summaryrefslogtreecommitdiff
path: root/lib/CodeGen
diff options
context:
space:
mode:
authorMichael Liao <michael.liao@intel.com>2012-10-24 04:14:18 +0000
committerMichael Liao <michael.liao@intel.com>2012-10-24 04:14:18 +0000
commit1a5cc710ee78bf55c3799b49c6fd37619d2dbb7b (patch)
treef06342afabab5fd60234e86c9aa727acd2807d36 /lib/CodeGen
parent991b6a22b6592b56ae890f30c478f835480e7cf5 (diff)
downloadllvm-1a5cc710ee78bf55c3799b49c6fd37619d2dbb7b.tar.gz
llvm-1a5cc710ee78bf55c3799b49c6fd37619d2dbb7b.tar.bz2
llvm-1a5cc710ee78bf55c3799b49c6fd37619d2dbb7b.tar.xz
Teach DAG combine to fold (buildvec (Xint2fp x)) to (Xint2fp (buildvec x))
- If more than 1 elemennts are defined and target supports the vectorized conversion, use the vectorized one instead to reduce the strength on conversion operation. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@166546 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/CodeGen')
-rw-r--r--lib/CodeGen/SelectionDAG/DAGCombiner.cpp71
1 files changed, 71 insertions, 0 deletions
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index da4e1bad06..8cc35bbdc2 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -271,6 +271,7 @@ namespace {
SDValue ReduceLoadOpStoreWidth(SDNode *N);
SDValue TransformFPLoadStorePair(SDNode *N);
SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
+ SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
SDValue GetDemandedBits(SDValue V, const APInt &Mask);
@@ -8467,6 +8468,72 @@ SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
return DAG.getNode(ISD::BITCAST, dl, VT, BV);
}
+SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
+ EVT VT = N->getValueType(0);
+
+ unsigned NumInScalars = N->getNumOperands();
+ DebugLoc dl = N->getDebugLoc();
+
+ EVT SrcVT = MVT::Other;
+ unsigned Opcode = ISD::DELETED_NODE;
+ unsigned NumDefs = 0;
+
+ for (unsigned i = 0; i != NumInScalars; ++i) {
+ SDValue In = N->getOperand(i);
+ unsigned Opc = In.getOpcode();
+
+ if (Opc == ISD::UNDEF)
+ continue;
+
+ // If all scalar values are floats and converted from integers.
+ if (Opcode == ISD::DELETED_NODE &&
+ (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
+ Opcode = Opc;
+ // If not supported by target, bail out.
+ if (TLI.getOperationAction(Opcode, VT) != TargetLowering::Legal &&
+ TLI.getOperationAction(Opcode, VT) != TargetLowering::Custom)
+ return SDValue();
+ }
+ if (Opc != Opcode)
+ return SDValue();
+
+ EVT InVT = In.getOperand(0).getValueType();
+
+ // If all scalar values are typed differently, bail out. It's chosen to
+ // simplify BUILD_VECTOR of integer types.
+ if (SrcVT == MVT::Other)
+ SrcVT = InVT;
+ if (SrcVT != InVT)
+ return SDValue();
+ NumDefs++;
+ }
+
+ // If the vector has just one element defined, it's not worth to fold it into
+ // a vectorized one.
+ if (NumDefs < 2)
+ return SDValue();
+
+ assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
+ && "Should only handle conversion from integer to float.");
+ assert(SrcVT != MVT::Other && "Cannot determine source type!");
+
+ EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
+ SmallVector<SDValue, 8> Opnds;
+ for (unsigned i = 0; i != NumInScalars; ++i) {
+ SDValue In = N->getOperand(i);
+
+ if (In.getOpcode() == ISD::UNDEF)
+ Opnds.push_back(DAG.getUNDEF(SrcVT));
+ else
+ Opnds.push_back(In.getOperand(0));
+ }
+ SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
+ &Opnds[0], Opnds.size());
+ AddToWorkList(BV.getNode());
+
+ return DAG.getNode(Opcode, dl, VT, BV);
+}
+
SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
unsigned NumInScalars = N->getNumOperands();
DebugLoc dl = N->getDebugLoc();
@@ -8480,6 +8547,10 @@ SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
if (V.getNode())
return V;
+ V = reduceBuildVecConvertToConvertBuildVec(N);
+ if (V.getNode())
+ return V;
+
// Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
// operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
// at most two distinct vectors, turn this into a shuffle node.