On Sat, Feb 23, 2019 at 10:05:40AM +0100, Jakub Jelinek wrote:
> And finally, for 1) the patch has a case for a dead .ABNORMAL_DISPATCHER,
> one that has only incoming edges but no successors. I didn't want to
> complicate or slow down the processing too much, so it is actually done
> only if .ABNORMAL_DISPATCHER has no outgoing edges, rather than if it has
> some, but all of them are in the end to non-visited basic blocks.
> This means when we for 5) or 4) above remove all such bbs as dead in one
> cfg cleanup, we remove .ABNORMAL_DISPATCHER only in the following cfg
> cleanup.
Thinking about the above, I think we could handle that cheaply in the same
cleanup cfg rather than in the next one. Don't know right now if we can
have at most one .ABNORMAL_DISPATCHER in a function or more (I vaguely
remember we disallow inlining in some of the ab cases like setjmp,
non-local goto etc.), if there can be at most one, we could just make
the .ABNORMAL_DISPATCHER visited normally, but additionally note it in some
basic_block pointer, if we can have more than one, we could have auto_vec of
them, and then after the loop just go through that one or more
.ABNORMAL_DISPATCHER bbs and if they don't have any successors visited,
unset the visited bit for them as well.
Shall I modify the patch for that?
> 2019-02-23 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/89280
> * tree-cfgcleanup.c (maybe_dead_abnormal_edge_p,
> builtin_setjmp_setup_bb): New functions.
> (cleanup_control_flow_pre): Ignore maybe_dead_abnormal_edge_p edges.
> When visiting __builtin_setjmp_setup block, queue in special
> setjmp_vec vector edges from .ABNORMAL_DISPATCHER to corresponding
> __builtin_setjmp_receiver.
>
> * gcc.c-torture/compile/pr89280.c: New test.
> * gcc.dg/torture/pr57147-2.c: Don't expect a setjmp after noreturn
> function. Skip the test for -O0.
Jakub
