Right now, the compiler has no special code to optimize pow (x, 0.75) into
something like sqrt (x) * sqrt (sqrt (x)) under -ffast-math, nor pow (x, 0.25)
into sqrt (sqrt (x)). On machines with a builtin sqrt instruction, it is often
times faster to do the calculations using sqrt than calling the pow function.
In particular, x**0.75 shows up in the bwaves spec 2006 benchmark.
For sqrt (sqrt (x)) vs pow (x, 0.25), I see:
IBM power6: 9.2 times faster
Intel core2 laptop: 5 times faster
AMD K8 system: 3.2 times faster
For sqrt (x) * sqrt (sqrt (x)) vs. pow (x, 0.75), I see:
IBM power6: 6.1 times faster
Intel core2 laptop: 3.4 times faster
AMD K8 system: 2.3 times faster
In addition, the compiler optimizes sqrt (sqrt (x)) into pow (x, 0.25), and
similar optimizations.
This should be fixed by adding a hook to say what the relative speed of sqrt
and cbrt vs. pow, so that the backend can control whether or not this
optimization should be done. By default, the optimization is probably only
useful if -Os is used on machines that have a hardware sqrt instruction.
--
Summary: Compiler could optimize pow (x, 0.75) into sqrt (x) *
sqrt (sqrt (x))
Product: gcc
Version: 4.5.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: rtl-optimization
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=42694
