aboutsummaryrefslogtreecommitdiff
path: root/lib/Target/IA64/IA64ISelPattern.cpp
diff options
context:
space:
mode:
authorDuraid Madina <duraid@octopus.com.au>2006-01-19 14:14:11 +0000
committerDuraid Madina <duraid@octopus.com.au>2006-01-19 14:14:11 +0000
commit95f59e1e768db173d0b3e8184e862732b9fac9e5 (patch)
treef9ff76b3565b119b3cbc5e70767e3cee8f5f17a4 /lib/Target/IA64/IA64ISelPattern.cpp
parentc5f247089c7e9d281c37c4cc08cb6a5792d72a69 (diff)
BOOM!
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25460 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target/IA64/IA64ISelPattern.cpp')
-rw-r--r--lib/Target/IA64/IA64ISelPattern.cpp2417
1 files changed, 0 insertions, 2417 deletions
diff --git a/lib/Target/IA64/IA64ISelPattern.cpp b/lib/Target/IA64/IA64ISelPattern.cpp
deleted file mode 100644
index 3a1a413caf..0000000000
--- a/lib/Target/IA64/IA64ISelPattern.cpp
+++ /dev/null
@@ -1,2417 +0,0 @@
-//===-- IA64ISelPattern.cpp - A pattern matching inst selector for IA64 ---===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file was developed by Duraid Madina and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-//
-// This file defines a pattern matching instruction selector for IA64.
-//
-//===----------------------------------------------------------------------===//
-
-#include "IA64.h"
-#include "IA64InstrBuilder.h"
-#include "IA64RegisterInfo.h"
-#include "IA64MachineFunctionInfo.h"
-#include "llvm/Constants.h" // FIXME: REMOVE
-#include "llvm/Function.h"
-#include "llvm/CodeGen/MachineConstantPool.h" // FIXME: REMOVE
-#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineFrameInfo.h"
-#include "llvm/CodeGen/SelectionDAG.h"
-#include "llvm/CodeGen/SelectionDAGISel.h"
-#include "llvm/CodeGen/SSARegMap.h"
-#include "llvm/Target/TargetData.h"
-#include "llvm/Target/TargetLowering.h"
-#include "llvm/Support/MathExtras.h"
-#include "llvm/ADT/Statistic.h"
-#include <set>
-#include <map>
-#include <algorithm>
-using namespace llvm;
-
-//===----------------------------------------------------------------------===//
-// IA64TargetLowering - IA64 Implementation of the TargetLowering interface
-namespace {
- class IA64TargetLowering : public TargetLowering {
- int VarArgsFrameIndex; // FrameIndex for start of varargs area.
-
- //int ReturnAddrIndex; // FrameIndex for return slot.
- unsigned GP, SP, RP; // FIXME - clean this mess up
- public:
-
- unsigned VirtGPR; // this is public so it can be accessed in the selector
- // for ISD::RET down below. add an accessor instead? FIXME
-
- IA64TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
-
- // register class for general registers
- addRegisterClass(MVT::i64, IA64::GRRegisterClass);
-
- // register class for FP registers
- addRegisterClass(MVT::f64, IA64::FPRegisterClass);
-
- // register class for predicate registers
- addRegisterClass(MVT::i1, IA64::PRRegisterClass);
-
- setOperationAction(ISD::BRCONDTWOWAY , MVT::Other, Expand);
- setOperationAction(ISD::BRTWOWAY_CC , MVT::Other, Expand);
- setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
-
- setSetCCResultType(MVT::i1);
- setShiftAmountType(MVT::i64);
-
- setOperationAction(ISD::EXTLOAD , MVT::i1 , Promote);
-
- setOperationAction(ISD::ZEXTLOAD , MVT::i1 , Expand);
-
- setOperationAction(ISD::SEXTLOAD , MVT::i1 , Expand);
- setOperationAction(ISD::SEXTLOAD , MVT::i8 , Expand);
- setOperationAction(ISD::SEXTLOAD , MVT::i16 , Expand);
- setOperationAction(ISD::SEXTLOAD , MVT::i32 , Expand);
-
- setOperationAction(ISD::FREM , MVT::f32 , Expand);
- setOperationAction(ISD::FREM , MVT::f64 , Expand);
-
- setOperationAction(ISD::UREM , MVT::f32 , Expand);
- setOperationAction(ISD::UREM , MVT::f64 , Expand);
-
- setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
- setOperationAction(ISD::MEMSET , MVT::Other, Expand);
- setOperationAction(ISD::MEMCPY , MVT::Other, Expand);
-
- setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
- setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
-
- // We don't support sin/cos/sqrt
- setOperationAction(ISD::FSIN , MVT::f64, Expand);
- setOperationAction(ISD::FCOS , MVT::f64, Expand);
- setOperationAction(ISD::FSQRT, MVT::f64, Expand);
- setOperationAction(ISD::FSIN , MVT::f32, Expand);
- setOperationAction(ISD::FCOS , MVT::f32, Expand);
- setOperationAction(ISD::FSQRT, MVT::f32, Expand);
-
- //IA64 has these, but they are not implemented
- setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
- setOperationAction(ISD::CTLZ , MVT::i64 , Expand);
- setOperationAction(ISD::ROTL , MVT::i64 , Expand);
- setOperationAction(ISD::ROTR , MVT::i64 , Expand);
- setOperationAction(ISD::BSWAP, MVT::i64 , Expand); // mux @rev
- // FIXME: implement mulhs (xma.h) and mulhu (xma.hu)
- setOperationAction(ISD::MULHS , MVT::i64 , Expand);
- setOperationAction(ISD::MULHU , MVT::i64 , Expand);
-
- // We don't have line number support yet.
- setOperationAction(ISD::LOCATION, MVT::Other, Expand);
- setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
- setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
-
- // Not implemented yet.
- setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
- setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
-
- computeRegisterProperties();
-
- addLegalFPImmediate(+0.0);
- addLegalFPImmediate(+1.0);
- addLegalFPImmediate(-0.0);
- addLegalFPImmediate(-1.0);
- }
-
- /// LowerArguments - This hook must be implemented to indicate how we should
- /// lower the arguments for the specified function, into the specified DAG.
- virtual std::vector<SDOperand>
- LowerArguments(Function &F, SelectionDAG &DAG);
-
- /// LowerCallTo - This hook lowers an abstract call to a function into an
- /// actual call.
- virtual std::pair<SDOperand, SDOperand>
- LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
- bool isTailCall, SDOperand Callee, ArgListTy &Args,
- SelectionDAG &DAG);
-
- virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
- Value *VAListV, SelectionDAG &DAG);
- virtual std::pair<SDOperand,SDOperand>
- LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
- const Type *ArgTy, SelectionDAG &DAG);
-
- void restoreGP_SP_RP(MachineBasicBlock* BB)
- {
- BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
- BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
- BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
- }
-
- void restoreSP_RP(MachineBasicBlock* BB)
- {
- BuildMI(BB, IA64::MOV, 1, IA64::r12).addReg(SP);
- BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
- }
-
- void restoreRP(MachineBasicBlock* BB)
- {
- BuildMI(BB, IA64::MOV, 1, IA64::rp).addReg(RP);
- }
-
- void restoreGP(MachineBasicBlock* BB)
- {
- BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(GP);
- }
-
- };
-}
-
-
-std::vector<SDOperand>
-IA64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
- std::vector<SDOperand> ArgValues;
-
- //
- // add beautiful description of IA64 stack frame format
- // here (from intel 24535803.pdf most likely)
- //
- MachineFunction &MF = DAG.getMachineFunction();
- MachineFrameInfo *MFI = MF.getFrameInfo();
-
- GP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
- SP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
- RP = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
-
- MachineBasicBlock& BB = MF.front();
-
- unsigned args_int[] = {IA64::r32, IA64::r33, IA64::r34, IA64::r35,
- IA64::r36, IA64::r37, IA64::r38, IA64::r39};
-
- unsigned args_FP[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
- IA64::F12,IA64::F13,IA64::F14, IA64::F15};
-
- unsigned argVreg[8];
- unsigned argPreg[8];
- unsigned argOpc[8];
-
- unsigned used_FPArgs = 0; // how many FP args have been used so far?
-
- unsigned ArgOffset = 0;
- int count = 0;
-
- for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I)
- {
- SDOperand newroot, argt;
- if(count < 8) { // need to fix this logic? maybe.
