| Message ID | 20180802165837.23379-1-siddhesh@sourceware.org |
|---|---|
| State | New |
| Headers | show |
| Series | [aarch64,v2] Fix falkor pipeline description for dup<q> | expand |
On Thu, Aug 02, 2018 at 11:58:37AM -0500, Siddhesh Poyarekar wrote: > There was a typo in the pipeline description where DUP was assigned to > the vector pipes for quad mode ops when it really only uses the VTOG > pipes. Fixing this does not show any noticeable difference in > performance (there's a very small bump of 1.7% in x264 but that's > about it) in my tests but is the more precise description of operations > for falkor. > > Bootstrapped and tested with --with-cpu=falkor to confirm that there > are no regressions. OK. Thanks, James > > Siddhesh > > Changes from v1: > > - Fixed up ChangeLog entry > > - Replaced falkor_gtov,falkor_gtov with falkor_gtov*2 > > * config/aarch64/falkor.md (falkor_am_1_vxvy_vxvy): Move > neon_dup_q to... > (falkor_am_1_gtov_gtov): ... a new insn reservation. >
On 08/03/2018 12:02 AM, James Greenhalgh wrote: > On Thu, Aug 02, 2018 at 11:58:37AM -0500, Siddhesh Poyarekar wrote: >> There was a typo in the pipeline description where DUP was assigned to >> the vector pipes for quad mode ops when it really only uses the VTOG >> pipes. Fixing this does not show any noticeable difference in >> performance (there's a very small bump of 1.7% in x264 but that's >> about it) in my tests but is the more precise description of operations >> for falkor. >> >> Bootstrapped and tested with --with-cpu=falkor to confirm that there >> are no regressions. > > OK. Thanks, may I backport this to gcc 8.x as well given that it has zero impact on anything except falkor? We are targeting gcc 8.x as the toolchain for falkor and so would like to at least get all falkor specific optimizations and fixes back into the gcc 8.x branch. Siddhesh
On Thu, Aug 02, 2018 at 09:02:05PM -0500, Siddhesh Poyarekar wrote: > On 08/03/2018 12:02 AM, James Greenhalgh wrote: > > On Thu, Aug 02, 2018 at 11:58:37AM -0500, Siddhesh Poyarekar wrote: > >> There was a typo in the pipeline description where DUP was assigned to > >> the vector pipes for quad mode ops when it really only uses the VTOG > >> pipes. Fixing this does not show any noticeable difference in > >> performance (there's a very small bump of 1.7% in x264 but that's > >> about it) in my tests but is the more precise description of operations > >> for falkor. > >> > >> Bootstrapped and tested with --with-cpu=falkor to confirm that there > >> are no regressions. > > > > OK. > > Thanks, may I backport this to gcc 8.x as well given that it has zero > impact on anything except falkor? We are targeting gcc 8.x as the > toolchain for falkor and so would like to at least get all falkor > specific optimizations and fixes back into the gcc 8.x branch. This one is OK for backport. Thanks, James
diff --git a/gcc/config/aarch64/falkor.md b/gcc/config/aarch64/falkor.md index 5cbc5150ef3..45cbff93b24 100644 --- a/gcc/config/aarch64/falkor.md +++ b/gcc/config/aarch64/falkor.md @@ -322,6 +322,12 @@ (eq_attr "type" "neon_from_gp_q")) "falkor_gtov,falkor_gtov") +;; DUP does not use vector pipes in Q mode, only gtov+gtov. +(define_insn_reservation "falkor_am_1_gtov_gtov" 1 + (and (eq_attr "tune" "falkor") + (eq_attr "type" "neon_dup_q")) + "falkor_gtov*2") + ;; neon_to_gp_q is used for 32-bit ARM instructions that move 64-bits of data ;; so no use needed here. @@ -337,7 +343,7 @@ (define_insn_reservation "falkor_am_1_vxvy_vxvy" 1 (and (eq_attr "tune" "falkor") - (eq_attr "type" "neon_bsl_q,neon_dup_q,neon_ext_q,neon_move_q,neon_rev_q,neon_tbl1_q,neon_permute_q")) + (eq_attr "type" "neon_bsl_q,neon_ext_q,neon_move_q,neon_rev_q,neon_tbl1_q,neon_permute_q")) "falkor_vxvy+falkor_vxvy") (define_insn_reservation "falkor_am_2_vxvy" 2