[math, Tru64, V2] compex_test failures
Markus, there's a failure of math/complex_test on Tru64: http://tinyurl.com/zu33q I can't make much of it, so I'd suspect that there's some "make floating point more accurate" option that's passed in V1, and not passed in V2. Unfortunately, CVS does not contain tru64cxx71-006 toolset, so I can't compare the command lines myself. Any change you'll run the test with V1 and V2 (with -a -n, to get just commnad lines), compare the compile command lines and tell me if there are any significant differences, or just send the command lines to me? Thanks in advance, Volodya
Vladimir Prus wrote:
Markus, there's a failure of math/complex_test on Tru64:
I can't make much of it, so I'd suspect that there's some "make floating point more accurate" option that's passed in V1, and not passed in V2.
Unfortunately, CVS does not contain tru64cxx71-006 toolset, so I can't compare the command lines myself.
Any change you'll run the test with V1 and V2 (with -a -n, to get just commnad lines), compare the compile command lines and tell me if there are any significant differences, or just send the command lines to me?
I added the needed flag to the my local configuration just today. It's -ieee and needed to enforce IEEE conforming math operations. It's never been part of the basic toolset because usually the IEEE conformance of the CPU itself is good enough, and using this flags may incur a significant performance penalty. So I would recommend not to add it to the basic toolset. Strictly speaking it should be there for all math relevant boost libraries which depend on IEEE conforming math, but I always thoughr that anyone using this libraries on Alpha CPUs will know about the pitfalls. The tests passes locally with my current V2 toolset, and there will be a fresh non-incremental regression run over the weekend, therefore it should show up green soon. Markus
I added the needed flag to the my local configuration just today. It's -ieee and needed to enforce IEEE conforming math operations. It's never been part of the basic toolset because usually the IEEE conformance of the CPU itself is good enough, and using this flags may incur a significant performance penalty. So I would recommend not to add it to the basic toolset.
Strictly speaking it should be there for all math relevant boost libraries which depend on IEEE conforming math, but I always thoughr that anyone using this libraries on Alpha CPUs will know about the pitfalls.
I suspect the issue is that numeric_limits<T>::has_infinity is true, so the test attempts to verify that infinities are correctly handled, but in practice infinities trap without the -ieee flag. Is there a macro that tells you when infinities are correctly handled? John.
John Maddock wrote: [...]
I suspect the issue is that numeric_limits<T>::has_infinity is true, so the test attempts to verify that infinities are correctly handled, but in practice infinities trap without the -ieee flag. Is there a macro that tells you when infinities are correctly handled?
-ieee sets the macro _IEEE_FP. Markus
participants (3)
-
John Maddock -
Markus Schöpflin -
Vladimir Prus