https://bugs.kde.org/show_bug.cgi?id=406256
--- Comment #1 from Carl Love <c...@us.ibm.com> --- Created attachment 119258 --> https://bugs.kde.org/attachment.cgi?id=119258&action=edit Proposed fix for making the subnormal results track the VSCR[NJ] bit The attachment is a proposed fix for the issue. The Iops to implement the various vector floating point instructions map to a subset of the vector floating point instructions. The underlying host hardware needs to run with the VSCR[NJ] bit set to zero to generate the subnormal results. The guest state needs to then adjust the arguments/results of the vector floating point instructions as needed based on the guest setting of VSCR[NJ]. Hence the need to setup the host with VSCR[NJ] = 0. This was done in VEX/priv/host_ppc_defs.c, function getRRegUniverse_PPC (). The function is actually setting up the guest register state. The function is called once as part of the initialization process. That is the right time to set the host configuration. Ideally there would be a host initialization function that would be called for doing this but I don't see one. Given that there is not host specific function, I had to put the code into the guest setup function. I don't consider this to be the ideal place for the code. I would be interested in ideas on a more appropriate location for this code. ******* The host setup code requires the addition of the -mvsx and -maltivec command line options to be set for Power 7. These options are on by default when compiling for Power 8, 9. Hence the Makefile.all.am change. Finally, the current regression tests do not seem to cover the subnormal cases well enough. I created an explicit subnormal test which runs with the VSCR[NJ] bit set to zero and one. Again, would appreciate feedback on where best to do the host initialization. -- You are receiving this mail because: You are watching all bug changes.
