https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71064
Thomas Schwinge <tschwinge at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|nvptx offloading: "long |Offloading vs. 'long
|double" data type |double' data type
Keywords| |openmp
Target|nvptx |nvptx, GCN
Status|UNCONFIRMED |NEW
Last reconfirmed| |2025-08-07
Ever confirmed|0 |1
CC| |burnus at gcc dot gnu.org,
| |kcy at codesourcery dot com
--- Comment #2 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
Kwok Cheung Yeung, May 21, 2025 at 7:06 PM:
> [...] testcases for the <complex>, <numbers> and <cmath> headers running on
> the offload target.
>
> One problem I have noticed is with the std::nexttoward() function in cmath,
> which triggers the following error on NVPTX:
>
> lto1: fatal error: nvptx-none - 80-bit-precision floating-point numbers
> unsupported (mode 'XF')
>
> Looking at the definition in c_global:
>
> constexpr float
> nexttoward(float __x, long double __y)
> { return __builtin_nexttowardf(__x, __y); }
>
> constexpr long double
> nexttoward(long double __x, long double __y)
> { return __builtin_nexttowardl(__x, __y); }
>
> There is always at least one long double in the arguments, so this function
> cannot work on an architecture that doesn’t support XF mode. I have #ifdef’ed
> out the test for this function.
Tobias Burnus, May 21, 2025 at 7:45 PM:
>> There is always at least one long double in the arguments, so this function
>> cannot work on an architecture that doesn’t support XF mode
>
> … and likewise on systems where the host’s ‘long double' is TF (like on
> PowerPC, which has multiple types of 128bit floating-point numbers).
>
> On the other hand, it works on systems where the host’s ‘long double’ is the
> same as ‘double’ (= DF) (like on Arm → Grace[-Hopper])
>
> Workaround: use nextafter — which has the same type for ‘y' as for 'x’/return
> type, contrary to nexttoward.
For that, we've put into 'libgomp.c++/target-std__cmath.C':
#if 0
/* TODO Due to 'std::nexttoward' using 'long double to', this triggers a
'80-bit-precision floating-point numbers unsupported (mode ‘XF’)'
error
with x86_64 host and nvptx, GCN offload compilers, or
'128-bit-precision floating-point numbers unsupported (mode ‘TF’)'
error
with powerpc64le host and nvptx offload compiler, for example;
PR71064 'nvptx offloading: "long double" data type'.
It ought to work on systems where the host's 'long double' is the same
as
'double' ('DF'): aarch64, for example? */
next = std::nexttoward (x, y);
if (!(next > x && next < y))
return false;
#endif
