[llvm] [VPlan] Hoist predicated loads with complementary masks. (PR #168373)

[Florian Hahn via llvm-commits]

================
@@ -3974,6 +3974,152 @@ void VPlanTransforms::hoistInvariantLoads(VPlan &Plan) {
   }
 }
 
+// Returns the intersection of metadata from a group of loads.
+static VPIRMetadata getCommonLoadMetadata(ArrayRef<VPReplicateRecipe *> Loads) {
+  VPIRMetadata CommonMetadata = *Loads.front();
+  for (VPReplicateRecipe *Load : drop_begin(Loads))
+    CommonMetadata.intersect(*Load);
+  return CommonMetadata;
+}
+
+// Check if a load can be hoisted by verifying it doesn't alias with any stores
+// in blocks between FirstBB and LastBB using scoped noalias metadata.
+static bool canHoistLoadWithNoAliasCheck(VPReplicateRecipe *Load,
+                                         VPBasicBlock *FirstBB,
+                                         VPBasicBlock *LastBB) {
+  // Get the load's memory location and check if it aliases with any stores
+  // using scoped noalias metadata.
+  auto LoadLoc = vputils::getMemoryLocation(*Load);
+  if (!LoadLoc || !LoadLoc->AATags.Scope)
+    return false;
+
+  const AAMDNodes &LoadAA = LoadLoc->AATags;
+  for (VPBlockBase *Block = FirstBB; Block;
+       Block = Block->getSingleSuccessor()) {
+    // This function assumes a simple linear chain of blocks. If there are
+    // multiple successors, we would need more complex analysis.
+    assert(Block->getNumSuccessors() <= 1 &&
+           "Expected at most one successor in block chain");
+    auto *VPBB = cast<VPBasicBlock>(Block);
+    for (VPRecipeBase &R : *VPBB) {
+      if (R.mayWriteToMemory()) {
+        auto Loc = vputils::getMemoryLocation(R);
+        // Bail out if we can't get the location or if the scoped noalias
+        // metadata indicates potential aliasing.
+        if (!Loc || ScopedNoAliasAAResult::mayAliasInScopes(
+                        LoadAA.Scope, Loc->AATags.NoAlias))
+          return false;
+      }
+    }
+
+    if (Block == LastBB)
+      break;
+  }
+  return true;
+}
+
+void VPlanTransforms::hoistPredicatedLoads(VPlan &Plan, ScalarEvolution &SE,
+                                           const Loop *L) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPTypeAnalysis TypeInfo(Plan);
+  VPDominatorTree VPDT(Plan);
+
+  // Group predicated loads by their address SCEV.
+  MapVector<const SCEV *, SmallVector<VPReplicateRecipe *>> LoadsByAddress;
----------------
fhahn wrote:

I changed it to a DenseMap, thanks. Which group gets processed first below should not matter, as the inserting order should remain the same regardless of order (the new load will be inserted just before the first one of each group).


https://github.com/llvm/llvm-project/pull/168373