Memory savings with -fcow
B::C has now better support for copy-on-write (COW) strings with about 6% memory savings for 5.20 and 5.22.
The perl5.18 implementation for COW strings is totally broken as it
uses the COW REFCNT field within the string. You cannot ever come to a
true successful copy-on-write COW scheme. You cannot put the string
into the .rodata segment as with
static const char* pv = "foo"; it
needs to be outlined as
static char* pv = "foo\000\001";. The byte
behind the NUL delimiter is used as REFCNT byte, which prohibits its
use in multi-threading or embedded scenarios. In cperl I'm was working
on moving this counter to an extra field, but the 2 authors made it
impossible to write it in a maintainable way. I could easily seperate
the refcnt flag but I couldn't make it COW yet.
But even if the COW implementation in the libperl run-time is broken by design it still can be put into good use to store more strings statically than expected. The problem was that since 5.18 and with this COW feature binaries needed 20% more memory, as I couldn't save the strings statically anymore and had to allocate them dynamically.
In the first attempt I save some kilobytes memory by removing the IsCOW flag and store more strings statically.
But now I do the opposite. I set the
IsCOW flags on much more
strings since 5.20 and -O2, store it not as
const char* to be able
up update the cow refcnt, and rely in the automatic cow and uncow
functions in the runtime to move this static buffer to the heap when
being written to, and don't need to rely on
LEN=0 anymore, which
indicates a normal static string.
With a typical example of a medium sized module, Net::DNS::Resolver, 64bit not threaded, the memory usage is now as follows:
pcc -O0 -S -e'use Net::DNS::Resolver; my $res = Net::DNS::Resolver->new; $res->send("www.google.com"); print `ps -p $$ -O rss,vsz`' pcc -O3 -S -e'use Net::DNS::Resolver; my $res = Net::DNS::Resolver->new; $res->send("www.google.com"); print `ps -p $$ -O rss,vsz`' rss without -fcow: 12832 with -fcow : 12112 cperl : 12532
6% percent memory win for 5.22. Even better than with cperl.
The current distribution of .rodata, .data and dynamic heap strings with this example is as follows:
.rodata .data heap -fno-cow (-O0): 305 1945 1435 -fcow (-O3): 110 2225 1024 cperl -O3: 107 2112 1001
Thus with -O3 we traded 40% less dynamic strings for 3x less .ro strings, but 14% more static strings. With cperl the improvements are no so dramatic, as cperl already has much more static optimizations already.