Re: [musl] [PATCH] s390x: derive float_t from compiler or default to float

[Rich Felker <dalias@libc.org>]

On Tue, Dec 01, 2020 at 03:36:34PM +0100, Marius Hillenbrand wrote:
> Hi,
> 
> float_t should represent the type that is used to evaluate float
> expressions internally. On s390(x), float_t is currently set to double.
> In contrast, the isa supports single-precision float operations and
> compilers by default evaluate float in single precision, which violates
> the C standard (sections 5.2.4.2.2 and 7.12 in C11/C17). With
> -fexcess-precision=standard, gcc evaluates float in double precision,
> which aligns with the standard yet at the cost of added conversion
> instructions. To improve standards compliance, this patch changes the
> definition of float_t to be derived from the compiler's
> __FLT_EVAL_METHOD__.
> 
> The port of glibc to s390 incorrectly deferred to the generic
> definitions which, back then, tied float_t to double. Since then, this
> definition has been kept to avoid ABI changes, most recently in the
> refactoring of float_t into bits/flt-eval-method.h
> https://sourceware.org/legacy-ml/libc-alpha/2016-11/msg00903.html
> and the discussion around
> https://gcc.gnu.org/legacy-ml/gcc-patches/2016-09/msg02392.html
> musl apparently adopted the definition from glibc.
> 
> Given the performance overhead and reduced standards compliance, I have
> reevaluated cleaning up the special behavior on s390x. I found only two
> packages, ImageMagick and clucene, that use float_t in their API, out of
> >130k Debian source packages scanned. To avoid breaking ABI changes, I
> patched these packages to avoid their reliance on float_t (in
> ImageMagick since 7.0.10-39, patch in
> https://github.com/ImageMagick/ImageMagick/pull/2832 - patch for
> clucene in https://sourceforge.net/p/clucene/bugs/233).
> 
> gcc-11 will drop the special case to retrofit double
> precision behavior for -fexcess-precision=standard so that
> __FLT_EVAL_METHOD__ will be 0 on s390x in any scenario.
> https://gcc.gnu.org/pipermail/gcc-patches/2020-November/560224.html
> https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=a5dd6b69fcbe74c02d4821ac2daf2b8c9f819f6e
> 
> glibc 2.33 will most likely adopt the same behavior as in this patch, so
> that float_t will eventually be float on s390x in any scenario.
> https://sourceware.org/pipermail/libc-alpha/2020-November/120212.html
> 
> Testing with libc-test showed no regressions. Failing testcases
> src/math/lgammaf[_r].exe succeed with the patch.
> 
> Please review and consider merging this patch.

Thanks for the detailed report. To be clear, all models/ISA-levels
support the single-precision ops and future GCC will always use them
even with -fexcess-precision=standard, but old ones switch to using
double precision ops with -fexcess-precision=standard to meet the
contract of evaluating in (old definition of) float_t. Is this
correct?

Rich