On 27 January 2017 at 03:49, Martin Sebor <mse...@gmail.com> wrote:
> I committed the patch below to clean up the "mess."
>
> Thanks
> Martin
>
> Index: gcc/gimple-ssa-sprintf.c
> ===================================================================
> --- gcc/gimple-ssa-sprintf.c (revision 244957)
> +++ gcc/gimple-ssa-sprintf.c (working copy)
> @@ -1686,20 +1686,10 @@ format_floating (const directive &dir, tree arg)
> prec[0] = -1;
> if (prec[1] < 0)
> {
> -#ifdef HAVE_XFmode
> - /* When L is specified use long double, otherwise double. */
> - unsigned fmtprec
> - = (dir.modifier == FMT_LEN_L
> - ? REAL_MODE_FORMAT (XFmode)->p
> - : REAL_MODE_FORMAT (DFmode)->p);
> -#elif defined HAVE_DFmode
> - /* No long double support, use double precision for both. */
> - unsigned fmtprec = REAL_MODE_FORMAT (DFmode)->p;
> -#else
> - /* No long double or double support. */
> - unsigned fmtprec = 0;
> -#endif
> - /* The precision of the IEEE 754 double format is 53.
> + unsigned fmtprec
> + = REAL_MODE_FORMAT (TYPE_MODE (TREE_TYPE (arg)))->p;
> +
> + /* The precision of the IEEE 754 double format is 53.
> The precision of all other GCC binary double formats
> is 56 or less. */
> prec[1] = fmtprec <= 56 ? 13 : 15;
>
> Thanks
> Martin
>
I confirm the builds now succeed on my side (since r244957).
Thanks for the prompt fix.
Christophe.
> On 01/26/2017 06:12 PM, Andrew Pinski wrote:
>>
>> On Thu, Jan 26, 2017 at 3:17 PM, Martin Sebor <mse...@gmail.com> wrote:
>>>>
>>>> Hi,
>>>>
>>>> With this patch all my builds for aarch64/arm failed:
>>>> /gcc/gimple-ssa-sprintf.c: In function
>>>> ‘<unnamed>::fmtresult<unnamed>::format_floating(const<unnamed>::direc
>>>> tive&, tree_node*)’:
>>>> /gcc/gimple-ssa-sprintf.c:1643: error: ‘XFmode’ was not declared in this
>>>> scope
>>>>
>>>> Is this fixed later in the series?
>>>
>>>
>>>
>>> It isn't. I just reproduced it with an aarch64 cross-compiler so
>>> I hope to be able to fix it quickly.
>>
>>
>> This actually broke any non-x86 targeted GCC.
>>
>> Even the current code is wrong:
>> #ifdef HAVE_XFmode
>> /* When L is specified use long double, otherwise double. */
>> unsigned fmtprec
>> = (dir.modifier == FMT_LEN_L
>> ? REAL_MODE_FORMAT (XFmode)->p
>> : REAL_MODE_FORMAT (DFmode)->p);
>> #elif defined HAVE_DFmode
>> /* No long double support, use double precision for both. */
>> unsigned fmtprec = REAL_MODE_FORMAT (DFmode)->p;
>> #else
>> /* No long double or double support. */
>> unsigned fmtprec = 0;
>> #endif
>>
>>
>> Shouldn't this be based instead on long_double_type_node instead?
>> Because different targets might have an option to select different
>> modes.
>>
>> So it just becomes something like which simplifies the mess you have
>> currently:
>> /* When L is specified use long double, otherwise double. */
>> unsigned fmtprec
>> = (dir.modifier == FMT_LEN_L
>> ? REAL_MODE_FORMAT (TYPE_MODE (long_double_type_node))->p
>> : REAL_MODE_FORMAT (TYPE_MODE ((double_type_node))->p);
>>
>>
>> Thanks,
>> Andrew
>>
>>>
>>> Martin
>>>
>