-
- switch (getValueType(I->getType())) {
- default:
- std::cerr << "ERROR in LowerArgs: unknown type "
- << getValueType(I->getType()) << "\n";
- abort();
- case MVT::f32:
- // fixme? (well, will need to for weird FP structy stuff,
- // see intel ABI docs)
- case MVT::f64:
-//XXX BuildMI(&BB, IA64::IDEF, 0, args_FP[used_FPArgs]);
- MF.addLiveIn(args_FP[used_FPArgs]); // mark this reg as liveIn
- // floating point args go into f8..f15 as-needed, the increment
- argVreg[count] = // is below..:
- MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::f64));
- // FP args go into f8..f15 as needed: (hence the ++)
- argPreg[count] = args_FP[used_FPArgs++];
- argOpc[count] = IA64::FMOV;
- argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), argVreg[count],
- MVT::f64);
- if (I->getType() == Type::FloatTy)
- argt = DAG.getNode(ISD::FP_ROUND, MVT::f32, argt);
- break;
- case MVT::i1: // NOTE: as far as C abi stuff goes,
- // bools are just boring old ints
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- case MVT::i64:
-//XXX BuildMI(&BB, IA64::IDEF, 0, args_int[count]);
- MF.addLiveIn(args_int[count]); // mark this register as liveIn
- argVreg[count] =
- MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
- argPreg[count] = args_int[count];
- argOpc[count] = IA64::MOV;
- argt = newroot =
- DAG.getCopyFromReg(DAG.getRoot(), argVreg[count], MVT::i64);
- if ( getValueType(I->getType()) != MVT::i64)
- argt = DAG.getNode(ISD::TRUNCATE, getValueType(I->getType()),
- newroot);
- break;
- }
- } else { // more than 8 args go into the frame
- // Create the frame index object for this incoming parameter...
- ArgOffset = 16 + 8 * (count - 8);
- int FI = MFI->CreateFixedObject(8, ArgOffset);
-
- // Create the SelectionDAG nodes corresponding to a load
- //from this parameter
- SDOperand FIN = DAG.getFrameIndex(FI, MVT::i64);
- argt = newroot = DAG.getLoad(getValueType(I->getType()),
- DAG.getEntryNode(), FIN, DAG.getSrcValue(NULL));
- }
- ++count;
- DAG.setRoot(newroot.getValue(1));
- ArgValues.push_back(argt);
- }
-
-
- // Create a vreg to hold the output of (what will become)
- // the "alloc" instruction
- VirtGPR = MF.getSSARegMap()->createVirtualRegister(getRegClassFor(MVT::i64));
- BuildMI(&BB, IA64::PSEUDO_ALLOC, 0, VirtGPR);
- // we create a PSEUDO_ALLOC (pseudo)instruction for now
-
- BuildMI(&BB, IA64::IDEF, 0, IA64::r1);
-
- // hmm:
- BuildMI(&BB, IA64::IDEF, 0, IA64::r12);
- BuildMI(&BB, IA64::IDEF, 0, IA64::rp);
- // ..hmm.
-
- BuildMI(&BB, IA64::MOV, 1, GP).addReg(IA64::r1);
-
- // hmm:
- BuildMI(&BB, IA64::MOV, 1, SP).addReg(IA64::r12);
- BuildMI(&BB, IA64::MOV, 1, RP).addReg(IA64::rp);
- // ..hmm.
-
- unsigned tempOffset=0;
-
- // if this is a varargs function, we simply lower llvm.va_start by
- // pointing to the first entry
- if(F.isVarArg()) {
- tempOffset=0;
- VarArgsFrameIndex = MFI->CreateFixedObject(8, tempOffset);
- }
-
- // here we actually do the moving of args, and store them to the stack
- // too if this is a varargs function:
- for (int i = 0; i < count && i < 8; ++i) {
- BuildMI(&BB, argOpc[i], 1, argVreg[i]).addReg(argPreg[i]);
- if(F.isVarArg()) {
- // if this is a varargs function, we copy the input registers to the stack
- int FI = MFI->CreateFixedObject(8, tempOffset);
- tempOffset+=8; //XXX: is it safe to use r22 like this?
- BuildMI(&BB, IA64::MOV, 1, IA64::r22).addFrameIndex(FI);
- // FIXME: we should use st8.spill here, one day
- BuildMI(&BB, IA64::ST8, 1, IA64::r22).addReg(argPreg[i]);
- }
- }
-
- // Finally, inform the code generator which regs we return values in.
- // (see the ISD::RET: case down below)
- switch (getValueType(F.getReturnType())) {
- default: assert(0 && "i have no idea where to return this type!");
- case MVT::isVoid: break;
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- case MVT::i64:
- MF.addLiveOut(IA64::r8);
- break;
- case MVT::f32:
- case MVT::f64:
- MF.addLiveOut(IA64::F8);
- break;
- }
-
- return ArgValues;
-}
-
-std::pair<SDOperand, SDOperand>
-IA64TargetLowering::LowerCallTo(SDOperand Chain,
- const Type *RetTy, bool isVarArg,
- unsigned CallingConv, bool isTailCall,
- SDOperand Callee, ArgListTy &Args,
- SelectionDAG &DAG) {
-
- MachineFunction &MF = DAG.getMachineFunction();
-
- unsigned NumBytes = 16;
- unsigned outRegsUsed = 0;
-
- if (Args.size() > 8) {
- NumBytes += (Args.size() - 8) * 8;
- outRegsUsed = 8;
- } else {
- outRegsUsed = Args.size();
- }
-
- // FIXME? this WILL fail if we ever try to pass around an arg that
- // consumes more than a single output slot (a 'real' double, int128
- // some sort of aggregate etc.), as we'll underestimate how many 'outX'
- // registers we use. Hopefully, the assembler will notice.
- MF.getInfo<IA64FunctionInfo>()->outRegsUsed=
- std::max(outRegsUsed, MF.getInfo<IA64FunctionInfo>()->outRegsUsed);
-
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
- DAG.getConstant(NumBytes, getPointerTy()));
-
- std::vector<SDOperand> args_to_use;
- for (unsigned i = 0, e = Args.size(); i != e; ++i)
- {
- switch (getValueType(Args[i].second)) {
- default: assert(0 && "unexpected argument type!");
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- //promote to 64-bits, sign/zero extending based on type
- //of the argument
- if(Args[i].second->isSigned())
- Args[i].first = DAG.getNode(ISD::SIGN_EXTEND, MVT::i64,
- Args[i].first);
- else
- Args[i].first = DAG.getNode(ISD::ZERO_EXTEND, MVT::i64,
- Args[i].first);
- break;
- case MVT::f32:
- //promote to 64-bits
- Args[i].first = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Args[i].first);
- case MVT::f64:
- case MVT::i64:
- break;
- }
- args_to_use.push_back(Args[i].first);
- }
-
- std::vector<MVT::ValueType> RetVals;
- MVT::ValueType RetTyVT = getValueType(RetTy);
- if (RetTyVT != MVT::isVoid)
- RetVals.push_back(RetTyVT);
- RetVals.push_back(MVT::Other);
-
- SDOperand TheCall = SDOperand(DAG.getCall(RetVals, Chain,
- Callee, args_to_use), 0);
- Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
- Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
- DAG.getConstant(NumBytes, getPointerTy()));
- return std::make_pair(TheCall, Chain);
-}
-
-SDOperand
-IA64TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
- Value *VAListV, SelectionDAG &DAG) {
- // vastart just stores the address of the VarArgsFrameIndex slot.
- SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i64);
- return DAG.getNode(ISD::STORE, MVT::Other, Chain, FR,
- VAListP, DAG.getSrcValue(VAListV));
-}
-
-std::pair<SDOperand,SDOperand> IA64TargetLowering::
-LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
- const Type *ArgTy, SelectionDAG &DAG) {
-
- MVT::ValueType ArgVT = getValueType(ArgTy);
- SDOperand Val = DAG.getLoad(MVT::i64, Chain,
- VAListP, DAG.getSrcValue(VAListV));
- SDOperand Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), Val,
- DAG.getSrcValue(NULL));
- unsigned Amt;
- if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
- Amt = 8;
- else {
- assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
- "Other types should have been promoted for varargs!");
- Amt = 8;
- }
- Val = DAG.getNode(ISD::ADD, Val.getValueType(), Val,
- DAG.getConstant(Amt, Val.getValueType()));
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Val, VAListP, DAG.getSrcValue(VAListV));
- return std::make_pair(Result, Chain);
-}
-
-namespace {
-
- //===--------------------------------------------------------------------===//
- /// ISel - IA64 specific code to select IA64 machine instructions for
- /// SelectionDAG operations.
- ///
- class ISel : public SelectionDAGISel {
- /// IA64Lowering - This object fully describes how to lower LLVM code to an
- /// IA64-specific SelectionDAG.
- IA64TargetLowering IA64Lowering;
- SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform
- // for sdiv and udiv until it is put into the future
- // dag combiner
-
- /// ExprMap - As shared expressions are codegen'd, we keep track of which
- /// vreg the value is produced in, so we only emit one copy of each compiled
- /// tree.
