Re: [PATCH,rs6000] Add pass to optimize away xxpermdi's from vector computations

[Bill Schmidt <wschmidt at linux dot vnet dot ibm dot com>]

On Sun, 2014-08-17 at 14:52 -0400, David Edelsohn wrote:
> On Wed, Aug 13, 2014 at 7:14 PM, Bill Schmidt
> <wschmidt@linux.vnet.ibm.com> wrote:
> > Hi,
> >
> > This patch adds a PowerPC-specific pass just prior to the first cse RTL
> > pass.  The pass runs only when generating little-endian code for Power8
> > with VSX enabled, and for -O1 and up.  For this particular subtarget,
> > the use of the big-endian-biased vector load and store instructions
> > requires permutations to order vector elements for little endian.  To
> > reduce the overhead of these permutations, this pass looks for
> > computations for which the exact lanes in which computations are
> > performed does not matter, so long as the results are returned to
> > storage in the proper order.  For such computations we can remove the
> > xxpermdi's associated with the vector loads and stores.
> >
> > This patch relies on another patch posted today that converts a struct
> > used by the web pass into a base class that this patch can subclass.  If
> > it's determined that the other patch isn't appropriate, then this patch
> > will need modifications to duplicate the union-find logic.
> >
> > A complete description of the new pass appears in rs6000.c (search for
> > "Analyze vector computations").  That description also identifies some
> > remaining opportunities we can follow up with later.
> >
> > A number of new tests are added to verify that the pass works as
> > expected for some vectorized code samples.
> >
> > Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no
> > regressions.  Is this ok for trunk?
> >
> > Thanks,
> > Bill
> >
> >
> > [gcc]
> >
> > 2014-08-13  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
> >
> >         * config/rs6000/rs6000.c (context.h): New include.
> >         (tree-pass.h): Likewise.
> >         (make_pass_analyze_swaps): New decl.
> >         (rs6000_option_override): Register pass_analyze_swaps.
> >         (swap_web_entry): New subsclass of web_entry_base (df.h).
> >         (special_handling_values): New enum.
> >         (union_defs): New function.
> >         (union_uses): Likewise.
> >         (insn_is_load_p): Likewise.
> >         (insn_is_store_p): Likewise.
> >         (insn_is_swap_p): Likewise.
> >         (rtx_is_swappable_p): Likewise.
> >         (insn_is_swappable_p): Likewise.
> >         (chain_purpose): New enum.
> >         (chain_contains_only_swaps): New function.
> >         (mark_swaps_for_removal): Likewise.
> >         (swap_const_vector_halves): Likewise.
> >         (adjust_subreg_index): Likewise.
> >         (permute_load): Likewise.
> >         (permute_store): Likewise.
> >         (handle_special_swappables): Likewise.
> >         (replace_swap_with_copy): Likewise.
> >         (dump_swap_insn_table): Likewise.
> >         (rs6000_analyze_swaps): Likewise.
> >         (pass_data_analyze_swaps): New pass_data.
> >         (pass_analyze_swaps): New rtl_opt_pass.
> >         (make_pass_analyze_swaps): New function.
> >         * config/rs6000/rs6000.opt (moptimize-swaps): New option.
> >
> > [gcc/testsuite]
> >
> > 2014-08-13  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
> >
> >         * gcc.target/powerpc/swaps-p8-1.c: New test.
> >         * gcc.target/powerpc/swaps-p8-2.c: New test.
> >         * gcc.target/powerpc/swaps-p8-3.c: New test.
> >         * gcc.target/powerpc/swaps-p8-4.c: New test.
> >         * gcc.target/powerpc/swaps-p8-5.c: New test.
> >         * gcc.target/powerpc/swaps-p8-6.c: New test.
> >         * gcc.target/powerpc/swaps-p8-7.c: New test.
> >         * gcc.target/powerpc/swaps-p8-8.c: New test.
> >         * gcc.target/powerpc/swaps-p8-9.c: New test.
> >         * gcc.target/powerpc/swaps-p8-10.c: New test.
> >         * gcc.target/powerpc/swaps-p8-11.c: New test.
> >         * gcc.target/powerpc/swaps-p8-12.c: New test.
> 
> This looks okay, although I was hoping that other developers with more
> DF and web experience would double check.

Thanks for the review!

> Why are you specifically gating the pass on POWER8?

The problem is introduced in POWER8 (since LE isn't supported earlier),
and I hope that this will no longer be necessary for -mcpu=power9.  If
for some reason this doesn't turn out to be the case, we'll need to make
a change at that time, I suppose...

Thanks,
Bill

> 
> Thanks, David
>