[llvm] 7b58716 - [LoopFusion] Remove DT edge from Extiblock to ExitBlockSuc (#193641)

[via llvm-commits]

Author: Afonso Rafael (Gafanhoto)
Date: 2026-04-29T06:39:25-04:00
New Revision: 7b58716d96c3ae4c0c4e6f72e29b16137bb6224b

URL: https://github.com/llvm/llvm-project/commit/7b58716d96c3ae4c0c4e6f72e29b16137bb6224b
DIFF: https://github.com/llvm/llvm-project/commit/7b58716d96c3ae4c0c4e6f72e29b16137bb6224b.diff

LOG: [LoopFusion] Remove DT edge from Extiblock to ExitBlockSuc (#193641)

To remove the exit block, it cannot have successors, if this edge is not
removed, when applying the updates to the DT the following assertion
will appear:
"Assertion `Node->isLeaf() && "Node is not a leaf node."' failed"

This assertion does not always fail because before applying the updates
on the "GenericDomTreeContruction", "ApplyUpdates" function it runs
CalculateFromScratch on some situations:

    // Make unittests of the incremental algorithm work
    if (DT.DomTreeNodes.size() <= 100) {
      if (BUI.NumLegalized > DT.DomTreeNodes.size())
        CalculateFromScratch(DT, &BUI);
    } else if (BUI.NumLegalized > DT.DomTreeNodes.size() / 40)
      CalculateFromScratch(DT, &BUI);

Added: 
    llvm/test/Transforms/LoopFusion/pr193641.ll

Modified: 
    llvm/lib/Transforms/Scalar/LoopFuse.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Scalar/LoopFuse.cpp b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
index be543062670c6..6de3fc0174d7d 100644
--- a/llvm/lib/Transforms/Scalar/LoopFuse.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopFuse.cpp
@@ -1818,6 +1818,8 @@ struct LoopFuser {
         DominatorTree::Insert, FC0GuardBlock, FC1NonLoopBlock));
 
     if (FC0.Peeled) {
+      TreeUpdates.emplace_back(DominatorTree::UpdateType(
+          DominatorTree::Delete, FC0.ExitBlock, FC0ExitBlockSuccessor));
       // Remove the Block after the ExitBlock of FC0
       TreeUpdates.emplace_back(DominatorTree::UpdateType(
           DominatorTree::Delete, FC0ExitBlockSuccessor, FC1GuardBlock));

diff  --git a/llvm/test/Transforms/LoopFusion/pr193641.ll b/llvm/test/Transforms/LoopFusion/pr193641.ll
new file mode 100644
index 0000000000000..4ed8eb2e911ad
--- /dev/null
+++ b/llvm/test/Transforms/LoopFusion/pr193641.ll
@@ -0,0 +1,45 @@
+; REQUIRES: asserts
+; RUN: opt -passes=loop-fusion -loop-fusion-peel-max-count=2 -disable-output -stats < %s 2>&1 | FileCheck -check-prefix=STAT %s
+; STAT: 1 loop-fusion - Loops fused
+
+; Test reduced from the file diags_on_lat_aux_grid.F90 present in pop2,
+; a program from spec 2017.
+; Loop fusion should not crash when flushing the DomTreeUpdater due to
+; missed edge removals when deleting basic blocks.
+
+define void @diags_on_lat_aux_grid(i1 %cond, i1 %cond1, i1 %0, ptr noalias %b1, ptr noalias %b2) {
+.critedge:
+  br i1 %cond, label %common.ret, label %thread-pre-split
+
+common.ret:                                       ; preds = %.loopexit, %.critedge
+  ret void
+
+thread-pre-split:                                 ; preds = %.critedge
+  br i1 %cond1, label %2, label %.loopexit
+
+2:                                                ; preds = %thread-pre-split
+  br i1 %0, label %.lr.ph1333, label %._crit_edge1334
+
+.lr.ph1333:                                       ; preds = %.lr.ph1333, %2
+  %indvars.iv1386 = phi i64 [ %indvars.iv.next1387, %.lr.ph1333 ], [ 0, %2 ]
+  %3 = getelementptr [8 x i8], ptr %b1, i64 %indvars.iv1386
+  store double 0.000000e+00, ptr %3, align 8
+  %indvars.iv.next1387 = add i64 %indvars.iv1386, 1
+  %exitcond1389.not = icmp eq i64 %indvars.iv1386, 100
+  br i1 %exitcond1389.not, label %._crit_edge1334, label %.lr.ph1333
+
+._crit_edge1334:                                  ; preds = %.lr.ph1333, %2
+  br i1 %0, label %.lr.ph1336, label %.loopexit
+
+.lr.ph1336:                                       ; preds = %.lr.ph1336, %._crit_edge1334
+  %indvars.iv1390 = phi i64 [ %indvars.iv.next1391, %.lr.ph1336 ], [ 0, %._crit_edge1334 ]
+  %5 = getelementptr [8 x i8], ptr %b2, i64 %indvars.iv1390
+  store double 0.000000e+00, ptr %5, align 8
+  %indvars.iv.next1391 = add i64 %indvars.iv1390, 1
+  %exitcond1393.not = icmp eq i64 %indvars.iv.next1391, 99
+  br i1 %exitcond1393.not, label %.loopexit, label %.lr.ph1336
+
+.loopexit:                                        ; preds = %.lr.ph1336, %._crit_edge1334, %thread-pre-split
+  store volatile i32 0, ptr %b1, align 4
+  br label %common.ret
+}