Re: _Float16-related failures on x86_64-apple-darwin
On Fri, 24 Dec 2021, Hongtao Liu via Gcc wrote:
> gcc define __FLT_EVAL_METHOD__ according to
>
> builtin_define_with_int_value ("__FLT_EVAL_METHOD__",
> c_flt_eval_method (true));
>
> and guess we need to handle things like:
>
>/* GCC only supports one interchange type right now, _Float16. If
> we're evaluating _Float16 in 16-bit precision, then flt_eval_method
> will be FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16. */
> + if (x == FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16
> + && x == y)
> +return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT;
That's an incorrect change. excess_precision_mode_join has
straightforward, well-defined semantics documented in the comment above
the function and correcly implemented by it before this change; modifying
those semantics is not the appropriate way to address this issue.
fixincludes is the right place for a fix for this issue. There is a
plausible case for having an architecture-independent
__FLT_EVAL_METHOD___ macro that takes only values defined by
C99 (regardless of -fpermitted-flt-eval-methods), rather than using the
new C23 values such as 16, but if you did have such a macro you'd still
need to fixinclude the system headers - it would just affect exactly what
change fixincludes makes to those headers (if there were such a macro,
fixincludes could change the headers to use it).
--
Joseph S. Myers
jos...@codesourcery.com
Re: _Float16-related failures on x86_64-apple-darwin
Hi Joseph, > fixincludes is the right place for a fix for this issue. There is a > plausible case for having an architecture-independent > __FLT_EVAL_METHOD___ macro that takes only values defined by > C99 (regardless of -fpermitted-flt-eval-methods), rather than using the > new C23 values such as 16, but if you did have such a macro you'd still > need to fixinclude the system headers - it would just affect exactly what > change fixincludes makes to those headers (if there were such a macro, > fixincludes could change the headers to use it). There is still a difference. If we define a new macro __FLT_EVAL_METHOD_OLDSTYLE__ (however it is named), it means we can make a more robust fixinclude, using that macro. If we fixinclude right now to handle the value of 16, then we might have to update the fixinclude for any new value that comes along in the future. FX
Re: Need for __builtin_issignaling()
On Wed, 29 Dec 2021, FX via Gcc wrote: > Is it particularly hard to do? I came across a post in the list archives > from Joseph, who said it would be good to have. I’d be willing to try > and put something together, unless you think it’s a big project. Any > pointers as to how to start would be appreciated. Start by looking at Tamar Christina's patch for bugs 77925, 77926, 66462, which got reverted because of problems it caused. In particular, see my comment 18 in bug 66462 pointing to some of the reports of issues, and look at the relevant discussion in June 2017. I think the following comment 19 is incorrect (when it refers to June and November, those are June 2017 and November 2016, so the November version is an *older* one, but that comments is under the apprehension that it was a newer one). There's no need to use that patch as a starting point, but it may well be helpful to do so, or at least to get ideas from it. It didn't add __builtin_issignaling, but did add implementations of other related built-in functions based on bit-manipulation, and __builtin_issignaling would need to be implemented based on such bit-manipulation. Apart from avoiding the bugs in that patch, for __builtin_issignaling there isn't any other implementation approach to fall back on, and there isn't any defined external-linkage function to fall back on either. So it's strongly desirable to have a built-in function that works (is expanded inline) for *all* floating-point formats supported by GCC, not just some. (For formats not supporting signaling NaNs, it can trivially return 0 after evaluating the argument for its side effects.) Note that supporting all formats includes working for formats where integer arithmetic on a same-size integer isn't supported (TFmode on 32-bit architectures where there isn't TImode integer arithmetic, in particular), so you need to be careful about working correctly in that case. -- Joseph S. Myers jos...@codesourcery.com
gcc-9-20211230 is now available
Snapshot gcc-9-20211230 is now available on https://gcc.gnu.org/pub/gcc/snapshots/9-20211230/ and on various mirrors, see http://gcc.gnu.org/mirrors.html for details. This snapshot has been generated from the GCC 9 git branch with the following options: git://gcc.gnu.org/git/gcc.git branch releases/gcc-9 revision 218f5148de327759769468030496bd370102bd64 You'll find: gcc-9-20211230.tar.xzComplete GCC SHA256=615dd83371326b63f16eac8bb47ee686ad5fba6547a3518ec749234b6246e11b SHA1=0818b925d5fb31206ad17ce3ac784b24254df60c Diffs from 9-20211223 are available in the diffs/ subdirectory. When a particular snapshot is ready for public consumption the LATEST-9 link is updated and a message is sent to the gcc list. Please do not use a snapshot before it has been announced that way.
