Reputation: 10261
Does anyone know what the current state of valgrind's FPU emulation is? A search shows issues with inaccuracy going back over 10 years, but no real clarity on whether this is considered to be an issue or will ever be addressed.
In my case, I need the results of 32-bit (yes, 32-bit) floating-point ops to be exactly as defined in IEEE-754, for all rounding modes. However, valgrind is slightly off, and my program terminates when it detects errors. On x86 getting a correct 32-bit IEEE FPU requires SSE ops, which makes life difficult.
One option for valgrind might be to link in an external IEEE emulator.
Upvotes: 1
Views: 142
Reputation: 3807
x86 floating point emulation still has limitations. See e.g. http://www.valgrind.org/docs/manual/manual-core.html#manual-core.limits and https://bugs.kde.org/show_bug.cgi?id=197915
Upvotes: 2