RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-30 07:22:55 +01:00
Code Issues Projects Releases Wiki Activity
afd68f8f76
llvm-mirror/lib/Target/IA64/IA64ISelPattern.cpp
Chris Lattner cc8ae687e1 Update to use the new MathExtras.h support for log2 computation. 
Patch contributed by Jim Laskey!

llvm-svn: 22594
2005-08-02 19:26:06 +00:00

2425 lines
82 KiB
C++
Raw Blame History

//===-- 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::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::SREM , MVT::f32 , Expand);
setOperationAction(ISD::SREM , 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);
// 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);
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(argVreg[count],
getValueType(I->getType()), DAG.getRoot());
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(argVreg[count], MVT::i64, DAG.getRoot());
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);
};
}
/// 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;
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);
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 dyn_cast<RegSDNode>(Node)->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 = cast<ConstantPoolSDNode>(N)->getIndex();
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;
}
case ISD::DYNAMIC_STACKALLOC: {
// Generate both result values.
if (Result != 1)
ExprMap[N.getValue(1)] = 1; // Generate the token
else
Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
// FIXME: We are currently ignoring the requested alignment for handling
// greater than the stack alignment. This will need to be revisited at some
// point. Align = N.getOperand(2);
if (!isa<ConstantSDNode>(N.getOperand(2)) ||
cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) {
std::cerr << "Cannot allocate stack object with greater alignment than"
<< " the stack alignment yet!";
abort();
}
/*
Select(N.getOperand(0));
if (ConstantSDNode* CN = dyn_cast<ConstantSDNode>(N.getOperand(1)))
{
if (CN->getValue() < 32000)
{
BuildMI(BB, IA64::ADDIMM22, 2, IA64::r12).addReg(IA64::r12)
.addImm(-CN->getValue());
} else {
Tmp1 = SelectExpr(N.getOperand(1));
// Subtract size from stack pointer, thereby allocating some space.
BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
}
} else {
Tmp1 = SelectExpr(N.getOperand(1));
// Subtract size from stack pointer, thereby allocating some space.
BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
}
*/
Select(N.getOperand(0));
Tmp1 = SelectExpr(N.getOperand(1));
// Subtract size from stack pointer, thereby allocating some space.
BuildMI(BB, IA64::SUB, 2, IA64::r12).addReg(IA64::r12).addReg(Tmp1);
// Put a pointer to the space into the result register, by copying the
// stack pointer.
BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r12);
return Result;
}
case ISD::SELECT: {
Tmp1 = SelectExpr(N.getOperand(0)); //Cond
Tmp2 = SelectExpr(N.getOperand(1)); //Use if TRUE
Tmp3 = SelectExpr(N.getOperand(2)); //Use if FALSE
unsigned bogoResult;
switch (N.getOperand(1).getValueType()) {
default: assert(0 &&
"ISD::SELECT: 'select'ing something other than i1, i64 or f64!\n");
// for i1, we load the condition into an integer register, then
// conditionally copy Tmp2 and Tmp3 to Tmp1 in parallel (only one
// of them will go through, since the integer register will hold
// either 0 or 1)
case MVT::i1: {
bogoResult=MakeReg(MVT::i1);
// load the condition into an integer register
unsigned condReg=MakeReg(MVT::i64);
unsigned dummy=MakeReg(MVT::i64);
BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
BuildMI(BB, IA64::TPCADDIMM22, 2, condReg).addReg(dummy)
.addImm(1).addReg(Tmp1);
// initialize Result (bool) to false (hence UNC) and if
// the select condition (condReg) is false (0), copy Tmp3
BuildMI(BB, IA64::PCMPEQUNC, 3, bogoResult)
.addReg(condReg).addReg(IA64::r0).addReg(Tmp3);
// now, if the selection condition is true, write 1 to the
// result if Tmp2 is 1
BuildMI(BB, IA64::TPCMPNE, 3, Result).addReg(bogoResult)
.addReg(condReg).addReg(IA64::r0).addReg(Tmp2);
break;
}
// for i64/f64, we just copy Tmp3 and then conditionally overwrite it
// with Tmp2 if Tmp1 is true
case MVT::i64:
bogoResult=MakeReg(MVT::i64);
BuildMI(BB, IA64::MOV, 1, bogoResult).addReg(Tmp3);
BuildMI(BB, IA64::CMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
.addReg(Tmp1);
break;
case MVT::f64:
bogoResult=MakeReg(MVT::f64);
BuildMI(BB, IA64::FMOV, 1, bogoResult).addReg(Tmp3);
BuildMI(BB, IA64::CFMOV, 2, Result).addReg(bogoResult).addReg(Tmp2)
.addReg(Tmp1);
break;
}
return Result;
}
case ISD::Constant: {
unsigned depositPos=0;
unsigned depositLen=0;
switch (N.getValueType()) {
default: assert(0 && "Cannot use constants of this type!");
case MVT::i1: { // if a bool, we don't 'load' so much as generate
// the constant:
if(cast<ConstantSDNode>(N)->getValue()) // true:
BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
else // false:
BuildMI(BB, IA64::CMPNE, 2, Result).addReg(IA64::r0).addReg(IA64::r0);
return Result; // early exit
}
case MVT::i64: break;
}
int64_t immediate = cast<ConstantSDNode>(N)->getValue();
if(immediate==0) { // if the constant is just zero,
BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r0); // just copy r0
return Result; // early exit
}
if (immediate <= 8191 && immediate >= -8192) {
// if this constants fits in 14 bits, we use a mov the assembler will
// turn into: "adds rDest=imm,r0" (and _not_ "andl"...)
BuildMI(BB, IA64::MOVSIMM14, 1, Result).addSImm(immediate);
return Result; // early exit
}
if (immediate <= 2097151 && immediate >= -2097152) {
// if this constants fits in 22 bits, we use a mov the assembler will
// turn into: "addl rDest=imm,r0"
BuildMI(BB, IA64::MOVSIMM22, 1, Result).addSImm(immediate);
return Result; // early exit
}
/* otherwise, our immediate is big, so we use movl */
uint64_t Imm = immediate;
BuildMI(BB, IA64::MOVLIMM64, 1, Result).addImm64(Imm);
return Result;
}
case ISD::UNDEF: {
BuildMI(BB, IA64::IDEF, 0, Result);
return Result;
}
case ISD::GlobalAddress: {
GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal();
unsigned Tmp1 = MakeReg(MVT::i64);
BuildMI(BB, IA64::ADD, 2, Tmp1).addGlobalAddress(GV).addReg(IA64::r1);
BuildMI(BB, IA64::LD8, 1, Result).addReg(Tmp1);
return Result;
}
case ISD::ExternalSymbol: {
const char *Sym = cast<ExternalSymbolSDNode>(N)->getSymbol();
// assert(0 && "sorry, but what did you want an ExternalSymbol for again?");
BuildMI(BB, IA64::MOV, 1, Result).addExternalSymbol(Sym); // XXX
return Result;
}
case ISD::FP_EXTEND: {
Tmp1 = SelectExpr(N.getOperand(0));
BuildMI(BB, IA64::FMOV, 1, Result).addReg(Tmp1);
return Result;
}
case ISD::ZERO_EXTEND: {
Tmp1 = SelectExpr(N.getOperand(0)); // value
switch (N.getOperand(0).getValueType()) {
default: assert(0 && "Cannot zero-extend this type!");
case MVT::i8: Opc = IA64::ZXT1; break;
case MVT::i16: Opc = IA64::ZXT2; break;
case MVT::i32: Opc = IA64::ZXT4; break;
// we handle bools differently! :
case MVT::i1: { // if the predicate reg has 1, we want a '1' in our GR.
