Re: [PATCH 4/4] alpha: expect test-float32x-float64-div to fail

On 03/08/26 20:55, Matt Turner wrote:
> _Float32x and _Float64 are both binary64 on Alpha, so this narrowing
> divide is a plain divide and the hardware alone decides whether to signal
> underflow.
> 
> IEEE 754 determines tininess after rounding from the result rounded as if
> the exponent range were unbounded, while Alpha determines it from the
> delivered result.  The two differ for a quotient that is tiny but rounds
> up to the smallest normal, as in DBL_MIN / (1 + 2^-52) under a rounding
> mode that rounds away from zero: the binade below DBL_MIN has a finer
> spacing than the subnormals, so the unbounded rounding stays below
> DBL_MIN and the result is tiny, but the delivered result is DBL_MIN and
> looks normal.  Alpha signals no underflow for it.
> 
> Nothing in software can correct this.  The hardware detects no underflow,
> so no software completion trap is taken and the kernel emulation never
> runs, and as the operation is not really narrowing there is no wider
> intermediate for libm to examine.

LGTM, thanks.  I wonder if we should add a header like tininess.h to handle
this operation in a generic manner and avoid the xfail definition; but it
seems really alpha-specific.

Reviewed-by: Adhemerval Zanella  <[email protected]>

> ---
>  sysdeps/alpha/Makefile | 14 ++++++++++++++
>  1 file changed, 14 insertions(+)
> 
> diff --git ./sysdeps/alpha/Makefile ./sysdeps/alpha/Makefile
> index 60a369e255..faa59ab27a 100644
> --- ./sysdeps/alpha/Makefile
> +++ ./sysdeps/alpha/Makefile
> @@ -53,6 +53,20 @@ CFLAGS-s_lrint.c += -mieee-with-inexact
>  CFLAGS-test-misc.c += -mieee-with-inexact
>  # Avoid "conflicting types for built-in function" warnings
>  CFLAGS-s_isnan.c += -fno-builtin-isnanf
> +
> +# _Float32x and _Float64 are both binary64 on Alpha, so this narrowing
> +# divide is a plain divide and the hardware alone decides whether to
> +# signal underflow.  IEEE 754 determines tininess after rounding from the
> +# result rounded as if the exponent range were unbounded, but Alpha
> +# determines it from the delivered result.  For a quotient that is tiny
> +# but rounds up to the smallest normal -- DBL_MIN / (1 + 2^-52) under a
> +# rounding mode that rounds away from zero -- the unbounded rounding
> +# stays below DBL_MIN, because that binade has a finer spacing than the
> +# subnormals, while the delivered result is DBL_MIN and looks normal.
> +# Alpha raises no underflow for it.  No trap is taken, so the kernel
> +# emulation cannot correct this, and the operation has no wider
> +# intermediate for libm to examine.
> +test-xfail-test-float32x-float64-div = yes
>  endif
>  
>  # Build everything with full IEEE math support, and with dynamic rounding;