Re: Conformance issues to address after 0.9.12 release

[Rich Felker <dalias@aerifal.cx>]

On Mon, Aug 12, 2013 at 09:27:04PM +0200, Szabolcs Nagy wrote:
> diff --git a/src/math/i386/asin.s b/src/math/i386/asin.s
> index 932c754..7e49409 100644
> --- a/src/math/i386/asin.s
> +++ b/src/math/i386/asin.s
> @@ -2,7 +2,20 @@
>  .type asinf,@function
>  asinf:
>  	flds 4(%esp)
> -	jmp 1f
> +	mov 4(%esp),%eax
> +	add %eax,%eax
> +	cmp $0x01000000,%eax
> +	jae 1f
> +		# subnormal x, return x with underflow
> +	fnstsw %ax
> +	and $16,%ax
> +	jnz 2f
> +	push %eax
> +	fld %st(0)
> +	fmul %st(1)
> +	fstps (%esp)
> +	pop %eax
> +2:	ret

Why the push/pop? Is there a reason you can't just store over top of
the argument slot at 4(%esp) or even below the stack pointer at
-4(%esp)? The latter would be unsafe if you ever wanted to read back
the value since it could be clobbered asynchronously by a signal
handler, but you don't want/need to read it back anyway.

Rich