[Bug target/104581] [12 Regression] Huge compile-time regression building SPEC 2017 538.imagick_r with PGO

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104581

--- Comment #9 from rguenther at suse dot de <rguenther at suse dot de> ---
On Thu, 17 Feb 2022, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104581
> 
> --- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> Ah, you're right.
> So, can't it instead of the quadratic walk just compare DF_INSN_LUID?
> If it isn't right after df_analyze and some insns could have been added in
> between, it would need to maintain the luids somehow (perhaps e.g. in the way
> how we do it in tree-ssa-reassoc.cc, if we add insns, we must set their uid to
> either the previous or next insn's uid and then can do some IL walk, but only
> as long as the uid is the same, so unless everything in the bb changes it
> should be still cheap.

I think this "feature" needs to be better integrated with the 
mode-switching data-flow.  It's too much bolted-on and thus has
unnecessarily high complexity.

In particular this is the initial local analysis run of mode-switching
where it just computes mode_in/mode_out of a BB (but the target hooks
do not have the pass meta data available).

The first calls to the hook are from 

  if (targetm.mode_switching.entry && targetm.mode_switching.exit)
    {
      /* Split the edge from the entry block, so that we can note that
         there NORMAL_MODE is supplied.  */
      post_entry = split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN 
(cfun)));
      pre_exit = create_pre_exit (n_entities, entity_map, num_modes);

which even runs before df_analyze () so if the DF walk would be
triggered there it definitely looks bogus.  I also see nothing
adding the chain problem (I dn't remember which ones are added by
default though).