Andrew Pinski <quic_apin...@quicinc.com> writes:
> Instead of waiting to get combine/rtl optimizations fixed here. This fixes the
> builtins at the gimple level. It should provide for slightly faster compile
> time
> since we have a simplification earlier on.
Yeah, and it's more global than combine would be.
> Built and tested for aarch64-linux-gnu.
>
> gcc/ChangeLog:
>
> PR target/114522
> * config/aarch64/aarch64-builtins.cc (aarch64_fold_aes_op): New
> function.
> (aarch64_general_gimple_fold_builtin): Call aarch64_fold_aes_op for
> crypto_aese
> and crypto_aesd.
>
> Signed-off-by: Andrew Pinski <quic_apin...@quicinc.com>
> ---
> gcc/config/aarch64/aarch64-builtins.cc | 30 ++++++++++++++++++++++++++
> 1 file changed, 30 insertions(+)
>
> diff --git a/gcc/config/aarch64/aarch64-builtins.cc
> b/gcc/config/aarch64/aarch64-builtins.cc
> index 6d5479c2e44..ab67194575d 100644
> --- a/gcc/config/aarch64/aarch64-builtins.cc
> +++ b/gcc/config/aarch64/aarch64-builtins.cc
> @@ -4722,6 +4722,31 @@ aarch64_fold_combine (gcall *stmt)
> return gimple_build_assign (gimple_call_lhs (stmt), ctor);
> }
>
> +/* Fold a call to vaeseq_u8 and vaesdq_u8.
> + That is `vaeseq_u8 (x ^ y, 0)` gets folded
> + into `vaeseq_u8 (x, y)`.*/
> +static gimple *
> +aarch64_fold_aes_op (gcall *stmt)
> +{
> + tree arg0 = gimple_call_arg (stmt, 0);
> + tree arg1 = gimple_call_arg (stmt, 1);
> + if (integer_zerop (arg0))
> + arg0 = arg1;
> + else if (!integer_zerop (arg1))
> + return nullptr;
> + if (TREE_CODE (arg0) != SSA_NAME)
> + return nullptr;
> + if (!has_single_use (arg0))
> + return nullptr;
I was initially unsure about the single-use check, but in the end
I agree it's the right choice.
> + gimple *s = SSA_NAME_DEF_STMT (arg0);
> + if (!is_gimple_assign (s)
> + || gimple_assign_rhs_code (s) != BIT_XOR_EXPR)
> + return nullptr;
Very minor, but how about the slightly less LISPy:
auto *s = dyn_cast<gassign *> (SSA_NAME_DEF_STMT (arg0));
if (!s || gimple_assign_rhs_code (s) != BIT_XOR_EXPR)
return nullptr;
so that we're always accessing gimple_assign_* through a gassign.
OK with that change from my POV, but please give others a day or
so to comment.
Thanks,
Richard
> + gimple_call_set_arg (stmt, 0, gimple_assign_rhs1 (s));
> + gimple_call_set_arg (stmt, 1, gimple_assign_rhs2 (s));
> + return stmt;
> +}
> +
> /* Fold a call to vld1, given that it loads something of type TYPE. */
> static gimple *
> aarch64_fold_load (gcall *stmt, tree type)
> @@ -4983,6 +5008,11 @@ aarch64_general_gimple_fold_builtin (unsigned int
> fcode, gcall *stmt,
> gimple_call_set_lhs (new_stmt, gimple_call_lhs (stmt));
> break;
>
> + VAR1 (BINOPU, crypto_aese, 0, DEFAULT, v16qi)
> + VAR1 (BINOPU, crypto_aesd, 0, DEFAULT, v16qi)
> + new_stmt = aarch64_fold_aes_op (stmt);
> + break;
> +
> /* Lower sqrt builtins to gimple/internal function sqrt. */
> BUILTIN_VHSDF_DF (UNOP, sqrt, 2, FP)
> new_stmt = gimple_build_call_internal (IFN_SQRT,