unsigned dummy = MakeReg(MVT::i64);
// first load zero:
BuildMI(BB, IA64::MOV, 1, dummy).addReg(IA64::r0);
// ...then conditionally (PR:Tmp1) add 1:
BuildMI(BB, IA64::TPCADDIMM22, 2, Result).addReg(dummy)
.addImm(1).addReg(Tmp1);
return Result; // XXX early exit!
}
}
BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
return Result;
}
case ISD::SIGN_EXTEND: { // we should only have to handle i1 -> i64 here!!!
assert(0 && "hmm, ISD::SIGN_EXTEND: shouldn't ever be reached. bad luck!\n");
Tmp1 = SelectExpr(N.getOperand(0)); // value
switch (N.getOperand(0).getValueType()) {
default: assert(0 && "Cannot sign-extend this type!");
case MVT::i1: assert(0 && "trying to sign extend a bool? ow.\n");
Opc = IA64::SXT1; break;
// FIXME: for now, we treat bools the same as i8s
case MVT::i8: Opc = IA64::SXT1; break;
case MVT::i16: Opc = IA64::SXT2; break;
case MVT::i32: Opc = IA64::SXT4; break;
}
BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
return Result;
}
case ISD::TRUNCATE: {
// we use the funky dep.z (deposit (zero)) instruction to deposit bits
// of R0 appropriately.
switch (N.getOperand(0).getValueType()) {
default: assert(0 && "Unknown truncate!");
case MVT::i64: break;
}
Tmp1 = SelectExpr(N.getOperand(0));
unsigned depositPos, depositLen;
switch (N.getValueType()) {
default: assert(0 && "Unknown truncate!");
case MVT::i1: {
// if input (normal reg) is 0, 0!=0 -> false (0), if 1, 1!=0 ->true (1):
BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1)
.addReg(IA64::r0);
return Result; // XXX early exit!
}
case MVT::i8: depositPos=0; depositLen=8; break;
case MVT::i16: depositPos=0; depositLen=16; break;
case MVT::i32: depositPos=0; depositLen=32; break;
}
BuildMI(BB, IA64::DEPZ, 1, Result).addReg(Tmp1)
.addImm(depositPos).addImm(depositLen);
return Result;
}
/*
case ISD::FP_ROUND: {
assert (DestType == MVT::f32 && N.getOperand(0).getValueType() == MVT::f64 &&
"error: trying to FP_ROUND something other than f64 -> f32!\n");
Tmp1 = SelectExpr(N.getOperand(0));
BuildMI(BB, IA64::FADDS, 2, Result).addReg(Tmp1).addReg(IA64::F0);
// we add 0.0 using a single precision add to do rounding
return Result;
}
*/
// FIXME: the following 4 cases need cleaning
case ISD::SINT_TO_FP: {
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = MakeReg(MVT::f64);
unsigned dummy = MakeReg(MVT::f64);
BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
BuildMI(BB, IA64::FCVTXF, 1, dummy).addReg(Tmp2);
BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
return Result;
}
case ISD::UINT_TO_FP: {
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = MakeReg(MVT::f64);
unsigned dummy = MakeReg(MVT::f64);
BuildMI(BB, IA64::SETFSIG, 1, Tmp2).addReg(Tmp1);
BuildMI(BB, IA64::FCVTXUF, 1, dummy).addReg(Tmp2);
BuildMI(BB, IA64::FNORMD, 1, Result).addReg(dummy);
return Result;
}
case ISD::FP_TO_SINT: {
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = MakeReg(MVT::f64);
BuildMI(BB, IA64::FCVTFXTRUNC, 1, Tmp2).addReg(Tmp1);
BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
return Result;
}
case ISD::FP_TO_UINT: {
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = MakeReg(MVT::f64);
BuildMI(BB, IA64::FCVTFXUTRUNC, 1, Tmp2).addReg(Tmp1);
BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(Tmp2);
return Result;
}
case ISD::ADD: {
if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
N.getOperand(0).Val->hasOneUse()) { // if we can fold this add
// into an fma, do so:
// ++FusedFP; // Statistic
Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
Tmp3 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::FMA, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
return Result; // early exit
}
if(DestType != MVT::f64 && N.getOperand(0).getOpcode() == ISD::SHL &&
N.getOperand(0).Val->hasOneUse()) { // if we might be able to fold
// this add into a shladd, try:
ConstantSDNode *CSD = NULL;
if((CSD = dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) &&
(CSD->getValue() >= 1) && (CSD->getValue() <= 4) ) { // we can:
// ++FusedSHLADD; // Statistic
Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
int shl_amt = CSD->getValue();
Tmp3 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::SHLADD, 3, Result)
.addReg(Tmp1).addImm(shl_amt).addReg(Tmp3);
return Result; // early exit
}
}
//else, fallthrough:
Tmp1 = SelectExpr(N.getOperand(0));
if(DestType != MVT::f64) { // integer addition:
switch (ponderIntegerAdditionWith(N.getOperand(1), Tmp3)) {
case 1: // adding a constant that's 14 bits
BuildMI(BB, IA64::ADDIMM14, 2, Result).addReg(Tmp1).addSImm(Tmp3);
return Result; // early exit
} // fallthrough and emit a reg+reg ADD:
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
} else { // this is a floating point addition
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::FADD, 2, Result).addReg(Tmp1).addReg(Tmp2);
}
return Result;
}
case ISD::MUL: {
if(DestType != MVT::f64) { // TODO: speed!
if(N.getOperand(1).getOpcode() != ISD::Constant) { // if not a const mul
// boring old integer multiply with xma
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = SelectExpr(N.getOperand(1));
unsigned TempFR1=MakeReg(MVT::f64);
unsigned TempFR2=MakeReg(MVT::f64);
unsigned TempFR3=MakeReg(MVT::f64);
BuildMI(BB, IA64::SETFSIG, 1, TempFR1).addReg(Tmp1);
BuildMI(BB, IA64::SETFSIG, 1, TempFR2).addReg(Tmp2);
BuildMI(BB, IA64::XMAL, 1, TempFR3).addReg(TempFR1).addReg(TempFR2)
.addReg(IA64::F0);
BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TempFR3);
return Result; // early exit
} else { // we are multiplying by an integer constant! yay
return Reg = SelectExpr(BuildConstmulSequence(N)); // avert your eyes!
