[PING] Re: [PATCH, ARM]: rewrite NEON vector operation intrinsics without UNSPECs

[Sandra Loosemore]

Ping?

http://gcc.gnu.org/ml/gcc-patches/2010-05/msg02262.html

-Sandra


> 2010-05-28  Sandra Loosemore  <sandra@codesourcery.com>
> 
>     gcc/
>     * config/arm/arm.c (neon_vdup_constant): Expand into canonical RTL
>     instead of an unspec.
>     (neon_expand_vector_init): Likewise.
>     * config/arm/neon.md (UNSPEC_VCOMBINE): Delete.
>     (UNSPEC_VDUP_LANE): Delete.
>     (UNSPEC VDUP_N): Delete.
>     (UNSPEC_VGET_HIGH): Delete.
>     (UNSPEC_VGET_LANE): Delete.
>     (UNSPEC_VGET_LOW): Delete.
>     (UNSPEC_VMVN): Delete.
>     (UNSPEC_VSET_LANE): Delete.
>     (V_double_vector_mode): New.
>     (vec_set<mode>_internal): Make code emitted match that for the
>     corresponding intrinsics.
>     (vec_setv2di_internal): Likewise.
>     (vec_extractv2di): Correct error in register numbering.
>     (neon_vget_lanedi): Rewrite to expand into emit_move_insn.
>     (neon_vget_lanev2di): Rewrite to expand into vec_extractv2di.
>     (neon_vset_lane<mode>): Combine double and quad patterns and
>     expand into vec_set<mode>_internal instead of UNSPEC_VSET_LANE.
>     (neon_vset_lanedi): Rewrite to expand into emit_move_insn.
>     (neon_vdup_n<mode>): Rewrite RTL without unspec.
>     (neon_vdup_ndi): Rewrite as define_expand and use emit_move_insn.
>     (neon_vdup_nv2di): Rewrite RTL without unspec and merge with
>     with neon_vdup_lanev2di, adjusting the pattern from the latter
>     to be predicable for consistency.
>     (neon_vdup_lane<mode>_internal): New.
>     (neon_vdup_lane<mode>): Turn into a define_expand and rewrite
>     to avoid using an unspec.
>     (neon_vdup_lanedi): Rewrite RTL pattern to avoid unspec.
>     (neon_vdup_lanev2di): Turn into a define_expand.
>     (neon_vcombine): Rewrite pattern to eliminate UNPSEC_VCOMBINE.
>     (neon_vget_high<mode>): Replace with....
>     (neon_vget_highv16qi): New pattern using canonical RTL.
>     (neon_vget_highv8hi): Likewise.
>     (neon_vget_highv4si): Likewise.
>     (neon_vget_highv4sf): Likewise.
>     (neon_vget_highv2di): Likewise.
>     (neon_vget_low<mode>): Replace with....
>     (neon_vget_lowv16qi): New pattern using canonical RTL.
>     (neon_vget_lowv8hi): Likewise.
>     (neon_vget_lowv4si): Likewise.
>     (neon_vget_lowv4sf): Likewise.
>     (neon_vget_lowv2di): Likewise.
> 
>     * config/arm/neon.ml (Vget_lane): Add No_op attribute to suppress
>     test for this emitting vmov.
>     (Vset_lane): Likewise.
>     (Vdup_n): Likewise.
>     (Vmov_n): Likewise.
> 
>     * doc/arm-neon-intrinsics.texi: Regenerated.
> 
>     gcc/testsuite/
>     * gcc.target/arm/neon/vdup_ns64.c: Regenerated.
>     * gcc.target/arm/neon/vdup_nu64.c: Regenerated.
>     * gcc.target/arm/neon/vdupQ_ns64.c: Regenerated.
>     * gcc.target/arm/neon/vdupQ_nu64.c: Regenerated.
>     * gcc.target/arm/neon/vmov_ns64.c: Regenerated.
>     * gcc.target/arm/neon/vmov_nu64.c: Regenerated.
>     * gcc.target/arm/neon/vmovQ_ns64.c: Regenerated.
>     * gcc.target/arm/neon/vmovQ_nu64.c: Regenerated.
>     * gcc.target/arm/neon/vget_lanes64.c: Regenerated.
>     * gcc.target/arm/neon/vget_laneu64.c: Regenerated.
>     * gcc.target/arm/neon/vset_lanes64.c: Regenerated.
>     * gcc.target/arm/neon/vset_laneu64.c: Regenerated.
>     * gcc.target/arm/neon-vdup_ns64.c: New.
>     * gcc.target/arm/neon-vdup_nu64.c: New.
>     * gcc.target/arm/neon-vdupQ_ns64.c: New.
>     * gcc.target/arm/neon-vdupQ_nu64.c: New.
>     * gcc.target/arm/neon-vdupQ_lanes64.c: New.
>     * gcc.target/arm/neon-vdupQ_laneu64.c: New.
>     * gcc.target/arm/neon-vmov_ns64.c: New.
>     * gcc.target/arm/neon-vmov_nu64.c: New.
>     * gcc.target/arm/neon-vmovQ_ns64.c: New.
>     * gcc.target/arm/neon-vmovQ_nu64.c: New.
>     * gcc.target/arm/neon-vget_lanes64.c: New.
>     * gcc.target/arm/neon-vget_laneu64.c: New.
>     * gcc.target/arm/neon-vset_lanes64.c: New.
>     * gcc.target/arm/neon-vset_laneu64.c: New.