The current powerpc -mrecip can be improved in several areas.
1) The compiler generates a conditional test to deal with NaN, etc. even though
it is only used with -ffast-math, where the programmer declares that he/she is
not interested in Nan/infinity/etc. This causes the code to single thread the
approximation, and does not allow the power7 to issue two streams of FRSQRTES
and fixup in parallel.
2) Right now, -mrecip only enables single precision reciprocal square estimate.
The compiler should be modified so that it also supports double precision
reciprocal estimate on newer machines (power6/power7) that have a high enough
precision FRSQRTE instruction.
3) In addition to optimizing 1/sqrt(x), the compiler should be able to optimize
normal divisions using the FRE/FRES instructions. Machines with higher
precision FRE instructions should do fewer passes in the fixup than the older
machines.
4) On altivec and VSX systems, the compiler should enable auto vectorization of
the estimate instructions and fixup.
5) The compiler should provide more builtins to allow the user to access to
double reciprocal square root, as well as vector versions.
--
Summary: Improve powerpc -mrecip support
Product: gcc
Version: 4.6.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: target
AssignedTo: meissner at gcc dot gnu dot org
ReportedBy: meissner at gcc dot gnu dot org
GCC build triplet: powerpc64-unknown-linux-gnu
GCC host triplet: powerpc64-unknown-linux-gnu
GCC target triplet: powerpc64-unknown-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44218
