[llvm-commits] CVS: llvm/lib/Target/Alpha/AlphaISelPattern.cpp AlphaInstrInfo.td
Andrew Lenharth
alenhar2 at cs.uiuc.edu
Sat Apr 2 13:07:07 PST 2005
Changes in directory llvm/lib/Target/Alpha:
AlphaISelPattern.cpp updated: 1.75 -> 1.76
AlphaInstrInfo.td updated: 1.36 -> 1.37
---
Log message:
Try several things. 1) drop /i from FP ops 2) factor out FP to Int moves and provide 21264 support for those 3) match not 4) match ornot andnot xornot
---
Diffs of the changes: (+120 -94)
AlphaISelPattern.cpp | 189 ++++++++++++++++++++++++++++-----------------------
AlphaInstrInfo.td | 25 ++++--
2 files changed, 120 insertions(+), 94 deletions(-)
Index: llvm/lib/Target/Alpha/AlphaISelPattern.cpp
diff -u llvm/lib/Target/Alpha/AlphaISelPattern.cpp:1.75 llvm/lib/Target/Alpha/AlphaISelPattern.cpp:1.76
--- llvm/lib/Target/Alpha/AlphaISelPattern.cpp:1.75 Sat Apr 2 13:11:07 2005
+++ llvm/lib/Target/Alpha/AlphaISelPattern.cpp Sat Apr 2 15:06:51 2005
@@ -36,7 +36,7 @@
cl::opt<bool> EnableAlphaIDIV("enable-alpha-intfpdiv",
cl::desc("Use the FP div instruction for integer div when possible"),
cl::Hidden);
- cl::opt<bool> EnableAlpha("enable-alpha-ftoi",
+ cl::opt<bool> EnableAlphaFTOI("enable-alpha-ftoi",
cl::desc("Enable use of ftoi* and itof* instructions (ev6 and higher)"),
cl::Hidden);
}
@@ -333,6 +333,8 @@
void SelectAddr(SDOperand N, unsigned& Reg, long& offset);
void SelectBranchCC(SDOperand N);
+ void MoveFP2Int(unsigned src, unsigned dst, bool isDouble);
+ void MoveInt2FP(unsigned src, unsigned dst, bool isDouble);
};
}
@@ -376,6 +378,46 @@
}
}
+void ISel::MoveFP2Int(unsigned src, unsigned dst, bool isDouble)
+{
+ unsigned Opc;
+ if (EnableAlphaFTOI) {
+ Opc = isDouble ? Alpha::FTOIT : Alpha::FTOIS;
+ BuildMI(BB, Opc, 1, dst).addReg(src);
+ } else {
+ //The hard way:
+ // Spill the integer to memory and reload it from there.
+ unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
+ MachineFunction *F = BB->getParent();
+ int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, 8);
+
+ Opc = isDouble ? Alpha::STT : Alpha::STS;
+ BuildMI(BB, Opc, 3).addReg(src).addFrameIndex(FrameIdx).addReg(Alpha::F31);
+ Opc = isDouble ? Alpha::LDQ : Alpha::LDL;
+ BuildMI(BB, Alpha::LDQ, 2, dst).addFrameIndex(FrameIdx).addReg(Alpha::F31);
+ }
+}
+
+void ISel::MoveInt2FP(unsigned src, unsigned dst, bool isDouble)
+{
+ unsigned Opc;
+ if (EnableAlphaFTOI) {
+ Opc = isDouble?Alpha::ITOFT:Alpha::ITOFS;
+ BuildMI(BB, Opc, 1, dst).addReg(src);
+ } else {
+ //The hard way:
+ // Spill the integer to memory and reload it from there.
+ unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
+ MachineFunction *F = BB->getParent();
+ int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, 8);
+
+ Opc = isDouble ? Alpha::STQ : Alpha::STL;
+ BuildMI(BB, Opc, 3).addReg(src).addFrameIndex(FrameIdx).addReg(Alpha::F31);
+ Opc = isDouble ? Alpha::LDT : Alpha::LDS;
+ BuildMI(BB, Opc, 2, dst).addFrameIndex(FrameIdx).addReg(Alpha::F31);
+ }
+}
+
//Check to see if the load is a constant offset from a base register
void ISel::SelectAddr(SDOperand N, unsigned& Reg, long& offset)
{
@@ -615,16 +657,13 @@
else
{
Tmp1 = SelectExpr(N.getOperand(0)); //Cond
- // Spill the cond to memory and reload it from there.
- unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
- MachineFunction *F = BB->getParent();
- int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, 8);
- unsigned Tmp4 = MakeReg(MVT::f64);
- BuildMI(BB, Alpha::STQ, 3).addReg(Tmp1).addFrameIndex(FrameIdx).addReg(Alpha::F31);
- BuildMI(BB, Alpha::LDT, 2, Tmp4).addFrameIndex(FrameIdx).addReg(Alpha::F31);
- //now ideally, we don't have to do anything to the flag...
- // Get the condition into the zero flag.
- BuildMI(BB, Alpha::FCMOVEQ, 3, Result).addReg(TV).addReg(FV).addReg(Tmp4);
+ BuildMI(BB, Alpha::FCMOVEQ_INT, 3, Result).addReg(TV).addReg(FV).addReg(Tmp1);
+// // Spill the cond to memory and reload it from there.
+// unsigned Tmp4 = MakeReg(MVT::f64);
+// MoveIntFP(Tmp1, Tmp4, true);
+// //now ideally, we don't have to do anything to the flag...
+// // Get the condition into the zero flag.
+// BuildMI(BB, Alpha::FCMOVEQ, 3, Result).addReg(TV).addReg(FV).addReg(Tmp4);
return Result;
}
}
@@ -784,25 +823,9 @@
&& "only quads can be loaded from");
Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
Tmp2 = MakeReg(MVT::f64);
-
- //The hard way:
- // Spill the integer to memory and reload it from there.
- unsigned Size = MVT::getSizeInBits(MVT::i64)/8;
- MachineFunction *F = BB->getParent();
- int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, Size);
-
- BuildMI(BB, Alpha::STQ, 3).addReg(Tmp1).addFrameIndex(FrameIdx).addReg(Alpha::F31);
- BuildMI(BB, Alpha::LDT, 2, Tmp2).addFrameIndex(FrameIdx).addReg(Alpha::F31);
+ MoveInt2FP(Tmp1, Tmp2, true);
Opc = DestType == MVT::f64 ? Alpha::CVTQT : Alpha::CVTQS;
BuildMI(BB, Opc, 1, Result).addReg(Tmp2);
-
- //The easy way: doesn't work
- // //so these instructions are not supported on ev56
- // Opc = DestType == MVT::f64 ? Alpha::ITOFT : Alpha::ITOFS;
- // BuildMI(BB, Opc, 1, Tmp2).addReg(Tmp1);
- // Opc = DestType == MVT::f64 ? Alpha::CVTQT : Alpha::CVTQS;
- // BuildMI(BB, Opc, 1, Result).addReg(Tmp1);
-
return Result;
}
}
@@ -1082,38 +1105,27 @@
return Result+N.ResNo;
}
- case ISD::SIGN_EXTEND:
- abort();
-
case ISD::SIGN_EXTEND_INREG:
{
//do SDIV opt for all levels of ints
if (EnableAlphaIDIV && N.getOperand(0).getOpcode() == ISD::SDIV)
{
- Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
- Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
- unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
- MachineFunction *F = BB->getParent();
- int FrameIdxL = F->getFrameInfo()->CreateStackObject(Size, 8);
- int FrameIdxR = F->getFrameInfo()->CreateStackObject(Size, 8);
- int FrameIdxF = F->getFrameInfo()->CreateStackObject(Size, 8);
unsigned Tmp4 = MakeReg(MVT::f64);
unsigned Tmp5 = MakeReg(MVT::f64);
unsigned Tmp6 = MakeReg(MVT::f64);
unsigned Tmp7 = MakeReg(MVT::f64);
unsigned Tmp8 = MakeReg(MVT::f64);
unsigned Tmp9 = MakeReg(MVT::f64);
-
- BuildMI(BB, Alpha::STQ, 3).addReg(Tmp1).addFrameIndex(FrameIdxL).addReg(Alpha::F31);
- BuildMI(BB, Alpha::STQ, 3).addReg(Tmp2).addFrameIndex(FrameIdxR).addReg(Alpha::F31);
- BuildMI(BB, Alpha::LDT, 2, Tmp4).addFrameIndex(FrameIdxL).addReg(Alpha::F31);
- BuildMI(BB, Alpha::LDT, 2, Tmp5).addFrameIndex(FrameIdxR).addReg(Alpha::F31);
+
+ Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
+ MoveInt2FP(Tmp1, Tmp4, true);
+ MoveInt2FP(Tmp2, Tmp5, true);
BuildMI(BB, Alpha::CVTQT, 1, Tmp6).addReg(Tmp4);
BuildMI(BB, Alpha::CVTQT, 1, Tmp7).addReg(Tmp5);
BuildMI(BB, Alpha::DIVT, 2, Tmp8).addReg(Tmp6).addReg(Tmp7);
BuildMI(BB, Alpha::CVTTQ, 1, Tmp9).addReg(Tmp8);
- BuildMI(BB, Alpha::STT, 3).addReg(Tmp9).addFrameIndex(FrameIdxF).addReg(Alpha::F31);
- BuildMI(BB, Alpha::LDQ, 2, Result).addFrameIndex(FrameIdxF).addReg(Alpha::F31);
+ MoveFP2Int(Tmp9, Result, true);
return Result;
}
@@ -1362,30 +1374,7 @@
BuildMI(BB, Alpha::ADDQi, 2, Tmp4).addReg(Alpha::R31).addImm(1);
Opc = inv?Alpha::CMOVNEi_FP:Alpha::CMOVEQi_FP;
BuildMI(BB, Opc, 3, Result).addReg(Tmp4).addImm(0).addReg(Tmp3);
-
-// // Spill the FP to memory and reload it from there.
-// unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
-// MachineFunction *F = BB->getParent();
-// int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, 8);
-// unsigned Tmp4 = MakeReg(MVT::f64);
-// BuildMI(BB, Alpha::CVTTQ, 1, Tmp4).addReg(Tmp3);
-// BuildMI(BB, Alpha::STT, 3).addReg(Tmp4).addFrameIndex(FrameIdx).addReg(Alpha::F31);
-// unsigned Tmp5 = MakeReg(MVT::i64);
-// BuildMI(BB, Alpha::LDQ, 2, Tmp5).addFrameIndex(FrameIdx).addReg(Alpha::F31);
-
-// //now, set result based on Tmp5
-// //Set Tmp6 if fp cmp was false
-// unsigned Tmp6 = MakeReg(MVT::i64);
-// BuildMI(BB, Alpha::CMPEQ, 2, Tmp6).addReg(Tmp5).addReg(Alpha::R31);
-// //and invert
-// BuildMI(BB, Alpha::CMPEQ, 2, Result).addReg(Tmp6).addReg(Alpha::R31);
-
}
- // else
- // {
- // Node->dump();
- // assert(0 && "Not a setcc in setcc");
- // }
}
return Result;
}
@@ -1409,9 +1398,49 @@
//Most of the plain arithmetic and logic share the same form, and the same
//constant immediate test
- case ISD::AND:
case ISD::OR:
