[Bug tree-optimization/87025] [9 Regression] ICE in add_record, at optinfo-emit-json.cc:175

[dmalcolm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87025

--- Comment #7 from David Malcolm <dmalcolm at gcc dot gnu.org> ---
Author: dmalcolm
Date: Mon Nov 19 16:42:03 2018
New Revision: 266280

URL: https://gcc.gnu.org/viewcvs?rev=266280&root=gcc&view=rev
Log:
Fix -fsave-optimization-record ICE (PR tree-optimization/87025)

PR tree-optimization/87025 reports an ICE within
-fsave-optimization-record's optrecord_json_writer.

The issue is that dump_context::begin_scope creates an optinfo
of kind OPTINFO_KIND_SCOPE, but fails to call
dump_context::end_any_optinfo, so the optinfo for the scope remains
pending.

The JSON writer would normally push a JSON array for the "scope" optinfo
when the latter is emitted.  However, if a dump_* call happens that
doesn't flush the "scope" optinfo e.g. dump_printf (as opposed to
dump_printf_loc), that dump_ call is added to the pending optinfo, and
optinfo::handle_dump_file_kind changes the pending optinfo's m_kind
(e.g. to OPTINFO_KIND_NOTE).  Hence when the pending optinfo is
eventually emitted, it isn't OPTINFO_KIND_SCOPE anymore, and hence
the JSON writer doesn't create and push a JSON array for it, leading
to dump_context's view of scopes getting out-of-sync with that of
the JSON writer's.

Later, dump_context::end_scope unconditionally tries to pop the JSON scope
array, but no JSON scope array was added, leading to an assertion
failure (or crash).

The fix is to call dump_context::end_any_optinfo immediately after
creating the scope optinfo, so that it is emitted immediately, ensuring
that the JSON writer stays in-sync with the dump_context.

gcc/ChangeLog:
        PR tree-optimization/87025
        * dumpfile.c (dump_context::begin_scope): Call end_any_optinfo
        immediately after creating the scope optinfo.
        (selftest::test_pr87025): New function.
        (selftest::dumpfile_c_tests): Call it.
        * optinfo-emit-json.cc (optrecord_json_writer::pop_scope): Assert
        that we're not popping the top-level records array.
        * optinfo.cc (optinfo::handle_dump_file_kind): Assert that we're
        not changing the kind of a "scope" optinfo.

gcc/testsuite/ChangeLog:
        PR tree-optimization/87025
        * gcc.dg/pr87025.c: New test.


Added:
    trunk/gcc/testsuite/gcc.dg/pr87025.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/dumpfile.c
    trunk/gcc/optinfo-emit-json.cc
    trunk/gcc/optinfo.cc
    trunk/gcc/testsuite/ChangeLog