2nd attempt, fixing SSE4.1 issues and implementing feedback from duncan.
PR2957
ISD::VECTOR_SHUFFLE now stores an array of integers representing the shuffle
mask internal to the node, rather than taking a BUILD_VECTOR of ConstantSDNodes
as the shuffle mask. A value of -1 represents UNDEF.
In addition to eliminating the creation of illegal BUILD_VECTORS just to
represent shuffle masks, we are better about canonicalizing the shuffle mask,
resulting in substantially better code for some classes of shuffles.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@70225 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/CellSPU/SPUISelLowering.cpp b/lib/Target/CellSPU/SPUISelLowering.cpp
index c07e6d5..cef87e9 100644
--- a/lib/Target/CellSPU/SPUISelLowering.cpp
+++ b/lib/Target/CellSPU/SPUISelLowering.cpp
@@ -1670,9 +1670,9 @@
/// \note
/// SPUISD::SHUFB is eventually selected as Cell's <i>shufb</i> instructions.
static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
+ const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op);
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
- SDValue PermMask = Op.getOperand(2);
DebugLoc dl = Op.getDebugLoc();
if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
@@ -1703,39 +1703,40 @@
} else
assert(0 && "Unhandled vector type in LowerVECTOR_SHUFFLE");
- for (unsigned i = 0; i != PermMask.getNumOperands(); ++i) {
- if (PermMask.getOperand(i).getOpcode() != ISD::UNDEF) {
- unsigned SrcElt = cast<ConstantSDNode > (PermMask.getOperand(i))->getZExtValue();
+ for (unsigned i = 0; i != MaxElts; ++i) {
+ if (SVN->getMaskElt(i) < 0)
+ continue;
+
+ unsigned SrcElt = SVN->getMaskElt(i);
- if (monotonic) {
- if (SrcElt >= V2EltIdx0) {
- if (1 >= (++EltsFromV2)) {
- V2Elt = (V2EltIdx0 - SrcElt) << 2;
- }
- } else if (CurrElt != SrcElt) {
- monotonic = false;
+ if (monotonic) {
+ if (SrcElt >= V2EltIdx0) {
+ if (1 >= (++EltsFromV2)) {
+ V2Elt = (V2EltIdx0 - SrcElt) << 2;
}
-
- ++CurrElt;
+ } else if (CurrElt != SrcElt) {
+ monotonic = false;
}
- if (rotate) {
- if (PrevElt > 0 && SrcElt < MaxElts) {
- if ((PrevElt == SrcElt - 1)
- || (PrevElt == MaxElts - 1 && SrcElt == 0)) {
- PrevElt = SrcElt;
- if (SrcElt == 0)
- V0Elt = i;
- } else {
- rotate = false;
- }
- } else if (PrevElt == 0) {
- // First time through, need to keep track of previous element
+ ++CurrElt;
+ }
+
+ if (rotate) {
+ if (PrevElt > 0 && SrcElt < MaxElts) {
+ if ((PrevElt == SrcElt - 1)
+ || (PrevElt == MaxElts - 1 && SrcElt == 0)) {
PrevElt = SrcElt;
+ if (SrcElt == 0)
+ V0Elt = i;
} else {
- // This isn't a rotation, takes elements from vector 2
rotate = false;
}
+ } else if (PrevElt == 0) {
+ // First time through, need to keep track of previous element
+ PrevElt = SrcElt;
+ } else {
+ // This isn't a rotation, takes elements from vector 2
+ rotate = false;
}
}
}
@@ -1768,17 +1769,11 @@
unsigned BytesPerElement = EltVT.getSizeInBits()/8;
SmallVector<SDValue, 16> ResultMask;
- for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
- unsigned SrcElt;
- if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
- SrcElt = 0;
- else
- SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getZExtValue();
+ for (unsigned i = 0, e = MaxElts; i != e; ++i) {
+ unsigned SrcElt = SVN->getMaskElt(i) < 0 ? 0 : SVN->getMaskElt(i);
- for (unsigned j = 0; j < BytesPerElement; ++j) {
- ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
- MVT::i8));
- }
+ for (unsigned j = 0; j < BytesPerElement; ++j)
+ ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,MVT::i8));
}
SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,