Advogato blog for chromatic

Windows Patches Welcome; Windows Attitude Not

Wed, 14 May 2008 03:13:14 GMT

I just love hearing "You're doing it wrong, and you're not doing enough because you don't care about exactly the same things I care about. Also you have a bad attitude."

Going Faster By Doing Less Work

Wed, 14 May 2008 03:13:14 GMT

Today's Rakudo-building speedup is 27.79%.

(Okay, the other slow part of Rakudo builds 17.59% faster. Still.)

The profile showed that string allocation was a hotspot in the benchmark. In particular, the part of Parrot which allocates memory out of arenas spent a lot of time performing garbage collection. Every time you can avoid either allocating unnecessary memory or running a full garbage collection, you can improve performance.

Parrot r27484 adds one line of code (and one line of comment).

Every time mem_allocate() successfully allocates a new block, it increments a counter. Whenever the garbage collector performs a full run, it resets that counter to zero.

This patch performs a garbage collection run from mem_allocate() only if that counter is non-zero. That is, if the garbage collector has already run, it's already found as much free arena memory as possible. (This is not memory for PMCs or STRING headers; this is buffer memory.) Running the GC again won't find any more free buffer memory. In that case, skipping the GC run and allocating more memory from the OS gives the performance improvement.

I should note that my comments about avoiding memory allocation apply in the general case. Parrot's current GC has some limitations. The biggest is that it stops the world to mark and sweep everything. The new GC Andrew Whitworth will implement as part of the Google Summer of Code will fix that. As well, we have some ideas to improve the implementation such that the GC will become even less expensive now. Then we'll see algorithmic improvements that make even this 27.79% optimization seem small.

Doing Nothing Better in Perl 5

Sat, 10 May 2008 08:11:13 GMT

Perl 6 has three code placeholder operators, known affectionately as the "yada, yada, yada" operator (see List Prefix Precedence in Synopsis 3). It's a matter of (very sarcastic) public record how much I love writing, maintaining, and patching parsers, so I've just sent a very preliminary five-line patch to p5p to add support for ... to Perl 5.

--- perly.y~ 2008-05-09 17:47:35.000000000 -0700 +++ perly.y 2008-05-09 17:47:41.000000000 -0700 @@ -1227,6 +1227,11 @@ } | WORD | listop + | DOTDOT + { + $$ = newUNOP(OP_DIE, 0, + newSVOP(OP_CONST, 0, newSVpvs("Unimplemented"))) + } ; /* "my" declarations, with optional attributes */

Apply this to recentish bleadperl sources, run perl regen_perly.pl, rebuild, and now you can run programs such as:

sub foo { ... } foo();

And get an "Unimplemented at file line line." error message.

(Now everyone who complains that I don't code enough to match my talk, please punch yourself in the face.)

Perl 6 Design Minutes for 07 May 2008

Fri, 9 May 2008 03:13:19 GMT

The Perl 6 design team met by phone on 07 May 2008. Larry, Allison, Patrick, Will, Jerry, Nicholas, Jesse, and chromatic attended.

Allison:

spent my time this week slicing and dicing the exceptions implementation
replaced the old internals with the new system
checked that in yesterday
still a few failing tests in edge cases on the branch
did more work on the Parrot Foundation

I own an acre of Mars, we could incorporate there

Allison:

don't you own a cow in the Philippines?

yes, but that doesn't give me any governmental powers

Patrick:

isn't that worth a lot?

the peso is improving against the dollar

Jesse:

moving on...

Larry:

clear bill of health from my medical reports
hacking a lot against the standard grammar in my STD5 implementation
lots of refactoring
all of the various parameters that used to go through separately now go through as part of the Match object
including the "fate" and whether we're peeking at the longest token set
the longest token matcher works now
I threw out my old mechanism for gathering Match objects
it now creates the more-or-less correctly
lots of grammar tweaks, as suggested by Mitchell Charity
lots of refactoring of how logging works so that it doesn't always spew enormous quantities of information to the screen
I can actually run the parser quite quickly now, for some definition of quickly that approximates 2000 characters per second
matches symbols directly, rather than calling a rule, which is faster
does the backoff now on longest token matching
started refactoring the grammar on the assumption that I cna trust the longest token matcher
no longer any nofat rule
the longest token should match the fat arrow, if there is one
started refactoring the quoting rules to parse as if they were sublanguages
getting rid of extra rigamarole to recreate the other mechanism we already use for other languages
working out the linkage for switching in and out of sublanguages
how to get to the outer language from the inner language
calling into pure Perl from closures in a regex
or the host language if you're calling the regex from another language
nailed down the available methods for Match objects in the specs
giving a talk in Seattle on Friday at SPU
flying to Japan on Saturday

