float behaivior on ARMv8-A

Andrew Haley aph@redhat.com
Mon May 2 09:40:00 GMT 2016


On 01/05/16 18:46, N30dG wrote:
> The Problem is that Task's from my ARMv8-App-Version(tryed 32bit and 64bit)
> are mostly above the threshold of the Validator. Compiling the App with the
> same Settings for ARMv7 on the same Device delivers Valid results. I've
> tryed this with many different GCC-versions(4.9.3; 5.3.0; 6.0.1),
> lib-versions, with and without using of NEON. The result is always the same:
> Compiling the App with -march=ARMv7-A gives Valid results, Compiling with
> -march=ARMv8-A gives me mostly invalid results(not all).
> I've talked to a Project-developer from the Einstein-Project and he say's me
> that my ARMv8-results are slightly(not much) different from the ones
> compiled for ARMv7. This occours on AARCH32 and AARCH64 Versions. There must
> be somthink changed in the float-handling when compiled for ARMv8-A but I
> really don't know?

This is very difficult.  You're asking for someone to debug this, but it
will take some work to investigate it.  We don't know if the program is
correct or what the results should be.  Sure, it might be a GCC bug, but
the only way to find out is to debug it: to look at a few sample results,
see by how much they differ, and figure out why.  We know that this kind
of thing usually is not a GCC bug, but it might be.

Andrew.



More information about the Gcc-help mailing list