}
}
else { // floating point multiply
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::FMPY, 2, Result).addReg(Tmp1).addReg(Tmp2);
return Result;
}
}
case ISD::SUB: {
if(DestType == MVT::f64 && N.getOperand(0).getOpcode() == ISD::MUL &&
N.getOperand(0).Val->hasOneUse()) { // if we can fold this sub
// into an fms, do so:
// ++FusedFP; // Statistic
Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
Tmp3 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::FMS, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
return Result; // early exit
}
Tmp2 = SelectExpr(N.getOperand(1));
if(DestType != MVT::f64) { // integer subtraction:
switch (ponderIntegerSubtractionFrom(N.getOperand(0), Tmp3)) {
case 1: // subtracting *from* an 8 bit constant:
BuildMI(BB, IA64::SUBIMM8, 2, Result).addSImm(Tmp3).addReg(Tmp2);
return Result; // early exit
} // fallthrough and emit a reg+reg SUB:
Tmp1 = SelectExpr(N.getOperand(0));
BuildMI(BB, IA64::SUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
} else { // this is a floating point subtraction
Tmp1 = SelectExpr(N.getOperand(0));
BuildMI(BB, IA64::FSUB, 2, Result).addReg(Tmp1).addReg(Tmp2);
}
return Result;
}
case ISD::FABS: {
Tmp1 = SelectExpr(N.getOperand(0));
assert(DestType == MVT::f64 && "trying to fabs something other than f64?");
BuildMI(BB, IA64::FABS, 1, Result).addReg(Tmp1);
return Result;
}
case ISD::FNEG: {
assert(DestType == MVT::f64 && "trying to fneg something other than f64?");
if (ISD::FABS == N.getOperand(0).getOpcode()) { // && hasOneUse()?
Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
BuildMI(BB, IA64::FNEGABS, 1, Result).addReg(Tmp1); // fold in abs
} else {
Tmp1 = SelectExpr(N.getOperand(0));
BuildMI(BB, IA64::FNEG, 1, Result).addReg(Tmp1); // plain old fneg
}
return Result;
}
case ISD::AND: {
switch (N.getValueType()) {
default: assert(0 && "Cannot AND this type!");
case MVT::i1: { // if a bool, we emit a pseudocode AND
unsigned pA = SelectExpr(N.getOperand(0));
unsigned pB = SelectExpr(N.getOperand(1));
/* our pseudocode for AND is:
*
(pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
cmp.eq pTemp,p0 = r0,r0 // pTemp = NOT pB
;;
(pB) cmp.ne pTemp,p0 = r0,r0
;;
(pTemp)cmp.ne pC,p0 = r0,r0 // if (NOT pB) pC = 0
*/
unsigned pTemp = MakeReg(MVT::i1);
unsigned bogusTemp1 = MakeReg(MVT::i1);
unsigned bogusTemp2 = MakeReg(MVT::i1);
unsigned bogusTemp3 = MakeReg(MVT::i1);
unsigned bogusTemp4 = MakeReg(MVT::i1);
BuildMI(BB, IA64::PCMPEQUNC, 3, bogusTemp1)
.addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
BuildMI(BB, IA64::CMPEQ, 2, bogusTemp2)
.addReg(IA64::r0).addReg(IA64::r0);
BuildMI(BB, IA64::TPCMPNE, 3, pTemp)
.addReg(bogusTemp2).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
BuildMI(BB, IA64::TPCMPNE, 3, Result)
.addReg(bogusTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pTemp);
break;
}
// if not a bool, we just AND away:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64: {
Tmp1 = SelectExpr(N.getOperand(0));
switch (ponderIntegerAndWith(N.getOperand(1), Tmp3)) {
case 1: // ANDing a constant that is 2^n-1 for some n
switch (Tmp3) {
case 8: // if AND 0x00000000000000FF, be quaint and use zxt1
BuildMI(BB, IA64::ZXT1, 1, Result).addReg(Tmp1);
break;
case 16: // if AND 0x000000000000FFFF, be quaint and use zxt2
BuildMI(BB, IA64::ZXT2, 1, Result).addReg(Tmp1);
break;
case 32: // if AND 0x00000000FFFFFFFF, be quaint and use zxt4
BuildMI(BB, IA64::ZXT4, 1, Result).addReg(Tmp1);
break;
default: // otherwise, use dep.z to paste zeros
BuildMI(BB, IA64::DEPZ, 3, Result).addReg(Tmp1)
.addImm(0).addImm(Tmp3);
break;
}
return Result; // early exit
} // fallthrough and emit a simple AND:
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::AND, 2, Result).addReg(Tmp1).addReg(Tmp2);
}
}
return Result;
}
case ISD::OR: {
switch (N.getValueType()) {
default: assert(0 && "Cannot OR this type!");
case MVT::i1: { // if a bool, we emit a pseudocode OR
unsigned pA = SelectExpr(N.getOperand(0));
unsigned pB = SelectExpr(N.getOperand(1));
unsigned pTemp1 = MakeReg(MVT::i1);
/* our pseudocode for OR is:
*
pC = pA OR pB
-------------
(pA) cmp.eq.unc pC,p0 = r0,r0 // pC = pA
;;
(pB) cmp.eq pC,p0 = r0,r0 // if (pB) pC = 1
*/
BuildMI(BB, IA64::PCMPEQUNC, 3, pTemp1)
.addReg(IA64::r0).addReg(IA64::r0).addReg(pA);
BuildMI(BB, IA64::TPCMPEQ, 3, Result)
.addReg(pTemp1).addReg(IA64::r0).addReg(IA64::r0).addReg(pB);
break;
}
// if not a bool, we just OR away:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64: {
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::OR, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
}
}
return Result;
}
case ISD::XOR: {
switch (N.getValueType()) {
default: assert(0 && "Cannot XOR this type!");
case MVT::i1: { // if a bool, we emit a pseudocode XOR
unsigned pY = SelectExpr(N.getOperand(0));
unsigned pZ = SelectExpr(N.getOperand(1));
/* one possible routine for XOR is:
// Compute px = py ^ pz
// using sum of products: px = (py & !pz) | (pz & !py)
// Uses 5 instructions in 3 cycles.
