N.getOperand(0).getOpcode()) { // && hasOneUse()?
1391 Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
1392 BuildMI(BB, IA64::FNEGABS, 1, Result).addReg(Tmp1); // fold in abs
1394 Tmp1 = SelectExpr(N.getOperand(0));
1395 BuildMI(BB, IA64::FNEG, 1, Result).addReg(Tmp1); // plain old fneg
1402 switch (N.getValueType()) {
1403 default: assert(0 && "Cannot AND this type!");
1404 case MVT::i1: { // if a bool, we emit a pseudocode AND
1405 unsigned pA = SelectExpr(N.getOperand(0));
1406 unsigned pB = SelectExpr(N.getOperand(1));
1408 /* our pseudocode for AND is:
1410 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1411 cmp.eq pTemp,p0 = r0,r0 // pTemp = NOT pB
1413 (pB) cmp.ne pTemp,p0 = r0,r0
1415 (pTemp)cmp.ne pC,p0 = r0,r0 // if (NOT pB) pC = 0
1418 unsigned pTemp = MakeReg(MVT::i1);
1420 unsigned bogusTemp1 = MakeReg(MVT::i1);
1421 unsigned bogusTemp2 = MakeReg(MVT::i1);
1422 unsigned bogusTemp3 = MakeReg(MVT::i1);
1423 unsigned bogusTemp4 = MakeReg(MVT::i1);
1425 BuildMI(BB, IA64::PCMPEQUNC, 3, bogusTemp1)
1426 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1427 BuildMI(BB, IA64::CMPEQ, 2, bogusTemp2)
1428 .addReg(IA64::r0).addReg(IA64::r0);
1429 BuildMI(BB, IA64::TPCMPNE, 3, pTemp)
1430 .addReg(bogusTemp2).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1431 BuildMI(BB, IA64::TPCMPNE, 3, Result)
1432 .addReg(bogusTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pTemp);
1436 // if not a bool, we just AND away:
1441 Tmp1 = SelectExpr(N.getOperand(0));
1442 switch (ponderIntegerAndWith(N.getOperand(1), Tmp3)) {
1443 case 1: // ANDing a constant that is 2^n-1 for some n
1445 case 8: // if AND 0x00000000000000FF, be quaint and use zxt1
1446 BuildMI(BB, IA64::ZXT1, 1, Result).addReg(Tmp1);
1448 case 16: // if AND 0x000000000000FFFF, be quaint and use zxt2
1449 BuildMI(BB, IA64::ZXT2, 1, Result).addReg(Tmp1);
1451 case 32: // if AND 0x00000000FFFFFFFF, be quaint and use zxt4
1452 BuildMI(BB, IA64::ZXT4, 1, Result).addReg(Tmp1);
1454 default: // otherwise, use dep.z to paste zeros
1455 BuildMI(BB, IA64::DEPZ, 3, Result).addReg(Tmp1)
1456 .addImm(0).addImm(Tmp3);
1459 return Result; // early exit
1460 } // fallthrough and emit a simple AND:
1461 Tmp2 = SelectExpr(N.getOperand(1));
1462 BuildMI(BB, IA64::AND, 2, Result).addReg(Tmp1).addReg(Tmp2);
1469 switch (N.getValueType()) {
1470 default: assert(0 && "Cannot OR this type!");
1471 case MVT::i1: { // if a bool, we emit a pseudocode OR
1472 unsigned pA = SelectExpr(N.getOperand(0));
1473 unsigned pB = SelectExpr(N.getOperand(1));
1475 unsigned pTemp1 = MakeReg(MVT::i1);
1477 /* our pseudocode for OR is:
1483 (pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
1485 (pB) cmp.eq pC,p0 = r0,r0 // if (pB) pC = 1
1488 BuildMI(BB, IA64::PCMPEQUNC, 3, pTemp1)
1489 .addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
1490 BuildMI(BB, IA64::TPCMPEQ, 3, Result)
1491 .addReg(pTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
1494 // if not a bool, we just OR away:
1499 Tmp1 = SelectExpr(N.getOperand(0));
1500 Tmp2 = SelectExpr(N.getOperand(1));
1501 BuildMI(BB, IA64::OR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1509 switch (N.getValueType()) {
1510 default: assert(0 && "Cannot XOR this type!");
1511 case MVT::i1: { // if a bool, we emit a pseudocode XOR
1512 unsigned pY = SelectExpr(N.getOperand(0));
1513 unsigned pZ = SelectExpr(N.getOperand(1));
1515 /* one possible routine for XOR is:
1517 // Compute px = py ^ pz
1518 // using sum of products: px = (py & !pz) | (pz & !py)
1519 // Uses 5 instructions in 3 cycles.