Patrick:

spent a lot of time teaching this past week
cleared up now
mostly I've answered questions on mailing lists and IRC
I'm not always sure that I'm helpful, but I'm there
yesterday I worked on trying to get a bunch of little small things here and there
fixed up a few things in PCT internals
today I'm bringing PGE up to date with some of the latest changes in S05
these all help Rakudo and other languages in small ways
trying to clean out my backlog and clean up a bunch of RT tickets
I'll continue over the next couple of days
and blogging about it as I go

Jerry:

things are busy, mostly non-Parrot related stuff
submitted a ticket that I hope Patrick can close today

Patrick:

many languages depend on the old behavior, including Plumhead
I'm not certain about some of them

Jerry:

mostly otherwise answering questions on #parrot
making sure that things are set up for the real work phase of GSoC
making sure that students have their CLAs, if not commit bits
astonished to see how much work Jonathan is getting done in just two funded days
it's amazing to see how much a motivator money can be
I'd like to see more of it, hint hint

working on closing as many open Parrot bugs as possible
applying as many open patches as possible
should be able to help on the concurrency branch soon
otherwise preparing for the release
going to check on received CLAs this week

Nicholas:

found it curious that Perl 5.10 has the best state implementation of any language
wanted to steal tests from another implementation
had a discussion with Leon about SMOP
there's no real description of how all of these implementations fit together
Rakudo plus Parrot is a complete implementation
SMOP and kp6 fit together nicely

Jesse:

I started a wiki page on the Perl 6 wiki at Perl_6_Implementations

Patrick:

I don't know that it says how things fit together

Jesse:

I tried to encourage other people to contribute stuff
didn't get much uptake

Nicholas:

should we suggest to Daniel that he should help explain things?

Jesse:

that's more likely to get people contributing to it

Will:

there's definitely some confusion about it within the Grants Committee

Jerry:

SMOP has the highest documentation-to-line-of-code ratio of any implementation

Patrick:

it needs a good overview though

Nicholas:

I'll ask Daniel to explain more
especially its relationship to Parrot and Rakudo

Jerry:

it sounds like it could be an alternate runcore for Perl 5 as well

Jesse:

tried a few different things
decided to write a test for Rakudo
tried a simple arithmetic test pulled from Pugs
found that Rakudo didn't implement a function specified in the S29 draft
Patrick helped me write a couple of lines of code to implement it
then discovered that fudge didn't support try blocks in a specific way
Larry patched that
then found that incrementing an undefined value didn't work in Rakudo
that was the end of my day
I still need to write up my findings
how easy is it for someone without experience in Rakudo and its internals to pick things up and contribute something?
more difficult than I thought it might be, but it's getting more doable
it's important to understand how it might fail before trying to get people to do it
then I started trying to play with MAD on the weekend
found and fixed a bug in its XML
refactored it such that you can run MAD's tests in the core if you add a copy of XML::Parser to the core
it's not far enough yet, but it's a start

Nicholas:

is it going to be difficult to restructure the Parrot foundation from 501(c)(3) to 501(c)(6)?

Allison:

you can do pretty much the same thing
sponsors are on the board in a c6
they're only advisory in a c3
the sponsors we've talked to are mostly only interested in getting regular status reports and the like

Jesse:

is there any jumping around to transfer copyright to the new foundation?

Allison:

we'll do a copyright assignment from the Perl Foundation to the Parrot Foundation
all of the CLAs that went into the code up to the point of signover will be fine
but we'll essentially copy the Perl Foundation CLA to a Parrot Foundation CLA

Will:

do we need to contact committers who haven't signed a CLA?

Patrick:

where does Rakudo fall?

