Re: redundant bitmap_bit_p followed by bitmap_clear_bit [was: Re: [COMMITTED] Kill second order relations in the path solver.]

[Richard Sandiford <richard.sandiford@arm.com>]

Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> On Mon, Nov 1, 2021 at 10:02 PM Bernhard Reutner-Fischer via
> Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
>>
>> On Mon, 1 Nov 2021 15:21:03 +0100
>> Aldy Hernandez <aldyh@redhat.com> wrote:
>>
>> > I'm not convinced this makes the code clearer to read, especially if
>> > it's not on a critical path.  But if you feel strongly, please submit
>> > a patch ;-).
>>
>> No i don't feel strongly about it.
>> Compiling e.g. -O2 ira.o
>> # Overhead       Samples  Command  Shared Object  Symbol
>> # ........  ............  .......  .............  .........................
>> #
>>    100.00%          4197  cc1plus  cc1plus        [.] mark_reachable_blocks
>>    100.00%         22000  cc1plus  cc1plus        [.] path_oracle::killing_def
>> and the mark_elimination is reload.
>> So it's not just a handful of calls saved but some. And it would be
>> smaller code as it saves a call. Well maybe another day.
>
> Note that single bit set/clear are already implemented as test && set/clear.
> Note that unfortunately the sbitmap bitmap_set/clear_bit overloads do not
> return the previous state of the bit.

+1 that it would good if the sbitmap versions behaved like the bitmap
versions.  Always low-key bothered me that they didn't do this when I
hit it, but never got round to do anything about it...

Bitmap operations consistently show up high in the profiles, so IMO
using the return value of bitmap_set_bit and bitmap_clear_bit should be
the preferred style.  Not all uses are performance-critical, of course,
but code tends to get copied around.  Having all code do it the fast
way reduces the risk that slow code gets copied to code that needs
to be fast. :-)

> Maybe providing bitmap_test_and_set_bit () and
> bitmap_test_and_clear_bit () would be better (but note we currently
> return true when the bit changed, not when it was set).

Yeah, maybe it's just familiarity, but I find the:

  if (bitmap_set_bit (bb, i))
    ...something changed...

thing easier to follow than:

  if (!bitmap_test_and_set_bit (bb, i))
    ...something changed...

Thanks,
Richard