1521 (pz) cmp.eq.unc px = r0, r0 // px = pz
1522 (py) cmp.eq.unc pt = r0, r0 // pt = py
1525 (pt) cmp.ne.and px = r0, r0 // px = px & !pt (px = pz & !pt)
1526 (pz) cmp.ne.and pt = r0, r0 // pt = pt & !pz
1530 (pt) cmp.eq.or px = r0, r0 // px = px | pt
1532 *** Another, which we use here, requires one scratch GR. it is:
1534 mov rt = 0 // initialize rt off critical path
1538 (pz) cmp.eq.unc px = r0, r0 // px = pz
1539 (pz) mov rt = 1 // rt = pz
1542 (py) cmp.ne px = 1, rt // if (py) px = !pz
1544 .. these routines kindly provided by Jim Hull
1546 unsigned rt = MakeReg(MVT::i64);
1548 // these two temporaries will never actually appear,
1549 // due to the two-address form of some of the instructions below
1550 unsigned bogoPR = MakeReg(MVT::i1); // becomes Result
1551 unsigned bogoGR = MakeReg(MVT::i64); // becomes rt
1553 BuildMI(BB, IA64::MOV, 1, bogoGR).addReg(IA64::r0);
1554 BuildMI(BB, IA64::PCMPEQUNC, 3, bogoPR)
1555 .addReg(IA64::r0).addReg(IA64::r0).addReg(pZ);
1556 BuildMI(BB, IA64::TPCADDIMM22, 2, rt)
1557 .addReg(bogoGR).addImm(1).addReg(pZ);
1558 BuildMI(BB, IA64::TPCMPIMM8NE, 3, Result)
1559 .addReg(bogoPR).addImm(1).addReg(rt).addReg(pY);
1562 // if not a bool, we just XOR away:
1567 Tmp1 = SelectExpr(N.getOperand(0));
1568 Tmp2 = SelectExpr(N.getOperand(1));
1569 BuildMI(BB, IA64::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
1577 Tmp1 = SelectExpr(N.getOperand(0));
1578 BuildMI(BB, IA64::POPCNT, 1, Result).addReg(Tmp1);
1583 Tmp1 = SelectExpr(N.getOperand(0));
1584 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1585 Tmp2 = CN->getValue();
1586 BuildMI(BB, IA64::SHLI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1588 Tmp2 = SelectExpr(N.getOperand(1));
1589 BuildMI(BB, IA64::SHL, 2, Result).addReg(Tmp1).addReg(Tmp2);
1595 Tmp1 = SelectExpr(N.getOperand(0));
1596 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1597 Tmp2 = CN->getValue();
1598 BuildMI(BB, IA64::SHRUI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1600 Tmp2 = SelectExpr(N.getOperand(1));
1601 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1607 Tmp1 = SelectExpr(N.getOperand(0));
1608 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1609 Tmp2 = CN->getValue();
1610 BuildMI(BB, IA64::SHRSI, 2, Result).addReg(Tmp1).addImm(Tmp2);
1612 Tmp2 = SelectExpr(N.getOperand(1));
1613 BuildMI(BB, IA64::SHRS, 2, Result).addReg(Tmp1).addReg(Tmp2);
1623 Tmp1 = SelectExpr(N.getOperand(0));
1624 Tmp2 = SelectExpr(N.getOperand(1));
1628 if(DestType == MVT::f64) // XXX: we're not gonna be fed MVT::f32, are we?
1631 bool isModulus=false; // is it a division or a modulus?
