<br><br><div class="gmail_quote">On Sun, Feb 22, 2009 at 10:37 AM, <span dir="ltr"><<a href="mailto:bryce@kampjes.demon.co.uk">bryce@kampjes.demon.co.uk</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div class="Ih2E3d"><br>
Eliot Miranda writes:<br>
><br>
> But what I really think is that this is too low a level to worry about.<br>
> Much more important to focus on<br>
> - context to stack mapping<br>
> - in-line cacheing via a JIT<br>
> - exploiting multicore via Hydra<br>
> and beyond (e.g. speculative inlining)<br>
> than worrying about tiny micro-optimizations like this :)<br>
<br>
</div>If you're planning on adding speculative, I assume Self style dynamic,<br>
inlining won't that reduce the value of context to stack mapping?</blockquote><div><br></div><div>Not at all; in fact quite the reverse. Context to stack mapping allows one to retain contexts while having the VM execute efficient, stack-based code (i.e. using hardware call instructions). This in turn enables the entire adaptive optimizer, including the stack analyser and the bytecode-to-bytecode compiler/method inliner to be written in Smalltalk. The image level code can examine the run-time stack using contexts as their interface without having to understand native stack formats or different ISAs. The optimizer is therefore completely portable with all machine specificities confined to the underlying VM which is much simpler by virtue of not containing a sophisticated optimizer (which one would have to squeeze through Slang etc).</div>
<div><br></div><div>So for me, context-to-stack mapping is fundamental to implementing speculative inlining in Smalltalk.</div><div><br></div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
My view with Exupery is context caches should be left until after<br>
dynamic inlining as their value will depend on how well dynamic<br>
inlining reduces the number of sends.<br></blockquote><div><br></div><div>I know and I disagree. Dynamic inlining depends on collecting good type information, something that inline caches do well. In-line caches are efficiently implemented with native call instructions, either to method entry-points or PIC jump tables. Native call instructions mesh well with stacks. So context-to-stack mapping, for me, is a sensible enabling optimization for speculative inlining because it meshes well with inline caches.</div>
<div><br></div><div>Further, context-to-stack mapping is such a huge win that it'll be of benefit even if the VM is spending 90% of its time in inlined call-less code. We see a speedup of very nearly 2x (48% sticks in my head) for one non-micro tree walking benchmark from the computer language shootout. And this is in a very slow VM. In a faster VM context-to-stack mapping would be even more valuable, because it would save an even greater percentage of overall execution time.</div>
<div><br></div><div>Further still using call & return instructions as conventionally as possible meshes extremely well with current processor implementations which, because of the extensive use thereon of conventional stack-oriented language implementations, have done a great job optimizing call/return.</div>
<div><br></div><div>Further still, the current performance of call/return on contemporary processors, specifically prefetch across call & return (prefetch across return only possible if one sticks to the processor's expected stack organization of return addresses) renders call/return performance the same as jumps. So the benefits of inlining are no longer in eliminating call/return, but rather in eliminating dispatch, argument copying, etc. So inlining per se isn't of benefit. It can actually worsen instruction cache density. Analysis and elimination of dispatch is. So again context-to-stack mapping makes sense because it means the speculative inliner/adaptive optimizer doesn't have to focus on creating humongous methods or inlining accessos etc etc, and can focus on higher level optimizations like block removal (lambda lifting?), common subexpression elimination, and so on. </div>
<div><br></div><div>best</div><div>Eliot<br> </div></div><br>