[llvm] [LoopFusion] Validate loop structure before creating LoopCandidates (PR #192280)
Madhur Amilkanthwar via llvm-commits
llvm-commits at lists.llvm.org
Thu Apr 16 01:22:34 PDT 2026
https://github.com/madhur13490 updated https://github.com/llvm/llvm-project/pull/192280
>From 5730f94cc9ef9dc7ea67586447bba8df3d9307a6 Mon Sep 17 00:00:00 2001
From: Madhur Amilkanthwar <madhura at nvidia.com>
Date: Tue, 3 Mar 2026 23:19:46 -0800
Subject: [PATCH 1/2] [LoopFusion] Validate loop structure before creating
LoopCandidates
This patch add early check in LoopFuse.cpp
to skip loops missing preheader, header, exiting,
exit block or latch. It also makes `reportInvalidCandidate`
gracefully return false instead of asserting when Preheader
is null.
In other words this patch skips loops which lack the required structure.
Fixes #156670.
---
llvm/lib/Transforms/Scalar/LoopFuse.cpp | 12 +++-
.../test/Transforms/LoopFusion/indirect-br.ll | 59 +++++++++++++++++++
2 files changed, 70 insertions(+), 1 deletion(-)
create mode 100644 llvm/test/Transforms/LoopFusion/indirect-br.ll
diff --git a/llvm/lib/Transforms/Scalar/LoopFuse.cpp b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
index b577bf6584615..9c7500b0a290e 100644
--- a/llvm/lib/Transforms/Scalar/LoopFuse.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
@@ -363,7 +363,9 @@ struct FusionCandidate {
bool reportInvalidCandidate(Statistic &Stat) const {
using namespace ore;
- assert(L && Preheader && "Fusion candidate not initialized properly!");
+ assert(L && "Fusion candidate has null loop!");
+ if (!Preheader)
+ return false;
#if LLVM_ENABLE_STATS
++Stat;
ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, Stat.getName(),
@@ -573,6 +575,14 @@ struct LoopFuser {
/// Flow Equivalent sets, sorted by dominance.
void collectFusionCandidates(const LoopVector &LV) {
for (Loop *L : LV) {
+ // Skip loops that do not have the structure required for fusion (e.g.
+ // loops with indirectbr may lack a preheader). Creating a FusionCandidate
+ // for such loops can lead to reportInvalidCandidate being called with
+ // a null Preheader.
+ if (!L->getLoopPreheader() || !L->getHeader() || !L->getExitingBlock() ||
+ !L->getExitBlock() || !L->getLoopLatch())
+ continue;
+
TTI::PeelingPreferences PP =
gatherPeelingPreferences(L, SE, TTI, std::nullopt, std::nullopt);
FusionCandidate CurrCand(L, DT, &PDT, ORE, PP);
diff --git a/llvm/test/Transforms/LoopFusion/indirect-br.ll b/llvm/test/Transforms/LoopFusion/indirect-br.ll
new file mode 100644
index 0000000000000..2f95a3a6c4b6f
--- /dev/null
+++ b/llvm/test/Transforms/LoopFusion/indirect-br.ll
@@ -0,0 +1,59 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -S -passes=loop-fusion < %s | FileCheck %s
+; Test that loop-fusion does not crash on functions with indirectbr and
+; loops that lack a preheader (or other required loop structure).
