Dear <span name="Henrique Santos" class="gD">Henrique,<br><br>Thanks for reply, I think I saw this project before. We have an LLVM-based GPU toolchain, where variable size is usually one of the major performance factors.<br>
<br>- D.<br></span><br><div class="gmail_quote">2013/5/7 Henrique Santos <span dir="ltr"><<a href="mailto:hnsantos@gmx.com" target="_blank">hnsantos@gmx.com</a>></span><br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<span style="font-family:Verdana"><span style="font-size:12px"><div>A few colleagues of mine implemented an inter-procedural range analysis algorithm for LLVM 2.7 a few years back.</div><div>Their paper (see <a href="http://homepages.dcc.ufmg.br/~fernando/publications/papers/SBLP12_victor.pdf" target="_blank">http://homepages.dcc.ufmg.br/~fernando/publications/papers/SBLP12_victor.pdf</a>) states a 13-14% reduction in variable size</div>
<div>for the SPEC CPU2006 benchmarks and 36% for the Stanford benchmarks.</div><div>Check out <a href="http://code.google.com/p/range-analysis/" target="_blank">http://code.google.com/p/range-analysis/</a> if you're interested and feel free to email me if you have any questions.</div>
<div> </div><div>H.</div><div> <blockquote style="border-left:1px solid #ccc;padding-left:5px;margin-left:5px;margin-bottom:0px;margin-top:0px;margin-right:0px" type="cite"><p style="margin:0px;padding:0px"><span style="font-family:Verdana"><span style="font-size:12px">----- Original Message -----</span></span></p>
<p style="margin:0px;padding:0px"><span style="font-family:Verdana"><span style="font-size:12px">From: Dmitry Mikushin</span></span></p><p style="margin:0px;padding:0px"><span style="font-family:Verdana"><span style="font-size:12px">Sent: 05/07/13 02:48 AM</span></span></p>
<p style="margin:0px;padding:0px"><span style="font-family:Verdana"><span style="font-size:12px">To: LLVM Developers Mailing List</span></span></p><p style="margin:0px;padding:0px"><span style="font-family:Verdana"><span style="font-size:12px">Subject: [LLVMdev] Any LLVM passes for integers type compaction?</span></span></p>
<div><div class="h5"> <div><p style="margin:0px;padding:0px">Dear all,<br><br>I'm wondering if anybody is working on optimization passes that could use smaller integer types for variables that are known to have values within smaller type range?<br>
<br>Thanks,<br>- D.</p></div></div></div></blockquote><p style="margin:0px;padding:0px"> </p></div></span></span>
</blockquote></div><br>