1632 bool isSigned=false;
1634 switch(N.getOpcode()) {
1635 case ISD::SDIV: isModulus=false; isSigned=true; break;
1636 case ISD::UDIV: isModulus=false; isSigned=false; break;
1637 case ISD::SREM: isModulus=true; isSigned=true; break;
1638 case ISD::UREM: isModulus=true; isSigned=false; break;
1641 if(!isModulus && !isFP) { // if this is an integer divide,
1642 switch (ponderIntegerDivisionBy(N.getOperand(1), isSigned, Tmp3)) {
1643 case 1: // division by a constant that's a power of 2
1644 Tmp1 = SelectExpr(N.getOperand(0));
1645 if(isSigned) { // argument could be negative, so emit some code:
1646 unsigned divAmt=Tmp3;
1647 unsigned tempGR1=MakeReg(MVT::i64);
1648 unsigned tempGR2=MakeReg(MVT::i64);
1649 unsigned tempGR3=MakeReg(MVT::i64);
1650 BuildMI(BB, IA64::SHRS, 2, tempGR1)
1651 .addReg(Tmp1).addImm(divAmt-1);
1652 BuildMI(BB, IA64::EXTRU, 3, tempGR2)
1653 .addReg(tempGR1).addImm(64-divAmt).addImm(divAmt);
1654 BuildMI(BB, IA64::ADD, 2, tempGR3)
1655 .addReg(Tmp1).addReg(tempGR2);
1656 BuildMI(BB, IA64::SHRS, 2, Result)
1657 .addReg(tempGR3).addImm(divAmt);
1659 else // unsigned div-by-power-of-2 becomes a simple shift right:
1660 BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addImm(Tmp3);
1661 return Result; // early exit
1665 unsigned TmpPR=MakeReg(MVT::i1); // we need two scratch
1666 unsigned TmpPR2=MakeReg(MVT::i1); // predicate registers,
1667 unsigned TmpF1=MakeReg(MVT::f64); // and one metric truckload of FP regs.
1668 unsigned TmpF2=MakeReg(MVT::f64); // lucky we have IA64?
1669 unsigned TmpF3=MakeReg(MVT::f64); // well, the real FIXME is to have
1670 unsigned TmpF4=MakeReg(MVT::f64); // isTwoAddress forms of these
1671 unsigned TmpF5=MakeReg(MVT::f64); // FP instructions so we can end up with
1672 unsigned TmpF6=MakeReg(MVT::f64); // stuff like setf.sig f10=f10 etc.
1673 unsigned TmpF7=MakeReg(MVT::f64);
1674 unsigned TmpF8=MakeReg(MVT::f64);
1675 unsigned TmpF9=MakeReg(MVT::f64);
1676 unsigned TmpF10=MakeReg(MVT::f64);
1677 unsigned TmpF11=MakeReg(MVT::f64);
1678 unsigned TmpF12=MakeReg(MVT::f64);
1679 unsigned TmpF13=MakeReg(MVT::f64);
1680 unsigned TmpF14=MakeReg(MVT::f64);
1681 unsigned TmpF15=MakeReg(MVT::f64);
1683 // OK, emit some code:
1686 // first, load the inputs into FP regs.
1687 BuildMI(BB, IA64::SETFSIG, 1, TmpF1).addReg(Tmp1);
1688 BuildMI(BB, IA64::SETFSIG, 1, TmpF2).addReg(Tmp2);
1690 // next, convert the inputs to FP
1692 BuildMI(BB, IA64::FCVTXF, 1, TmpF3).addReg(TmpF1);
1693 BuildMI(BB, IA64::FCVTXF, 1, TmpF4).addReg(TmpF2);
1695 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF3).addReg(TmpF1);
1696 BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF4).addReg(TmpF2);
1699 } else { // this is an FP divide/remainder, so we 'leak' some temp
1700 // regs and assign TmpF3=Tmp1, TmpF4=Tmp2
1705 // we start by computing an approximate reciprocal (good to 9 bits?)