+
+define i32 @foo(ptr nocapture readonly %p) {
+; CHECK-LABEL: define i32 @foo(
+; CHECK-SAME: ptr readonly captures(none) [[P:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*]]:
+; CHECK-NEXT: [[TBL:%.*]] = alloca <2 x ptr>, align 16
+; CHECK-NEXT: store <2 x ptr> <ptr blockaddress(@foo, %[[INDIRECT:.*]]), ptr blockaddress(@foo, %[[END:.*]])>, ptr [[TBL]], align 16
+; CHECK-NEXT: br label %[[FOR_COND:.*]]
+; CHECK: [[FOR_COND]]:
+; CHECK-NEXT: [[CUR:%.*]] = phi ptr [ [[P]], %[[ENTRY]] ], [ [[NEXT:%.*]], %[[FOR_COND]] ]
+; CHECK-NEXT: [[TMP0:%.*]] = load i8, ptr [[CUR]], align 1
+; CHECK-NEXT: [[CMP:%.*]] = icmp eq i8 [[TMP0]], 7
+; CHECK-NEXT: [[NEXT]] = getelementptr inbounds i8, ptr [[CUR]], i64 1
+; CHECK-NEXT: br i1 [[CMP]], label %[[INDIRECT_PREHEADER:.*]], label %[[FOR_COND]]
+; CHECK: [[INDIRECT_PREHEADER]]:
+; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[NEXT]], align 1
+; CHECK-NEXT: [[IDX:%.*]] = sext i8 [[TMP1]] to i64
+; CHECK-NEXT: [[SLOT:%.*]] = getelementptr inbounds <2 x ptr>, ptr [[TBL]], i64 0, i64 [[IDX]]
+; CHECK-NEXT: [[TMP2:%.*]] = load ptr, ptr [[SLOT]], align 8
+; CHECK-NEXT: br label %[[INDIRECT]]
+; CHECK: [[INDIRECT]]:
+; CHECK-NEXT: indirectbr ptr [[TMP2]], [label %[[INDIRECT]], label %[[END]]]
+; CHECK: [[INDIRECT2:.*:]]
+; CHECK-NEXT: indirectbr ptr [[TMP2]], [label %[[INDIRECT]], label %[[END]]]
+; CHECK: [[END]]:
+; CHECK-NEXT: ret i32 0
+;
+entry:
+ %tbl = alloca <2 x ptr>, align 16
+ store <2 x ptr> <ptr blockaddress(@foo, %indirect), ptr blockaddress(@foo, %exit)>, ptr %tbl, align 16
+ br label %for.cond
+
+for.cond: ; preds = %for.cond, %entry
+ %cur = phi ptr [ %p, %entry ], [ %next, %for.cond ]
+ %val = load i8, ptr %cur, align 1
+ %cmp = icmp eq i8 %val, 7
+ %next = getelementptr inbounds i8, ptr %cur, i64 1
+ br i1 %cmp, label %indirect.preheader, label %for.cond
+
+indirect.preheader: ; preds = %for.cond
+ %raw = load i8, ptr %next, align 1
+ %idx = sext i8 %raw to i64
+ %slot = getelementptr inbounds <2 x ptr>, ptr %tbl, i64 0, i64 %idx
+ %dest = load ptr, ptr %slot, align 8
+ br label %indirect
+
+indirect:
+ indirectbr ptr %dest, [label %indirect, label %exit]
+
+indirect2:
+ indirectbr ptr %dest, [label %indirect, label %exit]
+
+exit:
+ ret i32 0
+}
>From b732bc5b0405bd7649b5ef3a2516da97d9b8c955 Mon Sep 17 00:00:00 2001
From: Madhur Amilkanthwar <madhura at nvidia.com>
Date: Thu, 16 Apr 2026 01:21:49 -0700
Subject: [PATCH 2/2] fixup! address review comments
---
llvm/lib/Transforms/Scalar/LoopFuse.cpp | 19 ++++++-------------
.../test/Transforms/LoopFusion/indirect-br.ll | 18 +++++++++---------
2 files changed, 15 insertions(+), 22 deletions(-)
diff --git a/llvm/lib/Transforms/Scalar/LoopFuse.cpp b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
index 9c7500b0a290e..be543062670c6 100644
--- a/llvm/lib/Transforms/Scalar/LoopFuse.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
@@ -363,14 +363,15 @@ struct FusionCandidate {
bool reportInvalidCandidate(Statistic &Stat) const {
using namespace ore;
- assert(L && "Fusion candidate has null loop!");
- if (!Preheader)
- return false;
+ ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, "InvalidCandidate",
+ L->getStartLoc(), L->getHeader())
+ << "Loop is not a candidate for fusion");
+
#if LLVM_ENABLE_STATS
++Stat;
ORE.emit(OptimizationRemarkAnalysis(DEBUG_TYPE, Stat.getName(),
- L->getStartLoc(), Preheader)
- << "[" << Preheader->getParent()->getName() << "]: "
+ L->getStartLoc(), L->getHeader())
+ << "[" << L->getHeader()->getParent()->getName() << "]: "
<< "Loop is not a candidate for fusion: " << Stat.getDesc());
#endif
return false;
@@ -575,14 +576,6 @@ struct LoopFuser {
/// Flow Equivalent sets, sorted by dominance.
void collectFusionCandidates(const LoopVector &LV) {
for (Loop *L : LV) {
- // Skip loops that do not have the structure required for fusion (e.g.