// cycle 1
(pz) cmp.eq.unc px = r0, r0 // px = pz
(py) cmp.eq.unc pt = r0, r0 // pt = py
;;
// cycle 2
(pt) cmp.ne.and px = r0, r0 // px = px & !pt (px = pz & !pt)
(pz) cmp.ne.and pt = r0, r0 // pt = pt & !pz
;;
} { .mmi
// cycle 3
(pt) cmp.eq.or px = r0, r0 // px = px | pt
*** Another, which we use here, requires one scratch GR. it is:
mov rt = 0 // initialize rt off critical path
;;
// cycle 1
(pz) cmp.eq.unc px = r0, r0 // px = pz
(pz) mov rt = 1 // rt = pz
;;
// cycle 2
(py) cmp.ne px = 1, rt // if (py) px = !pz
.. these routines kindly provided by Jim Hull
*/
unsigned rt = MakeReg(MVT::i64);
// these two temporaries will never actually appear,
// due to the two-address form of some of the instructions below
unsigned bogoPR = MakeReg(MVT::i1); // becomes Result
unsigned bogoGR = MakeReg(MVT::i64); // becomes rt
BuildMI(BB, IA64::MOV, 1, bogoGR).addReg(IA64::r0);
BuildMI(BB, IA64::PCMPEQUNC, 3, bogoPR)
.addReg(IA64::r0).addReg(IA64::r0).addReg(pZ);
BuildMI(BB, IA64::TPCADDIMM22, 2, rt)
.addReg(bogoGR).addImm(1).addReg(pZ);
BuildMI(BB, IA64::TPCMPIMM8NE, 3, Result)
.addReg(bogoPR).addImm(1).addReg(rt).addReg(pY);
break;
}
// if not a bool, we just XOR away:
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64: {
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
}
}
return Result;
}
case ISD::CTPOP: {
Tmp1 = SelectExpr(N.getOperand(0));
BuildMI(BB, IA64::POPCNT, 1, Result).addReg(Tmp1);
return Result;
}
case ISD::SHL: {
Tmp1 = SelectExpr(N.getOperand(0));
if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
Tmp2 = CN->getValue();
BuildMI(BB, IA64::SHLI, 2, Result).addReg(Tmp1).addImm(Tmp2);
} else {
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::SHL, 2, Result).addReg(Tmp1).addReg(Tmp2);
}
return Result;
}
case ISD::SRL: {
Tmp1 = SelectExpr(N.getOperand(0));
if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
Tmp2 = CN->getValue();
BuildMI(BB, IA64::SHRUI, 2, Result).addReg(Tmp1).addImm(Tmp2);
} else {
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addReg(Tmp2);
}
return Result;
}
case ISD::SRA: {
Tmp1 = SelectExpr(N.getOperand(0));
if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
Tmp2 = CN->getValue();
BuildMI(BB, IA64::SHRSI, 2, Result).addReg(Tmp1).addImm(Tmp2);
} else {
Tmp2 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::SHRS, 2, Result).addReg(Tmp1).addReg(Tmp2);
}
return Result;
}
case ISD::SDIV:
case ISD::UDIV:
case ISD::SREM:
case ISD::UREM: {
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = SelectExpr(N.getOperand(1));
bool isFP=false;
if(DestType == MVT::f64) // XXX: we're not gonna be fed MVT::f32, are we?
isFP=true;
bool isModulus=false; // is it a division or a modulus?
bool isSigned=false;
switch(N.getOpcode()) {
case ISD::SDIV: isModulus=false; isSigned=true; break;
case ISD::UDIV: isModulus=false; isSigned=false; break;
case ISD::SREM: isModulus=true; isSigned=true; break;
case ISD::UREM: isModulus=true; isSigned=false; break;
}
if(!isModulus && !isFP) { // if this is an integer divide,
switch (ponderIntegerDivisionBy(N.getOperand(1), isSigned, Tmp3)) {
case 1: // division by a constant that's a power of 2
Tmp1 = SelectExpr(N.getOperand(0));
if(isSigned) { // argument could be negative, so emit some code:
unsigned divAmt=Tmp3;
unsigned tempGR1=MakeReg(MVT::i64);
unsigned tempGR2=MakeReg(MVT::i64);
unsigned tempGR3=MakeReg(MVT::i64);
BuildMI(BB, IA64::SHRS, 2, tempGR1)
.addReg(Tmp1).addImm(divAmt-1);
BuildMI(BB, IA64::EXTRU, 3, tempGR2)
.addReg(tempGR1).addImm(64-divAmt).addImm(divAmt);
BuildMI(BB, IA64::ADD, 2, tempGR3)
.addReg(Tmp1).addReg(tempGR2);
BuildMI(BB, IA64::SHRS, 2, Result)
.addReg(tempGR3).addImm(divAmt);
}
else // unsigned div-by-power-of-2 becomes a simple shift right:
BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addImm(Tmp3);
return Result; // early exit
}
}
unsigned TmpPR=MakeReg(MVT::i1); // we need two scratch
unsigned TmpPR2=MakeReg(MVT::i1); // predicate registers,
unsigned TmpF1=MakeReg(MVT::f64); // and one metric truckload of FP regs.
unsigned TmpF2=MakeReg(MVT::f64); // lucky we have IA64?
unsigned TmpF3=MakeReg(MVT::f64); // well, the real FIXME is to have
unsigned TmpF4=MakeReg(MVT::f64); // isTwoAddress forms of these
unsigned TmpF5=MakeReg(MVT::f64); // FP instructions so we can end up with
unsigned TmpF6=MakeReg(MVT::f64); // stuff like setf.sig f10=f10 etc.
unsigned TmpF7=MakeReg(MVT::f64);
unsigned TmpF8=MakeReg(MVT::f64);
unsigned TmpF9=MakeReg(MVT::f64);
unsigned TmpF10=MakeReg(MVT::f64);
unsigned TmpF11=MakeReg(MVT::f64);
unsigned TmpF12=MakeReg(MVT::f64);
unsigned TmpF13=MakeReg(MVT::f64);
unsigned TmpF14=MakeReg(MVT::f64);
unsigned TmpF15=MakeReg(MVT::f64);
// OK, emit some code:
if(!isFP) {
// first, load the inputs into FP regs.
BuildMI(BB, IA64::SETFSIG, 1, TmpF1).addReg(Tmp1);
BuildMI(BB, IA64::SETFSIG, 1, TmpF2).addReg(Tmp2);
// next, convert the inputs to FP
if(isSigned) {
BuildMI(BB, IA64::FCVTXF, 1, TmpF3).addReg(TmpF1);
BuildMI(BB, IA64::FCVTXF, 1, TmpF4).addReg(TmpF2);
} else {
BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF3).addReg(TmpF1);
BuildMI(BB, IA64::FCVTXUFS1, 1, TmpF4).addReg(TmpF2);
}
} else { // this is an FP divide/remainder, so we 'leak' some temp
// regs and assign TmpF3=Tmp1, TmpF4=Tmp2
TmpF3=Tmp1;
TmpF4=Tmp2;
}
// we start by computing an approximate reciprocal (good to 9 bits?)