Allison:

still under the Perl Foundation
it doesn't move at all

Will:

do we want to split up the repository at that point?

Allison:

eventually, we'll want to do so anyway
it's not an urgent thing

Jerry:

what would it take to version a Perl6Regex frontend to PGE?
let grammars specify a version of the grammar

Patrick:

I did that before by having a separate compiler
you're talking about something a bit finer grained
I don't want to do that
as we get closer to 1.0, that'd be fine
I already have enough to do keeping up with the latest versions

Will:

I don't think we want to keep up old versions

Patrick:

I don't mind sticking to our deprecation cycle
I hadn't put the change from today into the deprecation list yet
we'll get to it in a couple of weeks

Jerry:

just trying to figure out how to push forward with changes to PGE without having to update every language in the repository

Patrick:

freeze S05?
not a great solution

Larry:

I heard that

Patrick:

the last few changes have been great
I'm not really serious about that

Larry:

some of them you even asked for

it's an advantage to have these languages in the repository
we can update them
but only if we can run the tests before and after and know that they pass

Will:

we might consider removing languages with failing tests and no recent updates
there are 17 grant proposals, some of them Perl 6-related
please comment on the TPF blog
it'll help

Jesse:

blog.perlfoundation.org

Good Error Messages are Important

Wed, 7 May 2008 06:09:33 GMT

Parrot r27355 was fun to write.

One of the persistent error messages Parrot emits for compiler writers is Null PMC access in invoke(). If you've had your hands deep in the guts of Parrot, you know what that means -- you tried to call a Sub PMC when you don't have a Sub PMC, you have no PMC. (If you don't know what that means, this entry is for you.)

Sometimes this means that there's a problem in Parrot. We've fixed almost all of those problems though, so the error usually comes from elsewhere. If you're writing a compiler, or running a compiler built on Parrot, the error usually means "You tried to call a function that doesn't exist."

Parrot's optimizer does something interesting at the end of compilation time. You've probably heard that Parrot's compiler, IMCC, translates PIR into PBC. That is, it turns source code into bytecode, which Parrot can either serialize t to disk or execute immediately. That bytecode is just a chunk of linear data in memory. It's not really a data structure. (Okay, it's a C array, but that doesn't make it a data structure.)

After IMCC has finished building a standalone chunk of bytecode, it performs a constant fixup phase. The notable part of this phase is that it edits the bytecode in place to replace all named invocations of functions known at fixup time with offset invocations.

The previous code looks something like:

invoke known_function null # padding null # padding

If IMCC has already seen known_function by this time, the direct invocation of known_function can continue. There's no runtime lookup necessary; all functions already compiled and ready are available in the bytecode.

If IMCC hasn't seen that function, runtime lookup is necessary, and so this function replaces the bytecode earlier with the equivalent of:

.local pmc func func = find_name 'unknown_function' invoke func

(I've simplified what actually happens slightly, because the concepts are more important than the details. Hopefully you see why the padding is necessary. If not, just imagine trying to splice additional opcodes into what may presumably be a lengthy C array -- like I said, barely a data structure.)

The problem with this second form occurs when find_name returns a NULL PMC, which it can legitimately do. In that case, the invoke opcode tries to invoke a NULL PMC and fails, and Parrot throws an exception saying "There's nothing here to invoke." There's the error message.

It's clear why that happens, but it's not useful. It would be more useful to see the name of the function you tried to invoke in the error message. Unfortunately, by the time Parrot calls the invoke op, that name is long gone.

My first idea was to rewrite the dynamic lookup form into something resembling:

.local pmc func func = find_name 'unknown_function' if defined func goto call_it die "Can't invoke 'unknown_function'" call_it: invoke func

Unfortunately, I didn't have the space in the bytecode stream to insert that many ops, and I had no desire to move chunks of memory around in that C array. I could have added more padding after an invocation, but to be fair I'm only mostly sure that it exists there in the first place.

I had room for one op with a destination PMC and a string constant argument. I added an experimental op called find_sub_not_null which does the same thing as find_name but throws an exception which includes the requested name if Parrot can't find a PMC of that name.

This isn't entirely an ideal situation. It's a special case op, and I prefer to remove ops where possible. It's also nearly code duplication, though it's effectively three lines of code in an op, which isn't awful. I still want to be able to perform these kinds of transformations in PBC itself, but we need a different way to generate PBC and perform op-level transformations in PIR before we can do this effectively.