- std::map<SDOperand, unsigned> ExprMap;
- std::set<SDOperand> LoweredTokens;
-
- public:
- ISel(TargetMachine &TM) : SelectionDAGISel(IA64Lowering), IA64Lowering(TM),
- ISelDAG(0) { }
-
- /// InstructionSelectBasicBlock - This callback is invoked by
- /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
- virtual void InstructionSelectBasicBlock(SelectionDAG &DAG);
-
- unsigned SelectExpr(SDOperand N);
- void Select(SDOperand N);
- // a dag->dag to transform mul-by-constant-int to shifts+adds/subs
- SDOperand BuildConstmulSequence(SDOperand N);
-
- const char *getPassName() const { return "IA64 Instruction Selector"; }
- };
-}
-
-/// InstructionSelectBasicBlock - This callback is invoked by SelectionDAGISel
-/// when it has created a SelectionDAG for us to codegen.
-void ISel::InstructionSelectBasicBlock(SelectionDAG &DAG) {
-
- // Codegen the basic block.
- ISelDAG = &DAG;
- Select(DAG.getRoot());
-
- // Clear state used for selection.
- ExprMap.clear();
- LoweredTokens.clear();
- ISelDAG = 0;
-}
-
-// strip leading '0' characters from a string
-void munchLeadingZeros(std::string& inString) {
- while(inString.c_str()[0]=='0') {
- inString.erase(0, 1);
- }
-}
-
-// strip trailing '0' characters from a string
-void munchTrailingZeros(std::string& inString) {
- int curPos=inString.length()-1;
-
- while(inString.c_str()[curPos]=='0') {
- inString.erase(curPos, 1);
- curPos--;
- }
-}
-
-// return how many consecutive '0' characters are at the end of a string
-unsigned int countTrailingZeros(std::string& inString) {
- int curPos=inString.length()-1;
- unsigned int zeroCount=0;
- // assert goes here
- while(inString.c_str()[curPos--]=='0') {
- zeroCount++;
- }
- return zeroCount;
-}
-
-// booth encode a string of '1' and '0' characters (returns string of 'P' (+1)
-// '0' and 'N' (-1) characters)
-void boothEncode(std::string inString, std::string& boothEncodedString) {
-
- int curpos=0;
- int replacements=0;
- int lim=inString.size();
-
- while(curpos<lim) {
- if(inString[curpos]=='1') { // if we see a '1', look for a run of them
- int runlength=0;
- std::string replaceString="N";
-
- // find the run length
- for(;inString[curpos+runlength]=='1';runlength++) ;
-
- for(int i=0; i<runlength-1; i++)
- replaceString+="0";
- replaceString+="1";
-
- if(runlength>1) {
- inString.replace(curpos, runlength+1, replaceString);
- curpos+=runlength-1;
- } else
- curpos++;
- } else { // a zero, we just keep chugging along
- curpos++;
- }
- }
-
- // clean up (trim the string, reverse it and turn '1's into 'P's)
- munchTrailingZeros(inString);
- boothEncodedString="";
-
- for(int i=inString.size()-1;i>=0;i--)
- if(inString[i]=='1')
- boothEncodedString+="P";
- else
- boothEncodedString+=inString[i];
-
-}
-
-struct shiftaddblob { // this encodes stuff like (x=) "A << B [+-] C << D"
- unsigned firstVal; // A
- unsigned firstShift; // B
- unsigned secondVal; // C
- unsigned secondShift; // D
- bool isSub;
-};
-
-/* this implements Lefevre's "pattern-based" constant multiplication,
- * see "Multiplication by an Integer Constant", INRIA report 1999-06
- *
- * TODO: implement a method to try rewriting P0N<->0PP / N0P<->0NN
- * to get better booth encodings - this does help in practice
- * TODO: weight shifts appropriately (most architectures can't
- * fuse a shift and an add for arbitrary shift amounts) */
-unsigned lefevre(const std::string inString,
- std::vector<struct shiftaddblob> &ops) {
- std::string retstring;
- std::string s = inString;
- munchTrailingZeros(s);
-
- int length=s.length()-1;
-
- if(length==0) {
- return(0);
- }
-
- std::vector<int> p,n;
-
- for(int i=0; i<=length; i++) {
- if (s.c_str()[length-i]=='P') {
- p.push_back(i);
- } else if (s.c_str()[length-i]=='N') {
- n.push_back(i);
- }
- }
-
- std::string t, u;
- int c = 0;
- bool f;
- std::map<const int, int> w;
-
- for(unsigned i=0; i<p.size(); i++) {
- for(unsigned j=0; j<i; j++) {
- w[p[i]-p[j]]++;
- }
- }
-
- for(unsigned i=1; i<n.size(); i++) {
- for(unsigned j=0; j<i; j++) {
- w[n[i]-n[j]]++;
- }
- }
-
- for(unsigned i=0; i<p.size(); i++) {