// note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
BuildMI(BB, IA64::FRCPAS1, 4)
.addReg(TmpF5, MachineOperand::Def)
.addReg(TmpPR, MachineOperand::Def)
.addReg(TmpF3).addReg(TmpF4);
if(!isModulus) { // if this is a divide, we worry about div-by-zero
unsigned bogusPR=MakeReg(MVT::i1); // won't appear, due to twoAddress
// TPCMPNE below
BuildMI(BB, IA64::CMPEQ, 2, bogusPR).addReg(IA64::r0).addReg(IA64::r0);
BuildMI(BB, IA64::TPCMPNE, 3, TmpPR2).addReg(bogusPR)
.addReg(IA64::r0).addReg(IA64::r0).addReg(TmpPR);
}
// now we apply newton's method, thrice! (FIXME: this is ~72 bits of
// precision, don't need this much for f32/i32)
BuildMI(BB, IA64::CFNMAS1, 4, TmpF6)
.addReg(TmpF4).addReg(TmpF5).addReg(IA64::F1).addReg(TmpPR);
BuildMI(BB, IA64::CFMAS1, 4, TmpF7)
.addReg(TmpF3).addReg(TmpF5).addReg(IA64::F0).addReg(TmpPR);
BuildMI(BB, IA64::CFMAS1, 4, TmpF8)
.addReg(TmpF6).addReg(TmpF6).addReg(IA64::F0).addReg(TmpPR);
BuildMI(BB, IA64::CFMAS1, 4, TmpF9)
.addReg(TmpF6).addReg(TmpF7).addReg(TmpF7).addReg(TmpPR);
BuildMI(BB, IA64::CFMAS1, 4,TmpF10)
.addReg(TmpF6).addReg(TmpF5).addReg(TmpF5).addReg(TmpPR);
BuildMI(BB, IA64::CFMAS1, 4,TmpF11)
.addReg(TmpF8).addReg(TmpF9).addReg(TmpF9).addReg(TmpPR);
BuildMI(BB, IA64::CFMAS1, 4,TmpF12)
.addReg(TmpF8).addReg(TmpF10).addReg(TmpF10).addReg(TmpPR);
BuildMI(BB, IA64::CFNMAS1, 4,TmpF13)
.addReg(TmpF4).addReg(TmpF11).addReg(TmpF3).addReg(TmpPR);
// FIXME: this is unfortunate :(
// the story is that the dest reg of the fnma above and the fma below
// (and therefore possibly the src of the fcvt.fx[u] as well) cannot
// be the same register, or this code breaks if the first argument is
// zero. (e.g. without this hack, 0%8 yields -64, not 0.)
BuildMI(BB, IA64::CFMAS1, 4,TmpF14)
.addReg(TmpF13).addReg(TmpF12).addReg(TmpF11).addReg(TmpPR);
if(isModulus) { // XXX: fragile! fixes _only_ mod, *breaks* div! !
BuildMI(BB, IA64::IUSE, 1).addReg(TmpF13); // hack :(
}
if(!isFP) {
// round to an integer
if(isSigned)
BuildMI(BB, IA64::FCVTFXTRUNCS1, 1, TmpF15).addReg(TmpF14);
else
BuildMI(BB, IA64::FCVTFXUTRUNCS1, 1, TmpF15).addReg(TmpF14);
} else {
BuildMI(BB, IA64::FMOV, 1, TmpF15).addReg(TmpF14);
// EXERCISE: can you see why TmpF15=TmpF14 does not work here, and
// we really do need the above FMOV? ;)
}
if(!isModulus) {
if(isFP) { // extra worrying about div-by-zero
unsigned bogoResult=MakeReg(MVT::f64);
// we do a 'conditional fmov' (of the correct result, depending
// on how the frcpa predicate turned out)
BuildMI(BB, IA64::PFMOV, 2, bogoResult)
.addReg(TmpF12).addReg(TmpPR2);
BuildMI(BB, IA64::CFMOV, 2, Result)
.addReg(bogoResult).addReg(TmpF15).addReg(TmpPR);
}
else {
BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(TmpF15);
}
} else { // this is a modulus
if(!isFP) {
// answer = q * (-b) + a
unsigned ModulusResult = MakeReg(MVT::f64);
unsigned TmpF = MakeReg(MVT::f64);
unsigned TmpI = MakeReg(MVT::i64);
BuildMI(BB, IA64::SUB, 2, TmpI).addReg(IA64::r0).addReg(Tmp2);
BuildMI(BB, IA64::SETFSIG, 1, TmpF).addReg(TmpI);
BuildMI(BB, IA64::XMAL, 3, ModulusResult)
.addReg(TmpF15).addReg(TmpF).addReg(TmpF1);
BuildMI(BB, IA64::GETFSIG, 1, Result).addReg(ModulusResult);
} else { // FP modulus! The horror... the horror....
assert(0 && "sorry, no FP modulus just yet!\n!\n");
}
}
return Result;
}
case ISD::SIGN_EXTEND_INREG: {
Tmp1 = SelectExpr(N.getOperand(0));
switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
default:
Node->dump();
assert(0 && "don't know how to sign extend this type");
break;
case MVT::i8: Opc = IA64::SXT1; break;
case MVT::i16: Opc = IA64::SXT2; break;
case MVT::i32: Opc = IA64::SXT4; break;
}
BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
return Result;
}
case ISD::SETCC: {
Tmp1 = SelectExpr(N.getOperand(0));
if (SetCCSDNode *SetCC = dyn_cast<SetCCSDNode>(Node)) {
if (MVT::isInteger(SetCC->getOperand(0).getValueType())) {
if(ConstantSDNode *CSDN =
dyn_cast<ConstantSDNode>(N.getOperand(1))) {
// if we are comparing against a constant zero
if(CSDN->getValue()==0)
Tmp2 = IA64::r0; // then we can just compare against r0
else
Tmp2 = SelectExpr(N.getOperand(1));
} else // not comparing against a constant
Tmp2 = SelectExpr(N.getOperand(1));
switch (SetCC->getCondition()) {
default: assert(0 && "Unknown integer comparison!");
case ISD::SETEQ:
BuildMI(BB, IA64::CMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETGT:
BuildMI(BB, IA64::CMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETGE:
BuildMI(BB, IA64::CMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETLT:
BuildMI(BB, IA64::CMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETLE:
BuildMI(BB, IA64::CMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETNE:
BuildMI(BB, IA64::CMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETULT:
BuildMI(BB, IA64::CMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETUGT:
BuildMI(BB, IA64::CMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETULE:
BuildMI(BB, IA64::CMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETUGE:
BuildMI(BB, IA64::CMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
}
}
else { // if not integer, should be FP. FIXME: what about bools? ;)
assert(SetCC->getOperand(0).getValueType() != MVT::f32 &&
"error: SETCC should have had incoming f32 promoted to f64!\n");
if(ConstantFPSDNode *CFPSDN =
dyn_cast<ConstantFPSDNode>(N.getOperand(1))) {
// if we are comparing against a constant +0.0 or +1.0
if(CFPSDN->isExactlyValue(+0.0))
Tmp2 = IA64::F0; // then we can just compare against f0
else if(CFPSDN->isExactlyValue(+1.0))
Tmp2 = IA64::F1; // or f1
else
Tmp2 = SelectExpr(N.getOperand(1));
} else // not comparing against a constant
Tmp2 = SelectExpr(N.getOperand(1));
switch (SetCC->getCondition()) {
default: assert(0 && "Unknown FP comparison!");
case ISD::SETEQ:
BuildMI(BB, IA64::FCMPEQ, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETGT:
BuildMI(BB, IA64::FCMPGT, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETGE:
BuildMI(BB, IA64::FCMPGE, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETLT:
BuildMI(BB, IA64::FCMPLT, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETLE:
BuildMI(BB, IA64::FCMPLE, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETNE:
BuildMI(BB, IA64::FCMPNE, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETULT:
BuildMI(BB, IA64::FCMPLTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETUGT:
BuildMI(BB, IA64::FCMPGTU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETULE:
BuildMI(BB, IA64::FCMPLEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
case ISD::SETUGE:
BuildMI(BB, IA64::FCMPGEU, 2, Result).addReg(Tmp1).addReg(Tmp2);
break;
}
}
}
else
assert(0 && "this setcc not implemented yet");
return Result;
}
case ISD::EXTLOAD:
case ISD::ZEXTLOAD:
case ISD::LOAD: {
// Make sure we generate both values.
