maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

Use vmladduhm to do v8i16 multiplies which is faster and simpler than doing 

even/odd halves.  Thanks to Nate telling me what's what.

llvm-svn: 27793
This commit is contained in:
Chris Lattner 2006-04-18 04:28:57 +00:00
parent d6d82aa889
commit 96d50487c9
1 changed files with 3 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
llvm/lib/Target/PowerPC/PPCISelLowering.cpp
View File

@ -1602,25 +1602,10 @@ static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
} else if (Op.getValueType() == MVT::v8i16) { } else if (Op.getValueType() == MVT::v8i16) {
SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1); SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
// Multiply the even 16-bit parts, producing 32-bit sums. SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleuh,
LHS, RHS, DAG, MVT::v4i32);
EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, EvenParts);
// Multiply the odd 16-bit parts, producing 32-bit sums. return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh, LHS, RHS, Zero, DAG);
LHS, RHS, DAG, MVT::v4i32);
OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, OddParts);
// Merge the results together.
std::vector<SDOperand> Ops;
for (unsigned i = 0; i != 4; ++i) {
Ops.push_back(DAG.getConstant(2*i+1, MVT::i16));
Ops.push_back(DAG.getConstant(2*i+1+8, MVT::i16));
}
return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v8i16, EvenParts, OddParts,
DAG.getNode(ISD::BUILD_VECTOR, MVT::v8i16, Ops));
} else if (Op.getValueType() == MVT::v16i8) { } else if (Op.getValueType() == MVT::v16i8) {
SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1); SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);