- for(unsigned j=0; j<n.size(); j++) {
- w[-abs(p[i]-n[j])]++;
- }
- }
-
- std::map<const int, int>::const_iterator ii;
- std::vector<int> d;
- std::multimap<int, int> sorted_by_value;
-
- for(ii = w.begin(); ii!=w.end(); ii++)
- sorted_by_value.insert(std::pair<int, int>((*ii).second,(*ii).first));
-
- for (std::multimap<int, int>::iterator it = sorted_by_value.begin();
- it != sorted_by_value.end(); ++it) {
- d.push_back((*it).second);
- }
-
- int int_W=0;
- int int_d;
-
- while(d.size()>0 && (w[int_d=d.back()] > int_W)) {
- d.pop_back();
- retstring=s; // hmmm
- int x=0;
- int z=abs(int_d)-1;
-
- if(int_d>0) {
-
- for(unsigned base=0; base<retstring.size(); base++) {
- if( ((base+z+1) < retstring.size()) &&
- retstring.c_str()[base]=='P' &&
- retstring.c_str()[base+z+1]=='P')
- {
- // match
- x++;
- retstring.replace(base, 1, "0");
- retstring.replace(base+z+1, 1, "p");
- }
- }
-
- for(unsigned base=0; base<retstring.size(); base++) {
- if( ((base+z+1) < retstring.size()) &&
- retstring.c_str()[base]=='N' &&
- retstring.c_str()[base+z+1]=='N')
- {
- // match
- x++;
- retstring.replace(base, 1, "0");
- retstring.replace(base+z+1, 1, "n");
- }
- }
-
- } else {
- for(unsigned base=0; base<retstring.size(); base++) {
- if( ((base+z+1) < retstring.size()) &&
- ((retstring.c_str()[base]=='P' &&
- retstring.c_str()[base+z+1]=='N') ||
- (retstring.c_str()[base]=='N' &&
- retstring.c_str()[base+z+1]=='P')) ) {
- // match
- x++;
-
- if(retstring.c_str()[base]=='P') {
- retstring.replace(base, 1, "0");
- retstring.replace(base+z+1, 1, "p");
- } else { // retstring[base]=='N'
- retstring.replace(base, 1, "0");
- retstring.replace(base+z+1, 1, "n");
- }
- }
- }
- }
-
- if(x>int_W) {
- int_W = x;
- t = retstring;
- c = int_d; // tofix
- }
-
- } d.pop_back(); // hmm
-
- u = t;
-
- for(unsigned i=0; i<t.length(); i++) {
- if(t.c_str()[i]=='p' || t.c_str()[i]=='n')
- t.replace(i, 1, "0");
- }
-
- for(unsigned i=0; i<u.length(); i++) {
- if(u[i]=='P' || u[i]=='N')
- u.replace(i, 1, "0");
- if(u[i]=='p')
- u.replace(i, 1, "P");
- if(u[i]=='n')
- u.replace(i, 1, "N");
- }
-
- if( c<0 ) {
- f=true;
- c=-c;
- } else
- f=false;
-
- int pos=0;
- while(u[pos]=='0')
- pos++;
-
- bool hit=(u[pos]=='N');
-
- int g=0;
- if(hit) {
- g=1;
- for(unsigned p=0; p<u.length(); p++) {
- bool isP=(u[p]=='P');
- bool isN=(u[p]=='N');
-
- if(isP)
- u.replace(p, 1, "N");
- if(isN)
- u.replace(p, 1, "P");
- }
- }
-
- munchLeadingZeros(u);
-
- int i = lefevre(u, ops);
-
- shiftaddblob blob;
-
- blob.firstVal=i; blob.firstShift=c;
- blob.isSub=f;
- blob.secondVal=i; blob.secondShift=0;
-
- ops.push_back(blob);
-
- i = ops.size();
-
- munchLeadingZeros(t);
-
- if(t.length()==0)
- return i;
-
- if(t.c_str()[0]!='P') {
- g=2;
- for(unsigned p=0; p<t.length(); p++) {
- bool isP=(t.c_str()[p]=='P');
- bool isN=(t.c_str()[p]=='N');
-
- if(isP)
- t.replace(p, 1, "N");
- if(isN)
- t.replace(p, 1, "P");
- }
- }
-
- int j = lefevre(t, ops);
-
- int trail=countTrailingZeros(u);
- blob.secondVal=i; blob.secondShift=trail;
-
- trail=countTrailingZeros(t);
- blob.firstVal=j; blob.firstShift=trail;
-
- switch(g) {
- case 0:
- blob.isSub=false; // first + second
- break;
- case 1:
- blob.isSub=true; // first - second
- break;
- case 2:
- blob.isSub=true; // second - first
- int tmpval, tmpshift;
- tmpval=blob.firstVal;
- tmpshift=blob.firstShift;
- blob.firstVal=blob.secondVal;
- blob.firstShift=blob.secondShift;
- blob.secondVal=tmpval;
- blob.secondShift=tmpshift;
- break;
- //assert
- }
-
- ops.push_back(blob);
- return ops.size();
-}
-
-SDOperand ISel::BuildConstmulSequence(SDOperand N) {
- //FIXME: we should shortcut this stuff for multiplies by 2^n+1
- // in particular, *3 is nicer as *2+1, not *4-1
- int64_t constant=cast<ConstantSDNode>(N.getOperand(1))->getValue();
-
- bool flippedSign;
- unsigned preliminaryShift=0;
-
- assert(constant != 0 && "erk, you're trying to multiply by constant zero\n");
-
- // first, we make the constant to multiply by positive
- if(constant<0) {
- constant=-constant;
- flippedSign=true;
- } else {
- flippedSign=false;
- }
-
- // next, we make it odd.
