While working on PR 119903, I noticed that there is code in 
replace_stmt_with_simplification
which makes sure that the comparison of a GIMPLE_COND does not throw
(non-call exceptions and trapping math) but the gimple verifier does not
verify this. So let's add it.

Bootstrapped and tested on x86_64-linux-gnu.

gcc/ChangeLog:

        * tree-cfg.cc (verify_gimple_cond): Error out if the comparison
        throws.

Signed-off-by: Andrew Pinski <quic_apin...@quicinc.com>
---
 gcc/tree-cfg.cc | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/gcc/tree-cfg.cc b/gcc/tree-cfg.cc
index 6a29a56ca9a..712bda1f8ca 100644
--- a/gcc/tree-cfg.cc
+++ b/gcc/tree-cfg.cc
@@ -5113,6 +5113,19 @@ verify_gimple_cond (gcond *stmt)
       return true;
     }
 
+  tree lhs = gimple_cond_lhs (stmt);
+
+  /* GIMPLE_CONDs condition may not throw.  */
+  if (flag_exceptions
+      && cfun->can_throw_non_call_exceptions
+      && operation_could_trap_p (gimple_cond_code (stmt),
+                                FLOAT_TYPE_P (TREE_TYPE (lhs)),
+                                false, NULL_TREE))
+    {
+      error ("gimple cond condition cannot throw");
+      return true;
+    }
+
   return verify_gimple_comparison (boolean_type_node,
                                   gimple_cond_lhs (stmt),
                                   gimple_cond_rhs (stmt),
-- 
2.43.0

Reply via email to