+ //Match Not
+ if (N.getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(N.getOperand(1))->isAllOnesValue())
+ {
+ Tmp1 = SelectExpr(N.getOperand(0));
+ BuildMI(BB, Alpha::ORNOT, 2, Result).addReg(Alpha::R31).addReg(Tmp1);
+ return Result;
+ }
+ //Fall through
+ case ISD::AND:
case ISD::XOR:
+ //Check operand(0) == Not
+ if (N.getOperand(0).getOpcode() == ISD::OR &&
+ N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(N.getOperand(0).getOperand(1))->isAllOnesValue())
+ {
+ switch(opcode) {
+ case ISD::AND: Opc = Alpha::BIC; break;
+ case ISD::OR: Opc = Alpha::ORNOT; break;
+ case ISD::XOR: Opc = Alpha::EQV; break;
+ }
+ Tmp1 = SelectExpr(N.getOperand(1));
+ Tmp2 = SelectExpr(N.getOperand(0).getOperand(0));
+ BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
+ return Result;
+ }
+ //Check operand(1) == Not
+ if (N.getOperand(1).getOpcode() == ISD::OR &&
+ N.getOperand(1).getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->isAllOnesValue())
+ {
+ switch(opcode) {
+ case ISD::AND: Opc = Alpha::BIC; break;
+ case ISD::OR: Opc = Alpha::ORNOT; break;
+ case ISD::XOR: Opc = Alpha::EQV; break;
+ }
+ Tmp1 = SelectExpr(N.getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(1).getOperand(0));
+ BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
+ return Result;
+ }
+ //Fall through
case ISD::SHL:
case ISD::SRL:
case ISD::SRA:
@@ -1512,25 +1541,15 @@
MVT::ValueType SrcType = N.getOperand(0).getValueType();
assert (SrcType == MVT::f32 || SrcType == MVT::f64);
Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
-
- //The hard way:
- // Spill the integer to memory and reload it from there.
- unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
- MachineFunction *F = BB->getParent();
- int FrameIdx = F->getFrameInfo()->CreateStackObject(Size, 8);
-
- //CVTTQ STT LDQ
- //CVTST CVTTQ STT LDQ
if (SrcType == MVT::f32)
- {
- Tmp2 = MakeReg(MVT::f64);
- BuildMI(BB, Alpha::CVTST, 1, Tmp2).addReg(Tmp1);
- Tmp1 = Tmp2;
- }
+ {
+ Tmp2 = MakeReg(MVT::f64);
+ BuildMI(BB, Alpha::CVTST, 1, Tmp2).addReg(Tmp1);
+ Tmp1 = Tmp2;
+ }
Tmp2 = MakeReg(MVT::f64);
BuildMI(BB, Alpha::CVTTQ, 1, Tmp2).addReg(Tmp1);
- BuildMI(BB, Alpha::STT, 3).addReg(Tmp2).addFrameIndex(FrameIdx).addReg(Alpha::F31);
- BuildMI(BB, Alpha::LDQ, 2, Result).addFrameIndex(FrameIdx).addReg(Alpha::F31);
+ MoveFP2Int(Tmp2, Result, true);
return Result;
}
Index: llvm/lib/Target/Alpha/AlphaInstrInfo.td
diff -u llvm/lib/Target/Alpha/AlphaInstrInfo.td:1.36 llvm/lib/Target/Alpha/AlphaInstrInfo.td:1.37
--- llvm/lib/Target/Alpha/AlphaInstrInfo.td:1.36 Thu Mar 31 15:24:05 2005
+++ llvm/lib/Target/Alpha/AlphaInstrInfo.td Sat Apr 2 15:06:51 2005
@@ -96,7 +96,10 @@
// "lda $RES,1($$31)\n\tfbne $COND, 42f\n\tbis $$31,$$31,$RES\n42:\n">;
//An even better improvement on the Int = SetCC(FP): SelectCC!
+//These are evil because they hide control flow in a MBB
+//really the ISel should emit multiple MBB
let isTwoAddress = 1 in {
+//Conditional move of an int based on a FP CC
def CMOVEQ_FP : PseudoInstAlpha<(ops GPRC:$RDEST, GPRC:$RSRC_F, GPRC:$RSRC_T, FPRC:$RCOND),
"fbne $RCOND, 42f\n\tbis $RSRC_T,$RSRC_T,$RDEST\n42:\n">;
def CMOVEQi_FP : PseudoInstAlpha<(ops GPRC:$RDEST, GPRC:$RSRC_F, u8imm:$L, FPRC:$RCOND),
@@ -106,7 +109,11 @@
"fbeq $RCOND, 42f\n\tbis $RSRC_T,$RSRC_T,$RDEST\n42:\n">;
def CMOVNEi_FP : PseudoInstAlpha<(ops GPRC:$RDEST, GPRC:$RSRC_F, u8imm:$L, FPRC:$RCOND),
"fbeq $RCOND, 42f\n\taddq $$31,$L,$RDEST\n42:\n">;
-
+//Conditional move of an FP based on a Int CC
+ def FCMOVEQ_INT : PseudoInstAlpha<(ops GPRC:$RDEST, GPRC:$RSRC_F, GPRC:$RSRC_T, FPRC:$RCOND),
+ "bne $RCOND, 42f\n\tcpys $RSRC_T,$RSRC_T,$RDEST\n42:\n">;
+ def FCMOVNE_INT : PseudoInstAlpha<(ops GPRC:$RDEST, GPRC:$RSRC_F, GPRC:$RSRC_T, FPRC:$RCOND),
+ "beq $RCOND, 42f\n\tcpys $RSRC_T,$RSRC_T,$RDEST\n42:\n">;
}
//***********************
@@ -380,14 +387,14 @@
def CPYSN : FPForm<0x17, 0x021, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "cpysn $RA,$RB,$RC">; //Copy sign negate
//Basic Floating point ops
-def ADDS : FPForm<0x16, 0x080, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "adds/sui $RA,$RB,$RC">; //Add S_floating
-def ADDT : FPForm<0x16, 0x0A0, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "addt/sui $RA,$RB,$RC">; //Add T_floating
-def SUBS : FPForm<0x16, 0x081, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "subs/sui $RA,$RB,$RC">; //Subtract S_floating
-def SUBT : FPForm<0x16, 0x0A1, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "subt/sui $RA,$RB,$RC">; //Subtract T_floating
-def DIVS : FPForm<0x16, 0x083, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "divs/sui $RA,$RB,$RC">; //Divide S_floating
-def DIVT : FPForm<0x16, 0x0A3, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "divt/sui $RA,$RB,$RC">; //Divide T_floating
-def MULS : FPForm<0x16, 0x082, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "muls/sui $RA,$RB,$RC">; //Multiply S_floating
-def MULT : FPForm<0x16, 0x0A2, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "mult/sui $RA,$RB,$RC">; //Multiply T_floating
+def ADDS : FPForm<0x16, 0x080, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "adds/su $RA,$RB,$RC">; //Add S_floating
+def ADDT : FPForm<0x16, 0x0A0, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "addt/su $RA,$RB,$RC">; //Add T_floating
+def SUBS : FPForm<0x16, 0x081, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "subs/su $RA,$RB,$RC">; //Subtract S_floating
+def SUBT : FPForm<0x16, 0x0A1, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "subt/su $RA,$RB,$RC">; //Subtract T_floating
+def DIVS : FPForm<0x16, 0x083, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "divs/su $RA,$RB,$RC">; //Divide S_floating
+def DIVT : FPForm<0x16, 0x0A3, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "divt/su $RA,$RB,$RC">; //Divide T_floating
+def MULS : FPForm<0x16, 0x082, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "muls/su $RA,$RB,$RC">; //Multiply S_floating
+def MULT : FPForm<0x16, 0x0A2, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "mult/su $RA,$RB,$RC">; //Multiply T_floating
def SQRTS : FPForm<0x14, 0x08B, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "sqrts $RA,$RB,$RC">; //Square root S_floating
def SQRTT : FPForm<0x14, 0x0AB, (ops FPRC:$RC, FPRC:$RA, FPRC:$RB), "sqrtt $RA,$RB,$RC">; //Square root T_floating
More information about the llvm-commits
mailing list