c++: Don't clear TREE_READONLY for -fmerge-all-constants for non-aggregates [PR107558]
authorJakub Jelinek <jakub@redhat.com>
Fri, 10 Mar 2023 19:38:13 +0000 (20:38 +0100)
committerJakub Jelinek <jakub@redhat.com>
Fri, 10 Mar 2023 19:38:13 +0000 (20:38 +0100)
commit60b6f5c0a334db3f8f6dffaf0b9aab42fd5c54a2
tree95ca6b2bfe6591b68ab19f0886659ab67a820ef9
parent991f9eb2da3a268b7b08346761fa0078ab55f506
c++: Don't clear TREE_READONLY for -fmerge-all-constants for non-aggregates [PR107558]

The following testcase ICEs, because OpenMP lowering for shared clause
on l variable with REFERENCE_TYPE creates POINTER_TYPE to REFERENCE_TYPE.
The reason is that the automatic variable has non-trivial construction
(reference to a lambda) and -fmerge-all-constants is on and so TREE_READONLY
isn't set - omp-low will handle automatic TREE_READONLY vars in shared
specially and only copy to the construct and not back, while !TREE_READONLY
are assumed to be changeable.
The PR91529 change rationale was that the gimplification can change
some non-addressable automatic variables to TREE_STATIC with
-fmerge-all-constants and therefore TREE_READONLY on them is undesirable.
But, the gimplifier does that only for aggregate variables:
  switch (TREE_CODE (type))
    {
    case RECORD_TYPE:
    case UNION_TYPE:
    case QUAL_UNION_TYPE:
    case ARRAY_TYPE:
and not for anything else.  So, I think clearing TREE_READONLY for
automatic integral or reference or pointer etc. vars for
-fmerge-all-constants only is unnecessary.

2023-03-10  Jakub Jelinek  <jakub@redhat.com>

PR c++/107558
* decl.cc (cp_finish_decl): Don't clear TREE_READONLY on
automatic non-aggregate variables just because of
-fmerge-all-constants.

* g++.dg/gomp/pr107558.C: New test.
gcc/cp/decl.cc
gcc/testsuite/g++.dg/gomp/pr107558.C [new file with mode: 0644]