[llvm] [IfConversion] Fix bug related to !HasFallThrough (PR #145471)

[Björn Pettersson via llvm-commits]

================
@@ -0,0 +1,114 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=thumbv7-apple-ios -run-pass=if-converter %s -o - | FileCheck %s
+
+# Testcase with unanalyzable branches (that may fallthrough) in the BB
+# following the diamond/triangle.
+
+# Goal here is to showcase a problem seen in the IfConverter when
+# AnalyzeBranch is indicating that the branches couldn't be analyzed. Problem
+# was originally seen for an out-of-tree target, and here we use ARM and a
+# MBB with two conditional jumps to make AnalyzeBranch return false.
+#
+# The problem was that if-converter when analyzing branches was using a
+# variable named HasFallThrough, to remember that an MBB could fallthrough to
+# the textual successor. When HasFallThrough is set we know that there are
+# fallthrough exits, but the opposite is not guaranteed. If
+# HasFallThrough==false there could still be fallthrough exists in situations
+# when analyzeBranch found unanalyzable branches. There were however a couple
+# of places in the code that checked !HasFallThrough assuming that it would
+# imply that there was no fallthrough exit.
+#
+# As a consequence we could end up merging blocks at the end of a converted
+# diamond/triangle and while doing that we messed up when fixing up the CFG
+# related to fallthrough edges. For the test cases below we incorrectly ended
+# up ended up with a fallthrough from the MBBs with two Bcc instructions to
----------------
bjope wrote:

Done!

https://github.com/llvm/llvm-project/pull/145471