replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Optimize some 64-bit multiplication by constants into two lea's or one lea + shl since imulq is slow (latency 5). e.g. 

x * 40
=>
shlq    $3, %rdi
leaq    (%rdi,%rdi,4), %rax

This has the added benefit of allowing more multiply to be folded into addressing mode. e.g.
a * 24 + b
=>
leaq    (%rdi,%rdi,2), %rax
leaq    (%rsi,%rax,8), %rax

llvm-svn: 67917
This commit is contained in:
Evan Cheng 2009-03-28 05:57:29 +00:00
parent c0c3dffa3d
commit fd81c73cde
4 changed files with 98 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
llvm/include/llvm/Target/TargetLowering.h
View File

@ -791,9 +791,10 @@ public:
bool isCalledByLegalizer() const { return CalledByLegalizer; }
void AddToWorklist(SDNode *N);
SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To);
SDValue CombineTo(SDNode *N, SDValue Res);
SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1);
SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To,
bool AddTo = true);
SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true);
SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true);
void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO);
};

16
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
View File

@ -93,14 +93,14 @@ namespace {
}
SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
bool AddTo = true);
bool AddTo = true);
SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
return CombineTo(N, &Res, 1, AddTo);
}
SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
bool AddTo = true) {
bool AddTo = true) {
SDValue To[] = { Res0, Res1 };
return CombineTo(N, To, 2, AddTo);
}
@ -293,19 +293,19 @@ void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
}
SDValue TargetLowering::DAGCombinerInfo::
CombineTo(SDNode *N, const std::vector<SDValue> &To) {
return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size());
CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
}
SDValue TargetLowering::DAGCombinerInfo::
CombineTo(SDNode *N, SDValue Res) {
return ((DAGCombiner*)DC)->CombineTo(N, Res);
CombineTo(SDNode *N, SDValue Res, bool AddTo) {
return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
}
SDValue TargetLowering::DAGCombinerInfo::
CombineTo(SDNode *N, SDValue Res0, SDValue Res1) {
return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1);
CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
}
void TargetLowering::DAGCombinerInfo::

71
llvm/lib/Target/X86/X86ISelLowering.cpp
View File

@ -826,6 +826,8 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
setTargetDAGCombine(ISD::SRA);
setTargetDAGCombine(ISD::SRL);
setTargetDAGCombine(ISD::STORE);
if (Subtarget->is64Bit())
setTargetDAGCombine(ISD::MUL);
computeRegisterProperties();
@ -8407,6 +8409,74 @@ static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
}
/// PerformMulCombine - Optimize a single multiply with constant into two
/// in order to implement it with two cheaper instructions, e.g.
/// LEA + SHL, LEA + LEA.
static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI) {
if (DAG.getMachineFunction().
getFunction()->hasFnAttr(Attribute::OptimizeForSize))
return SDValue();
if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
return SDValue();
MVT VT = N->getValueType(0);
if (VT != MVT::i64)
return SDValue();
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (!C)
return SDValue();
uint64_t MulAmt = C->getZExtValue();
if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
return SDValue();
uint64_t MulAmt1 = 0;
uint64_t MulAmt2 = 0;
if ((MulAmt % 9) == 0) {
MulAmt1 = 9;
MulAmt2 = MulAmt / 9;
} else if ((MulAmt % 5) == 0) {
MulAmt1 = 5;
MulAmt2 = MulAmt / 5;
} else if ((MulAmt % 3) == 0) {
MulAmt1 = 3;
MulAmt2 = MulAmt / 3;
}
if (MulAmt2 &&
(isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
DebugLoc DL = N->getDebugLoc();
if (isPowerOf2_64(MulAmt2) &&
!(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
// If second multiplifer is pow2, issue it first. We want the multiply by
// 3, 5, or 9 to be folded into the addressing mode unless the lone use
// is an add.
std::swap(MulAmt1, MulAmt2);
SDValue NewMul;
if (isPowerOf2_64(MulAmt1))
NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
else
NewMul = DAG.getNode(ISD::MUL, DL, VT, N->getOperand(0),
DAG.getConstant(MulAmt1, VT));
if (isPowerOf2_64(MulAmt2))
NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
else
NewMul = DAG.getNode(ISD::MUL, DL, VT, NewMul,
DAG.getConstant(MulAmt2, VT));
// Do not add new nodes to DAG combiner worklist.
DCI.CombineTo(N, NewMul, false);
}
return SDValue();
}
/// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
/// when possible.
static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
@ -8668,6 +8738,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
return PerformBuildVectorCombine(N, DAG, DCI, Subtarget, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
case ISD::SHL:
case ISD::SRA:
case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);

15
llvm/test/CodeGen/X86/imul-lea-2.ll Normal file
View File

@ -0,0 +1,15 @@
; RUN: llvm-as < %s | llc -march=x86-64 | grep lea | count 3
; RUN: llvm-as < %s | llc -march=x86-64 | grep shl | count 1
; RUN: llvm-as < %s | llc -march=x86-64 | not grep imul
define i64 @t1(i64 %a) nounwind readnone {
entry:
%0 = mul i64 %a, 81 ; <i64> [#uses=1]
ret i64 %0
}
define i64 @t2(i64 %a) nounwind readnone {
entry:
%0 = mul i64 %a, 40 ; <i64> [#uses=1]
ret i64 %0
}