[Bug c++/91165] [10 Regression] error: location references block not in block tree

[jason at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91165

--- Comment #8 from Jason Merrill <jason at gcc dot gnu.org> ---
Author: jason
Date: Mon Dec 16 23:25:08 2019
New Revision: 279447

URL: https://gcc.gnu.org/viewcvs?rev=279447&root=gcc&view=rev
Log:
        PR c++/91165 - verify_gimple ICE with cached constexpr.

It seems we need to unshare even non-CONSTRUCTOR expressions that we are
going to stick in the constexpr_call_table, so we don't end up sharing the
same e.g. ADDR_EXPR between two different functions.  I now think I
understand why unsharing CONSTRUCTOR arguments was improving memory
performance: separating the arguments from the caller function allows the
caller function to be GC'd better.  But it occurs to me that we don't need
to unshare until we decide that we're evaluating and caching this call, so
we can avoid the CONSTRUCTOR unshare/free pair for tentative arguments.
Freeing the tentative TREE_VEC still seems worth doing, so free_bindings
isn't going away entirely.

        * constexpr.c (cxx_bind_parameters_in_call): Don't unshare.
        (cxx_eval_call_expression): Unshare all args if we're caching.

Added:
    trunk/gcc/testsuite/g++.dg/cpp0x/constexpr-string2.C
Modified:
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/constexpr.c