if (Result != 1)
ExprMap[N.getValue(1)] = 1; // Generate the token
else
Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType());
bool isBool=false;
if(opcode == ISD::LOAD) { // this is a LOAD
switch (Node->getValueType(0)) {
default: assert(0 && "Cannot load this type!");
case MVT::i1: Opc = IA64::LD1; isBool=true; break;
// FIXME: for now, we treat bool loads the same as i8 loads */
case MVT::i8: Opc = IA64::LD1; break;
case MVT::i16: Opc = IA64::LD2; break;
case MVT::i32: Opc = IA64::LD4; break;
case MVT::i64: Opc = IA64::LD8; break;
case MVT::f32: Opc = IA64::LDF4; break;
case MVT::f64: Opc = IA64::LDF8; break;
}
} else { // this is an EXTLOAD or ZEXTLOAD
MVT::ValueType TypeBeingLoaded =
cast<VTSDNode>(Node->getOperand(3))->getVT();
switch (TypeBeingLoaded) {
default: assert(0 && "Cannot extload/zextload this type!");
// FIXME: bools?
case MVT::i8: Opc = IA64::LD1; break;
case MVT::i16: Opc = IA64::LD2; break;
case MVT::i32: Opc = IA64::LD4; break;
case MVT::f32: Opc = IA64::LDF4; break;
}
}
SDOperand Chain = N.getOperand(0);
SDOperand Address = N.getOperand(1);
if(Address.getOpcode() == ISD::GlobalAddress) {
Select(Chain);
unsigned dummy = MakeReg(MVT::i64);
unsigned dummy2 = MakeReg(MVT::i64);
BuildMI(BB, IA64::ADD, 2, dummy)
.addGlobalAddress(cast<GlobalAddressSDNode>(Address)->getGlobal())
.addReg(IA64::r1);
BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
if(!isBool)
BuildMI(BB, Opc, 1, Result).addReg(dummy2);
else { // emit a little pseudocode to load a bool (stored in one byte)
// into a predicate register
assert(Opc==IA64::LD1 && "problem loading a bool");
unsigned dummy3 = MakeReg(MVT::i64);
BuildMI(BB, Opc, 1, dummy3).addReg(dummy2);
// we compare to 0. true? 0. false? 1.
BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
}
} else if(ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) {
Select(Chain);
IA64Lowering.restoreGP(BB);
unsigned dummy = MakeReg(MVT::i64);
BuildMI(BB, IA64::ADD, 2, dummy).addConstantPoolIndex(CP->getIndex())
.addReg(IA64::r1); // CPI+GP
if(!isBool)
BuildMI(BB, Opc, 1, Result).addReg(dummy);
else { // emit a little pseudocode to load a bool (stored in one byte)
// into a predicate register
assert(Opc==IA64::LD1 && "problem loading a bool");
unsigned dummy3 = MakeReg(MVT::i64);
BuildMI(BB, Opc, 1, dummy3).addReg(dummy);
// we compare to 0. true? 0. false? 1.
BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
}
} else if(Address.getOpcode() == ISD::FrameIndex) {
Select(Chain); // FIXME ? what about bools?
unsigned dummy = MakeReg(MVT::i64);
BuildMI(BB, IA64::MOV, 1, dummy)
.addFrameIndex(cast<FrameIndexSDNode>(Address)->getIndex());
if(!isBool)
BuildMI(BB, Opc, 1, Result).addReg(dummy);
else { // emit a little pseudocode to load a bool (stored in one byte)
// into a predicate register
assert(Opc==IA64::LD1 && "problem loading a bool");
unsigned dummy3 = MakeReg(MVT::i64);
BuildMI(BB, Opc, 1, dummy3).addReg(dummy);
// we compare to 0. true? 0. false? 1.
BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy3).addReg(IA64::r0);
}
} else { // none of the above...
Select(Chain);
Tmp2 = SelectExpr(Address);
if(!isBool)
BuildMI(BB, Opc, 1, Result).addReg(Tmp2);
else { // emit a little pseudocode to load a bool (stored in one byte)
// into a predicate register
assert(Opc==IA64::LD1 && "problem loading a bool");
unsigned dummy = MakeReg(MVT::i64);
BuildMI(BB, Opc, 1, dummy).addReg(Tmp2);
// we compare to 0. true? 0. false? 1.
BuildMI(BB, IA64::CMPNE, 2, Result).addReg(dummy).addReg(IA64::r0);
}
}
return Result;
}
case ISD::CopyFromReg: {
if (Result == 1)
Result = ExprMap[N.getValue(0)] =
MakeReg(N.getValue(0).getValueType());
SDOperand Chain = N.getOperand(0);
Select(Chain);
unsigned r = dyn_cast<RegSDNode>(Node)->getReg();
if(N.getValueType() == MVT::i1) // if a bool, we use pseudocode
BuildMI(BB, IA64::PCMPEQUNC, 3, Result)
.addReg(IA64::r0).addReg(IA64::r0).addReg(r);
// (r) Result =cmp.eq.unc(r0,r0)
else
BuildMI(BB, IA64::MOV, 1, Result).addReg(r); // otherwise MOV
return Result;
}
case ISD::TAILCALL:
case ISD::CALL: {
Select(N.getOperand(0));
// The chain for this call is now lowered.
ExprMap.insert(std::make_pair(N.getValue(Node->getNumValues()-1), 1));
//grab the arguments
std::vector<unsigned> argvregs;
for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
argvregs.push_back(SelectExpr(N.getOperand(i)));
// see section 8.5.8 of "Itanium Software Conventions and
// Runtime Architecture Guide to see some examples of what's going
// on here. (in short: int args get mapped 1:1 'slot-wise' to out0->out7,
// while FP args get mapped to F8->F15 as needed)
unsigned used_FPArgs=0; // how many FP Args have been used so far?
