Re: [Tinycc-devel] A double-to-float conversion bug

[grischka]

On 31.07.2023 13:42, Herman ten Brugge via Tinycc-devel wrote:
> I agree with your comments above. The size is incorrect.
> I could change gfunc_sret in x86_64-gen.c and then calculate the size in 
> tccgen.c
> as you suggested. But I am not sure regsize is set correctly all the time.
>
> I like this better:
> --- a/tccgen.c
> +++ b/tccgen.c
> @@ -6142,7 +6142,7 @@ special_math_val:
>                          space.  Assume register size is power of 2. */
>                       if (regsize > align)
>                         align = regsize;
> -                   loc &= -align;
> +                   size = (size + regsize - 1) & -regsize;
>                       loc = (loc - size) & -align;
>                       addr = loc;
>                       offset = 0;
>
> The size should be a multiple of regsize.
> What is your opinion?

Hi,

yes, that seems to make sense.  Then again I think that align should
additionally depend on just 'ret_align', as in

      if (ret_align > align)
           align = ret_align;


> I was working on a better fix for riscv and reverted the patch from Yao Zi by 
> accident.
> I will reapply that change.

Well, as I was trying to say, that change wasn't a good one either.

Anyway, I just applied some more extensive fix for both.  Hope it does work
now.  Also yet another pragma_once_normalize patch (which you may or may
not want to comment ;)

-- gr

>      Herman