When scanning an array of float values for minima and maxima, among other
tasks, the compiler will correctly use "minss" instruction, however it does
this by:
movaps %samplereg,%tempreg
minss %minreg,%tempreg
movaps %tempreg,%minreg
This could be done simply as:
minss %samplereg,%minreg
without a need for a temporary register, and associated delay slot shadows.
Code is roughly:
{
float minreg = ...
float maxreg = ...
float sumreg = 0.0;
int sumcount = 0;
loopconstruct {
float samplereg = source[idx];
minreg = (minreg > samplereg) ? samplereg : minreg;
maxreg = (maxreg < samplereg) ? samplereg : maxreg;
sumreg += samplereg;
++sumcount;
}
// ... use results
}
Oddly, addition to sumreg is done without above mentioned register shuffles.
Furthermore, math-library function fminf()/fmaxf() (and fmin()/fmax() for
double) would benefit from map to intrinsic minss/maxss processing. Now they
cause math library calls, where they are implemented as minss/maxss.
Another optimization adventure would be to be able to unroll that loop, and use
packed float values in xmm registers to do up to 4 operations in parallel.
minreg/maxreg/sumreg could be described at C level as:
float minreg[4];
and code would have an explicit loop from 0 to 3 processing sample sets.
--
Summary: floating point optimizations needlessly shuffle xmm
registers
Product: gcc
Version: 4.4.2
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: matti dot aarnio--gcc-bugs at zmailer dot org
GCC build triplet: x86_64-redhat-linux
GCC host triplet: x86_64-redhat-linux
GCC target triplet: x86_64-redhat-linux
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42682