There are always tradeoffs, though. Doing this check in C is slightly faster than doing it in PIR. The standard Perl 5 rule of optimization applies even in Parrot -- the fewer ops, the mostly faster you can go. As well, I was able to improve the warning message today, rather than at some point in the future when we have better PBC optimization possibilities.

After all, I can always remove this op in the future.

What are You Going to Do, SMS at Me?

Sun, 4 May 2008 01:09:33 GMT

For years, I've thought that the only thing sillier than complaining about disaffection and how the world really should work differently in IRC was to sign silly Internet petitions about it. Now I realize that people who feel compelled to register their righteous indignation in 141 characters of chatspeak matter least.

Here's a quarter, kid. Go buy a postcard.

What a GC Bug Looks Like in Parrot

Mon, 28 Apr 2008 08:12:11 GMT

Every so often someone reports weird behavior in #parrot, and someone says "Hey, that looks like a GC bug!"

Most of them aren't. (Most of them lately seem to be that we're changing the way bytecode works, and we don't have all of the dependencies for all of the generated PBC files correct, so you have to run make realclean and rebuild.)

While adding the vtable override cache the other day, I did create a forehead-slapper of GC bug, but I caught it before I checked in the code. How did I know it was a GC bug? Easy.

The Class PMC itself contains pointers to several other PMCs and GC entities, including the name of the class and its corresponding namespace. I added a pointer to a Hash PMC which maps the names of vtable overrides to Sub PMCs.

I remember thinking at the time "Hey, it's just a cache. I don't have to mark it during the mark phase explicitly. All of the Subs it refers to will stay alive as long as their namespaces live. That's easy."

When I ran the tests, I saw a weird error about not being able to perform a keyed index PMC lookup on a Key PMC. I set a breakpoint on the real_exception function (which reports these kinds of errors) in the debugger, and the backtrace showed that the cause of the call was my cache lookup function.

"That's weird," I thought. Then I realized what I had done.

My line of thinking was correct in that I don't have to mark all of the PMCs contained in the cache PMC. They're already reachable from the rootset through the namespace. The GC won't collect them.

The problem is that the cache itself -- the Hash PMC -- is only reachable through the Class PMC. Unless it gets marked as live, the GC will reclaim its header and put it on the free list again.

The Class PMC still has a pointer to that header, but the next PMC allocated from the GC which uses that header will overwrite the PMC's information, effectively morphing my lovely cache into something else. In this case, my Hash PMC turned into a Key PMC.

Usually they're not this obvious, but I've gone through all of the PMCs in Parrot to make sure they mark their contained GCable entities appropriately.

Perl 6 Design Minutes for 23 April 2008

Thu, 24 Apr 2008 21:17:45 GMT

The Perl 6 design team met by phone on 23 April 2008. Larry, Allison, Patrick, Jerry, Will, Jesse, Nicholas, and chromatic attended.

Jerry:

we're in a period where everyone's trying to break Parrot
they're adding new features and accidentally breaking thing
but they're fixing it
it's a good part of the cycle
people fix it
we don't have a build farm, so we can't test everywhere before committing to trunk

Patrick:

I thought that was the point of the release cycle

Jerry:

some people have suggested that we always keep trunk building and passing tests
but we don't have the means to do that
especially when we're playing with config
moving on, the big news is that TPF has six slots in Google's Summer of Code
one of them is fleshing out the Perl 6 test suite
we've needed someone to spearhead that
having a funded new contributor is wonderful
two Parrot-related projects
one is generating NCI stubs
Kevin Tew, a long-time contributor
the other is the incremental GC specified in the PDD
that's Andrew Whitworth
there's also an ASF project for integrating the GC from Apache Harmony into Parrot
they've wanted to release it as a standalone library
Parrot's the first test of a standalone system

Nicholas:

nice that it doesn't count against TPF's slot list

Jesse:

and it's nice visibility for Parrot from another group

Jerry:

I finally have six weeks of no plans to travel
should be able to devote more time to Parrot and Rakudo
looking forward to that

