RPCS3/llvm
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm.git synced 2024-12-28 07:05:03 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix bogus division-by-power-of-2 (was wrong for negative input, adds extr insn)

Browse Source 
fix hack in division (clean up frcpa instruction)


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21153 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Duraid Madina 2005-04-08 10:01:48 +00:00
parent bf70b838e7
commit 6dcceb5ecb
2 changed files with 25 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
lib/Target/IA64/IA64ISelPattern.cpp
View File

@ -1122,9 +1122,21 @@ pC = pA OR pB
switch (ponderIntegerDivisionBy(N.getOperand(1), isSigned, Tmp3)) { switch (ponderIntegerDivisionBy(N.getOperand(1), isSigned, Tmp3)) {
case 1: // division by a constant that's a power of 2 case 1: // division by a constant that's a power of 2
Tmp1 = SelectExpr(N.getOperand(0)); Tmp1 = SelectExpr(N.getOperand(0));
if(isSigned) // becomes a shift right: if(isSigned) { // argument could be negative, so emit some code:
BuildMI(BB, IA64::SHRS, 2, Result).addReg(Tmp1).addImm(Tmp3); unsigned divAmt=Tmp3;
else 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); BuildMI(BB, IA64::SHRU, 2, Result).addReg(Tmp1).addImm(Tmp3);
return Result; // early exit return Result; // early exit
} }
@ -1171,10 +1183,11 @@ pC = pA OR pB
} }
// we start by computing an approximate reciprocal (good to 9 bits?) // we start by computing an approximate reciprocal (good to 9 bits?)
// note, this instruction writes _both_ TmpF5 (answer) and tmpPR (predicate) // note, this instruction writes _both_ TmpF5 (answer) and TmpPR (predicate)
// FIXME: or at least, it should!! BuildMI(BB, IA64::FRCPAS1, 4)
BuildMI(BB, IA64::FRCPAS1FLOAT, 2, TmpF5).addReg(TmpF3).addReg(TmpF4); .addReg(TmpF5, MachineOperand::Def)
BuildMI(BB, IA64::FRCPAS1PREDICATE, 2, TmpPR).addReg(TmpF3).addReg(TmpF4); .addReg(TmpPR, MachineOperand::Def)
.addReg(TmpF3).addReg(TmpF4);
if(!isModulus) { // if this is a divide, we worry about div-by-zero if(!isModulus) { // if this is a divide, we worry about div-by-zero
unsigned bogusPR=MakeReg(MVT::i1); // won't appear, due to twoAddress unsigned bogusPR=MakeReg(MVT::i1); // won't appear, due to twoAddress

13
lib/Target/IA64/IA64InstrInfo.td
View File

@ -120,6 +120,9 @@ def SHRS : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src1, GR:$src2),
def SHRSI : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src1, s21imm:$imm), def SHRSI : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src1, s21imm:$imm),
"shr $dst = $src1, $imm;;">; "shr $dst = $src1, $imm;;">;
def EXTRU : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src1, u6imm:$imm1, u6imm:$imm2),
"extr.u $dst = $src1, $imm1, $imm2;;">;
def DEPZ : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src1, u6imm:$imm1, u6imm:$imm2), "dep.z $dst = $src1, $imm1, $imm2;;">; def DEPZ : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src1, u6imm:$imm1, u6imm:$imm2), "dep.z $dst = $src1, $imm1, $imm2;;">;
def SXT1 : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src), "sxt1 $dst = $src;;">; def SXT1 : AForm<0x03, 0x0b, (ops GR:$dst, GR:$src), "sxt1 $dst = $src;;">;
@ -258,14 +261,8 @@ def CFNMAS1 : AForm<0x03, 0x0b,
(ops FP:$dst, FP:$src1, FP:$src2, FP:$src3, PR:$qp), (ops FP:$dst, FP:$src1, FP:$src2, FP:$src3, PR:$qp),
"($qp) fnma.s1 $dst = $src1, $src2, $src3;;">; "($qp) fnma.s1 $dst = $src1, $src2, $src3;;">;
// FIXME: we 'explode' FRCPA (which should write two registers) into two def FRCPAS1 : AForm<0x03, 0x0b, (ops FP:$dstFR, PR:$dstPR, FP:$src1, FP:$src2),
// operations that write one each. this is a waste, and is also destroying "frcpa.s1 $dstFR, $dstPR = $src1, $src2;;">;
// f127. not cool.
def FRCPAS1FLOAT : AForm<0x03, 0x0b, (ops FP:$dst, FP:$src1, FP:$src2),
"frcpa.s1 $dst , p0 = $src1, $src2;;">;
// XXX: this _will_ break things: (f127)
def FRCPAS1PREDICATE : AForm<0x03, 0x0b, (ops PR:$dst, FP:$src1, FP:$src2),
"frcpa.s1 f127 , $dst = $src1, $src2;; // XXX FIXME!!!!">;
def XMAL : AForm<0x03, 0x0b, (ops FP:$dst, FP:$src1, FP:$src2, FP:$src3), def XMAL : AForm<0x03, 0x0b, (ops FP:$dst, FP:$src1, FP:$src2, FP:$src3),
"xma.l $dst = $src1, $src2, $src3;;">; "xma.l $dst = $src1, $src2, $src3;;">;