[DAGCombiner] use narrow load to avoid vector extract

If we have (extract_subvector(load wide vector)) with no other users, 
that can just be (load narrow vector). This is intentionally conservative.
Follow-ups may loosen the one-use constraint to account for the extract cost
or just remove the one-use check.

The memop chain updating is based on code that already exists multiple times
in x86 lowering, so that should be pulled into a helper function as a follow-up.

Background: this is a potential improvement noticed via regressions caused by
making x86's peekThroughBitcasts() not loop on consecutive bitcasts (see 
comments in D33137).

Differential Revision: https://reviews.llvm.org/D33578

llvm-svn: 304072

1 files changed, 50 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 6928ad6bef0..23a302f3e56 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -14555,6 +14555,52 @@ static SDValue narrowExtractedVectorBinOp(SDNode *Extract, SelectionDAG &DAG) {
   return DAG.getBitcast(VT, NarrowBinOp);
 }
 
+/// If we are extracting a subvector from a wide vector load, convert to a
+/// narrow load to eliminate the extraction:
+/// (extract_subvector (load wide vector)) --> (load narrow vector)
+static SDValue narrowExtractedVectorLoad(SDNode *Extract, SelectionDAG &DAG) {
+  // TODO: Add support for big-endian. The offset calculation must be adjusted.
+  if (DAG.getDataLayout().isBigEndian())
+    return SDValue();
+
+  // TODO: The one-use check is overly conservative. Check the cost of the
+  // extract instead or remove that condition entirely.
+  auto *Ld = dyn_cast<LoadSDNode>(Extract->getOperand(0));
+  auto *ExtIdx = dyn_cast<ConstantSDNode>(Extract->getOperand(1));
+  if (!Ld || !Ld->hasOneUse() || Ld->isVolatile() || !ExtIdx)
+    return SDValue();
+
+  // The narrow load will be offset from the base address of the old load if
+  // we are extracting from something besides index 0 (little-endian).
+  EVT VT = Extract->getValueType(0);
+  SDLoc DL(Extract);
+  SDValue BaseAddr = Ld->getOperand(1);
+  unsigned Offset = ExtIdx->getZExtValue() * VT.getScalarType().getStoreSize();
+
+  // TODO: Use "BaseIndexOffset" to make this more effective.
+  SDValue NewAddr = DAG.getMemBasePlusOffset(BaseAddr, Offset, DL);
+  MachineFunction &MF = DAG.getMachineFunction();
+  MachineMemOperand *MMO = MF.getMachineMemOperand(Ld->getMemOperand(), Offset,
+                                                   VT.getStoreSize());
+  SDValue NewLd = DAG.getLoad(VT, DL, Ld->getChain(), NewAddr, MMO);
+
+  // The new load must have the same position as the old load in terms of memory
+  // dependency. Create a TokenFactor for Ld and NewLd and update uses of Ld's
+  // output chain to use that TokenFactor.
+  // TODO: This code is based on a similar sequence in x86 lowering. It should
+  // be moved to a helper function, so it can be shared and reused.
+  if (Ld->hasAnyUseOfValue(1)) {
+    SDValue OldChain = SDValue(Ld, 1);
+    SDValue NewChain = SDValue(NewLd.getNode(), 1);
+    SDValue TokenFactor = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
+                                      OldChain, NewChain);
+    DAG.ReplaceAllUsesOfValueWith(OldChain, TokenFactor);
+    DAG.UpdateNodeOperands(TokenFactor.getNode(), OldChain, NewChain);
+  }
+
+  return NewLd;
+}
+
 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
   EVT NVT = N->getValueType(0);
   SDValue V = N->getOperand(0);
@@ -14563,6 +14609,10 @@ SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
   if (V.isUndef())
     return DAG.getUNDEF(NVT);
 
+  if (TLI.isOperationLegalOrCustomOrPromote(ISD::LOAD, NVT))
+    if (SDValue NarrowLoad = narrowExtractedVectorLoad(N, DAG))
+      return NarrowLoad;
+
   // Combine:
   //    (extract_subvec (concat V1, V2, ...), i)
   // Into: