I mean, it certainly *seems* reasonable a 15% hit could come from pipelining changes or cache behavior or something. I don't think alignment would really be a huge issue; post-Nehalem I believe non-aligned writes/reads are extremely cheap. Non-intuitive behavior can totally happen too: I've seen cases of adding instructions to a loop which speeds things up (e.g. by taking the extra step, you may mitigate a dependency stall, which massively helps pipelining across the loop body etc.)
Nicolas, can I ask what benchmark you're looking at? And what performance tools are you using, Intels'? If you're on Linux, the 'perf' tool on a modern kernel can be used to quickly get an overview of how many cache misses/hits your process has, how many pipeline stalls occur, etc. You can then use it to drill down a bit into the assembly that's problematic. That might not give you an exact culprit (it could be many changes and accumulative hits,) but it's a start. On Wed, Jun 19, 2013 at 10:43 AM, Nicolas Frisby <[email protected]> wrote: > I'm also seeing performance regressions in the shootout benchmarks that I > can't identify in the asm. The new asm looks better but performs worse, with > a ~15% slowdown. > > I fired up the performance counters in my CPU and the free Intel code for > inspecting them showed that my CPU utilization took about a 10% hit, even > while executing fewer total instructions. > > 1) Jan, perhaps we're seeing the same sort of behavior — the shootout > benchmarks have extremely hot loops (hundreds of millions of iterations > IIRC). I used ticky profiling too, and saw no suspicious changes in any > counters. > > 2) Dear Low-level Gurus: How feasible is it that a ~15% slowdown in a > program with a very hot loop is due to incidentally inhibiting some caching > behavior (instr? data?)? Or perhaps effecting alignment? FTR my CPU is a > Core i7-2620M, Sandy Bridge. > > Thanks all. > > On Wed, Jun 19, 2013 at 9:27 AM, Jan Stolarek <[email protected]> > wrote: >> >> > If it's not sorted out, can you open a ticket, put in the relevant info >> > (so >> > we don't need to look at the email trail), and we can tackle it when you >> > get here. >> Currently there's a temporary workaround: I'm using new folding rules for >> all primitive types, >> except for Integer, in which case I left the old folding rules unchanged. >> This of course should >> be modified to make all rules uniform, but for now it at least passes >> validation. I didn't fill >> the ticket, because the bug does not exist yet :) It only manifests itself >> in my patches, which >> have not been applied yet. I'll add all the information from this >> discussion to my github fork of >> GHC and then move it to Trac once the bug makes it to HEAD. >> >> What worries me more about my patches is the performance regression in >> kahan, because I see no >> obvious differences in the generated assembly. >> >> Janek >> >> > >> > Simon >> > >> > -----Original Message----- >> > From: [email protected] [mailto:[email protected]] >> > On >> > Behalf Of Jan Stolarek Sent: 20 May 2013 12:35 >> > To: Ian Lynagh >> > Cc: [email protected] >> > Subject: Re: Integer constant folding in the presence of new primops >> > >> > > If you remove everything but the quotInteger test from >> > > integerConstantFolding and compile with -ddump-rule-rewrites then >> > > you'll see that the eqInteger rule fires before quotInteger. This is >> > > presumably comparing against 0, as the definition of quot for Integer >> > > (in GHC.Real) is >> > > _ `quot` 0 = divZeroError >> > > n `quot` d = n `quotInteger` d >> > >> > Yes, I noticed these two rules firing together - perhaps that's the >> > explanation why. I created a small program for testing: >> > >> > main = print quotInt >> > quotInt :: Integer >> > quotInt = 100063 `quot` 156 >> > >> > I noticed that when I define eqInteger wrapper to be NOINLINE, the call >> > to >> > quot is translated to Core as: >> > >> > Main.quotInt = >> > GHC.Real.$fIntegralInteger_$cquot >> > (__integer 100063) (__integer 156) >> > >> > but when I change the wrapper to INLINE I get: >> > >> > Main.quotInt = >> > GHC.Real.$fNumRatio_$cquot <-------- NumRatio instead of >> > IntegralInteger (__integer 100063) (__integer 156) >> > >> > All rule firing happens later (I used -ddump-simpl-iterations >> > -ddump-rule-firings), except that for $fNumRatio_$cquot the quot rules >> > don't fire. >> > >> > > Do you also still have eqInteger wired in? It sounds like you might >> > > have given them both the same unique? >> > >> > No, they didn't have the same unique. I modified the existing rules to >> > work >> > on the new primops and ignore their wrappers. At the moment I reverted >> > these changes so that I can make progress and leave this problem for >> > later. >> > >> > Janek >> > >> > _______________________________________________ >> > ghc-devs mailing list >> > [email protected] >> > http://www.haskell.org/mailman/listinfo/ghc-devs >> >> >> >> _______________________________________________ >> ghc-devs mailing list >> [email protected] >> http://www.haskell.org/mailman/listinfo/ghc-devs > > > > _______________________________________________ > ghc-devs mailing list > [email protected] > http://www.haskell.org/mailman/listinfo/ghc-devs > -- Regards, Austin - PGP: 4096R/0x91384671 _______________________________________________ ghc-devs mailing list [email protected] http://www.haskell.org/mailman/listinfo/ghc-devs
