[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;
----------------
fhahn wrote:

I think this would mean we would not check the BB of the last load any longer, but there may be problematic stores between the start of the block and the load (future extension could be to only check recipes in the range starting at First, ending at Last).

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