- for(; (constant%2==0); preliminaryShift++)
- constant>>=1;
-
- //OK, we have a positive, odd number of 64 bits or less. Convert it
- //to a binary string, constantString[0] is the LSB
- char constantString[65];
- for(int i=0; i<64; i++)
- constantString[i]='0'+((constant>>i)&0x1);
- constantString[64]=0;
-
- // now, Booth encode it
- std::string boothEncodedString;
- boothEncode(constantString, boothEncodedString);
-
- std::vector<struct shiftaddblob> ops;
- // do the transformation, filling out 'ops'
- lefevre(boothEncodedString, ops);
-
- assert(ops.size() < 80 && "constmul code has gone haywire\n");
- SDOperand results[80]; // temporary results (of adds/subs of shifts)
-
- // now turn 'ops' into DAG bits
- for(unsigned i=0; i<ops.size(); i++) {
- SDOperand amt = ISelDAG->getConstant(ops[i].firstShift, MVT::i64);
- SDOperand val = (ops[i].firstVal == 0) ? N.getOperand(0) :
- results[ops[i].firstVal-1];
- SDOperand left = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
- amt = ISelDAG->getConstant(ops[i].secondShift, MVT::i64);
- val = (ops[i].secondVal == 0) ? N.getOperand(0) :
- results[ops[i].secondVal-1];
- SDOperand right = ISelDAG->getNode(ISD::SHL, MVT::i64, val, amt);
- if(ops[i].isSub)
- results[i] = ISelDAG->getNode(ISD::SUB, MVT::i64, left, right);
- else
- results[i] = ISelDAG->getNode(ISD::ADD, MVT::i64, left, right);
- }
-
- // don't forget flippedSign and preliminaryShift!
- SDOperand shiftedresult;
- if(preliminaryShift) {
- SDOperand finalshift = ISelDAG->getConstant(preliminaryShift, MVT::i64);
- shiftedresult = ISelDAG->getNode(ISD::SHL, MVT::i64,
- results[ops.size()-1], finalshift);
- } else { // there was no preliminary divide-by-power-of-2 required
- shiftedresult = results[ops.size()-1];
- }
-
- SDOperand finalresult;
- if(flippedSign) { // if we were multiplying by a negative constant:
- SDOperand zero = ISelDAG->getConstant(0, MVT::i64);
- // subtract the result from 0 to flip its sign
- finalresult = ISelDAG->getNode(ISD::SUB, MVT::i64, zero, shiftedresult);
- } else { // there was no preliminary multiply by -1 required
- finalresult = shiftedresult;
- }
-
- return finalresult;
-}
-
-/// ponderIntegerDivisionBy - When handling integer divides, if the divide
-/// is by a constant such that we can efficiently codegen it, this
-/// function says what to do. Currently, it returns 0 if the division must
-/// become a genuine divide, and 1 if the division can be turned into a
-/// right shift.
-static unsigned ponderIntegerDivisionBy(SDOperand N, bool isSigned,
- unsigned& Imm) {
- if (N.getOpcode() != ISD::Constant) return 0; // if not a divide by
- // a constant, give up.
-
- int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
-
- if (isPowerOf2_64(v)) { // if a division by a power of two, say so
- Imm = Log2_64(v);
- return 1;
- }
-
- return 0; // fallthrough
-}
-
-static unsigned ponderIntegerAndWith(SDOperand N, unsigned& Imm) {
- if (N.getOpcode() != ISD::Constant) return 0; // if not ANDing with
- // a constant, give up.
-
- int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
-
- if (isMask_64(v)) { // if ANDing with ((2^n)-1) for some n
- Imm = Log2_64(v) + 1;
- return 1; // say so
- }
-
- return 0; // fallthrough
-}
-
-static unsigned ponderIntegerAdditionWith(SDOperand N, unsigned& Imm) {
- if (N.getOpcode() != ISD::Constant) return 0; // if not adding a
- // constant, give up.
- int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
-
- if (v <= 8191 && v >= -8192) { // if this constants fits in 14 bits, say so
- Imm = v & 0x3FFF; // 14 bits
- return 1;
- }
- return 0; // fallthrough
-}
-
-static unsigned ponderIntegerSubtractionFrom(SDOperand N, unsigned& Imm) {
- if (N.getOpcode() != ISD::Constant) return 0; // if not subtracting a
- // constant, give up.
- int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
-
- if (v <= 127 && v >= -128) { // if this constants fits in 8 bits, say so
- Imm = v & 0xFF; // 8 bits
- return 1;
- }
- return 0; // fallthrough
-}
-
-unsigned ISel::SelectExpr(SDOperand N) {
- unsigned Result;
- unsigned Tmp1, Tmp2, Tmp3;
- unsigned Opc = 0;
- MVT::ValueType DestType = N.getValueType();
-
- unsigned opcode = N.getOpcode();
-
- SDNode *Node = N.Val;
- SDOperand Op0, Op1;
-
- if (Node->getOpcode() == ISD::CopyFromReg)
- // Just use the specified register as our input.
- return cast<RegisterSDNode>(Node->getOperand(1))->getReg();
-
- unsigned &Reg = ExprMap[N];
- if (Reg) return Reg;
-
- if (N.getOpcode() != ISD::CALL && N.getOpcode() != ISD::TAILCALL)
- Reg = Result = (N.getValueType() != MVT::Other) ?
- MakeReg(N.getValueType()) : 1;
- else {
- // If this is a call instruction, make sure to prepare ALL of the result
- // values as well as the chain.
- if (Node->getNumValues() == 1)
- Reg = Result = 1; // Void call, just a chain.
- else {
- Result = MakeReg(Node->getValueType(0));
- ExprMap[N.getValue(0)] = Result;
- for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i)
- ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i));
- ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1;
- }
- }
-
- switch (N.getOpcode()) {
- default:
- Node->dump();
- assert(0 && "Node not handled!\n");
-
- case ISD::FrameIndex: {
- Tmp1 = cast<FrameIndexSDNode>(N)->getIndex();
- BuildMI(BB, IA64::MOV, 1, Result).addFrameIndex(Tmp1);
- return Result;
- }
-
- case ISD::ConstantPool: {
- Tmp1 = BB->getParent()->getConstantPool()->
- getConstantPoolIndex(cast<ConstantPoolSDNode>(N)->get());
- IA64Lowering.restoreGP(BB); // FIXME: do i really need this?
- BuildMI(BB, IA64::ADD, 2, Result).addConstantPoolIndex(Tmp1)
- .addReg(IA64::r1);
- return Result;
- }
-
- case ISD::ConstantFP: {
- Tmp1 = Result; // Intermediate Register
- if (cast<ConstantFPSDNode>(N)->getValue() < 0.0 ||
- cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
- Tmp1 = MakeReg(MVT::f64);
-
- if (cast<ConstantFPSDNode>(N)->isExactlyValue(+0.0) ||
- cast<ConstantFPSDNode>(N)->isExactlyValue(-0.0))
- BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F0); // load 0.0
- else if (cast<ConstantFPSDNode>(N)->isExactlyValue(+1.0) ||
- cast<ConstantFPSDNode>(N)->isExactlyValue(-1.0))
- BuildMI(BB, IA64::FMOV, 1, Tmp1).addReg(IA64::F1); // load 1.0
- else
- assert(0 && "Unexpected FP constant!");
- if (Tmp1 != Result)
- // we multiply by +1.0, negate (this is FNMA), and then add 0.0
- BuildMI(BB, IA64::FNMA, 3, Result).addReg(Tmp1).addReg(IA64::F1)
- .addReg(IA64::F0);
- return Result;<