From Fedora Project Wiki

Revision as of 22:46, 6 November 2018 by Bcotton (talk | contribs) (Re-raised it to FESCo for voting)


New 128-bit IEEE long double ABI for IBM 64-bit POWER LE

Summary

Transition IBM 64-bit POWER LE systems to the new 128-bit IEEE long double ABI.

Owner

  • Name: Carlos O'Donell
  • Email: carlos@redhat.com
  • Release notes owner: carlos@redhat.com
  • Release notes ticket:

Current status

  • Targeted release: Fedora 30
  • Last updated: 2018-11-06
  • Tracker bug:

Detailed Description

IBM has designed a new long double ABI that adheres to the 128-bit IEEE format. This format is more standard than the existing AIX double-double or IBM long double (2 grouped 64-bit doubles) which has discontinuous mantissas and is difficult for developers to use. In Fedora 30 the plan is to switch to the new ABI for long double, while still supporting old applications via compatibility symbols. Newly compiled applications use either the old or new ABI but not a mix of both. Changes are required in the core C libraries, and the compiler and the compiler runtimes including the C++ standard libraries. Therefore there is coordination required across the core toolchain componenents e.g. gcc, binutils, glibc, gdb (to debug the new types).

Benefit to Fedora

Fedora developers will be using a standard 128-bit IEEE format for long double instead of the non-standard double-double AIX format which has a discontinuous mantissa and multiple representations for the same value.

Scope

The change is relatively limited in that not many packages use the long double floating point ABI. The double floating point ABI is much more used, but not long double. It is estimated that few packages use long double directly, and those packages will need to be rebuilt in order to use the new ABI. This rebuilding can be targetted by analyzing which packages have long double usage in their debug information and rebuilding just those packages. However, we plan to just use the existing mass rebuild for glibc 2.29 to handle this issue.

  • Proposal owners: Transition glibc to float128 format for long double for IBM ppc64le. Transition gcc to the default for long double. Implement support for the new long double format in libstdc++. Ensure gdb can handle the new types.
  • Other developers: Developers need to ensure that rawhide is stable and ready for the Fedora 30 branch.
  • Release engineering: A mass rebuild request has been filed for the parent system-wide change to upgrade glibc to 2.29
    #7907
  • Policies and guidelines: The policies and guidelines do not need to be updated.
  • Trademark approval: Not needed for this change

Upgrade/compatibility impact

The library and language runtimes are backwards compatible with the version shipped in Fedora.

We fully expect to fix all packaging changes in Fedora Rawhide first when everything is ready.

How To Test

The GNU C Library has its own testsuite, which is run during the package build and examined by the glibc developers before being uploaded. This test suite has 2500+ tests that run to verify the correct operation of the library. In the future we'll also be running the microbenchmark to look for performance regressions as well as behavioural ones.

Specific testing for 128-bit IEEE long double ABI will be carried out by the glibc testsuite. Integration smoke testing will be carried out by the glibc developers to make sure new applications are built with the correct defaults and work as expected.

Specific testing for 128-bit IEEE long double ABI will be carried out by the gcc testsuite.

Specific smoke testing will be carried out using gdb to read and write the new types.

User Experience

Users will see a new 128-bit floating point ABI, but this will largely be transparent to them. On POWER hardware that supports 128-bit long double in hardware the compiler will use the hardware transparently to accelerate floating point operations, otherwise software floating point emulation will be used.

Dependencies

This change requires coordination of glibc and gcc to change the compiler defaults and build the compiler language runtimes correctly. Also gdb must be able to support the new type to make the process of transition seamless.

Contingency Plan

  • Contingency mechanism: Ship glibc 2.28 instead of glibc 2.29, or ship glibc 2.29 without this feature if it is not ready.
  • Contingency deadline: Final mass rebuild before Fedora release.
  • Blocks release? Upgrading glibc does block the release. We should not ship without the float128 ABI change.

Documentation

The glibc/gcc manual contain the documentation for the release and don't need any more additional work.

Release Notes

  • The ppc64le architecture changed the format of the long double type to binary128. (Previously, a pair of two doubles was used.)