1706 // note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
1707 BuildMI(BB, IA64::FRCPAS1, 4)
1708 .addReg(TmpF5, MachineOperand::Def)
1709 .addReg(TmpPR, MachineOperand::Def)
1710 .addReg(TmpF3).addReg(TmpF4);
1712 if(!isModulus) { // if this is a divide, we worry about div-by-zero
1713 unsigned bogusPR=MakeReg(MVT::i1); // won't appear, due to twoAddress
1715 BuildMI(BB, IA64::CMPEQ, 2, bogusPR).addReg(IA64::r0).addReg(IA64::r0);
1716 BuildMI(BB, IA64::TPCMPNE, 3, TmpPR2).addReg(bogusPR)
1717 .addReg(IA64::r0).addReg(IA64::r0).addReg(TmpPR);
1720 // now we apply newton's method, thrice! (FIXME: this is ~72 bits of
1721 // precision, don't need this much for f32/i32)
1722 BuildMI(BB, IA64::CFNMAS1, 4, TmpF6)
1723 .addReg(TmpF4).addReg(TmpF5).addReg(IA64::F1).addReg(TmpPR);
1724 BuildMI(BB, IA64::CFMAS1, 4, TmpF7)
1725 .addReg(TmpF3).addReg(TmpF5).addReg(IA64::F0).addReg(TmpPR);
1726 BuildMI(BB, IA64::CFMAS1, 4, TmpF8)
1727 .addReg(TmpF6).addReg(TmpF6).addReg(IA64::F0).addReg(TmpPR);
1728 BuildMI(BB, IA64::CFMAS1, 4, TmpF9)
1729 .addReg(TmpF6).addReg(TmpF7).addReg(TmpF7).addReg(TmpPR);
1730 BuildMI(BB, IA64::CFMAS1, 4,TmpF10)
1731 .addReg(TmpF6).addReg(TmpF5).addReg(TmpF5).addReg(TmpPR);
1732 BuildMI(BB, IA64::CFMAS1, 4,TmpF11)
1733 .addReg(TmpF8).addReg(TmpF9).addReg(TmpF9).addReg(TmpPR);
1734 BuildMI(BB, IA64::CFMAS1, 4,TmpF12)
1735 .addReg(TmpF8).addReg(TmpF10).addReg(TmpF10).addReg(TmpPR);
1736 BuildMI(BB, IA64::CFNMAS1, 4,TmpF13)
1737 .addReg(TmpF4).addReg(TmpF11).addReg(TmpF3).addReg(TmpPR);
1739 // FIXME: this is unfortunate :(
1740 // the story is that the dest reg of the fnma above and the fma below
1741 // (and therefore possibly the src of the fcvt.fx[u] as well) cannot
1742 // be the same register, or this code breaks if the first argument is
1743 // zero. (e.g. without this hack, 0%8 yields -64, not 0.)
1744 BuildMI(BB, IA64::CFMAS1, 4,TmpF14)
1745 .addReg(TmpF13).addReg(TmpF12).addReg(TmpF11).addReg(TmpPR);
1747 if(isModulus) { // XXX: fragile! fixes _only_ mod, *breaks* div! !
1748 BuildMI(BB, IA64::IUSE, 1).addReg(TmpF13); // hack :(
1752 // round to an integer
1754 BuildMI(BB, IA64::FCVTFXTRUNCS1, 1, TmpF15).addReg(TmpF14);
1756 BuildMI(BB, IA64::FCVTFXUTRUNCS1, 1, TmpF15).addReg(TmpF14);
1758 BuildMI(BB, IA64::FMOV, 1, TmpF15).addReg(TmpF14);
1759 // EXERCISE: can you see why TmpF15=TmpF14 does not work here, and
1760 // we really do need the above FMOV? ;)
1764 if(isFP) { // extra worrying about div-by-zero
1765 unsigned bogoResult=MakeReg(MVT::f64);
1767 // we do a 'conditional fmov' (of the correct result, depending
1768 // on how the frcpa predicate turned out)
1769 BuildMI(BB, IA64::PFMOV, 2, bogoResult)
1770 .addReg(TmpF12).addReg(TmpPR2);
1771 BuildMI(BB, IA64::CFMOV, 2, Result)
1772 .addReg(bogoResult).addReg(TmpF15).addReg(TmpPR);
1775 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TmpF15);
1777 } else { // this is a modulus
1779 // answer = q * (-b) + a
1780 unsigned ModulusResult = MakeReg(MVT::f64);
1781 unsigned TmpF = MakeReg(MVT::f64);
1782 unsigned TmpI = MakeReg(MVT::i64);
1784 BuildMI(BB, IA64::SUB, 2, TmpI).addReg(IA64::r0).addReg(Tmp2);
1785 BuildMI(BB, IA64::SETFSIG, 1, TmpF).addReg(TmpI);
1786 BuildMI(BB, IA64::XMAL, 3, ModulusResult)
1787 .addReg(TmpF15).addReg(TmpF).addReg(TmpF1);
1788 BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(ModulusResult);
1789 } else { // FP modulus! The horror... the horror....