// in reg args
for(int i = 0, e = std::min(8, (int)argvregs.size()); i < e; ++i)
{
unsigned intArgs[] = {IA64::out0, IA64::out1, IA64::out2, IA64::out3,
IA64::out4, IA64::out5, IA64::out6, IA64::out7 };
unsigned FPArgs[] = {IA64::F8, IA64::F9, IA64::F10, IA64::F11,
IA64::F12, IA64::F13, IA64::F14, IA64::F15 };
switch(N.getOperand(i+2).getValueType())
{
default: // XXX do we need to support MVT::i1 here?
Node->dump();
N.getOperand(i).Val->dump();
std::cerr << "Type for " << i << " is: " <<
N.getOperand(i+2).getValueType() << std::endl;
assert(0 && "Unknown value type for call");
case MVT::i64:
BuildMI(BB, IA64::MOV, 1, intArgs[i]).addReg(argvregs[i]);
break;
case MVT::f64:
BuildMI(BB, IA64::FMOV, 1, FPArgs[used_FPArgs++])
.addReg(argvregs[i]);
// FIXME: we don't need to do this _all_ the time:
BuildMI(BB, IA64::GETFD, 1, intArgs[i]).addReg(argvregs[i]);
break;
}
}
//in mem args
for (int i = 8, e = argvregs.size(); i < e; ++i)
{
unsigned tempAddr = MakeReg(MVT::i64);
switch(N.getOperand(i+2).getValueType()) {
default:
Node->dump();
N.getOperand(i).Val->dump();
std::cerr << "Type for " << i << " is: " <<
N.getOperand(i+2).getValueType() << "\n";
assert(0 && "Unknown value type for call");
case MVT::i1: // FIXME?
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
BuildMI(BB, IA64::ADDIMM22, 2, tempAddr)
.addReg(IA64::r12).addImm(16 + (i - 8) * 8); // r12 is SP
BuildMI(BB, IA64::ST8, 2).addReg(tempAddr).addReg(argvregs[i]);
break;
case MVT::f32:
case MVT::f64:
BuildMI(BB, IA64::ADDIMM22, 2, tempAddr)
.addReg(IA64::r12).addImm(16 + (i - 8) * 8); // r12 is SP
BuildMI(BB, IA64::STF8, 2).addReg(tempAddr).addReg(argvregs[i]);
break;
}
}
// build the right kind of call. if we can branch directly, do so:
if (GlobalAddressSDNode *GASD =
dyn_cast<GlobalAddressSDNode>(N.getOperand(1)))
{
BuildMI(BB, IA64::BRCALL, 1).addGlobalAddress(GASD->getGlobal(),true);
IA64Lowering.restoreGP_SP_RP(BB);
} else
if (ExternalSymbolSDNode *ESSDN =
dyn_cast<ExternalSymbolSDNode>(N.getOperand(1)))
{ // FIXME : currently need this case for correctness, to avoid
// "non-pic code with imm relocation against dynamic symbol" errors
BuildMI(BB, IA64::BRCALL, 1)
.addExternalSymbol(ESSDN->getSymbol(), true);
IA64Lowering.restoreGP_SP_RP(BB);
}
else { // otherwise we need to get the function descriptor
// load the branch target (function)'s entry point and
// GP, then branch
Tmp1 = SelectExpr(N.getOperand(1));
unsigned targetEntryPoint=MakeReg(MVT::i64);
unsigned targetGPAddr=MakeReg(MVT::i64);
unsigned currentGP=MakeReg(MVT::i64);
// b6 is a scratch branch register, we load the target entry point
// from the base of the function descriptor
BuildMI(BB, IA64::LD8, 1, targetEntryPoint).addReg(Tmp1);
BuildMI(BB, IA64::MOV, 1, IA64::B6).addReg(targetEntryPoint);
// save the current GP:
BuildMI(BB, IA64::MOV, 1, currentGP).addReg(IA64::r1);
/* TODO: we need to make sure doing this never, ever loads a
* bogus value into r1 (GP). */
// load the target GP (which is at mem[functiondescriptor+8])
BuildMI(BB, IA64::ADDIMM22, 2, targetGPAddr)
.addReg(Tmp1).addImm(8); // FIXME: addimm22? why not postincrement ld
BuildMI(BB, IA64::LD8, 1, IA64::r1).addReg(targetGPAddr);
// and then jump: (well, call)
BuildMI(BB, IA64::BRCALL, 1).addReg(IA64::B6);
// and finally restore the old GP
BuildMI(BB, IA64::MOV, 1, IA64::r1).addReg(currentGP);
IA64Lowering.restoreSP_RP(BB);
}
switch (Node->getValueType(0)) {
default: assert(0 && "Unknown value type for call result!");
case MVT::Other: return 1;
case MVT::i1:
BuildMI(BB, IA64::CMPNE, 2, Result)
.addReg(IA64::r8).addReg(IA64::r0);
break;
case MVT::i8:
case MVT::i16:
case MVT::i32:
case MVT::i64:
BuildMI(BB, IA64::MOV, 1, Result).addReg(IA64::r8);
break;
case MVT::f64:
BuildMI(BB, IA64::FMOV, 1, Result).addReg(IA64::F8);
break;
}
return Result+N.ResNo;
}
} // <- uhhh XXX
return 0;
}
void ISel::Select(SDOperand N) {
unsigned Tmp1, Tmp2, Opc;
unsigned opcode = N.getOpcode();
if (!LoweredTokens.insert(N).second)
return; // Already selected.
SDNode *Node = N.Val;
switch (Node->getOpcode()) {
default:
Node->dump(); std::cerr << "\n";
assert(0 && "Node not handled yet!");
case ISD::EntryToken: return; // Noop
case ISD::TokenFactor: {
for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
Select(Node->getOperand(i));
return;
}
case ISD::CopyToReg: {
Select(N.getOperand(0));
Tmp1 = SelectExpr(N.getOperand(1));
Tmp2 = cast<RegSDNode>(N)->getReg();
if (Tmp1 != Tmp2) {
if(N.getValueType() == MVT::i1) // if a bool, we use pseudocode
BuildMI(BB, IA64::PCMPEQUNC, 3, Tmp2)
.addReg(IA64::r0).addReg(IA64::r0).addReg(Tmp1);
// (Tmp1) Tmp2 = cmp.eq.unc(r0,r0)
else
BuildMI(BB, IA64::MOV, 1, Tmp2).addReg(Tmp1);
// XXX is this the right way 'round? ;)
}
return;
}
case ISD::RET: {
/* what the heck is going on here:
<_sabre_> ret with two operands is obvious: chain and value
<camel_> yep
<_sabre_> ret with 3 values happens when 'expansion' occurs
<_sabre_> e.g. i64 gets split into 2x i32
<camel_> oh right
<_sabre_> you don't have this case on ia64
<camel_> yep
<_sabre_> so the two returned values go into EAX/EDX on ia32
<camel_> ahhh *memories*
<_sabre_> :)
<camel_> ok, thanks :)
<_sabre_> so yeah, everything that has a side effect takes a 'token chain'
<_sabre_> this is the first operand always
<_sabre_> these operand often define chains, they are the last operand
<_sabre_> they are printed as 'ch' if you do DAG.dump()
*/
switch (N.getNumOperands()) {
default:
assert(0 && "Unknown return instruction!");
case 2:
Select(N.getOperand(0));
Tmp1 = SelectExpr(N.getOperand(1));
switch (N.getOperand(1).getValueType()) {
default: assert(0 && "All other types should have been promoted!!");
// FIXME: do I need to add support for bools here?
