Re: [PATCH 3/4] Add ILP32 support to aarch64

[Richard Henderson <rth at twiddle dot net>]

On 08/08/2017 11:10 AM, Steve Ellcey wrote:
> @@ -44,6 +66,32 @@ __CONCATX(__,FUNC) (ITYPE x)
>  {
>    OTYPE result;
>    ITYPE temp;
> +
> +#if IREG_SIZE == 64 && OREG_SIZE == 32
> +  if (__builtin_fabs (x) > INT32_MAX)
> +    {
> +      /* Converting large values to a 32 bit int may cause the frintx/fcvtza
> +	 sequence to set both FE_INVALID and FE_INEXACT.  To avoid this
> +         we save and restore the FE and only set one or the other.  */
> +
> +      fenv_t env;
> +      int feflags;
> +
> +      libc_feholdexcept (&env);
> +      asm ( "frintx" "\t%" IREGS "1, %" IREGS "2\n\t"
> +	    "fcvtzs" "\t%" OREGS "0, %" IREGS "1"
> +	    : "=r" (result), "=w" (temp) : "w" (x) );
> +      feflags = libc_fetestexcept (FE_INVALID | FE_INEXACT);
> +      libc_fesetenv (&env);
> +
> +      if (feflags & FE_INVALID)
> +	feraiseexcept (FE_INVALID);
> +      else if (feflags & FE_INEXACT)
> +	feraiseexcept (FE_INEXACT);
> +
> +      return result;
> +  }
> +#endif

Surely it is simply better to do the conversion in one step, getting the proper
flags set the first time.  Like so.


r~

Attachment: zz
Description: Text document