1790 assert(0 && "sorry, no FP modulus just yet!\n!\n");
1797 case ISD::SIGN_EXTEND_INREG: {
1798 Tmp1 = SelectExpr(N.getOperand(0));
1799 switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
1802 assert(0 && "don't know how to sign extend this type");
1804 case MVT::i8: Opc = IA64::SXT1; break;
1805 case MVT::i16: Opc = IA64::SXT2; break;
1806 case MVT::i32: Opc = IA64::SXT4; break;
1808 BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
1813 Tmp1 = SelectExpr(N.getOperand(0));
1814 ISD::CondCode CC = cast<CondCodeSDNode>(Node->getOperand(2))->get();
1815 if (MVT::isInteger(N.getOperand(0).getValueType())) {
1817 if(ConstantSDNode *CSDN =
1818 dyn_cast<ConstantSDNode>(N.getOperand(1))) {
1819 // if we are comparing against a constant zero
1820 if(CSDN->getValue()==0)
1821 Tmp2 = IA64::r0; // then we can just compare against r0
1823 Tmp2 = SelectExpr(N.getOperand(1));
1824 } else // not comparing against a constant
1825 Tmp2 = SelectExpr(N.getOperand(1));
1828 default: assert(0 && "Unknown integer comparison!");
1830 BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1833 BuildMI(BB, IA64::CMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1836 BuildMI(BB, IA64::CMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1839 BuildMI(BB, IA64::CMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1842 BuildMI(BB, IA64::CMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1845 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1848 BuildMI(BB, IA64::CMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1851 BuildMI(BB, IA64::CMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1854 BuildMI(BB, IA64::CMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1857 BuildMI(BB, IA64::CMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1860 } else { // if not integer, should be FP.
1861 assert(N.getOperand(0).getValueType() != MVT::f32 &&
1862 "error: SETCC should have had incoming f32 promoted to f64!\n");
1864 if(ConstantFPSDNode *CFPSDN =
1865 dyn_cast<ConstantFPSDNode>(N.getOperand(1))) {
1867 // if we are comparing against a constant +0.0 or +1.0
1868 if(CFPSDN->isExactlyValue(+0.0))
1869 Tmp2 = IA64::F0; // then we can just compare against f0
1870 else if(CFPSDN->isExactlyValue(+1.0))
1871 Tmp2 = IA64::F1; // or f1
1873 Tmp2 = SelectExpr(N.getOperand(1));
1874 } else // not comparing against a constant
1875 Tmp2 = SelectExpr(N.getOperand(1));
1878 default: assert(0 && "Unknown FP comparison!");
1880 BuildMI(BB, IA64::FCMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
1883 BuildMI(BB, IA64::FCMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1886 BuildMI(BB, IA64::FCMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1889 BuildMI(BB, IA64::FCMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
1892 BuildMI(BB, IA64::FCMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1895 BuildMI(BB, IA64::FCMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
1898 BuildMI(BB, IA64::FCMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1901 BuildMI(BB, IA64::FCMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1904 BuildMI(BB, IA64::FCMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1907 BuildMI(BB, IA64::FCMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
1917 // Make sure we generate both values.
1919 ExprMap[N.getValue(1)] = 1; // Generate the token
1921 Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
1925 if(opcode == ISD::LOAD) { // this is a LOAD
1926 switch (Node->getValueType(0)) {
1927 default: assert(0 && "Cannot load this type!");
1928 case MVT::i1: Opc = IA64::LD1; isBool=true; break;
1929 // FIXME: for now, we treat bool loads the same as i8 loads */
1930 case MVT::i8: Opc = IA64::LD1; break;
1931 case MVT::i16: Opc = IA64::LD2; break;
1932 case MVT::i32: Opc = IA64::LD4; break;
1933 case MVT::i64: Opc = IA64::LD8; break;
1935 case MVT::f32: Opc = IA64::LDF4; break;
1936 case MVT::f64: Opc = IA64::LDF8; break;
1938 } else { // this is an EXTLOAD or ZEXTLOAD
1939 MVT::ValueType TypeBeingLoaded =
1940 cast<VTSDNode>(Node->getOperand(3))->getVT();
1941 switch (TypeBeingLoaded) {
1942 default: assert(0 && "Cannot extload/zextload this type!");
1944 case MVT::i8: Opc = IA64::LD1; break;
1945 case MVT::i16: Opc = IA64::LD2; break;
1946 case MVT::i32: Opc = IA64::LD4; break;
1947 case MVT::f32: Opc = IA64::LDF4; break;
1951 SDOperand Chain = N.getOperand(0);
1952 SDOperand Address = N.getOperand(1);
1954 if(Address.getOpcode() == ISD::GlobalAddress) {
1956 unsigned dummy = MakeReg(MVT::i64);
1957 unsigned dummy2 = MakeReg(MVT::i64);
1958 BuildMI(BB, IA64::ADD, 2, dummy)
1959 .addGlobalAddress(cast<GlobalAddressSDNode>(Address)->getGlobal())
1961 BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
1963 BuildMI(BB, Opc, 1, Result).addReg(dummy2);
1964 else { // emit a little pseudocode to load a bool (stored in one byte)
1965 // into a predicate register
1966 assert(Opc==IA64::LD1 && "problem loading a bool");
1967 unsigned dummy3 = MakeReg(MVT::i64);
1968 BuildMI(BB, Opc, 1, dummy3).addReg(dummy2);
1969 // we compare to 0. true? 0. false? 1.