Larry:

getting some hacking in on my two pet bugaboos, the longest token matcher and match object generator
I refactored the matcher
it still uses TGE
instead of lumping all of the expect term possible tokens (that is, all of them) into one bucket it separates them into buckets based on their first letter
it's a one-level tree
we can build a much smaller DFA for the regexes that start with that letter
it caches that, of course
can get an instant reject if the next token can't possibly start with that letter
also flattened out all of the rules such that the list of tokens is easy to read
if the first probe with the DFA engine fails, I can take that small set of tokens that start with the same letter, run all of those rules, and sort from longest to shortest
preserves the token matching order without building huge DFA structures
as a backoff strategy, that will scale pretty well
refactored the parameter passing on the matcher side (STD 5)
instead of passing an initial array of random things, I have parameters
constructs match objects more correctly
in the sense that it gets all the information it's supposed to have
also has some attachments where it shouldn't have
I'm cleaning that up
that should scale pretty well

Jerry:

is there a drop in memory usage?

Larry:

I haven't measured
I'm sure that not feeding 800 regexes to TRE at once will make it allocate 17 megabytes on the stack
it might still be allocating too much for some of the larger things
I'm still aiming for correct, as opposed to fast
just trying to bear fast in mind
the longest token matcher now returns a linked list of states
not a string
should be a lot faster; easier to cache
functionally it's the same as before
one of those things you don't even have to measure to know it'll be faster
trying to avoid the bugaboo of premature optimization by doing what I know will be efficient to begin with
all the while trying to make the thing work
it has a good chance of being pretty speedy
my talk in Tokyo will be about all of the places where the current grammar allows extensibility
it'd be nice to be able to demonstrate some of that

Allison:

getting work done again
launched the Strings PDD
list of tasks for concurrency that I'm breaking down into smaller pieces
may post what I have now, and leave other people to break them down

Patrick:

are they parallelizable? :)

Allison:

many of them are
there are some bigger things, like switching the exception system over to the event handler
otherwise, just life stuff

Patrick:

had paying work come up this past week, so not a lot of actual coding
need to type the milestones document
it's all in my head
managed to remove a lot of unused code thanks to chromatic's post about possible optimizations
mostly just cleanup
but helped me figure out things which will feed into my redesign of PGE for longest token matching
should be able to return to direct Rakduo hacking later this week

Will:

various Parrot cleanups
TPF quarterly grant proposals are due at the end of the month
haven't seen anything come in yet

Allison:

they're queued in an RT queue
I don't know if grant members have access to that queue

Will:

we do
please, get your proposal in now, sooner than later
that goes for you on the call as well as people reading the minutes

mostly spent the past week optimizing Parrot and Rakudo
looks like it's the building speed is twice as fast as when I started
runtime is faster too, but the optimization is compilation time
found some infelicities that need more design thought
but I'm happy to put these improvements in now and take them out later when the design improves around them
hope to start adding new features again soon

Patrick:

most of the test execution time is in compilation
how useful would it be to compile Rakudo to standalone PIR?

I'd find it useful
but I'd find about ten things useful with all I work on
so not a blocker at the moment

Jesse:

how far will that get you toward native executables?

Patrick:

the existing trick for building perl6 would work
but it's not the same

if it takes an hour or two, it would help me with debugging and profiling
if it takes more time, it's not that important

Patrick:

we have to figure out runtime deployment issues for the Perl 6 runtime library

Will:

we could add the requirement to run from languages/perl6/ right now

that's fine by me for now

Patrick:

that's an afternoon job, not too bad
what do we need to do to get the Perl 6 and Parrot pages up to date?

Will:

I'll work on that

Nicholas:

why is C99 useful to Parrot and the compiler tools?

front-end parsing for C header files to build NCI declarations automatically
the backend is pretty easy, that's thunk generation
the front-end keeps people from having to write boilerplate code by hand
generate the front-end once, where you have the headers, and then you can run the generated code anywhere even if you don't have the headers

Jesse:

how does that compare to Python's ctypes?

as I understand it, they have the nice backend stuff
not so sure about the front-end
my P5NCI is nicer, if incomplete
just haven't had time to work on it...

