[Bug c++/82115] [8 Regression] ICE on (valid) C++11 code: Segmentation fault signal terminated program cc1plus

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82115

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jason at gcc dot gnu.org,
                   |                            |nathan at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Anyway, if value_dependent_expression_p needs to recurse on DECL_INITIAL that
can contain arbitrary stuff, including the VAR_DECL with that DECL_INITIAL in
it, we need to avoid the recursion.
"it is a constant with literal type and is initialized with an expression that
is value-dependent."
is what applies here.
So do we need some hash-map that will track VAR_DECLs  on whose DECL_INITIAL
we've already recursed?  The problem is that value_dependent_expression_p calls
type_dependent_expression_p and vice versa, so it is unclear when that hash-map
should be saved/restored.  Even if it is not possible to reach similar infinite
recursion through both of those functions (i.e. when we could save/restore the
hash-map in "toplevel" value_dependent_expression_p call and use
value_dependent_expression_p_1 recursing to itself from it, another question is
how to do that efficiently; how common are VAR_DECLs on which we recurse on
DECL_INITIAL?  How many there are on average during one top-level
value_dependent_expression_p call?  E.g. if the common cases are 0 or 1 times,
perhaps we could have next to the hash-map a single tree which we'd compare and
only create hash-map if seeing another one.