On Thu, Dec 19, 2013 at 1:40 PM, Matt Turner <matts...@gmail.com> wrote:
> total instructions in shared programs: 1520829 -> 1511991 (-0.58%)
> instructions in affected programs: 559725 -> 550887 (-1.58%)
> GAINED: 6
> LOST: 9
> ---
> src/mesa/drivers/dri/i965/Makefile.sources | 1 +
> src/mesa/drivers/dri/i965/brw_fs.cpp | 37 +-
> src/mesa/drivers/dri/i965/brw_fs.h | 1 +
> .../drivers/dri/i965/brw_fs_value_numbering.cpp | 398
> +++++++++++++++++++++
> 4 files changed, 435 insertions(+), 2 deletions(-)
> create mode 100644 src/mesa/drivers/dri/i965/brw_fs_value_numbering.cpp
>
> diff --git a/src/mesa/drivers/dri/i965/Makefile.sources
> b/src/mesa/drivers/dri/i965/Makefile.sources
> index 43f152e..36d8261 100644
> --- a/src/mesa/drivers/dri/i965/Makefile.sources
> +++ b/src/mesa/drivers/dri/i965/Makefile.sources
> @@ -63,6 +63,7 @@ i965_FILES = \
> brw_fs_peephole_predicated_break.cpp \
> brw_fs_reg_allocate.cpp \
> brw_fs_sel_peephole.cpp \
> + brw_fs_value_numbering.cpp \
> brw_fs_vector_splitting.cpp \
> brw_fs_visitor.cpp \
> brw_gs.c \
> diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp
> b/src/mesa/drivers/dri/i965/brw_fs.cpp
> index e4a4737..08837da 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
> @@ -3285,9 +3285,11 @@ fs_visitor::run()
> remove_dead_constants();
> setup_pull_constants();
>
> - bool progress;
> + bool progress, cp_progress, vn_progress;
> do {
> progress = false;
> + cp_progress = false;
> + vn_progress = false;
>
> compact_virtual_grfs();
>
> @@ -3295,16 +3297,47 @@ fs_visitor::run()
>
> progress = opt_algebraic() || progress;
> progress = opt_cse() || progress;
> - progress = opt_copy_propagate() || progress;
> progress = opt_peephole_predicated_break() || progress;
> + cp_progress = opt_copy_propagate();
There is a tab here, according to:
git show mattst88/register-coalesce~5|grep -P "\+(\s*\t\s*)"
> progress = dead_code_eliminate() || progress;
> progress = dead_code_eliminate_local() || progress;
> progress = opt_peephole_sel() || progress;
> progress = dead_control_flow_eliminate(this) || progress;
> + vn_progress = opt_value_numbering();
Could something like this help with the fighting, and remove the need
for the 'fixup' block below?
if (progress || !cp_progress)
vn_progress = opt_value_numbering();
-Jordan
> progress = register_coalesce() || progress;
> progress = compute_to_mrf() || progress;
> +
> + /* Value numbering rewrites
> + *
> + * MOV g5, 1.0F
> + * MOV g6, 1.0F
> + *
> + * into
> + *
> + * MOV g5, 1.0F
> + * MOV g6, g5
> + *
> + * but constant propagation undoes this. If these were the only two
> + * passes to make progress, they're just fighting.
> + */
> + if (!progress && (cp_progress != vn_progress)) {
> + progress = true;
> + }
> } while (progress);
>
> + /* Copy propagation and value numbering were fighting, which means that
> + * the last changes made by value numbering weren't useful, so rerun
> + * copy propagation and the rest of the optimizations after it, modulo
> + * value numbering.
> + */
> + if (cp_progress && vn_progress) {
> + opt_copy_propagate();
> + dead_code_eliminate();
> + dead_code_eliminate_local();
> + register_coalesce();
> + compute_to_mrf();
> + }
> +
> lower_uniform_pull_constant_loads();
>
> assign_curb_setup();
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
