Re: [patch][cprop.c] Clean up hash table building

[Richard Guenther <richard dot guenther at gmail dot com>]

On Thu, Mar 31, 2011 at 12:39 AM, Steven Bosscher <stevenb.gcc@gmail.com> wrote:
> Hi,
>
> This is the first cleanup for cprop.c. These cleanups are only
> possible now, thanks to splitting the CPROP code out from gcse.c
>
> The first change is to not treat unfolded conditions as constant in
> gcse_constant_p(). This never happens because:
> - during hash table building any such condition should have been
> simplified by any prior pass
> - during cprop the expression table is not updated because the
> patterns are unshared
>
> This causes no changes in code generation for all cc1-i files on
> x86_64-unknown-linux-gnu. I could add an assert to make sure this kind
> of condition really never happens, but IMHO it's not worth it and the
> bug would be elsewhere anyway.
>
> The second change is to remove oprs_available_p(). After the
> gcse_constant_p() cleanup, we do not have to traverse the whole
> pattern of a SET candidate for CPROP, because we know that SET_DEST is
> a REG and SET_SRC is a REG or a shareable constant. Therefore a simple
> check to see if the registers involved were set in the block is
> sufficient, and a regset can be used instead of the last_set table.
> See reg_available_p().
>
> The third change is to compute_hash_table_work(), which now traverses
> the insns in each basic block just once (instead of twice), in reverse
> order to record all registers set between BB_END and the current insn.
> This change allows further simplify hash_scan_set() which now doesn't
> have to look at INSN+1 anymore (saving another half insns stream
> traversal by avoiding next_nonnote_nondebug_insn()).
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu. OK?

Ok.

Thanks,
Richard.

> Ciao!
> Steven
>