Re: [PATCH] amdgcn, libm: fix infinite loop

Hi Andrew!

On 2025-08-07T13:09:41+0000, Andrew Stubbs <[email protected]> wrote:
> The end condition on this loop, unlike all the other similar loops, is
> "i >= 0", which is a problem because "i <<= 1" can go negative and then zero if
> you continue shifting, and so back to true, again.  This isn't a problem for
> the loop in the scalar implementation, but it means we need to mask the shift
> in the vector implementation.
>
> This fixes GCC PR#121392.

Thanks, I confirm this fixes the issue.

I suggest you 'git push', or I can do it for you if you don't have access
to newlib Git -- or, I'll be happy to sponsor you to get access, as
maintainer of the GCN parts in newlib.
(<https://sourceware.org/cgi-bin/pdw/ps_form.cgi> has the instructions
for the latter.)


Grüße
 Thomas


>  newlib/libm/machine/amdgcn/v64sf_fmod.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/newlib/libm/machine/amdgcn/v64sf_fmod.c b/newlib/libm/machine/amdgcn/v64sf_fmod.c
> index 7302420ad..b62b81929 100644
> --- a/newlib/libm/machine/amdgcn/v64sf_fmod.c
> +++ b/newlib/libm/machine/amdgcn/v64sf_fmod.c
> @@ -70,8 +70,11 @@ DEF_VS_MATH_FUNC (v64sf, fmodf, v64sf x, v64sf y)
>    v64si iy;
>    VECTOR_IF (hy < 0x00800000, cond)	// subnormal y
>      iy = VECTOR_INIT (-126);
> -    for (v64si i = (hy << 8); !ALL_ZEROES_P (cond & (i >= 0)); i <<= 1)
> -      VECTOR_COND_MOVE (iy, iy - 1, cond & (i >= 0));
> +    for (v64si i = (hy << 8); !ALL_ZEROES_P (cond & (i >= 0)); /* i <<= 1 */)
> +      {
> +	VECTOR_COND_MOVE (iy, iy - 1, cond & (i >= 0));
> +	VECTOR_COND_MOVE (i, i << 1, cond & (i >= 0));
> +      }
>    VECTOR_ELSE (cond)
>      VECTOR_COND_MOVE (iy, (hy >> 23) - 127, cond);
>    VECTOR_ENDIF
> -- 
> 2.50.0