Re: [RFC PATCH] tcg: allow a target to request canonicalization of SUBI to ADDI

[Richard Henderson]

On 10/25/23 11:59, Paolo Bonzini wrote:
> On x86, this is more efficient because it enables generation of
> more LEA instructions.
>
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>   include/tcg/tcg.h     |  4 ++++
>   tcg/i386/tcg-target.h |  2 ++
>   tcg/tcg-op.c          | 20 ++++++++++++++++----
>   3 files changed, 22 insertions(+), 4 deletions(-)
>
> diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h
> index a9282cdcc60..48e5aeef173 100644
> --- a/include/tcg/tcg.h
> +++ b/include/tcg/tcg.h
> @@ -109,6 +109,10 @@ typedef uint64_t TCGRegSet;
>   #define TCG_TARGET_HAS_sub2_i32         1
>   #endif
>   
> +#ifndef TCG_TARGET_PREFERS_addi
> +#define TCG_TARGET_PREFERS_addi         0
> +#endif

I would rather do this unconditionally.

Many of the tcg backends do this manually, e.g. s390x:

    case INDEX_op_sub_i64:
        a0 = args[0], a1 = args[1], a2 = args[2];
        if (const_args[2]) {
            a2 = -a2;
            goto do_addi_64;
        } else {
            tcg_out_insn(s, RRFa, SGRK, a0, a1, a2);
        }
        break;

While we could do something similar for i386, it would be better to not require such hoops 
in each backend.

We would also want to perform this transformation in optimize.c.


r~