gimplify: Clear TREE_READONLY on automatic vars being stored into [PR104529]
authorJakub Jelinek <jakub@redhat.com>
Fri, 4 Mar 2022 14:14:59 +0000 (15:14 +0100)
committerJakub Jelinek <jakub@redhat.com>
Fri, 4 Mar 2022 14:14:59 +0000 (15:14 +0100)
commitc85aaf2cbe9da50e23655a8082a37166adf4c0f7
tree740b795d72aff22cb9e36d1692806144095cd7fe
parent07395f19dff610f03d1b1d30c8bd640f610c45dc
gimplify: Clear TREE_READONLY on automatic vars being stored into [PR104529]

The following testcase regressed when SRA started punting on stores to
TREE_READONLY vars.  We document that:
"In a VAR_DECL, PARM_DECL or FIELD_DECL, or any kind of ..._REF node,
nonzero means it may not be the lhs of an assignment."
so the SRA change looks desirable.  On the other side, at least in this
testcase the TREE_READONLY is set there intentionally from the
PR85873 fix, because gimplify_init_constructor itself uses TREE_READONLY
on the object to determine if it can perform promotion to static const
or not.

So, similarly to other spots in the gimplifier where we also clear
TREE_READONLY when we emit IL that stores into the object, this
does the same in gimplify_init_constructor, but in the way so that
the TREE_READONLY test for the promotion to static const keeps working
and doesn't change anything for notify_temp_creation mode, which doesn't
emit any IL, just tests if it would need a temporary or not.

This keeps PR85873 testcase working as before and fixes this regression.

2022-03-04  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/104529
* gimplify.cc (gimplify_init_constructor): Clear TREE_READONLY
on automatic objects which will be runtime initialized.

* g++.dg/tree-ssa/pr104529.C: New test.
gcc/gimplify.cc
gcc/testsuite/g++.dg/tree-ssa/pr104529.C [new file with mode: 0644]