1970 BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
1972 } else if(ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
1974 IA64Lowering.restoreGP(BB);
1975 unsigned dummy = MakeReg(MVT::i64);
1976 BuildMI(BB, IA64::ADD, 2, dummy).addConstantPoolIndex(CP->getIndex())
1977 .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 = dyn_cast<RegSDNode>(Node)->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(1));
2213 Tmp2 = cast<RegSDNode>(N)->getReg();
2216 if(N.getValueType() == MVT::i1) // if a bool, we use pseudocode
2217 BuildMI(BB, IA64::PCMPEQUNC, 3, Tmp2)
2218 .addReg(IA64::r0).addReg(IA64::r0).addReg(Tmp1);
2219 // (Tmp1) Tmp2 = cmp.eq.unc(r0,r0)
2221 BuildMI(BB, IA64::MOV, 1, Tmp2).addReg(Tmp1);
2222 // XXX is this the right way 'round? ;)
2229 /* what the heck is going on here:
2231 <_sabre_> ret with two operands is obvious: chain and value
2233 <_sabre_> ret with 3 values happens when 'expansion' occurs
2234 <_sabre_> e.g. i64 gets split into 2x i32
2236 <_sabre_> you don't have this case on ia64
2238 <_sabre_> so the two returned values go into EAX/EDX on ia32
2239 <camel_> ahhh *memories*
2241 <camel_> ok, thanks :)
2242 <_sabre_> so yeah, everything that has a side effect takes a 'token chain'
2243 <_sabre_> this is the first operand always
2244 <_sabre_> these operand often define chains, they are the last operand
2245 <_sabre_> they are printed as 'ch' if you do DAG.dump()
2248 switch (N.getNumOperands()) {
2250 assert(0 && "Unknown return instruction!");
2252 Select(N.getOperand(0));
2253 Tmp1 = SelectExpr(N.getOperand(1));
2254 switch (N.getOperand(1).getValueType()) {
2255 default: assert(0 && "All other types should have been promoted!!");
2256 // FIXME: do I need to add support for bools here?
2257 // (return '0' or '1' r8, basically...)
