[llvm] [NaryReassociate] Teach NaryReassociate about UniformityAnalysis (PR #175167)

[Sameer Sahasrabuddhe via llvm-commits]

================
@@ -379,14 +396,32 @@ NaryReassociatePass::tryReassociateGEPAtIndex(GetElementPtrInst *GEP,
 
     Value *LHS = AO->getOperand(0), *RHS = AO->getOperand(1);
     // IndexToSplit = LHS + RHS.
-    if (auto *NewGEP = tryReassociateGEPAtIndex(GEP, I, LHS, RHS, IndexedType))
-      return NewGEP;
-    // Symmetrically, try IndexToSplit = RHS + LHS.
-    if (LHS != RHS) {
-      if (auto *NewGEP =
+    // tryReassociateGEPAtIndex(GEP, I, LHS, RHS, ...) looks for a dominating
+    // GEP with LHS as index, then creates: NewGEP = existingGEP + RHS * scale.
+    // So the RHS becomes the "remaining" index calculation.
+
+    // When LHS == RHS, both call orders are identical, so only try once.
+    if (LHS == RHS)
+      return tryReassociateGEPAtIndex(GEP, I, LHS, RHS, IndexedType);
+
+    // When uniformity analysis is available, prefer the remaining calculation
+    // to be uniform, keeping uniform computations grouped together.
+    // Default order tries LHS first (RHS as remainder). If LHS is uniform and
+    // RHS is divergent, try RHS first so uniform LHS becomes the remainder.
+    if (UI && UI->isUniform(LHS) && !UI->isUniform(RHS)) {
----------------
ssahasra wrote:

Just a fly-by note, need not be addressed in this PR: We should probably rename "isUniform" to "isUniformDef" so that any one inserting the call has to think about which one they want.

https://github.com/llvm/llvm-project/pull/175167