// (return '0' or '1' r8, basically...)
//
// FIXME: need to round floats - 80 bits is bad, the tester
// told me so
case MVT::i64:
// we mark r8 as live on exit up above in LowerArguments()
BuildMI(BB, IA64::MOV, 1, IA64::r8).addReg(Tmp1);
break;
case MVT::f64:
// we mark F8 as live on exit up above in LowerArguments()
BuildMI(BB, IA64::FMOV, 1, IA64::F8).addReg(Tmp1);
}
break;
case 1:
Select(N.getOperand(0));
break;
}
// before returning, restore the ar.pfs register (set by the 'alloc' up top)
BuildMI(BB, IA64::MOV, 1).addReg(IA64::AR_PFS).addReg(IA64Lowering.VirtGPR);
BuildMI(BB, IA64::RET, 0); // and then just emit a 'ret' instruction
return;
}
case ISD::BR: {
Select(N.getOperand(0));
MachineBasicBlock *Dest =
cast<BasicBlockSDNode>(N.getOperand(1))->getBasicBlock();
BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(IA64::p0).addMBB(Dest);
// XXX HACK! we do _not_ need long branches all the time
return;
}
case ISD::ImplicitDef: {
Select(N.getOperand(0));
BuildMI(BB, IA64::IDEF, 0, cast<RegSDNode>(N)->getReg());
return;
}
case ISD::BRCOND: {
MachineBasicBlock *Dest =
cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
Select(N.getOperand(0));
Tmp1 = SelectExpr(N.getOperand(1));
BuildMI(BB, IA64::BRLCOND_NOTCALL, 1).addReg(Tmp1).addMBB(Dest);
// XXX HACK! we do _not_ need long branches all the time
return;
}
case ISD::EXTLOAD:
case ISD::ZEXTLOAD:
case ISD::SEXTLOAD:
case ISD::LOAD:
case ISD::TAILCALL:
case ISD::CALL:
case ISD::CopyFromReg:
case ISD::DYNAMIC_STACKALLOC:
SelectExpr(N);
return;
case ISD::TRUNCSTORE:
case ISD::STORE: {
Select(N.getOperand(0));
Tmp1 = SelectExpr(N.getOperand(1)); // value
bool isBool=false;
if(opcode == ISD::STORE) {
switch (N.getOperand(1).getValueType()) {
default: assert(0 && "Cannot store this type!");
case MVT::i1: Opc = IA64::ST1; isBool=true; break;
// FIXME?: for now, we treat bool loads the same as i8 stores */
case MVT::i8: Opc = IA64::ST1; break;
case MVT::i16: Opc = IA64::ST2; break;
case MVT::i32: Opc = IA64::ST4; break;
case MVT::i64: Opc = IA64::ST8; break;
case MVT::f32: Opc = IA64::STF4; break;
case MVT::f64: Opc = IA64::STF8; break;
}
} else { // truncstore
switch(cast<VTSDNode>(Node->getOperand(4))->getVT()) {
default: assert(0 && "unknown type in truncstore");
case MVT::i1: Opc = IA64::ST1; isBool=true; break;
//FIXME: DAG does not promote this load?
case MVT::i8: Opc = IA64::ST1; break;
case MVT::i16: Opc = IA64::ST2; break;
case MVT::i32: Opc = IA64::ST4; break;
case MVT::f32: Opc = IA64::STF4; break;
}
}
if(N.getOperand(2).getOpcode() == ISD::GlobalAddress) {
unsigned dummy = MakeReg(MVT::i64);
unsigned dummy2 = MakeReg(MVT::i64);
BuildMI(BB, IA64::ADD, 2, dummy)
.addGlobalAddress(cast<GlobalAddressSDNode>
(N.getOperand(2))->getGlobal()).addReg(IA64::r1);
BuildMI(BB, IA64::LD8, 1, dummy2).addReg(dummy);
if(!isBool)
BuildMI(BB, Opc, 2).addReg(dummy2).addReg(Tmp1);
else { // we are storing a bool, so emit a little pseudocode
// to store a predicate register as one byte
assert(Opc==IA64::ST1);
unsigned dummy3 = MakeReg(MVT::i64);
unsigned dummy4 = MakeReg(MVT::i64);
BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
.addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
BuildMI(BB, Opc, 2).addReg(dummy2).addReg(dummy4);
}
} else if(N.getOperand(2).getOpcode() == ISD::FrameIndex) {
// FIXME? (what about bools?)
unsigned dummy = MakeReg(MVT::i64);
BuildMI(BB, IA64::MOV, 1, dummy)
.addFrameIndex(cast<FrameIndexSDNode>(N.getOperand(2))->getIndex());
BuildMI(BB, Opc, 2).addReg(dummy).addReg(Tmp1);
} else { // otherwise
Tmp2 = SelectExpr(N.getOperand(2)); //address
if(!isBool)
BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(Tmp1);
else { // we are storing a bool, so emit a little pseudocode
// to store a predicate register as one byte
assert(Opc==IA64::ST1);
unsigned dummy3 = MakeReg(MVT::i64);
unsigned dummy4 = MakeReg(MVT::i64);
BuildMI(BB, IA64::MOV, 1, dummy3).addReg(IA64::r0);
BuildMI(BB, IA64::TPCADDIMM22, 2, dummy4)
.addReg(dummy3).addImm(1).addReg(Tmp1); // if(Tmp1) dummy=0+1;
BuildMI(BB, Opc, 2).addReg(Tmp2).addReg(dummy4);
}
}
return;
}
case ISD::CALLSEQ_START:
case ISD::CALLSEQ_END: {
Select(N.getOperand(0));
Tmp1 = cast<ConstantSDNode>(N.getOperand(1))->getValue();
Opc = N.getOpcode() == ISD::CALLSEQ_START ? IA64::ADJUSTCALLSTACKDOWN :
IA64::ADJUSTCALLSTACKUP;
BuildMI(BB, Opc, 1).addImm(Tmp1);
return;
}
return;
}
assert(0 && "GAME OVER. INSERT COIN?");
}
/// createIA64PatternInstructionSelector - This pass converts an LLVM function
/// into a machine code representation using pattern matching and a machine
/// description file.
///
FunctionPass *llvm::createIA64PatternInstructionSelector(TargetMachine &TM) {
return new ISel(TM);
}
Reference in New Issue View Git Blame Copy Permalink