[llvm] [LoopFusion] Perform cheaper dependency check first (NFCI) (PR #228718)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Ehsan Amiri (amehsan)

<details>
<summary>Changes</summary>

If two stores write the same SSA value to memory, we can skip DA checks. This change of order has a significant impact on compile time of SingleSource/Benchmarks/Misc/himenobmtxpa.c in test suite when fusion runs with no cost model.

Assisted by Claude Code.

---
Full diff: https://github.com/llvm/llvm-project/pull/228718.diff


1 Files Affected:

- (modified) llvm/lib/Transforms/Scalar/LoopFuse.cpp (+3-3) 


``````````diff
diff --git a/llvm/lib/Transforms/Scalar/LoopFuse.cpp b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
index 20b05ce6a1302..c819ea49246d3 100644
--- a/llvm/lib/Transforms/Scalar/LoopFuse.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
@@ -1193,9 +1193,6 @@ struct LoopFuser {
       LLVM_DEBUG(dbgs() << "Check dep: " << I0 << " vs " << I1 << "\n");
     }
 #endif
-    auto DepResult = DI.depends(&I0, &I1);
-    if (!DepResult)
-      return true;
     // If two stores write the same SSA value, fusion is safe regardless of
     // aliasing - writing the same value twice is idempotent.
     if (isa<StoreInst>(I0) && isa<StoreInst>(I1)) {
@@ -1204,6 +1201,9 @@ struct LoopFuser {
       if (S0->getValueOperand() == S1->getValueOperand())
         return true;
     }
+    auto DepResult = DI.depends(&I0, &I1);
+    if (!DepResult)
+      return true;
 #ifndef NDEBUG
     if (VerboseFusionDebugging) {
       LLVM_DEBUG(dbgs() << "DA res: "; DepResult->dump(dbgs());

``````````

</details>


https://github.com/llvm/llvm-project/pull/228718