On Thu, Dec 11, 2014 at 06:06:55PM +0100, Jan Hubicka wrote:
> Hi,
> in firefox .optimized dumps one can see few places where __builtin_unreachable
> is called (as a result of devirtualization code proving the code path to be
> undefined). There is usually some argument setup for the parameters of
> __builtin_unreachable that are dead. This patch makes it somewhat better
> so now we get:
> <bb 30>:
>
> # prephitmp_222 = PHI <_52(27), pretmp_245(29)>
>
> _57 = prephitmp_222 + 2;
>
> pool_40(D)->ptr = _57;
>
> __builtin_unreachable ();
>
>
> Why DSE does not eliminate the stores prior noreturn const function?
>
> Bootstrapped/regtested x86_64-linux, OK?
>
> Honza
> * tree-ssa-dce.c (mark_stmt_if_obviously_necessary): Remove dead
> parameters
> of BUILT_IN_UNREACHABLE
Shouldn't this be done when you actually change the call to
__builtin_unreachable ()? I mean, __builtin_unreachable () has no
arguments, so leaving any arguments there is broken IL, even if you clean it
up during the next DCE.
> --- tree-ssa-dce.c (revision 218610)
> +++ tree-ssa-dce.c (working copy)
> @@ -250,6 +250,15 @@ mark_stmt_if_obviously_necessary (gimple
> case BUILT_IN_ALLOCA:
> case BUILT_IN_ALLOCA_WITH_ALIGN:
> return;
> + case BUILT_IN_UNREACHABLE:
> + /* All parameters of BUILT_IN_UNREACHABLE are dead. Remove them
> + from the stmt, so we can remove their definitions. */
> + if (gimple_call_num_args (stmt))
> + {
> + gimple_set_num_ops (stmt, 3);
> + update_stmt (stmt);
> + }
> + break;
>
> default:;
> }
Jakub
