Re: [PATCH] Fix devirtualization ICE (PR tree-optimization/59622)

[Jakub Jelinek <jakub at redhat dot com>]

On Mon, Dec 30, 2013 at 05:56:04PM +0100, Richard Biener wrote:
> Jakub Jelinek <jakub@redhat.com> wrote:
> >Setting gimple_call_set_fndecl of a random method call to
> >__bultin_unreachable can't work properly, the method call probably
> >passes in arguments that __builtin_unreachable doesn't expect, and if
> >the method call has LHS, we'll ICE, because __builtin_unreachable
> >doesn't
> >have a LHS.  Fixed by not doing anything for fold_stmt_inplace and
> >for fold_stmt just adding a new __buitin_unreachable () call before the
> >method call, DCE will DTRT then and we don't have to bother with adding
> >some dummy stmt to set up the lhs etc.
> >
> >Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
> 
> You could re-set the number of arguments and the LHS though.

While we have gimple_set_num_ops, I'd find that very ugly and apparently
nothing in the compiler does that for calls right now.

> But with a LHS and a not dead consumer both approaches will probably ice.
> You'd need to replace all uses with a default def.

Also, I think it could be quite surprising if fold_stmt changed other stmts
(users of lhs), plus the lhs doesn't have to be in SSA form yet.
One would need to also unlink vdefs and release the vdef SSA_NAME, and not
sure if fold_stmt_inplace callers would be even prepared for the stmt to
become noreturn, and no longer throwing, etc.

What would be the advantages of doing all that over the posted patch?

	Jakub