[Bug analyzer/109094] [13 Regression] ICE in -fanalyzer seen in qemu's target/i386/tcg/translate.c

["cvs-commit at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109094

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by David Malcolm <dmalcolm@gcc.gnu.org>:

https://gcc.gnu.org/g:430d7d88c1a123d787f529dbc29e6632c6e556fb

commit r13-6749-g430d7d88c1a123d787f529dbc29e6632c6e556fb
Author: David Malcolm <dmalcolm@redhat.com>
Date:   Sat Mar 18 12:48:01 2023 -0400

    analyzer: fix ICE on certain longjmp calls [PR109094]

    PR analyzer/109094 reports an ICE in the analyzer seen on qemu's
    target/i386/tcg/translate.c

    The issue turned out to be that when handling a longjmp, the code
    to pop the frames was generating an svalue for the result_decl of any
    popped frame that had a non-void return type (and discarding it) leading
    to "uninit" poisoned_svalue_diagnostic instances being saved since the
    result_decl is only set by the greturn stmt.  Later, when checking the
    feasibility of the path to these diagnostics, m_check_expr was evaluated
    in the context of the frame of the longjmp, leading to an attempt to
    evaluate the result_decl of each intervening frames whilst in the
    context of the topmost frame, leading to an assertion failure in
    frame_region::get_region_for_local here:

    919             case RESULT_DECL:
    920               gcc_assert (DECL_CONTEXT (expr) == m_fun->decl);
    921               break;

    This patch updates the analyzer's longjmp implementation so that it
    doesn't attempt to generate svalues for the result_decls when popping
    frames, fixing the assertion failure (and presumably fixing "uninit"
    false positives in a release build).

    gcc/analyzer/ChangeLog:
            PR analyzer/109094
            * region-model.cc (region_model::on_longjmp): Pass false for
            new "eval_return_svalue" param of pop_frame.
            (region_model::pop_frame): Add new "eval_return_svalue" param and
            use it to suppress the call to get_rvalue on the result when
            needed by on_longjmp.
            * region-model.h (region_model::pop_frame): Add new
            "eval_return_svalue" param.

    gcc/testsuite/ChangeLog:
            PR analyzer/109094
            * gcc.dg/analyzer/setjmp-pr109094.c: New test.

    Signed-off-by: David Malcolm <dmalcolm@redhat.com>