This is the mail archive of the mailing list for the glibc project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: underflow exceptions

On Fri, 25 May 2012, Marek Polacek wrote:

> I see one failure on x86_64:
> testing double (without inline functions)
> Failure: fma (-0x1.d2eaed6e8e9d3p-979, -0x1.4e066c62ac9ddp-63, -0x0.9245e6b003454p-1022) == -0x0.9245c09c5fb5dp-1022: Exception
> "Underflow" not set

This is probably an instance of the same issue as in bug 14152, but for 
fma.  I suggest updating the summary of that bug to reflect that the issue 
applies to this fma test as well, and changing UNDERFLOW_EXCEPTION to 
UNDERFLOW_EXCEPTION_OK in the test (with a comment above it referring to 
bug 14152).

I generally do libm testing with --disable-multi-arch so that it's 
predictable what version of a function is used and to be sure that the 
generic versions are properly fixed.  In this case, I suspect you had 
multi-arch enabled, but a processor without FMA4 instructions, and this 
combination meant your testing ended up using the dbl-64 version of 
s_fma.c whereas mine used the ldbl-96 one (and this issue of missing 
underflow exceptions may only arise when the operation is done in the same 
internal precision as the final result).

Joseph S. Myers

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]