[llvm] 778d3c8 - [NFC] Partial reduce test to demonstrate regression post commit #cc9c64d (#162681)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 14 23:51:39 PDT 2025
Author: Sushant Gokhale
Date: 2025-10-14T23:51:36-07:00
New Revision: 778d3c8cccbc6ab72cea7f8fb4fd680b8f679e86
URL: https://github.com/llvm/llvm-project/commit/778d3c8cccbc6ab72cea7f8fb4fd680b8f679e86
DIFF: https://github.com/llvm/llvm-project/commit/778d3c8cccbc6ab72cea7f8fb4fd680b8f679e86.diff
LOG: [NFC] Partial reduce test to demonstrate regression post commit #cc9c64d (#162681)
We have seen performance regression for several instances of the Numba
benchmark, with some ranging around 70%, on Neoverse-v2 post #158641.
The mentioned case is short reproducer of the same. See
https://godbolt.org/z/j9Mj5WM7c for the IR differences.. A future patch
will address this.
Added:
Modified:
llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce.ll
Removed:
################################################################################
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce.ll b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce.ll
index 3c2ae1c74d171..ba5eed15c9107 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce.ll
@@ -1093,6 +1093,123 @@ exit:
ret i32 %add.lcssa
}
+define i64 @sext_reduction_i32_to_i64(ptr %arr, i64 %n) #1 {
+; CHECK-INTERLEAVE1-LABEL: define i64 @sext_reduction_i32_to_i64(
+; CHECK-INTERLEAVE1-SAME: ptr [[ARR:%.*]], i64 [[N:%.*]]) #[[ATTR2]] {
+; CHECK-INTERLEAVE1-NEXT: entry:
+; CHECK-INTERLEAVE1-NEXT: [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N]], i64 1)
+; CHECK-INTERLEAVE1-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-INTERLEAVE1-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 1
+; CHECK-INTERLEAVE1-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX]], [[TMP1]]
+; CHECK-INTERLEAVE1-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-INTERLEAVE1: vector.ph:
+; CHECK-INTERLEAVE1-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-INTERLEAVE1-NEXT: [[TMP3:%.*]] = mul nuw i64 [[TMP2]], 2
+; CHECK-INTERLEAVE1-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[UMAX]], [[TMP3]]
+; CHECK-INTERLEAVE1-NEXT: [[N_VEC:%.*]] = sub i64 [[UMAX]], [[N_MOD_VF]]
+; CHECK-INTERLEAVE1-NEXT: br label [[VECTOR_BODY:%.*]]
+; CHECK-INTERLEAVE1: vector.body:
+; CHECK-INTERLEAVE1-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVE1-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVE1-NEXT: [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[ARR]], i64 [[INDEX]]
+; CHECK-INTERLEAVE1-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 2 x i32>, ptr [[TMP4]], align 4
+; CHECK-INTERLEAVE1-NEXT: [[TMP5:%.*]] = sext <vscale x 2 x i32> [[WIDE_LOAD]] to <vscale x 2 x i64>
+; CHECK-INTERLEAVE1-NEXT: [[TMP6]] = add <vscale x 2 x i64> [[VEC_PHI]], [[TMP5]]
+; CHECK-INTERLEAVE1-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
+; CHECK-INTERLEAVE1-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-INTERLEAVE1-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
+; CHECK-INTERLEAVE1: middle.block:
+; CHECK-INTERLEAVE1-NEXT: [[TMP8:%.*]] = call i64 @llvm.vector.reduce.add.nxv2i64(<vscale x 2 x i64> [[TMP6]])
+; CHECK-INTERLEAVE1-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[UMAX]], [[N_VEC]]
+; CHECK-INTERLEAVE1-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
+; CHECK-INTERLEAVE1: scalar.ph:
+;
+; CHECK-INTERLEAVED-LABEL: define i64 @sext_reduction_i32_to_i64(
+; CHECK-INTERLEAVED-SAME: ptr [[ARR:%.*]], i64 [[N:%.*]]) #[[ATTR2]] {
+; CHECK-INTERLEAVED-NEXT: entry:
+; CHECK-INTERLEAVED-NEXT: [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N]], i64 1)
+; CHECK-INTERLEAVED-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-INTERLEAVED-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2
+; CHECK-INTERLEAVED-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX]], [[TMP1]]
+; CHECK-INTERLEAVED-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-INTERLEAVED: vector.ph:
+; CHECK-INTERLEAVED-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-INTERLEAVED-NEXT: [[TMP3:%.*]] = mul nuw i64 [[TMP2]], 4
+; CHECK-INTERLEAVED-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[UMAX]], [[TMP3]]
+; CHECK-INTERLEAVED-NEXT: [[N_VEC:%.*]] = sub i64 [[UMAX]], [[N_MOD_VF]]
+; CHECK-INTERLEAVED-NEXT: br label [[VECTOR_BODY:%.*]]
