Re: [PATCH 3/14] Add new optabs for reducing vectors to scalars

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

On Thu, Sep 18, 2014 at 1:54 PM, Alan Lawrence <alan.lawrence@arm.com> wrote:
> These match their corresponding tree codes, by taking a vector and returning
> a scalar; this is more architecturally neutral than the (somewhat loosely
> defined) previous optab that took a vector and returned a vector with the
> result in the least significant bits (i.e. element 0 for little-endian or
> N-1 for bigendian). However, the old optabs are preserved so as not to break
> existing backends, so clients check for both old + new optabs.
>
> Bootstrap, check-gcc and check-g++ on x86_64-none-linux-gnu.
> aarch64.exp + vect.exp on aarch64{,_be}-none-elf.
> (of course at this point in the series all these are using the old optab +
> migration path.)

scalar_reduc_to_vector misses a comment.

I wonder if at the end we wouldn't transition all backends and then
renaming reduc_*_scal_optab back to reduc_*_optab makes sense.

The optabs have only one mode - I wouldn't be surprised if an ISA
invents for example v4si -> di reduction?  So do we want to make
reduc_plus_scal_optab a little bit more future proof (maybe there
is already an ISA that supports this kind of reduction?).

Otherwise the patch looks good to me.

Thanks,
Richard.

> gcc/ChangeLog:
>
>         * doc/md.texi (Standard Names): Add reduc_(plus,[us](min|max))|scal
>         optabs, and note in reduc_[us](plus|min|max) to prefer the former.
>
>         * expr.c (expand_expr_real_2): Use reduc_..._scal if available, fall
>         back to old reduc_... + BIT_FIELD_REF only if not.
>
>         * optabs.c (optab_for_tree_code): for REDUC_(MAX,MIN,PLUS)_EXPR,
>         return the reduce-to-scalar (reduc_..._scal) optab.
>         (scalar_reduc_to_vector): New.
>
>         * optabs.def (reduc_smax_scal_optab, reduc_smin_scal_optab,
>         reduc_plus_scal_optab, reduc_umax_scal_optab,
> reduc_umin_scal_optab):
>         New.
>
>         * optabs.h (scalar_reduc_to_vector): Declare.
>
>         * tree-vect-loop.c (vectorizable_reduction): Look for optabs
> reducing
>         to either scalar or vector.