[llvm] [LICM] Fold associative binary ops to promote code hoisting (PR #81608)

[Ricardo Jesus via llvm-commits]

================
@@ -2778,6 +2780,75 @@ static bool hoistMulAddAssociation(Instruction &I, Loop &L,
   return true;
 }
 
+/// Reassociate general associative binary expressions of the form
+///
+/// 1. "(LV op C1) op C2" ==> "LV op (C1 op C2)"
+///
+/// where op is an associative binary op, LV is a loop variant, and C1 and C2
+/// are loop invariants.
+///
+/// TODO: This can be extended to more cases such as
+/// 2. "C1 op (C2 op LV)" ==> "(C1 op C2) op LV"
+/// 3. "(C1 op LV) op C2" ==> "LV op (C1 op C2)" if op is commutative
+/// 4. "C1 op (LV op C2)" ==> "(C1 op C2) op LV" if op is commutative
+static bool hoistBOAssociation(Instruction &I, Loop &L,
+                               ICFLoopSafetyInfo &SafetyInfo,
+                               MemorySSAUpdater &MSSAU, AssumptionCache *AC,
+                               DominatorTree *DT) {
+  if (!isa<BinaryOperator>(I))
+    return false;
+
+  Instruction::BinaryOps Opcode = dyn_cast<BinaryOperator>(&I)->getOpcode();
+  BinaryOperator *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
+
+  auto ClearSubclassDataAfterReassociation = [](Instruction &I) {
+    FPMathOperator *FPMO = dyn_cast<FPMathOperator>(&I);
+    if (!FPMO) {
+      I.clearSubclassOptionalData();
+      return;
+    }
+
+    FastMathFlags FMF = I.getFastMathFlags();
+    I.clearSubclassOptionalData();
+    I.setFastMathFlags(FMF);
+  };
+
+  if (I.isAssociative()) {
+    // Transform: "(LV op C1) op C2" ==> "LV op (C1 op C2)"
+    if (Op0 && Op0->getOpcode() == Opcode) {
+      Value *LV = Op0->getOperand(0);
+      Value *C1 = Op0->getOperand(1);
+      Value *C2 = I.getOperand(1);
+
+      if (L.isLoopInvariant(LV) || !L.isLoopInvariant(C1) ||
+          !L.isLoopInvariant(C2))
+        return false;
+
+      bool singleUseOp0 = Op0->hasOneUse();
+
+      // Conservatively clear all optional flags since they may not be
+      // preserved by the reassociation, but preserve fast-math flags where
+      // applicable,
+      ClearSubclassDataAfterReassociation(I);
+
+      auto *Preheader = L.getLoopPreheader();
+      assert(Preheader && "Loop is not in simplify form?");
+      IRBuilder<> Builder(Preheader->getTerminator());
+      Value *V = Builder.CreateBinOp(Opcode, C1, C2, "invariant.op");
+      I.setOperand(0, LV);
+      I.setOperand(1, V);
+
+      // Note: (LV op CV1) might not be erased if it has more than one use.
+      if (singleUseOp0)
----------------
rj-jesus wrote:

In general I'd agree, but in this case even if it's not one-use the transform will most likely help performance since the new op gets hoisted out of the loop. This is especially true if the folded op is in the loop's critical path, which we have seen in a few RAJAPerf and other kernels.

https://github.com/llvm/llvm-project/pull/81608