Wait, does that mean the corresponding constants in base for 7.8.3 are busted? On Saturday, July 19, 2014, Jost Berthold < berthold@mathematik.uni-marburg.de> wrote:

Message: 2

...

Date: Sat, 19 Jul 2014 15:00:17 +1000 From: Kyle Van Berendonck To: "ghc-devs@haskell.org" Subject: Thread status constants Message-ID: Content-Type: text/plain; charset="utf-8"

Hi,

I found these: https://github.com/ghc/ghc/blob/5f3c5384df59717ca8013c5df8d1f6 5692867825/includes/rts/Constants.h#L194

They go only 0-14, so there's some long chains of branches and stuff in hot paths that could be cleaned up into single &-masked branches by changing these into a set of flags.

And then I saw these: https://github.com/ghc/ghc/blob/master/libraries/base/ GHC/Conc/Sync.lhs#L483

Where does 16 and 17 come from -- I couldn't find them in the header files anywhere?

Kyle

Hi Kyle,

I had a look at these constants just recently.

About 16 and 17, see here: https://github.com/ghc/ghc/blob/master/rts/PrimOps.cmm#L853

About the others:The block reasons received two updates just recently. Turns out that the eventlog (GHC's post-mortem tracing for threadscope et al.) expects these states to have particular values (those in includes/rts/Constants.h, consider them a fixed ABI for the time being). See here for the story: https://ghc.haskell.org/trac/ghc/ticket/9003

NB: Darn... base still uses GHC-7.8.2 constants. Thanks for the pointer!

About your proposal to use bit masks:

Did you find some of those hot paths? I believe (believe, unchecked...) that there are not too many places where a complete switch over all possible reasons is done (apart from debug output maybe); the typical use is rather to compare to one single constant than to all. So the performance might not increase too much from optimisations. Again, just a hunch, not verified.

If the RTS contains code specialised to the current constants, these places need to be kept symbolic, to avoid future breakage when new states are added.

A fix which removes the tight implicit dependency of ghc-events from the constants in includes/rts/Constants.h is pending, btw.

/ Jost

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

Yeah, it's not like a big deal or anything. Here's all the places which grep told me use chains of || (or long switch statements which can't be turned into tables) to compare these values: ghc/rts//win32/AsyncIO.c:280 ghc/rts//Threads.c:286 ghc/rts//Threads.c:257 ghc/rts//sm/Scav.c:73 ghc/rts//sm/Sanity.c:521 ghc/rts//sm/Compact.c:436 ghc/rts//Schedule.c:2617 ghc/rts//Schedule.c:2605 ghc/rts//Schedule.c:946 ghc/rts//RaiseAsync.h:54 ghc/rts//RaiseAsync.c:416 I didn't include all the ones in the profiler because they probably don't get hot. Kyle On Sun, Jul 20, 2014 at 4:43 AM, Jost Berthold < berthold@mathematik.uni-marburg.de> wrote:

Message: 2

...

Date: Sat, 19 Jul 2014 15:00:17 +1000 From: Kyle Van Berendonck To: "ghc-devs@haskell.org" Subject: Thread status constants Message-ID: Content-Type: text/plain; charset="utf-8"

Hi,

I found these: https://github.com/ghc/ghc/blob/5f3c5384df59717ca8013c5df8d1f6 5692867825/includes/rts/Constants.h#L194

They go only 0-14, so there's some long chains of branches and stuff in hot paths that could be cleaned up into single &-masked branches by changing these into a set of flags.

And then I saw these: https://github.com/ghc/ghc/blob/master/libraries/base/ GHC/Conc/Sync.lhs#L483

Where does 16 and 17 come from -- I couldn't find them in the header files anywhere?

Kyle

Hi Kyle,

I had a look at these constants just recently.

About 16 and 17, see here: https://github.com/ghc/ghc/blob/master/rts/PrimOps.cmm#L853

About the others:The block reasons received two updates just recently. Turns out that the eventlog (GHC's post-mortem tracing for threadscope et al.) expects these states to have particular values (those in includes/rts/Constants.h, consider them a fixed ABI for the time being). See here for the story: https://ghc.haskell.org/trac/ghc/ticket/9003

NB: Darn... base still uses GHC-7.8.2 constants. Thanks for the pointer!

About your proposal to use bit masks:

Did you find some of those hot paths? I believe (believe, unchecked...) that there are not too many places where a complete switch over all possible reasons is done (apart from debug output maybe); the typical use is rather to compare to one single constant than to all. So the performance might not increase too much from optimisations. Again, just a hunch, not verified.

If the RTS contains code specialised to the current constants, these places need to be kept symbolic, to avoid future breakage when new states are added.

A fix which removes the tight implicit dependency of ghc-events from the constants in includes/rts/Constants.h is pending, btw.

/ Jost