- // loops with indirectbr may lack a preheader). Creating a FusionCandidate
- // for such loops can lead to reportInvalidCandidate being called with
- // a null Preheader.
- if (!L->getLoopPreheader() || !L->getHeader() || !L->getExitingBlock() ||
- !L->getExitBlock() || !L->getLoopLatch())
- continue;
-
TTI::PeelingPreferences PP =
gatherPeelingPreferences(L, SE, TTI, std::nullopt, std::nullopt);
FusionCandidate CurrCand(L, DT, &PDT, ORE, PP);
diff --git a/llvm/test/Transforms/LoopFusion/indirect-br.ll b/llvm/test/Transforms/LoopFusion/indirect-br.ll
index 2f95a3a6c4b6f..688338bfb77fc 100644
--- a/llvm/test/Transforms/LoopFusion/indirect-br.ll
+++ b/llvm/test/Transforms/LoopFusion/indirect-br.ll
@@ -8,25 +8,25 @@ define i32 @foo(ptr nocapture readonly %p) {
; CHECK-SAME: ptr readonly captures(none) [[P:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*]]:
; CHECK-NEXT: [[TBL:%.*]] = alloca <2 x ptr>, align 16
-; CHECK-NEXT: store <2 x ptr> <ptr blockaddress(@foo, %[[INDIRECT:.*]]), ptr blockaddress(@foo, %[[END:.*]])>, ptr [[TBL]], align 16
+; CHECK-NEXT: store <2 x ptr> <ptr blockaddress(@foo, %[[INDIRECT:.*]]), ptr blockaddress(@foo, %[[EXIT:.*]])>, ptr [[TBL]], align 16
; CHECK-NEXT: br label %[[FOR_COND:.*]]
; CHECK: [[FOR_COND]]:
; CHECK-NEXT: [[CUR:%.*]] = phi ptr [ [[P]], %[[ENTRY]] ], [ [[NEXT:%.*]], %[[FOR_COND]] ]
-; CHECK-NEXT: [[TMP0:%.*]] = load i8, ptr [[CUR]], align 1
-; CHECK-NEXT: [[CMP:%.*]] = icmp eq i8 [[TMP0]], 7
+; CHECK-NEXT: [[VAL:%.*]] = load i8, ptr [[CUR]], align 1
+; CHECK-NEXT: [[CMP:%.*]] = icmp eq i8 [[VAL]], 7
; CHECK-NEXT: [[NEXT]] = getelementptr inbounds i8, ptr [[CUR]], i64 1
; CHECK-NEXT: br i1 [[CMP]], label %[[INDIRECT_PREHEADER:.*]], label %[[FOR_COND]]
; CHECK: [[INDIRECT_PREHEADER]]:
-; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[NEXT]], align 1
-; CHECK-NEXT: [[IDX:%.*]] = sext i8 [[TMP1]] to i64
+; CHECK-NEXT: [[RAW:%.*]] = load i8, ptr [[NEXT]], align 1
+; CHECK-NEXT: [[IDX:%.*]] = sext i8 [[RAW]] to i64
; CHECK-NEXT: [[SLOT:%.*]] = getelementptr inbounds <2 x ptr>, ptr [[TBL]], i64 0, i64 [[IDX]]
-; CHECK-NEXT: [[TMP2:%.*]] = load ptr, ptr [[SLOT]], align 8
+; CHECK-NEXT: [[DEST:%.*]] = load ptr, ptr [[SLOT]], align 8
; CHECK-NEXT: br label %[[INDIRECT]]
; CHECK: [[INDIRECT]]:
-; CHECK-NEXT: indirectbr ptr [[TMP2]], [label %[[INDIRECT]], label %[[END]]]
+; CHECK-NEXT: indirectbr ptr [[DEST]], [label %[[INDIRECT]], label %[[EXIT]]]
; CHECK: [[INDIRECT2:.*:]]
-; CHECK-NEXT: indirectbr ptr [[TMP2]], [label %[[INDIRECT]], label %[[END]]]
-; CHECK: [[END]]:
+; CHECK-NEXT: indirectbr ptr [[DEST]], [label %[[INDIRECT]], label %[[EXIT]]]
+; CHECK: [[EXIT]]:
; CHECK-NEXT: ret i32 0
;
entry:
More information about the llvm-commits
mailing list