Project Description

Valgrind currently lacks support for 80-bit extended precision floating point arithmetic on x86_64 (long double). The current implementation uses 64-bit arithmetic instead, which means that programs running on x86_64 with "long double" give different results when running under Valgrind and natively. People were asking for support for this since like forever, but to date nobody went on implementing it. It's a project for a year, unless somebody has time to work on it full time, but one has to start somewhere. :-)

Goal for this Hackweek

• figure out what needs to be changed in the current implementation
• figure out what are the possibilities of testing it
• compile a list of instructions to be implemented
• write a test program exercising every instruction to be implemented individually
  • must demonstrate no loss of accuracy
  • example: none/tests/amd64/sse4-64.c
• clone valgrind repo and start commiting

The result of the hackweek should be the test case and some writeup on what else needs to happen.

Resources

• https://bugs.kde.org/show_bug.cgi?id=197915
• https://sourceware.org/git/?p=valgrind.git
• https://en.wikipedia.org/wiki/X87
• https://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-software-developer-vol-1-manual.pdf
• https://refspecs.linuxbase.org/elf/x86_64-SysV-psABI.pdf
• https://valgrind.org/docs/manual/manual-core.html#manual-core.limits
• https://en.wikibooks.org/wiki/X86Assembly/FloatingPoint