[Bug middle-end/59630] [4.7/4.8/4.9 Regression] ICE converting the return type of a builtin function

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59630

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with my r182761 but say:
_Bool foo (int x)
{
  _Bool (*f)(int) = __builtin_abs;
  return f(x);
}
ICEs at -O2 already since the gimple_call_fntype introduction in r172310.
Since the r172310 change we also without any warnings happily inline:
inline int fn1 (int x, int y) { return x + y; }
int
fn2 (void)
{
  int (*fn) (long long) = (void *) fn1;
  return fn (5LL);
}
inline int fn3 (long long x) { return x; }
int
fn4 (void)
{
  int (*fn) (int x, int y) = (void *) fn3;
  return fn (5, 6);
}
(well, for fn1 into fn2 we complain about uninitialized y).
This is also related to PR59622.  Even with gimple_call_fntype, I'd say we
should never propagate an indirect call into a gimple_call_fndecl if it doesn't
satisfy some minimal consistency predicate (something along the lines of what I
wrote in PR59622, basically if the stmt has lhs, then the return types need to
be compatible, and TYPE_ARG_TYPES should be compatible (perhaps an exception
can be made for the C () vs. (void) arguments).  Because if the conversion
between fn return type and/or arguments types is non-useless, I'm afraid tons
of issues can happen everywhere.  At minimum we should disallow it if the
fndecl we want to propagate is a builtin, and we should not inline it, or
perhaps clone.