Re: [PATCH 1/2] middle-end: If supported, use __builtin_*. [PR126625]

On Sun, Aug 9, 2026 at 8:03 AM Kael Andrew Alonzo Franco
<[email protected]> wrote:
>
> hwint.cc reflect_hwi uses a naive for loop approach to
> emulate __builtin_bitreverse64 ().
> This is slow compared to using __builtin_bitreverse64 () plus bitshift.
> This is useful for bootstrapping GCC since r17-523.
> Also add assert on BITWIDTH <= 64.
>
> For reflect_hwi and ira-color.cc, add STAGE0_CXX_HAS_BUILTIN
> in system.h to test for specific builtins.
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu.
>
>         PR middle-end/126625
>
> gcc/ChangeLog:
>
>         * hwint.cc (reflect_hwi): Use STAGE0_CXX_HAS_BUILTIN
>         for __builtin_bitreverse64.
>         * ira-color.cc (HAS_SMUL_OVERFLOW): Remove
>         and use STAGE0_CXX_HAS_BUILTIN instead.
>         * system.h (STAGE0_CXX_HAS_BUILTIN): New macro.

I can't think of a better name for this macro so ok,

>
> Signed-off-by: Kael Andrew Franco <[email protected]>
> ---
>  gcc/hwint.cc     | 13 ++++++++++++-
>  gcc/ira-color.cc |  7 +------
>  gcc/system.h     |  8 ++++++++
>  3 files changed, 21 insertions(+), 7 deletions(-)
>
> diff --git a/gcc/hwint.cc b/gcc/hwint.cc
> index f3b3e7b8408..da7bc05f085 100644
> --- a/gcc/hwint.cc
> +++ b/gcc/hwint.cc
> @@ -189,12 +189,21 @@ least_common_multiple (HOST_WIDE_INT a, HOST_WIDE_INT b)
>    return mul_hwi (abs_hwi (a) / gcd (a, b), abs_hwi (b));
>  }
>
> -/* Reflect (reverse) the bits of a given VALUE within a specified BITWIDTH.  */
> +/* Reflect (reverse) the bits of a given VALUE within a specified BITWIDTH <= 64.  */
>
>  unsigned HOST_WIDE_INT
>  reflect_hwi (unsigned HOST_WIDE_INT value, unsigned bitwidth)
>  {
> +  if (bitwidth == 0)
> +    return 0;
> +
> +  gcc_checking_assert (bitwidth <= 64);
> +
> +#if STAGE0_CXX_HAS_BUILTIN (bitreverse64)
> +  return __builtin_bitreverse64 (value) >> (64 - bitwidth);
> +#else
>    unsigned HOST_WIDE_INT reflected_value = 0;
> +
>    /* Loop through each bit in the specified BITWIDTH.  */
>    for (size_t i = 0; i < bitwidth; i++)
>      {
> @@ -204,5 +213,7 @@ reflect_hwi (unsigned HOST_WIDE_INT value, unsigned bitwidth)
>        reflected_value |= (value & 1);
>        value >>= 1;
>      }
> +
>    return reflected_value;
> +#endif
>  }
> diff --git a/gcc/ira-color.cc b/gcc/ira-color.cc
> index 2b389b5aab1..866d5b6bec3 100644
> --- a/gcc/ira-color.cc
> +++ b/gcc/ira-color.cc
> @@ -3154,14 +3154,9 @@ setup_allocno_priorities (ira_allocno_t *consideration_allocnos, int n)
>        ira_assert (mult >= 0);
>        mult *= ira_reg_class_max_nregs[ALLOCNO_CLASS (a)][ALLOCNO_MODE (a)];
>        diff = ALLOCNO_MEMORY_COST (a) - ALLOCNO_CLASS_COST (a);
> -#ifdef __has_builtin
> -#if __has_builtin(__builtin_smul_overflow)
> -#define HAS_SMUL_OVERFLOW
> -#endif
> -#endif
>        /* Multiplication can overflow for very large functions.
>          Check the overflow and constrain the result if necessary: */
> -#ifdef HAS_SMUL_OVERFLOW
> +#if STAGE0_CXX_HAS_BUILTIN (smul_overflow)
>        if (__builtin_smul_overflow (mult, diff, &priority)
>           || priority < -INT_MAX)
>         priority = diff >= 0 ? INT_MAX : -INT_MAX;
> diff --git a/gcc/system.h b/gcc/system.h
> index f0fa062d469..08fefd6f054 100644
> --- a/gcc/system.h
> +++ b/gcc/system.h
> @@ -22,6 +22,14 @@ along with GCC; see the file COPYING3.  If not see
>  #ifndef GCC_SYSTEM_H
>  #define GCC_SYSTEM_H
>
> +/* True if __builtin_* () is supported.
> +   This is done for optimizing GCC itself.  */
> +#ifdef __has_builtin
> +# define STAGE0_CXX_HAS_BUILTIN(NAME) __has_builtin (__builtin_ ## NAME)
> +#else
> +# define STAGE0_CXX_HAS_BUILTIN(NAME) 0
> +#endif
> +
>  /* Define this so that inttypes.h defines the PRI?64 macros even
>     when compiling with a C++ compiler.  Define it here so in the
>     event inttypes.h gets pulled in by another header it is already
> --
> 2.55.0
>