[llvm] [LV] Use vscale for tuning to improve branch weight estimates (PR #144733)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-vectorizers

Author: David Sherwood (david-arm)

<details>
<summary>Changes</summary>

In addBranchWeightToMiddleTerminator we attempt to add branch weights to the middle block terminator. We pessimistically assume vscale=1, whereas we can improve the estimate by using the value of vscale used for tuning.

---
Full diff: https://github.com/llvm/llvm-project/pull/144733.diff


4 Files Affected:

- (modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+4-2) 
- (modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp (+4-2) 
- (modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.h (+2-1) 
- (modified) llvm/test/Transforms/LoopVectorize/AArch64/check-prof-info.ll (+1-1) 


``````````diff
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index f1470fd1f7314..3445cfa355e8a 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7330,9 +7330,11 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
   VPlanTransforms::runPass(VPlanTransforms::unrollByUF, BestVPlan, BestUF,
                            OrigLoop->getHeader()->getContext());
   VPlanTransforms::runPass(VPlanTransforms::materializeBroadcasts, BestVPlan);
-  if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator()))
+  if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())) {
+    std::optional<unsigned> VScale = CM.getVScaleForTuning();
     VPlanTransforms::runPass(VPlanTransforms::addBranchWeightToMiddleTerminator,
-                             BestVPlan, BestVF);
+                             BestVPlan, BestVF, VScale);
+  }
   VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE);
   VPlanTransforms::simplifyRecipes(BestVPlan, *Legal->getWidestInductionType());
   VPlanTransforms::narrowInterleaveGroups(
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 05a0e15f9a199..f474ee94ec5e5 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -3213,8 +3213,8 @@ void VPlanTransforms::narrowInterleaveGroups(VPlan &Plan, ElementCount VF,
 
 /// Add branch weight metadata, if the \p Plan's middle block is terminated by a
 /// BranchOnCond recipe.
-void VPlanTransforms::addBranchWeightToMiddleTerminator(VPlan &Plan,
-                                                        ElementCount VF) {
+void VPlanTransforms::addBranchWeightToMiddleTerminator(
+    VPlan &Plan, ElementCount VF, std::optional<unsigned> VScale) {
   VPBasicBlock *MiddleVPBB = Plan.getMiddleBlock();
   auto *MiddleTerm =
       dyn_cast_or_null<VPInstruction>(MiddleVPBB->getTerminator());
@@ -3226,6 +3226,8 @@ void VPlanTransforms::addBranchWeightToMiddleTerminator(VPlan &Plan,
          "must have a BranchOnCond");
   // Assume that `TripCount % VectorStep ` is equally distributed.
   unsigned VectorStep = Plan.getUF() * VF.getKnownMinValue();
+  if (VF.isScalable() && VScale.has_value())
+    VectorStep *= *VScale;
   assert(VectorStep > 0 && "trip count should not be zero");
   MDBuilder MDB(Plan.getScalarHeader()->getIRBasicBlock()->getContext());
   MDNode *BranchWeights =
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index 5a03bdb7c6882..32c4c48275121 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -237,7 +237,8 @@ struct VPlanTransforms {
 
   /// Add branch weight metadata, if the \p Plan's middle block is terminated by
   /// a BranchOnCond recipe.
-  static void addBranchWeightToMiddleTerminator(VPlan &Plan, ElementCount VF);
+  static void addBranchWeightToMiddleTerminator(VPlan &Plan, ElementCount VF,
+                                                std::optional<unsigned> VScale);
 };
 
 } // namespace llvm
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/check-prof-info.ll b/llvm/test/Transforms/LoopVectorize/AArch64/check-prof-info.ll
index 9435c544fc812..1f619898ea788 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/check-prof-info.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/check-prof-info.ll
@@ -92,7 +92,7 @@ for.cond.cleanup:                                 ; preds = %for.body
 ; CHECK-V1-IC1: [[LOOP1]] = distinct !{[[LOOP1]], [[META2:![0-9]+]], [[META3:![0-9]+]]}
 ; CHECK-V1-IC1: [[META2]] = !{!"llvm.loop.isvectorized", i32 1}
 ; CHECK-V1-IC1: [[META3]] = !{!"llvm.loop.unroll.runtime.disable"}
-; CHECK-V1-IC1: [[PROF4]] = !{!"branch_weights", i32 1, i32 3}
+; CHECK-V1-IC1: [[PROF4]] = !{!"branch_weights", i32 1, i32 7}
 ; CHECK-V1-IC1: [[PROF5]] = !{!"branch_weights", i32 0, i32 0}
 ; CHECK-V1-IC1: [[LOOP6]] = distinct !{[[LOOP6]], [[META3]], [[META2]]}
 ;.

``````````

</details>


https://github.com/llvm/llvm-project/pull/144733