Re: [PATCH] rs6000: Fix huge compile time on thunks (PR64580)

[David Edelsohn <dje dot gcc at gmail dot com>]

On Thu, Jan 29, 2015 at 7:38 PM, Segher Boessenkool
<segher@kernel.crashing.org> wrote:
> The problem in this PR is that rs6000_stack_info spends an enormous amount
> of time when compiling thunks.  This turns out to be a loop in
> compute_vrsave_mask that loops (almost) 4G times because the loop counter
> is unsigned and the reversed loop is not written in a way that can handle
> crtl->args.info.vregno == 0.  Make it a forward loop free of such traps,
> which is easier to read anyway.
>
> But, vregno should never be 0 if rs6000_stack_info is called; it means
> that various data is not initialised for this function.  Which is correct,
> this is a thunk: rs6000_stack_info should not be called for thunks.  Add
> an assert for that.
>
> rs6000_output_function_prologue (the thing outputting asm stuff) calls
> it though.  Don't do the savres thing for thunks (they never need it).
> Factor the relevant code to a new function while at it.
>
> rs6000_output_function_epilogue already guards properly with is_thunk,
> and the assert didn't fire on a bootstrap+testrun, so it seems this was
> the only place the mistake happened.
>
>
> Bootstrapped and tested as usual; okay for mainline?  And release branches
> if it happens there (the PR doesn't say yet, it's not marked as regression
> either, looking at history it seems to go way way back)?
>
>
> Segher
>
>
> 2015-01-29  Segher Boessenkool  <segher@kernel.crashing.org>
>
> gcc/
>         PR target/64580
>         * config.rs6000/rs6000.c (compute_vrsave_mask): Reverse loop order.
>         (rs6000_stack_info): Add assert.
>         (rs6000_output_savres_externs): New function, split off from...
>         (rs6000_output_function_prologue): ... here.  Do not call it for
>         thunks.

Okay.

Thanks, David