[PATCH 0/2] IBM Z: Prepare cleanup of float express precision

Marius Hillenbrand mhillen@linux.ibm.com
Wed Nov 25 17:06:21 GMT 2020


gcc has special behavior for FLT_EVAL_METHOD on s390x that causes
performance impact in some scenarios and fails to align with float_t wrt
the C standard in others. These two patches prepare gcc for a cleanup to
get rid of that special case, to improve standard compliance and avoid
the overhead.

On s390 today, float_t is defined as double while gcc by default sets
FLT_EVAL_METHOD to 0 and evaluates float expressions in
single-precision. To mitigate that mismatch, with -std=c99 gcc emits
double precision instructions for evaluating float expressions -- at the
cost of additional conversion instructions. Earlier discussions favored
this behavior to maintain ABI compatibility and compliance with the C
standard (that is, for -std=c99), see the discussion around
https://gcc.gnu.org/legacy-ml/gcc-patches/2016-09/msg02392.html Given
the performance overhead, I have reevaluated the impact of cleaning up
the special behavior and changing float_t into float on s390, and now
think that option to be favorable.

The reason for float_t being defined as double is that the port of glibc
to s390 deferred to the generic definition, which back then defaulted to
double. Since then, that definition has not been changed to avoid
breaking ABIs that use float_t. I found only two affected packages,
clucene and ImageMagick, out of >130k Debian packages scanned, and
prepared patches to avoid impact. ImageMagick's ABI has become
independent of float_t on s390 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/.

To smoothen the transition, the first patch makes gcc's behavior
configurable at compile-time with the flag
--enable-s390-excess-float-precision. When the flag is enabled, gcc
maintains the current behavior. By default, configure will test glibc's
behavior: if glibc ties float_t to double, configure will enable the
flag and maintain the current behavior.  Otherwise, it will disable the
flag and drop the special behavior.

Bootstrapped and regtested on s390x, both with a conventional and a
modified glibc. Bootstrapped and regtested on x86-64. Inspected the
generated headers on both targets.

Marius Hillenbrand (2):
  IBM Z: Configure excess precision for float at compile-time
  gcc/testsuite/s390: Add test cases for float_t

 gcc/config/s390/s390.c                    | 27 ++++++++++----
 gcc/configure.ac                          | 45 +++++++++++++++++++++++
 gcc/doc/install.texi                      | 10 +++++
 gcc/testsuite/gcc.target/s390/float_t-1.c | 15 ++++++++
 gcc/testsuite/gcc.target/s390/float_t-2.c | 13 +++++++
 5 files changed, 103 insertions(+), 7 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/s390/float_t-1.c
 create mode 100644 gcc/testsuite/gcc.target/s390/float_t-2.c


More information about the Gcc-patches mailing list