Nicholas Clark's "Improve Perl 5" grant work
Tue, 07-Mar-2006 by
Curtis "Ovid" Poe
edit post
<![CDATA[
<p>I have just received the first report for Nicholas Clark's grant. So far, things are going rather well. Read on to see what Nicholas has done.</p>
<hr/>
<p>I've completed the planned typeglob (PVGV) structure shrinking. This turned
out to be somewhat more complex than I expected. Specifically:</p>
<ul>
<li>Discovered that PVGVs use magic to perform stringification. Removed this
need for magic, saving 24 bytes.</li>
<li>Discovered that pads use PVGVs for storing C<our> variable names. Changed
C<our> variables to be stored in PVMGs</li>
<li>Moved the GP pointer to the SV head structure, saving 1 level of
dereferencing (and 4 bytes)</li>
<li>Managed to merge in all the other fields of PVGVs into the fixed part of SV
structure common with all the smaller SVs. (saves 16 bytes)</li>
<li>Change from storing the GV's name in a string to using a sharing the hash
key structure used by the symbol table holding this typeglob. (saves the
length of the typeglob name) [sizes for systems with 4 byte pointers and longs,
such as x86 and default PPC]</li>
</ul>
<p>I think this roughly halves the size of PVGVs. For an example, the
<tt>installman</tt> script in the perl source tree uses over 1200 PVGVs, so
this saves about 50K.</p>
<p>PVLVs had no magic associated, but benefit from all the other savings.</p>
<p>I've completed the planned PVCV and PVFM shrinking. Both are now about 25%
smaller (eg 20 bytes). <tt>installman</tt> uses 600 PVCVs, so this saves about
12K</p>
]]>
Comments (1)
Is anything known about performance improvements yet? I'm curious if there's extra speed in these changes as well as memory savings.