Jesse:

if you put in for a TPF grant, that would be very useful

get me a clone first, and you have a deal

Jerry:

Allison, we talked about implementing return
that requires tying in exceptions to the concurrency scheduler?

Allison:

yes
just not implemented yet
when we did exceptions, we didn't have concurrency
so it's on the top of my list to tie them together

Will:

Tcl's already using exceptions to handle return, break, and continue

Allison:

right now, you can't have an exception handler which is a full subroutine

Patrick:

I'm not sure we need one for that feature
every subroutine block decorated as such in PAST puts an exception handler in the block
if any nested block throws a return type exception, it grabs the arguments, does what it has to, and then does a Parrot return

Allison:

if that's what you need, go ahead and do it
I thought there are some features you didn't have yet

Patrick:

I thought there might be some opcodes I needed, like handled
but we can do something now
might not be completely optimal
but it's just packaging things up now
I have something I think will work
it's not trivial, but I'm 80% confident
just a matter of sitting down and doing it

Allison:

the concurrency stuff will be there before the next release
might not want to roll it in before the release
but it'll be there soon

Patrick:

I want to get return in for the April 15 milestone we're behind on

Jerry:

have you put in tickets for the breakdown of specific tasks?

Allison:

I've never done tickets for that
just sent mail to the list of the tasks
handed them out to people as they volunteered

can you put them on a wiki page?
some of the other committers wanted that

Allison:

I can do that

Refcounting Isn't All Bad

Tue, 22 Apr 2008 07:12:34 GMT

A lot of people complain about reference counting as a memory management strategy in Perl 5, and they have a point. Unless you're very careful, reference counting spreads its tendrils throughout your system. Unless you're very disciplined, you'll forget to increase a reference count somewhere, and you'll end up collecting things in the wrong places.

The paper A Unified Theory of Garbage Collection discusses the most popular modern GC strategies to demonstrate how reference counting and tracing garbage collection are roughly equivalent, especially as concerns about performance, correctness, and conservativeness increase.

I explain this because I added reference counting to a small part of Parrot today as an optimization -- 11.73% on my Rakudo-building benchmark.

Parrot has an internal data structure called a stack. This represents a particular environment within the call graph. If you think of a normal program flow, where function A calls function B calls function C which returns to B and returns to A, you'll have a sense of a normal call stack. Parrot's stack is similar, except that we use continuation passing style (CPS), where it's not a stack, it's a graph, and control flow doesn't have to be that linear.

There's plenty of literature on CPS (including Luke Palmer's how to get a free sandwich with continuations, but the important idea is that function A sets a bookmark inside and passes that bookmark along to function B. B does the same thing when it calls C. At any point, you can look up any bookmark you happen to have and magically teleport (okay, it's just a jump instruction in your processor -- someday I'll explain how very low level languages have eval and very high level languages have eval and it's just the poor jerks in the middle who can't do anything useful without feeding the world's pointiest programming language into their AbstractFactoryFactoryFactories) to the point of that bookmark.

Ranty story short, our stack is a mostly-acyclic graph represented as a linked list, and stack entries -- or chunks -- are garbage collectable entities.

If you've even sat next to one of my "Let's Make Parrot Go Faster!" entries on a bus somewhere, or heard it playing softly in an elevator, you know that one good way to make Parrot go faster is to use fewer garbage collectable entities. One way to make an O(2n) algorithm faster is to make it an O(log n) algorithm. Another way is to make n smaller, and that's easier for now, especially because all I have to do is watch over Andrew Whitworth who's going to write a nice new garbage collector and keep track of Senaka Fernando, who's going to connect the nice garbage collector from Apache Harmony into Parrot.

One nice feature of reference counting (ah, now you see the connection!) is that you can recycle dead objects as soon as you know they're dead. One sad misfeature of a conservative garbage collector is that dead bodies stack up for a while until you can mulch them. One horrible thing about a stop-the-world full mark-and-sweep garbage collector (and try to guess what Parrot has at the moment) is that if you're out of memory in one of the arenas you care about, you have to mark all of the living GCable entities in your system and sweep all of the pools in all of the arenas to find everything that's alive and to recycle everything that isn't.

If you want to go fast, you want to avoid this until it's absolutely necessary.

