When possible, custom lower 32-bit SINT_TO_FP to this:
_foo2:
extsw r2, r3
std r2, -8(r1)
lfd f0, -8(r1)
fcfid f0, f0
frsp f1, f0
blr
instead of this:
_foo2:
lis r2, ha16(LCPI2_0)
lis r4, 17200
xoris r3, r3, 32768
stw r3, -4(r1)
stw r4, -8(r1)
lfs f0, lo16(LCPI2_0)(r2)
lfd f1, -8(r1)
fsub f0, f1, f0
frsp f1, f0
blr
This speeds up Misc/pi from 2.44s->2.09s with LLC and from 3.01->2.18s
with llcbeta (16.7% and 38.1% respectively).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@26943 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index 4af94b3..d29ec20 100644
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -140,6 +140,7 @@
// They also have instructions for converting between i64 and fp.
setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
+ setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
// To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
} else {
@@ -222,6 +223,8 @@
case PPCISD::SRL: return "PPCISD::SRL";
case PPCISD::SRA: return "PPCISD::SRA";
case PPCISD::SHL: return "PPCISD::SHL";
+ case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
+ case PPCISD::STD_32: return "PPCISD::STD_32";
case PPCISD::CALL: return "PPCISD::CALL";
case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
}
@@ -302,15 +305,41 @@
Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
return Bits;
}
- case ISD::SINT_TO_FP: {
- assert(MVT::i64 == Op.getOperand(0).getValueType() &&
- "Unhandled SINT_TO_FP type in custom expander!");
- SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
- SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
- if (MVT::f32 == Op.getValueType())
- FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
- return FP;
- }
+ case ISD::SINT_TO_FP:
+ if (Op.getOperand(0).getValueType() == MVT::i64) {
+ SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
+ SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
+ if (Op.getValueType() == MVT::f32)
+ FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
+ return FP;
+ } else {
+ assert(Op.getOperand(0).getValueType() == MVT::i32 &&
+ "Unhandled SINT_TO_FP type in custom expander!");
+ // Since we only generate this in 64-bit mode, we can take advantage of
+ // 64-bit registers. In particular, sign extend the input value into the
+ // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
+ // then lfd it and fcfid it.
+ MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
+ int FrameIdx = FrameInfo->CreateStackObject(8, 8);
+ SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
+
+ SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
+ Op.getOperand(0));
+
+ // STD the extended value into the stack slot.
+ SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
+ DAG.getEntryNode(), Ext64, FIdx,
+ DAG.getSrcValue(NULL));
+ // Load the value as a double.
+ SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, DAG.getSrcValue(NULL));
+
+ // FCFID it and return it.
+ SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
+ if (Op.getValueType() == MVT::f32)
+ FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
+ return FP;
+ }
+
case ISD::SELECT_CC: {
// Turn FP only select_cc's into fsel instructions.
if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
@@ -1106,27 +1135,30 @@
default: break;
case ISD::SINT_TO_FP:
if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
- // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
- // We allow the src/dst to be either f32/f64, but force the intermediate
- // type to be i64.
- if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT &&
- N->getOperand(0).getValueType() == MVT::i64) {
-
- SDOperand Val = N->getOperand(0).getOperand(0);
- if (Val.getValueType() == MVT::f32) {
- Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
+ if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
+ // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
+ // We allow the src/dst to be either f32/f64, but the intermediate
+ // type must be i64.
+ if (N->getOperand(0).getValueType() == MVT::i64) {
+ SDOperand Val = N->getOperand(0).getOperand(0);
+ if (Val.getValueType() == MVT::f32) {
+ Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
+ DCI.AddToWorklist(Val.Val);
+ }
+
+ Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
DCI.AddToWorklist(Val.Val);
- }
-
- Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
- DCI.AddToWorklist(Val.Val);
- Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
- DCI.AddToWorklist(Val.Val);
- if (N->getValueType(0) == MVT::f32) {
- Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
+ Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
DCI.AddToWorklist(Val.Val);
+ if (N->getValueType(0) == MVT::f32) {
+ Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
+ DCI.AddToWorklist(Val.Val);
+ }
+ return Val;
+ } else if (N->getOperand(0).getValueType() == MVT::i32) {
+ // If the intermediate type is i32, we can avoid the load/store here
+ // too.
}
- return Val;
}
}
break;