[SimplifyCFG] Try to merge edge block when threading (PR55765)

When threading, we always create a new block for the threaded edge
(even if the edge is not critical), which will later get folded back
into the predecessor if possible. Depending on precise processing
order, this separate block may break the detection of trivial
cycles in the threading code, which normally avoids infinite
threading of loops. Explicitly merge the created edge block into
the predecessor to avoid this.

Fixes llvm#55765.

Differential Revision: https://reviews.llvm.org/D127216

Author: nikic

llvm/lib/Transforms/Utils/SimplifyCFG.cpp

@@ -3145,6 +3145,12 @@ FoldCondBranchOnValueKnownInPredecessorImpl(BranchInst *BI, DomTreeUpdater *DTU,
       DTU->applyUpdates(Updates);
     }
 
+    // For simplicity, we created a separate basic block for the edge. Merge
+    // it back into the predecessor if possible. This not only avoids
+    // unnecessary SimplifyCFG iterations, but also makes sure that we don't
+    // bypass the check for trivial cycles above.
+    MergeBlockIntoPredecessor(EdgeBB, DTU);
+
     // Signal repeat, simplifying any other constants.
     return None;
   }

llvm/test/Transforms/SimplifyCFG/pr55765.ll

@@ -0,0 +1,63 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -simplifycfg < %s | FileCheck %s
+
+; This used to infinitely thread between loop and loop.latch without reaching a
+; fixed point.
+
+declare void @dummy()
+
+define i32 @main(i1 %c1, i1 %c2, i32 %y) {
+; CHECK-LABEL: @main(
+; CHECK-NEXT:    br i1 [[C1:%.*]], label [[EXIT:%.*]], label [[LOOP_PRE_PREHEADER:%.*]]
+; CHECK:       loop.pre.preheader:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[Y:%.*]], -1
+; CHECK-NEXT:    br i1 [[CMP]], label [[LOOP_PREHEADER:%.*]], label [[EXIT]]
+; CHECK:       loop.preheader:
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp eq i32 [[Y]], 0
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    br i1 [[C1]], label [[LOOP2:%.*]], label [[LOOP]]
+; CHECK:       loop.latch:
+; CHECK-NEXT:    br i1 [[CMP]], label [[LOOP]], label [[EXIT]]
+; CHECK:       loop2:
+; CHECK-NEXT:    br i1 [[CMP2]], label [[JOIN:%.*]], label [[IF:%.*]]
+; CHECK:       if:
+; CHECK-NEXT:    call void @dummy()
+; CHECK-NEXT:    br label [[JOIN]]
+; CHECK:       join:
+; CHECK-NEXT:    br i1 [[C2:%.*]], label [[LOOP2]], label [[LOOP_LATCH:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret i32 0
+;
+  br i1 %c1, label %exit, label %loop.pre.preheader
+
+loop.pre.preheader:
+  %cmp = icmp sgt i32 %y, -1
+  br i1 %cmp, label %loop.preheader, label %exit
+
+loop.preheader:
+  %cmp2 = icmp eq i32 %y, 0
+  br label %loop
+
+loop:
+  br i1 %c1, label %loop2, label %loop.latch
+
+loop.latch:
+  br i1 %cmp, label %loop, label %exit
+
+loop2:
+  br i1 %cmp2, label %join, label %if
+
+if:
+  call void @dummy()
+  br label %join
+
+join:
+  br i1 %c2, label %loop2, label %loop.latch
+
+exit:
+  ret i32 0
+
+; uselistorder directives
+  uselistorder label %loop2, { 1, 0 }
+}