[PATCH] PowerPC: Extend fpu fenv operations to operate on 64-bit FPSCR

[Ryan S. Arnold]

Greetings,

The included patch provides the following changes to extend the FPSCR
(floating point status and control register) to 64-bits for PowerPC
POWER6 and POWER6x:

o For PowerPC in general it redefines fpu_control_t from an unsigned
long int to an unsigned long long int so that the entire 64-bit FPSCR on
POWER6 can be saved/restored.  On non-POWER6 hardware the high-order
32-bits are simply discarded/ignored for all operations.

  It is necessary to change it for all PowerPC because the loader is
built as the default PowerPC arch (i.e. non-POWER6) and uses the
rtld_global_ro struct which contains an fpu_control_t and therefore must
be able to load a libc built for POWER6 or non-POWER6 (which also uses
the rtld_global_ro struct).

  Any difference in size of fpu_control_t between the loader and libc
will cause libc to access rtld_global_ro struct members at the wrong
offsets from where the structure was populated by the loader.

  The impact is minimal since currently the fpu_control_t is initially
populated from a double anyway and the fpu_control_t type is not used
dynamically in any external function and thus, no symbol versioning is
required.

o Provides a conditional implementation of _FPU_GETCW and _FPU_SETCW
which operate on either the entire 64-bit FPSCR on POWER6[x] or simply
the low-order 32-bits of the fpu_control_t on non-POWER6[x].

o Provides granular control over the high order 32 bits of the FPSCR
that aren't reserved.

o Provides an overridden sysdeps/powerpc/math/test-fpucw that verifies
that the high-order word is saved to the FPSCR and restored using the
_FPU_[GET|SET]CW macros.  It also verifies that there are no deleterious
effects on non-POWER6 systems.

o Provides conditional definitions of the PowerPC fenv helper macros
(fesetenv_register & fegetenv_register) which are used by the fenv
functions to set/get the contents of the FPSCR.

o Provides sysdeps/powerpc/math/test-powerpc-fenv which tests the
high-order word save/restore using the fe[set|get]env_register
functions.

o Replaces the FPSCR restore with a macro which will provide a 64-bit
restore on POWER6 for swapcontext and setcontext.

o Replaces some of the magic number fenv masks in the fenv functions
with a define that better describes what the magic number means.

I successfully executed build and runtime tests (make check) on POWER5
(32-bit FPSCR) as well as POWER6 (64-bit FPSCR).

Ryan S. Arnold
IBM Linux Technology Center
Linux Toolchain Development
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 64bit-fpscr.patch
Type: text/x-patch
Size: 28059 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20071022/db4d3f3b/attachment.bin>