two things: 1)evilness reduction patch, reduces the number of instructions hiding in the small jump CC moving code. 2)implement div of small data types by FP DIV (also reduces evilness and should speed things up)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@20756 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/Alpha/AlphaISelPattern.cpp b/lib/Target/Alpha/AlphaISelPattern.cpp
index 0e0e0ac..d25c36e 100644
--- a/lib/Target/Alpha/AlphaISelPattern.cpp
+++ b/lib/Target/Alpha/AlphaISelPattern.cpp
@@ -679,10 +679,10 @@
}
return Result;
+ case ISD::SDIV:
case ISD::MUL:
case ISD::ADD:
case ISD::SUB:
- case ISD::SDIV:
switch( opcode ) {
case ISD::MUL: Opc = DestType == MVT::f64 ? Alpha::MULT : Alpha::MULS; break;
case ISD::ADD: Opc = DestType == MVT::f64 ? Alpha::ADDT : Alpha::ADDS; break;
@@ -1050,9 +1050,39 @@
case ISD::SIGN_EXTEND_INREG:
{
+ //do SDIV opt for all levels of ints
+ if (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(Tmp1).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);
+ 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, 3).addReg(Result).addFrameIndex(FrameIdxF).addReg(Alpha::F31);
+ return Result;
+ }
+
//Alpha has instructions for a bunch of signed 32 bit stuff
if( dyn_cast<MVTSDNode>(Node)->getExtraValueType() == MVT::i32)
- {
+ {
switch (N.getOperand(0).getOpcode()) {
case ISD::ADD:
case ISD::SUB:
@@ -1291,8 +1321,12 @@
BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2);
//now arrange for Result (int) to have a 1 or 0
- Opc = inv?Alpha::CC2INT_INV:Alpha::CC2INT;
- BuildMI(BB, Opc, 1, Result).addReg(Tmp3);
+ unsigned Tmp4 = MakeReg(MVT::i64);
+ 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);
+// Opc = inv?Alpha::CC2INT_INV:Alpha::CC2INT;
+// BuildMI(BB, Opc, 1, Result).addReg(Tmp3);
// // Spill the FP to memory and reload it from there.
// unsigned Size = MVT::getSizeInBits(MVT::f64)/8;
@@ -1412,9 +1446,9 @@
return Result;
}
+ case ISD::SDIV:
case ISD::UREM:
case ISD::SREM:
- case ISD::SDIV:
case ISD::UDIV:
//FIXME: alpha really doesn't support any of these operations,
// the ops are expanded into special library calls with
@@ -1470,6 +1504,7 @@
case ISD::SELECT:
{
+ //FIXME: look at parent to decide if intCC can be folded, or if setCC(FP) and can save stack use
Tmp1 = SelectExpr(N.getOperand(0)); //Cond
Tmp2 = SelectExpr(N.getOperand(1)); //Use if TRUE
Tmp3 = SelectExpr(N.getOperand(2)); //Use if FALSE
diff --git a/lib/Target/Alpha/AlphaInstrInfo.td b/lib/Target/Alpha/AlphaInstrInfo.td
index 9d4b91d..1ac6650 100644
--- a/lib/Target/Alpha/AlphaInstrInfo.td
+++ b/lib/Target/Alpha/AlphaInstrInfo.td
@@ -88,10 +88,24 @@
}
//This is an improvement on the old style setcc (FP)
-def CC2INT_INV : PseudoInstAlpha<(ops GPRC:$RES, FPRC:$COND),
- "lda $RES,1($$31)\n\tfbeq $COND, 42f\n\tbis $$31,$$31,$RES\n42:\n">;
-def CC2INT : PseudoInstAlpha<(ops GPRC:$RES, FPRC:$COND),
- "lda $RES,1($$31)\n\tfbne $COND, 42f\n\tbis $$31,$$31,$RES\n42:\n">;
+//def CC2INT_INV : PseudoInstAlpha<(ops GPRC:$RES, FPRC:$COND),
+// "lda $RES,1($$31)\n\tfbeq $COND, 42f\n\tbis $$31,$$31,$RES\n42:\n">;
+//def CC2INT : PseudoInstAlpha<(ops GPRC:$RES, FPRC:$COND),
+// "lda $RES,1($$31)\n\tfbne $COND, 42f\n\tbis $$31,$$31,$RES\n42:\n">;
+
+//An even better improvement on the Int = SetCC(FP): SelectCC!
+let isTwoAddress = 1 in {
+ def CMOVEQ_FP : PseudoInstAlpha<(ops GPRC:$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),
+ "fbne $RCOND, 42f\n\taddi $$31,$L,$RDEST\n42:\n">;
+
+ def CMOVNE_FP : PseudoInstAlpha<(ops GPRC:$RDEST, GPRC:$RSRC_F, GPRC:$RSRC_T, FPRC:$RCOND),
+ "fbeq $RCOND, 42f\n\tbis $RSRC_T,$RSRC_T,$RDEST\n42:\n">;
+ def CMOVNEi_FP : PseudoInstAlpha<(ops GPRC:$RDEST, GPRC:$RSRC_F, u8imm:$L, FPRC:$RCOND),
+ "fbeq $RCOND, 42f\n\taddi $$31,$L,$RDEST\n42:\n">;
+
+}
//***********************
//Real instructions