Some additional remarks:
The "memset" in smallobject.c is not necessary on Linux. mmap() (which
obviously gets called for memalign - at least for this arena size)
does clear the memory. We need some tests, from which size memory is
cleard for malloc and memalign.
I tossed the memset for now and saved ~450.000 L2-misses or ~0.2 s.
I did some optimizations in list.c to avoid generating sparse lists:
When its clear that the whole list gets filled the programmer/compiler
shall insert a
set P0, I0 # set size of list
before setting the first element. The rules, which grow_type is chosen
when are straightened and better documented (in list.c).
So with these two refinements, I have new numbers for some stress tests:
stress stress1 stress2 life
CVS 1.00 1.44 721
SPMC 0.60 12.0 1.50 793
my current 0.33 8.8 1.24 800
perl 5.8.0 0.6 12.0 2.41
stress1 does 10 times the (10+20+20) allocations of 200.000 elements.
I'll check it in soon.
stress tests are in seconds, life test is generations/sec, -O3
compiled parrot, JIT runtime (-P isn't slower here), i386/linux.
SPMC ... the patch with smaller (-8 bytes) PMC
current ... additionally DOD flags moved to arena + above
Have fun,
leo
Acknowledgments: valgrind and ccache are really great tools. Get them
if you don't already have them.
