RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-24 11:42:57 +01:00
Code Issues Projects Releases Wiki Activity

Some cheap DAG combine goodness for multiplication with a particular constant.

Browse Source 
This can be extended later on to handle more "complex" constants.

llvm-svn: 103881
This commit is contained in:
Anton Korobeynikov 2010-05-15 18:16:59 +00:00
parent 4f8b48d8c6
commit 314ccc5501
2 changed files with 101 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
lib/Target/ARM/ARMISelLowering.cpp
View File

@ -463,6 +463,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
// ARMISD::VMOVRRD - No need to call setTargetDAGCombine
setTargetDAGCombine(ISD::ADD);
setTargetDAGCombine(ISD::SUB);
setTargetDAGCombine(ISD::MUL);
setStackPointerRegisterToSaveRestore(ARM::SP);
setSchedulingPreference(SchedulingForRegPressure);
@ -3584,6 +3585,75 @@ static SDValue PerformSUBCombine(SDNode *N,
return SDValue();
}
static SDValue PerformMULCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
const ARMSubtarget *Subtarget) {
SelectionDAG &DAG = DCI.DAG;
if (Subtarget->isThumb1Only())
return SDValue();
if (DAG.getMachineFunction().
getFunction()->hasFnAttr(Attribute::OptimizeForSize))
return SDValue();
if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
return SDValue();
EVT VT = N->getValueType(0);
if (VT != MVT::i32)
return SDValue();
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (!C)
return SDValue();
uint64_t MulAmt = C->getZExtValue();
unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
ShiftAmt = ShiftAmt & (32 - 1);
SDValue V = N->getOperand(0);
DebugLoc DL = N->getDebugLoc();
SDValue NewAdd;
// FIXME: Handle arbitrary powers of 2.
switch (MulAmt >> ShiftAmt) {
case 3: // 2 + 1
NewAdd = DAG.getNode(ISD::ADD, DL, VT,
V, DAG.getNode(ISD::SHL, DL, VT,
V, DAG.getConstant(1, MVT::i32)));
break;
case 5: // 4 + 1
NewAdd = DAG.getNode(ISD::ADD, DL, VT,
V, DAG.getNode(ISD::SHL, DL, VT,
V, DAG.getConstant(2, MVT::i32)));
break;
case 7: // 8 - 1
NewAdd = DAG.getNode(ISD::SUB, DL, VT,
DAG.getNode(ISD::SHL, DL, VT,
V, DAG.getConstant(3, MVT::i32)),
V);
break;
case 9: // 8 + 1
NewAdd = DAG.getNode(ISD::ADD, DL, VT,
V, DAG.getNode(ISD::SHL, DL, VT,
V, DAG.getConstant(3, MVT::i32)));
break;
default: return SDValue();
}
if (ShiftAmt != 0) {
SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, NewAdd,
DAG.getConstant(ShiftAmt, MVT::i32));
// Do not add new nodes to DAG combiner worklist.
DCI.CombineTo(N, NewShift, false);
return SDValue();
}
// Do not add new nodes to DAG combiner worklist.
DCI.CombineTo(N, NewAdd, false);
return SDValue();
}
/// PerformVMOVRRDCombine - Target-specific dag combine xforms for
/// ARMISD::VMOVRRD.
static SDValue PerformVMOVRRDCombine(SDNode *N,
@ -3970,6 +4040,7 @@ SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
default: break;
case ISD::ADD: return PerformADDCombine(N, DCI);
case ISD::SUB: return PerformSUBCombine(N, DCI);
case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
case ISD::SHL:

34
test/CodeGen/ARM/mul_const.ll
View File

@ -1,17 +1,43 @@
; RUN: llc < %s -march=arm | FileCheck %s
define i32 @t1(i32 %v) nounwind readnone {
define i32 @t9(i32 %v) nounwind readnone {
entry:
; CHECK: t1:
; CHECK: t9:
; CHECK: add r0, r0, r0, lsl #3
%0 = mul i32 %v, 9
ret i32 %0
}
define i32 @t2(i32 %v) nounwind readnone {
define i32 @t7(i32 %v) nounwind readnone {
entry:
; CHECK: t2:
; CHECK: t7:
; CHECK: rsb r0, r0, r0, lsl #3
%0 = mul i32 %v, 7
ret i32 %0
}
define i32 @t5(i32 %v) nounwind readnone {
entry:
; CHECK: t5:
; CHECK: add r0, r0, r0, lsl #2
%0 = mul i32 %v, 5
ret i32 %0
}
define i32 @t3(i32 %v) nounwind readnone {
entry:
; CHECK: t3:
; CHECK: add r0, r0, r0, lsl #1
%0 = mul i32 %v, 3
ret i32 %0
}
define i32 @t12288(i32 %v) nounwind readnone {
entry:
; CHECK: t12288:
; CHECK: add r0, r0, r0, lsl #1
; CHECK: mov r0, r0, lsl #12
%0 = mul i32 %v, 12288
ret i32 %0
}