[llvm] [LV] Relax OutOfLoopUses check in `getMinMaxRecurrence()` (PR #189906)
Madhur Amilkanthwar via llvm-commits
llvm-commits at lists.llvm.org
Wed Apr 22 21:10:03 PDT 2026
https://github.com/madhur13490 updated https://github.com/llvm/llvm-project/pull/189906
>From 1bec987e81b366a2c06ee425bd965108f7fb7ff0 Mon Sep 17 00:00:00 2001
From: Madhur Amilkanthwar <madhura at nvidia.com>
Date: Wed, 1 Apr 2026 00:44:22 -0700
Subject: [PATCH 1/2] [LV] Relax OutOfLoopUses check in `getMinMaxRecurrence()`
For more context of this patch, please see #182473.
By relaxing, `OutOfLoopUses` check, we can recognize more
case for minmax reduction and thus vectorize.
---
llvm/lib/Analysis/IVDescriptors.cpp | 4 +-
.../minmax-extra-outside-users.ll | 77 +++++++++++++++++++
2 files changed, 78 insertions(+), 3 deletions(-)
create mode 100644 llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll
diff --git a/llvm/lib/Analysis/IVDescriptors.cpp b/llvm/lib/Analysis/IVDescriptors.cpp
index 185943e0a9d2e..fa92ef6d8cb7c 100644
--- a/llvm/lib/Analysis/IVDescriptors.cpp
+++ b/llvm/lib/Analysis/IVDescriptors.cpp
@@ -363,14 +363,12 @@ static RecurrenceDescriptor getMinMaxRecurrence(PHINode *Phi, Loop *TheLoop,
// Validate chain entries and collect stores from chain entries and
// intermediate ops.
SmallVector<StoreInst *> Stores;
- unsigned OutOfLoopUses = 0;
for (Value *V : Chain) {
for (User *U : V->users()) {
if (Chain.contains(U))
continue;
auto *I = dyn_cast<Instruction>(U);
- if (!I || (!TheLoop->contains(I) &&
- (V != BackedgeValue || ++OutOfLoopUses > 1)))
+ if (!I || (!TheLoop->contains(I) && V != BackedgeValue))
return {};
if (!TheLoop->contains(I))
continue;
diff --git a/llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll b/llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll
new file mode 100644
index 0000000000000..667080665a702
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll
@@ -0,0 +1,77 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -p loop-vectorize -force-vector-width=4 -S %s | FileCheck %s
+
+define i32 @smin_three_exit_users(ptr %src, i64 %n) {
+; CHECK-LABEL: define i32 @smin_three_exit_users(
+; CHECK-SAME: ptr [[SRC:%.*]], i64 [[N:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 4
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
+; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ splat (i32 2147483647), %[[VECTOR_PH]] ], [ [[TMP2:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i64 [[INDEX]]
+; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 4
+; CHECK-NEXT: [[TMP1:%.*]] = icmp slt <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
+; CHECK-NEXT: [[TMP2]] = select <4 x i1> [[TMP1]], <4 x i32> [[VEC_PHI]], <4 x i32> [[WIDE_LOAD]]
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: [[TMP4:%.*]] = call i32 @llvm.vector.reduce.smin.v4i32(<4 x i32> [[TMP2]])
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP4]], %[[MIDDLE_BLOCK]] ], [ 2147483647, %[[ENTRY]] ]
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[MIN:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[MIN_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i64 [[IV]]
+; CHECK-NEXT: [[V:%.*]] = load i32, ptr [[GEP]], align 4
+; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[MIN]], [[V]]
+; CHECK-NEXT: [[MIN_NEXT]] = select i1 [[CMP]], i32 [[MIN]], i32 [[V]]
+; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[R0:%.*]] = phi i32 [ [[MIN_NEXT]], %[[LOOP]] ], [ [[TMP4]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[R1:%.*]] = phi i32 [ [[MIN_NEXT]], %[[LOOP]] ], [ [[TMP4]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[R2:%.*]] = phi i32 [ [[MIN_NEXT]], %[[LOOP]] ], [ [[TMP4]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[SUM0:%.*]] = add i32 [[R0]], [[R1]]
+; CHECK-NEXT: [[SUM1:%.*]] = add i32 [[SUM0]], [[R2]]
+; CHECK-NEXT: ret i32 [[SUM1]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %min = phi i32 [ 2147483647, %entry ], [ %min.next, %loop ]
+ %gep = getelementptr inbounds i32, ptr %src, i64 %iv
+ %v = load i32, ptr %gep, align 4
+ %cmp = icmp slt i32 %min, %v
+ %min.next = select i1 %cmp, i32 %min, i32 %v
+ %iv.next = add nuw nsw i64 %iv, 1
+ %ec = icmp eq i64 %iv.next, %n
+ br i1 %ec, label %exit, label %loop
+
+exit:
+ %r0 = phi i32 [ %min.next, %loop ]
+ %r1 = phi i32 [ %min.next, %loop ]
+ %r2 = phi i32 [ %min.next, %loop ]
+ %sum0 = add i32 %r0, %r1
+ %sum1 = add i32 %sum0, %r2
+ ret i32 %sum1
+}
+;.
+; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
+;.
>From 9988c52c5d2472ad916b043851deb6673f0577ca Mon Sep 17 00:00:00 2001
From: Madhur Amilkanthwar <madhura at nvidia.com>
Date: Wed, 22 Apr 2026 21:08:28 -0700
Subject: [PATCH 2/2] fixup! address review comments
---
.../minmax-extra-outside-users.ll | 77 +++++++++++++++++--
1 file changed, 70 insertions(+), 7 deletions(-)
diff --git a/llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll b/llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll
index 667080665a702..1c58544bea456 100644
--- a/llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll
+++ b/llvm/test/Transforms/LoopVectorize/minmax-extra-outside-users.ll
@@ -1,4 +1,4 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
; RUN: opt -p loop-vectorize -force-vector-width=4 -S %s | FileCheck %s
define i32 @smin_three_exit_users(ptr %src, i64 %n) {
@@ -69,9 +69,72 @@ exit:
%sum1 = add i32 %sum0, %r2
ret i32 %sum1
}
-;.
-; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
-; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
-; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
-; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
-;.
+
+define i32 @smax_three_exit_users(ptr %src, i64 %n) {
+; CHECK-LABEL: define i32 @smax_three_exit_users(
+; CHECK-SAME: ptr [[SRC:%.*]], i64 [[N:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 4
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
+; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ splat (i32 -2147483648), %[[VECTOR_PH]] ], [ [[TMP2:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i64 [[INDEX]]
+; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 4
+; CHECK-NEXT: [[TMP1:%.*]] = icmp sgt <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
+; CHECK-NEXT: [[TMP2]] = select <4 x i1> [[TMP1]], <4 x i32> [[VEC_PHI]], <4 x i32> [[WIDE_LOAD]]
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP3]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: [[TMP4:%.*]] = call i32 @llvm.vector.reduce.smax.v4i32(<4 x i32> [[TMP2]])
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP4]], %[[MIDDLE_BLOCK]] ], [ -2147483648, %[[ENTRY]] ]
+; CHECK-NEXT: br label %[[LOOP:.*]]
+; CHECK: [[LOOP]]:
+; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[MAX:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[MAX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i64 [[IV]]
+; CHECK-NEXT: [[V:%.*]] = load i32, ptr [[GEP]], align 4
+; CHECK-NEXT: [[CMP:%.*]] = icmp sgt i32 [[MAX]], [[V]]
+; CHECK-NEXT: [[MAX_NEXT]] = select i1 [[CMP]], i32 [[MAX]], i32 [[V]]
+; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: [[R0:%.*]] = phi i32 [ [[MAX_NEXT]], %[[LOOP]] ], [ [[TMP4]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[R1:%.*]] = phi i32 [ [[MAX_NEXT]], %[[LOOP]] ], [ [[TMP4]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[R2:%.*]] = phi i32 [ [[MAX_NEXT]], %[[LOOP]] ], [ [[TMP4]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: [[SUM0:%.*]] = add i32 [[R0]], [[R1]]
+; CHECK-NEXT: [[SUM1:%.*]] = add i32 [[SUM0]], [[R2]]
+; CHECK-NEXT: ret i32 [[SUM1]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %max = phi i32 [ -2147483648, %entry ], [ %max.next, %loop ]
+ %gep = getelementptr inbounds i32, ptr %src, i64 %iv
+ %v = load i32, ptr %gep, align 4
+ %cmp = icmp sgt i32 %max, %v
+ %max.next = select i1 %cmp, i32 %max, i32 %v
+ %iv.next = add nuw nsw i64 %iv, 1
+ %ec = icmp eq i64 %iv.next, %n
+ br i1 %ec, label %exit, label %loop
+
+exit:
+ %r0 = phi i32 [ %max.next, %loop ]
+ %r1 = phi i32 [ %max.next, %loop ]
+ %r2 = phi i32 [ %max.next, %loop ]
+ %sum0 = add i32 %r0, %r1
+ %sum1 = add i32 %sum0, %r2
+ ret i32 %sum1
+}
More information about the llvm-commits
mailing list