On 21/04/15 09:44, Kyrill Tkachov wrote:
> Hi all,
>
> We were missing the patterns for the zero-extend versions of the
> negated-logic ops, bic,orn,eon
> leading to redundant zero-extends being generated for code like:
>
> unsigned long
> bar (unsigned int a, unsigned int b)
> {
> return a ^ ~b;
> }
>
> unsigned long
> bar2 (unsigned int a, unsigned int b)
> {
> return a & ~b;
> }
>
>
> With this patch for the above we can generate:
> bar:
> eon w0, w1, w0
> ret
>
> bar2:
> bic w0, w0, w1
> ret
>
>
> instead of:
> bar:
> eon w0, w1, w0
> uxtw x0, w0
> ret
>
> bar2:
> bic w0, w0, w1
> uxtw x0, w0
> ret
>
>
> Bootstrapped and tested on aarch64-linux.
> Ok for trunk?
>
> Thanks,
> Kyrill
>
> 2015-04-21 Kyrylo Tkachov <[email protected]>
>
> * config/aarch64/aarch64.md (*<NLOGICAL:optab>_one_cmplsidi3_ze):
> New pattern.
> (*xor_one_cmplsidi3_ze): Likewise.
>
> aarch64-ze-logic.patch
>
>
> commit 8ff7aaaa6787ce2674b918e1e6ed8b09cafb6b7a
> Author: Kyrylo Tkachov <[email protected]>
> Date: Mon Mar 2 16:20:10 2015 +0000
>
> [AArch64] Add zero_extend variants of logical+not ops
>
> diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
> index 4aa8f5c..1a7f888 100644
> --- a/gcc/config/aarch64/aarch64.md
> +++ b/gcc/config/aarch64/aarch64.md
> @@ -3058,6 +3058,26 @@ (define_insn "*<NLOGICAL:optab>_one_cmpl<mode>3"
> (set_attr "simd" "*,yes")]
> )
>
> +(define_insn "*<NLOGICAL:optab>_one_cmplsidi3_ze"
> + [(set (match_operand:DI 0 "register_operand" "=r")
> + (zero_extend:DI
> + (NLOGICAL:SI (not:SI (match_operand:SI 1 "register_operand" "r"))
> + (match_operand:SI 2 "register_operand" "r"))))]
> + ""
> + "<NLOGICAL:nlogical>\\t%w0, %w2, %w1"
> + [(set_attr "type" "logic_reg")]
> +)
> +
> +(define_insn "*xor_one_cmplsidi3_ze"
> + [(set (match_operand:DI 0 "register_operand" "=r")
> + (zero_extend:DI
> + (not:SI (xor:SI (match_operand:SI 1 "register_operand" "r")
> + (match_operand:SI 2 "register_operand" "r")))))]
> + ""
> + "eon\\t%w0, %w1, %w2"
> + [(set_attr "type" "logic_reg")]
> +)
> +
I would have thought combine ought to know how to canonicalize this last
case into the form supported above. That helps if one of the operands
is a constant, since then you can eliminate the NOT entirely.
Anyway, that's probably best held for a follow-up.
OK.
R.