+; CHECK-INTERLEAVED: vector.body:
+; CHECK-INTERLEAVED-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP10:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT: [[VEC_PHI1:%.*]] = phi <vscale x 2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[VECTOR_BODY]] ]
+; CHECK-INTERLEAVED-NEXT: [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[ARR]], i64 [[INDEX]]
+; CHECK-INTERLEAVED-NEXT: [[TMP5:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-INTERLEAVED-NEXT: [[TMP6:%.*]] = shl nuw i64 [[TMP5]], 1
+; CHECK-INTERLEAVED-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[TMP4]], i64 [[TMP6]]
+; CHECK-INTERLEAVED-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 2 x i32>, ptr [[TMP4]], align 4
+; CHECK-INTERLEAVED-NEXT: [[WIDE_LOAD2:%.*]] = load <vscale x 2 x i32>, ptr [[TMP7]], align 4
+; CHECK-INTERLEAVED-NEXT: [[TMP8:%.*]] = sext <vscale x 2 x i32> [[WIDE_LOAD]] to <vscale x 2 x i64>
+; CHECK-INTERLEAVED-NEXT: [[TMP9:%.*]] = sext <vscale x 2 x i32> [[WIDE_LOAD2]] to <vscale x 2 x i64>
+; CHECK-INTERLEAVED-NEXT: [[TMP10]] = add <vscale x 2 x i64> [[VEC_PHI]], [[TMP8]]
+; CHECK-INTERLEAVED-NEXT: [[TMP11]] = add <vscale x 2 x i64> [[VEC_PHI1]], [[TMP9]]
+; CHECK-INTERLEAVED-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
+; CHECK-INTERLEAVED-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-INTERLEAVED-NEXT: br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
+; CHECK-INTERLEAVED: middle.block:
+; CHECK-INTERLEAVED-NEXT: [[BIN_RDX:%.*]] = add <vscale x 2 x i64> [[TMP11]], [[TMP10]]
+; CHECK-INTERLEAVED-NEXT: [[TMP13:%.*]] = call i64 @llvm.vector.reduce.add.nxv2i64(<vscale x 2 x i64> [[BIN_RDX]])
+; CHECK-INTERLEAVED-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[UMAX]], [[N_VEC]]
+; CHECK-INTERLEAVED-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
+; CHECK-INTERLEAVED: scalar.ph:
+;
+; CHECK-MAXBW-LABEL: define i64 @sext_reduction_i32_to_i64(
+; CHECK-MAXBW-SAME: ptr [[ARR:%.*]], i64 [[N:%.*]]) #[[ATTR2]] {
+; CHECK-MAXBW-NEXT: entry:
+; CHECK-MAXBW-NEXT: [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 [[N]], i64 1)
+; CHECK-MAXBW-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-MAXBW-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 1
+; CHECK-MAXBW-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX]], [[TMP1]]
+; CHECK-MAXBW-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK-MAXBW: vector.ph:
+; CHECK-MAXBW-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-MAXBW-NEXT: [[TMP3:%.*]] = mul nuw i64 [[TMP2]], 2
+; CHECK-MAXBW-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[UMAX]], [[TMP3]]
+; CHECK-MAXBW-NEXT: [[N_VEC:%.*]] = sub i64 [[UMAX]], [[N_MOD_VF]]
+; CHECK-MAXBW-NEXT: br label [[VECTOR_BODY:%.*]]
+; CHECK-MAXBW: vector.body:
+; CHECK-MAXBW-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; CHECK-MAXBW-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 2 x i64> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
+; CHECK-MAXBW-NEXT: [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[ARR]], i64 [[INDEX]]
+; CHECK-MAXBW-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 2 x i32>, ptr [[TMP4]], align 4
+; CHECK-MAXBW-NEXT: [[TMP5:%.*]] = sext <vscale x 2 x i32> [[WIDE_LOAD]] to <vscale x 2 x i64>
+; CHECK-MAXBW-NEXT: [[TMP6]] = add <vscale x 2 x i64> [[VEC_PHI]], [[TMP5]]
+; CHECK-MAXBW-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
+; CHECK-MAXBW-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-MAXBW-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP21:![0-9]+]]
+; CHECK-MAXBW: middle.block:
+; CHECK-MAXBW-NEXT: [[TMP8:%.*]] = call i64 @llvm.vector.reduce.add.nxv2i64(<vscale x 2 x i64> [[TMP6]])
+; CHECK-MAXBW-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[UMAX]], [[N_VEC]]
+; CHECK-MAXBW-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
+; CHECK-MAXBW: scalar.ph:
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %acc = phi i64 [ 0, %entry ], [ %add, %loop ]
+ %gep = getelementptr inbounds i32, ptr %arr, i64 %iv
+ %load = load i32, ptr %gep
+ %sext = sext i32 %load to i64
+ %add = add i64 %acc, %sext
+ %iv.next = add i64 %iv, 1
+ %cmp = icmp ult i64 %iv.next, %n
+ br i1 %cmp, label %loop, label %exit
+
+exit:
+ ret i64 %add
+}
+
!0 = distinct !{!0, !1}
!1 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}
More information about the llvm-commits
mailing list