[AMDGPU] Enable expensive unroll trip count.

carlobertolli · carlobertolli · commit 6d35135d8271 · 2025-10-28T15:46:37.000-05:00
This patch enables unrolling innermost loop of Pytorch reduce function in Normalization.cuh.
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -271,6 +271,8 @@ void AMDGPUTTIImpl::getUnrollingPreferences(
     if (L->isInnermost() && BB->size() < UnrollMaxBlockToAnalyze)
       UP.MaxIterationsCountToAnalyze = 32;
   }
+
+  UP.AllowExpensiveTripCount = true;
 }
 
 void AMDGPUTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,

Original file line number	Diff line number	Diff line change
`@@ -271,6 +271,8 @@ void AMDGPUTTIImpl::getUnrollingPreferences(`
`271`	`271`	`if (L->isInnermost() && BB->size() < UnrollMaxBlockToAnalyze)`
`272`	`272`	`UP.MaxIterationsCountToAnalyze = 32;`
`273`	`273`	`}`
	`274`	`+`
	`275`	`+ UP.AllowExpensiveTripCount = true;`
`274`	`276`	`}`
`275`	`277`
`276`	`278`	`void AMDGPUTTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,`