Re: [PATCH], PowerPC IEEE 128-bit patch #8 (libgcc bits, revised)

[Joseph Myers <joseph@codesourcery.com>]

On Fri, 13 Nov 2015, Michael Meissner wrote:

> This patch should address the following IEEE problems with the libgcc emulator
> with exceptions, etc.  Steve is working on patches to libdfp for Decimal
> conversions.

Are the libgcc2.c functions for KFmode and the corresponding complex mode 
KCmode (__divkc3 __mulkc3 __powikf2) also planned for GCC 6?

> 	* config/rs6000/sfp-machine.h (_FP_W_TYPE_SIZE): Update
> 	defintiions to use 64-bit types for 64-bit systems or 32-bit VSX
> 	systems with IEEE 128-bit floating point support.
> 	(_FP_W_TYPE): Likewise.
> 	(_FP_WS_TYPE): Likewise.
> 	(_FP_I_TYPE): Likewise.
> 	(TItype): Add defintions for IEEE 128-bit floating point support.
> 	(UTItype): Likewise.
> 	(TI_BITS): Likewise.
> 	(_FP_MUL_MEAT_Q): Use 64-bit types on 64-bit systems and for
> 	32-bit VSX systems with IEEE 128-bit floating point.
> 	(_FP_DIV_MEAT_Q): Likewise.
> 	(_FP_NANFRAC_Q): Likewise.

This ChangeLog entry doesn't seem to have been updated for the exceptions 
/ rounding modes support.

> +#define ISA_BIT(x) (1 << (63 - x))

I think all the new sfp-machine.h code for integration with hardware 
exceptions and rounding modes needs to be conditional on hard-float + 
FPRs, so it doesn't affect building for SFmode / DFmode for soft-float / 
e500.

> +  /* First do the unordered comparison.  */
> +  FP_CMP_UNORD_Q (r, A, B, 1);
> +  if (r)
> +    r2 = PPC_UNORDERED;
> +
> +  else
> +    {
> +      /* If the value is not unordered, do the normal comparison.  */
> +      FP_CMP_Q (r, A, B, 2, 2);
> +      if (r == CMP_INVALID)
> +	FP_SET_EXCEPTION (FP_EX_INVALID);
> +
> +      r2 = ((r < CMP_LOW || r > CMP_HIGH)
> +	    ? PPC_UNORDERED
> +	    : ppc_cr_map[r-CMP_LOW]);

This may be correct, but it's overly complicated.  All the handling in the 
"else" case for exceptions / unordered is dead code, since in that case 
you know that the values aren't unordered.  And you can more simply do

  FP_CMP_Q (r, A, B, 2, 1);

at which point "invalid" has been raised if appropriate (for signaling 
NaNs but not for other arguments), and r contains one of the values -1, 0, 
1, 2, which you can map directly to the appropriate return value.

-- 
Joseph S. Myers
joseph@codesourcery.com