On August 21, 2015 6:42:15 PM GMT+02:00, Marek Polacek <pola...@redhat.com>
wrote:
>On Fri, Aug 21, 2015 at 06:23:09PM +0200, Richard Biener wrote:
>> >> Yes, but gimple_call_noreturn_p is false on __builtin_trap.
>That's
>> >quite
>> >> confusing... but flags_from_decl_or_type really returns 0 for
>> >__builtin_trap.
>> >Well, if that's intentional (and offhand I have no idea if it is),
>then
>> >
>> >you could emit a __builtin_trap followed by a __builtin_unreachable.
>>
>> But maybe go is non-call-exceptions and that makes a difference?
>
>I suppose that's the case. In Makefile.am I see
>AM_CFLAGS = -fexceptions -fnon-call-exceptions -fplan9-extensions
>But I'm not clear on how this could make a difference wrt whether
>__builtin_trap is volatile (thus ECF_NORETURN).
Not sure either. Needs to be investigated. BTW, built-in trap is also nothrow
AFAIR.
Richard.
>This is a patch to also generate __builtin_unreachable, just for the
>record.
>
>Bootstrapped/regtested on x86_64-linux.
>
>2015-08-21 Marek Polacek <pola...@redhat.com>
>
> PR tree-optimization/67284
> * gimple-ssa-isolate-paths.c (insert_trap): Also emit
> __builtin_unreachable.
>
>diff --git gcc/gimple-ssa-isolate-paths.c
>gcc/gimple-ssa-isolate-paths.c
>index ca2322d..06263d3 100644
>--- gcc/gimple-ssa-isolate-paths.c
>+++ gcc/gimple-ssa-isolate-paths.c
>@@ -97,15 +97,18 @@ insert_trap (gimple_stmt_iterator *si_p, tree op)
> = gimple_build_call (builtin_decl_explicit (BUILT_IN_TRAP), 0);
> gimple_seq seq = NULL;
> gimple_seq_add_stmt (&seq, new_stmt);
>+ new_stmt
>+ = gimple_build_call (builtin_decl_explicit
>(BUILT_IN_UNREACHABLE), 0);
>+ gimple_seq_add_stmt (&seq, new_stmt);
>
> /* If we had a NULL pointer dereference, then we want to insert the
>- __builtin_trap after the statement, for the other cases we want
>- to insert before the statement. */
>+ __builtin_trap/__builtin_unreachable pair after the statement,
>for
>+ the other cases we want to insert before the statement. */
> if (walk_stmt_load_store_ops (stmt, (void *)op,
> check_loadstore,
> check_loadstore))
> {
>- gsi_insert_after (si_p, seq, GSI_NEW_STMT);
>+ gsi_insert_seq_after (si_p, seq, GSI_NEW_STMT);
> if (stmt_ends_bb_p (stmt))
> {
> split_block (gimple_bb (stmt), stmt);
>@@ -113,7 +116,7 @@ insert_trap (gimple_stmt_iterator *si_p, tree op)
> }
> }
> else
>- gsi_insert_before (si_p, seq, GSI_NEW_STMT);
>+ gsi_insert_seq_before (si_p, seq, GSI_NEW_STMT);
>
> split_block (gimple_bb (new_stmt), new_stmt);
> *si_p = gsi_for_stmt (stmt);
>@@ -194,7 +197,7 @@ isolate_path (basic_block bb, basic_block
>duplicate,
>
>/* If we did not run to the end of DUPLICATE, then SI points to STMT
>and
> SI2 points to the duplicate of STMT in DUPLICATE. Insert a trap
>- before SI2 and remove SI2 and all trailing statements. */
>+ before SI2 and split the block. */
> if (!gsi_end_p (si2))
> {
> if (ret_zero)
>
> Marek
