[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Access to arm box

> Is it correct
>  that arm does not support setting the FPU rounding mode via fesetround()? At
>  least, <bits/fenv.h> does not define FE_UPWARD, FE_DOWNWARD and FE_TOWARDZERO.

It says in that file:

/* The ARM FPU basically only supports round-to-nearest.  Other rounding
   modes exist, but you have to encode them in the actual instruction.  */
#define FE_TONEAREST    0

- that's talking about the ancient "FPA" unit that a few early ARM
CPUs had, and which is supported in linux on all modern CPUs by
emulating the illegal instructions in the kernel.

>  While at it I also builded the package on armel. <bits/fenv.h> defines all four
>  FE_* constants, but fegetround() always returns FE_TONEAREST. Is this a bug/not
>  supported/...? (I didn't check the rounding mode by a computation yet.)

The armel version of the file says:

/* VFP supports all of the four defined rounding modes.  */

... but the implementaion you were using has no FPU, so uses the
soft-float library coded in ARM assembler in glibc. It could be that
either it doesn't implement rounding modes or the set/get function is
broken (just guessing).

I'm afraid I don't have any units with a VFP FPU here to test on. If
you can make a testcase that would be interesting. however there are
also other FPUs around associated with ARM processors, such as the
buggy Cirrus Maverick Crunch unit, so if you are in a position to test
for macro existence and possibly also run "configure" tests to see
whether the implementation actually works, rather than relying on
system-specific knowledge, that might result in more robust software.


Reply to: