Thanks for the prompt replies. Please read no further, PEBKAC: I was on an old compiler on the machine I was running on (I had updated the compiler on my desktop, but clearly not the Pi I was testing this on - now re-building with latest).
In truth this was a problem I came across a few weeks ago and meant to post about it at the time, but was busy - now clearing my task backlog before Christmas prompted this post. So I had to try and reproduce the issue upon request. Upon doing so, I actually got the opposite behaviour: that adding an if true/if false around code was faster than commenting. This made me think that I had something basic wrong! So yes, I agree that I must be changing things like how the code was allocated in the cache and other micro-changes. I'm going to take this as an I have pushed that algo far enough and I need to look for improvements elsewhere when I get to this last few percent stage that I was here. Regards & Apologies Chris On Tuesday, 19 December 2017 11:40:36 UTC, Jan Mercl wrote: > > On Tue, Dec 19, 2017 at 12:05 PM Chris Hopkins <cbeho...@gmail.com > <javascript:>> wrote: > > ... > > > was slower than: > > ... > > Without more information the conclusion that the optimization is not there > cannot be established. Inspect the real produced machine code. I guess > you'll find that the machine code differs in cache-line alignment only, and > that's the cause of the observed effect. > > Even changing whitespace in a program can potentially move the code in the > binary. > > > So I assume the Go compiler doesn't try and do this analysis in a > Compile time vs Execution time tradeoff. > > I think this kind of DCE was in at the date of the first public release. > > > Anyway Hope that is useful to someone. It has certainly changed the way > I will write my (profiled) inner loops. > > I don't think that's a good idea. The bug, if any, in the compiler should > be fixed. OTOH, adjusting code to a particular compiler/version/whitespace > will hurt as soon as anything of that changes. > > Generally, benchmarking short inner loops is often unreliable enough to > render it unusable. > > -- > > -j > -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