Here's the interesting thing about stack chunks: Parrot stores them in one place in one function (stack_push) and unstores them in one place in one function (stack_pop). Again, the names are really bad because it's a graph and not a stack, but you get the point.

I noticed this when I profiled the Rakudo-building benchmark again and saw that stack_push was one of the most expensive inclusive calls, apart from the garbage collector. It calls a function which calls a function which calls a function which asks for a new stack chunk from the GC. When there are no more free stack chunks, the GC runs again, and that eats up precious cycles.

After a few minutes of thinking, I realized that the stack_push/stack_pop pair formed a boundary around a stack chunk's lifespan. Popping a chunk off of the stack meant that nothing else needed it. That chunk is immediately identifiably a dead object, suitable for recycling.

Rather than waiting for a full GC run to find every live object in the system and then run across this dead object and recycle it, I added a couple of lines of code to recycle it automatically then and there.

There was my immediate 11.73% speedup.

There was also a problem in the Continuation PMC tests; I noticed this only after I patted myself on the back for moving the bottleneck in my benchmark elsewhere. I'm sure you'll see it if you think for a moment, especially about how I kept writing that the stack isn't a stack, it's a graph implemented as a linked list.

The problem is that there's not a single linked list. If you take a continuation at various points in the call graph, you can have several linked lists that all share the same tail. If you exhaust one of those call chains, you may pop a stack chunk that's part of another call graph elsewhere. Your reference to it goes away, but it's still live elsewhere, so if you recycle it then and there, you may end up reusing the chunk and storing inappropriate data in it while something else expects it to stay pristine.

If you're very clever, you've figured out the solution.

I added a reference count to the stack chunk structure, incremented it on stack_push, decremented it on stack_pop, and recycled the chunk in the latter function only if the refcount is zero.

Is this a long-term solution? Probably not; I posted my initial investigation to the parrot-porters list, and we had a nice discussion on alternate techniques. We'll have to refactor the code more heavily once we find a better approach, but for now, I'll live with this optimization.

(Seneca Cunningham found what appears to be a crash related to this, but I think I've found the culprit and fixed it by adding another reference count pair to the Continuation PMC. See? I told you reference counting made itself pervasive.)

It's Cheaper at Compile Time

Mon, 21 Apr 2008 07:14:18 GMT

I've long used part of the Rakudo-building process as my benchmark for Parrot optimizations. It doesn't exercise all of Parrot by any means, but it pushes the garbage collector, the string system, and the object system pretty hard. It's also the longest part of the build process anywhere, so improving performance there has a dramatic effect on productivity.

I've long known that it appends to and concatenates strings very frequently. The corresponding Parrot functions always show up near the top of the cost list from Callgrind. I've spent a few hours here and there trying to speed them up, but short of rethinking them entirely, I've never found a good solution.

We don't (yet) have a PIR-level profiler like Callgrind (but I'm working on it). Sometimes I can read through the code and see things that might be expensive, knowing what eventually gets called and where. On a whim, I browsed through the generated code for Rakudo to find concatenation. One idiom jumped out at me:

$S0 = concat "PIR", ':"'

This line of code catenates two constant strings into one, and assigns the result to the virtual string register $S0. It's not terribly inefficient; the two constant strings are in the constant table in the bytecode, so they don't get recreated on execution. However, the result is always going to be a constant string.

There's no reason the optimizer couldn't rewrite that line of code into:

$S0 = 'PIR:"'

It didn't. Now it does.

The optimizer already knew how to optimize away constant addition, subtraction, and other math operations. All I had to do was tell the opcode generator not to generate the concat_s_sc_sc op (concatenate two string constants and put the results in a string register -- the destination register always comes first in the argument list for an opcode), add the concat operator to the list of rewritable three-register opcodes, and add a case to the rewriter where the destination register is a string register.

The result is a modest 5.5% speedup in my benchmark. Part of the gain is avoiding run-time concatenation, but part of it is also using fewer GCable entities. The less garbage you create, the less garbage you have to collect and the less frequently you have to collect it. 5.5% isn't a huge gain, but it's a step in the right direction. Every second we shave off of the hack-compile-test cycle is an improvement, and these optimizations help all Parrot programs as well.

(Are these entries interesting? I seem to get more feedback when I complain about various Technology Distortion Fields.)