Re: [PATCH qemu v14 00/15] Add tail agnostic behavior for rvv instructions

[Alistair Francis]

On Tue, May 3, 2022 at 10:20 AM ~eopxd <eopxd@git.sr.ht> wrote:
>
> According to v-spec, tail agnostic behavior can be either kept as
> undisturbed or set elements' bits to all 1s. To distinguish the
> difference of tail policies, QEMU should be able to simulate the tail
> agnostic behavior as "set tail elements' bits to all 1s". An option
> 'rvv_ta_all_1s' is added to enable the behavior, it is default as
> disabled.
>
> There are multiple possibility for agnostic elements according to
> v-spec. The main intent of this patch-set tries to add option that
> can distinguish between tail policies. Setting agnostic elements to
> all 1s makes things simple and allow QEMU to express this.
>
> We may explore other possibility of agnostic behavior by adding
> other options in the future. Please understand that this patch-set
> is limited.
>
> v2 updates:
> - Addressed comments from Weiwei Li
> - Added commit tail agnostic on load / store instructions (which
>   I forgot to include into the patch-set)
>
> v3 updates:
> - Missed the very 1st commit, adding it back
>
> v4 updates:
> - Renamed vlmax to total_elems
> - Deal with tail element when vl_eq_vlmax == true
>
> v5 updates:
> - Let `vext_get_total_elems` take `desc` and `esz`
> - Utilize `simd_maxsz(desc)` to get `vlenb`
> - Fix alignments to code
>
> v6 updates:
> - Fix `vext_get_total_elems`
>
> v7 updates:
> - Reuse `max_elems` for vector load / store helper functions. The
>   translation sets desc's `lmul` to `min(1, lmul)`, making
>   `vext_max_elems` equivalent to `vext_get_total_elems`.
>
> v8 updates:
> - Simplify `vext_set_elems_1s`, don't need `vext_set_elems_1s_fns`
> - Fix `vext_get_total_elems`, it should derive upon EMUL instead
>   of LMUL
>
> v9 updates:
> - Let instructions that is tail agnostic regardless of vta respect the
>   option and not the vta.
>
> v10 updates:
> - Correct range to set element to 1s for load instructions
>
> v11 updates:
> - Separate addition of option 'rvv_ta_all_1s' as a new (last) commit
> - Add description to show intent of the option in first commit for the
>   optional tail agnostic behavior
> - Tag WeiWei as Reviewed-by for all commits
> - Tag Alistair as Reviewed-by for commit 01, 02
> - Tag Alistair as Acked-by for commit 03
>
> v12 updates:
> - Add missing space in WeiWei's "Reviewed-by" tag
>
> v13 updates:
> - Fix tail agnostic for vext_ldst_us. The function operates on input
>   parameter 'evl' rather than 'env->vl'.
> - Fix tail elements for vector segment load / store instructions
>   A vector segment load / store instruction may contain fractional
>   lmul with nf * lmul > 1. The rest of the elements in the last
>   register should be treated as tail elements.
> - Fix tail agnostic length for instructions with mask destination
>   register. Instructions with mask destination register should have
>   'vlen - vl' tail elements.
>
> v14 updates:
> - Pass lmul information to into vector helper function.
>   `vext_get_total_elems` needs it.
>
> eopXD (15):
>   target/riscv: rvv: Prune redundant ESZ, DSZ parameter passed
>   target/riscv: rvv: Rename ambiguous esz
>   target/riscv: rvv: Early exit when vstart >= vl
>   target/riscv: rvv: Add tail agnostic for vv instructions
>   target/riscv: rvv: Add tail agnostic for vector load / store
>     instructions
>   target/riscv: rvv: Add tail agnostic for vx, vvm, vxm instructions
>   target/riscv: rvv: Add tail agnostic for vector integer shift
>     instructions
>   target/riscv: rvv: Add tail agnostic for vector integer comparison
>     instructions
>   target/riscv: rvv: Add tail agnostic for vector integer merge and move
>     instructions
>   target/riscv: rvv: Add tail agnostic for vector fix-point arithmetic
>     instructions
>   target/riscv: rvv: Add tail agnostic for vector floating-point
>     instructions
>   target/riscv: rvv: Add tail agnostic for vector reduction instructions
>   target/riscv: rvv: Add tail agnostic for vector mask instructions
>   target/riscv: rvv: Add tail agnostic for vector permutation
>     instructions
>   target/riscv: rvv: Add option 'rvv_ta_all_1s' to enable optional tail
>     agnostic behavior

Thanks for the patches, do you mind rebasing them on master?

Alistair

>
>  target/riscv/cpu.c                      |    1 +
>  target/riscv/cpu.h                      |    2 +
>  target/riscv/cpu_helper.c               |    2 +
>  target/riscv/insn_trans/trans_rvv.c.inc |  179 +++
>  target/riscv/internals.h                |    6 +-
>  target/riscv/translate.c                |    4 +
>  target/riscv/vector_helper.c            | 1558 ++++++++++++++---------
>  7 files changed, 1135 insertions(+), 617 deletions(-)
>
> --
> 2.34.2
>