2259 // FIXME: need to round floats - 80 bits is bad, the tester
2262 // we mark r8 as live on exit up above in LowerArguments()
2263 BuildMI(BB, IA64::MOV, 1, IA64::r8).addReg(Tmp1);
2266 // we mark F8 as live on exit up above in LowerArguments()
2267 BuildMI(BB, IA64::FMOV, 1, IA64::F8).addReg(Tmp1);
2271 Select(N.getOperand(0));
2274 // before returning, restore the ar.pfs register (set by the 'alloc' up top)
2275 BuildMI(BB, IA64::MOV, 1).addReg(IA64::AR_PFS).addReg(IA64Lowering.VirtGPR);
2276 BuildMI(BB, IA64::RET, 0); // and then just emit a 'ret' instruction
2281 Select(N.getOperand(0));
2282 MachineBasicBlock *Dest =
2283 cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
2284 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(IA64::p0).addMBB(Dest);
2285 // XXX HACK! we do _not_ need long branches all the time
2289 case ISD::ImplicitDef: {
2290 Select(N.getOperand(0));
2291 BuildMI(BB, IA64::IDEF, 0, cast<RegSDNode>(N)->getReg());
2296 MachineBasicBlock *Dest =
2297 cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
2299 Select(N.getOperand(0));
2300 Tmp1 = SelectExpr(N.getOperand(1));
2301 BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(Tmp1).addMBB(Dest);
2302 // XXX HACK! we do _not_ need long branches all the time
2312 case ISD::CopyFromReg:
2313 case ISD::DYNAMIC_STACKALLOC:
2317 case ISD::TRUNCSTORE:
2319 Select(N.getOperand(0));
2320 Tmp1 = SelectExpr(N.getOperand(1)); // value
2324 if(opcode == ISD::STORE) {
2325 switch (N.getOperand(1).getValueType()) {
2326 default: assert(0 && "Cannot store this type!");
2327 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2328 // FIXME?: for now, we treat bool loads the same as i8 stores */
2329 case MVT::i8: Opc = IA64::ST1; break;
2330 case MVT::i16: Opc = IA64::ST2; break;
2331 case MVT::i32: Opc = IA64::ST4; break;
2332 case MVT::i64: Opc = IA64::ST8; break;
2334 case MVT::f32: Opc = IA64::STF4; break;
2335 case MVT::f64: Opc = IA64::STF8; break;
2337 } else { // truncstore
2338 switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
2339 default: assert(0 && "unknown type in truncstore");
2340 case MVT::i1: Opc = IA64::ST1; isBool=true; break;
2341 //FIXME: DAG does not promote this load?
2342 case MVT::i8: Opc = IA64::ST1; break;
2343 case MVT::i16: Opc = IA64::ST2; break;
2344 case MVT::i32: Opc = IA64::ST4; break;
2345 case MVT::f32: Opc = IA64::STF4; break;
2349 if(N.getOperand(2).getOpcode() == ISD::GlobalAddress) {
2350 unsigned dummy = MakeReg(MVT::i64);
2351 unsigned dummy2 = MakeReg(MVT::i64);
2352 BuildMI(BB, IA64::ADD, 2, dummy)
2353 .addGlobalAddress(cast<GlobalAddressSDNode>
2354 (N.getOperand(2))->getGlobal()).addReg(IA64::r1);
2355 BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
2358 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(Tmp1);
2359 else { // we are storing a bool, so emit a little pseudocode
2360 // to store a predicate register as one byte
2361 assert(Opc==IA64::ST1);
2362 unsigned dummy3 = MakeReg(MVT::i64);
2363 unsigned dummy4 = MakeReg(MVT::i64);
2364 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2365 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2366 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2367 BuildMI(BB, Opc, 2).addReg(dummy2).addReg(dummy4);
2369 } else if(N.getOperand(2).getOpcode() == ISD::FrameIndex) {
2371 // FIXME? (what about bools?)
2373 unsigned dummy = MakeReg(MVT::i64);
2374 BuildMI(BB, IA64::MOV, 1, dummy)
2375 .addFrameIndex(cast<FrameIndexSDNode>(N.getOperand(2))->getIndex());
2376 BuildMI(BB, Opc, 2).addReg(dummy).addReg(Tmp1);
2377 } else { // otherwise
2378 Tmp2 = SelectExpr(N.getOperand(2)); //address
2380 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(Tmp1);
2381 else { // we are storing a bool, so emit a little pseudocode
2382 // to store a predicate register as one byte
2383 assert(Opc==IA64::ST1);
2384 unsigned dummy3 = MakeReg(MVT::i64);
2385 unsigned dummy4 = MakeReg(MVT::i64);
2386 BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
2387 BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
2388 .addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
2389 BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(dummy4);
2395 case ISD::CALLSEQ_START:
2396 case ISD::CALLSEQ_END: {
2397 Select(N.getOperand(0));
2398 Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
2400 Opc = N.getOpcode() == ISD::CALLSEQ_START ? IA64::ADJUSTCALLSTACKDOWN :
2401 IA64::ADJUSTCALLSTACKUP;
2402 BuildMI(BB, Opc, 1).addImm(Tmp1);
2408 assert(0 && "GAME OVER. INSERT COIN?");
2412 /// createIA64PatternInstructionSelector - This pass converts an LLVM function
2413 /// into a machine code representation using pattern matching and a machine
2414 /// description file.
2416 FunctionPass *llvm::createIA64PatternInstructionSelector(TargetMachine &TM) {
2417 return new ISel(TM);