[Bug middle-end/69156] [6 Regression] gcc.dg/pr59630.c: ICE: verify_gimple failed: non-trivial conversion at assignment with -fno-tree-ccp

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69156

--- Comment #5 from rguenther at suse dot de <rguenther at suse dot de> ---
On Fri, 8 Jan 2016, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69156
> 
> --- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> (In reply to Richard Biener from comment #3)
> > (In reply to Jakub Jelinek from comment #2)
> > > Created attachment 37266 [details]
> > > gcc6-pr69156.patch
> > > 
> > > Untested fix.  Not sure if it is the best spot though.  And the testcase is
> > > really nasty.
> > 
> > Hmm, but this means things were mismatched in the first place (earlier) and
> > thus this should be checked somewhere else?
> 
> This happens in gimple-match-head.c:
> 714                   tree decl = TREE_OPERAND (fn, 0);
> 715                   if (DECL_BUILT_IN_CLASS (decl) != BUILT_IN_NORMAL
> 716                       || !gimple_builtin_call_types_compatible_p (stmt,
> decl))
> 717                     return false;
> and the problem is that gimple_builtin_call_types_compatible_p is extremely
> lose (basically, for return type or arguments considers all integral types
> equal, or all pointer types equal, or all RECORD_TYPEs equal etc.).
> Bet for this purpose (but don't about all the others) it might be better to
> check useless_type_conversion_p equality instead.
> I guess the question is when is a mere validate_type instead of
> types_compatible_p useful.

Ick.  That's the code "copied" from builtins.c.  I suppose we should
simply exchange the validate_type calls with useless_type_conversion_p ()
(in the correct order).  That should be coarse enough these days.