Re: [PATCH] match.pd: rewrite select to branchless expression

[Richard Biener <richard.guenther@gmail.com>]

On Tue, Nov 8, 2022 at 9:02 PM Michael Collison <collison@rivosinc.com> wrote:
>
> This patches transforms (cond (and (x , 0x1) == 0), y, (z op y)) into
> (-(and (x , 0x1)) & z ) op y, where op is a '^' or a '|'. It also
> transforms (cond (and (x , 0x1) != 0), (z op y), y ) into (-(and (x ,
> 0x1)) & z ) op y.
>
> Matching this patterns allows GCC to generate branchless code for one of
> the functions in coremark.
>
> Bootstrapped and tested on x86 and RISC-V. Okay?
>
> Michael.
>
> 2022-11-08  Michael Collison  <collison@rivosinc.com>
>
>      * match.pd ((cond (and (x , 0x1) == 0), y, (z op y) )
>      -> (-(and (x , 0x1)) & z ) op y)
>
> 2022-11-08  Michael Collison  <collison@rivosinc.com>
>
>      * gcc.dg/tree-ssa/branchless-cond.c: New test.
>
> ---
>   gcc/match.pd                                  | 22 ++++++++++++++++
>   .../gcc.dg/tree-ssa/branchless-cond.c         | 26 +++++++++++++++++++
>   2 files changed, 48 insertions(+)
>   create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index 194ba8f5188..722f517ac6d 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -3486,6 +3486,28 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>     (cond (le @0 integer_zerop@1) (negate@2 @0) integer_zerop@1)
>     (max @2 @1))
>
> +/* (cond (and (x , 0x1) == 0), y, (z ^ y) ) -> (-(and (x , 0x1)) & z )
> ^ y */

Please write the match as a C expression in the comment, as present
it's a weird mix.  So x & 0x1 == 0 ? y : z <op> y -> (-(typeof(y))(x &
0x1) & z) <op> y

> +(for op (bit_xor bit_ior)
> + (simplify
> +  (cond (eq (bit_and @0 integer_onep@1)
> +            integer_zerop)
> +        @2
> +        (op:c @3 @2))
> +  (if (INTEGRAL_TYPE_P (type)
> +       && (INTEGRAL_TYPE_P (TREE_TYPE (@0))))
> +       (op (bit_and (negate (convert:type (bit_and @0 @1))) @3) @2))))

Since you are literally keeping (bit_and @0 @1) and not matching @0 with
anything I suspect you could instead use

 (simplify (cond (eq zero_one_valued_p@0 integer_zerop) ...

eventually extending that to cover bit_and with one.  Do you need to guard
this against 'type' being a signed/unsigned 1-bit precision integer?

> +
> +/* (cond (and (x , 0x1) != 0), (z ^ y), y ) -> (-(and (x , 0x1)) & z )
> ^ y */
> +(for op (bit_xor bit_ior)
> + (simplify
> +  (cond (ne (bit_and @0 integer_onep@1)
> +            integer_zerop)
> +    (op:c @3 @2)
> +        @2)
> +  (if (INTEGRAL_TYPE_P (type)
> +       && (INTEGRAL_TYPE_P (TREE_TYPE (@0))))
> +       (op (bit_and (negate (convert:type (bit_and @0 @1))) @3) @2))))
> +
>   /* Simplifications of shift and rotates.  */
>
>   (for rotate (lrotate rrotate)
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c
> b/gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c
> new file mode 100644
> index 00000000000..68087ae6568
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/branchless-cond.c
> @@ -0,0 +1,26 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-optimized" } */
> +
> +int f1(unsigned int x, unsigned int y, unsigned int z)
> +{
> +  return ((x & 1) == 0) ? y : z ^ y;
> +}
> +
> +int f2(unsigned int x, unsigned int y, unsigned int z)
> +{
> +  return ((x & 1) != 0) ? z ^ y : y;
> +}
> +
> +int f3(unsigned int x, unsigned int y, unsigned int z)
> +{
> +  return ((x & 1) == 0) ? y : z | y;
> +}
> +
> +int f4(unsigned int x, unsigned int y, unsigned int z)
> +{
> +  return ((x & 1) != 0) ? z | y : y;
> +}
> +
> +/* { dg-final { scan-tree-dump-times " -" 4 "optimized" } } */
> +/* { dg-final { scan-tree-dump-times " & " 8 "optimized" } } */
> +/* { dg-final { scan-tree-dump-not "if" "optimized" } } */
> --
> 2.34.1
>
>
>
>