Archive for junho \29\UTC 2010

Immix on GHC Summer of Code weekly report #7

29/6/2010 (terça-feira)

My project.

This post assumes that the reader has read my last post.

This week I had less time to work on my project than the last ones, because it was kind of the last week of the semester for some disciplines. This was unfortunate from a point of view, but it’s good in another, because now I’ve finished most of my disciplines and can focus more. This is also the reason why the weekly report is slightly delayed.

Most of my work this week was continuing the investigation of the segfault caused by my implementation of allocation in lines. The most interesting thing I discovered came after a suggestion from my menthor: to use +RTS -DS to turn on the sanity checker. I runned the code with my allocation implementation and the sanity checker and got a segfault. I runned again only with the code to free memory in lines and I got also a segfault. Then I runned without all my patches, using sweep, and I got the segfault. So it seems that there is something wrong with sweep, and now I’m investigating this new segfault.

I’m using the bernouilli program from nofib to test, running with 148 as a parameter and, naturally, +RTS -w -DS passed to the RunTime System. The output of gdb:

Current directory is /home/marcot/trabalho/livre/ghc/nofib/imaginary/bernouilli/
GNU gdb (GDB) 7.1-debian
Copyright (C) 2010 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-linux-gnu".
For bug reporting instructions, please see:
Reading symbols from /home/marcot/trabalho/livre/ghc/nofib/imaginary/bernouilli/Main...done.
(gdb) r 148 +RTS -w -DS
Starting program: /home/marcot/trabalho/livre/ghc/nofib/imaginary/bernouilli/Main 148 +RTS -w -DS
[Thread debugging using libthread_db enabled]

Program received signal SIGSEGV, Segmentation fault.
0x00000000006309cb in LOOKS_LIKE_INFO_PTR_NOT_NULL (p=12297829382473034410) at includes/rts/storage/ClosureMacros.h:225
(gdb) where
#0  0x00000000006309cb in LOOKS_LIKE_INFO_PTR_NOT_NULL (p=12297829382473034410) at includes/rts/storage/ClosureMacros.h:225
#1  0x0000000000630a16 in LOOKS_LIKE_INFO_PTR (p=12297829382473034410) at includes/rts/storage/ClosureMacros.h:230
#2  0x0000000000630a4b in LOOKS_LIKE_CLOSURE_PTR (p=0x7ffff6c84062) at includes/rts/storage/ClosureMacros.h:235
#3  0x0000000000631427 in checkClosure (p=0x7ffff6853a08) at rts/sm/Sanity.c:320
#4  0x00000000006319ef in checkHeap (bd=0x7ffff68014c0) at rts/sm/Sanity.c:479
#5  0x000000000063222f in checkSanity (check_heap=rtsTrue) at rts/sm/Sanity.c:686
#6  0x000000000062dfdb in GarbageCollect (force_major_gc=rtsFalse, gc_type=0, cap=0x8d2ec0) at rts/sm/GC.c:768
#7  0x0000000000620431 in scheduleDoGC (cap=0x8d2ec0, task=0x8f5080, force_major=rtsFalse) at rts/Schedule.c:1420
#8  0x000000000061fa2c in schedule (initialCapability=0x8d2ec0, task=0x8f5080) at rts/Schedule.c:539
#9  0x0000000000620c77 in scheduleWaitThread (tso=0x7ffff6c80000, ret=0x0, cap=0x8d2ec0) at rts/Schedule.c:1902
#10 0x000000000065762b in rts_evalLazyIO (cap=0x8d2ec0, p=0x89d8b0, ret=0x0) at rts/RtsAPI.c:495
#11 0x000000000061d3db in real_main () at rts/RtsMain.c:66
#12 0x000000000061d4ca in hs_main (argc=5, argv=0x7fffffffe6d8, main_init=0x406558 <__stginit_ZCMain>, main_closure=0x89d8b0) at rts/RtsMain.c:115
#13 0x00007ffff6fbcabd in __libc_start_main (main=<value optimized out>, argc=<value optimized out>, ubp_av=<value optimized out>, init=<value optimized out>, fini=<value optimized out>, rtld_fini=<value optimized out>, stack_end=0x7fffffffe6c8) at libc-start.c:222
#14 0x0000000000403a69 in _start ()

I started the investigation with the most obvious test: to remove the call to sweep() in rts/sm/GC.c.

  if (major_gc && oldest_gen->mark) {
      if (oldest_gen->compact) 
      // else
      //     sweep(oldest_gen);

The result was the same, same segfault in the same place. So I decided to do it a little stronger and avoid the blocks getting the BF_MARKED flag, so that they are not even marked.

                if (!(bd->flags & BF_FRAGMENTED)) {
                    // bd->flags |= BF_MARKED;

This one worked. The code in sweep() is not commented, but it’s irrelevant, since it ignores blocks that don’t have this flag. My third try was to comment the places where the BF_MARKED flag is read, to see each one was causing the segfault. I got a list of places to search with grep, and there weren’t a lot of them.

$ grep BF_MARKED rts/sm/*.c
rts/sm/Compact.c:       if (bd->flags & BF_MARKED)
rts/sm/Evac.c:  if ((bd->flags & (BF_LARGE | BF_MARKED | BF_EVACUATED)) != 0) {
rts/sm/Evac.c:        if (bd->flags & BF_MARKED) {
rts/sm/GCAux.c:    if ((bd->flags & BF_MARKED) && is_marked((P_)q,bd)) {
rts/sm/GC.c:                    if (!(bd->flags & BF_MARKED))
rts/sm/GC.c:                        // time, so reset the BF_MARKED flags.
rts/sm/GC.c:                        // compact.  (search for BF_MARKED above).
rts/sm/GC.c:                        bd->flags &= ~BF_MARKED;
rts/sm/GC.c:                // Also at this point we set the BF_MARKED flag
rts/sm/GC.c:                // BF_MARKED is always unset, except during GC
rts/sm/GC.c:                    bd->flags |= BF_MARKED;
rts/sm/Sweep.c:        if (!(bd->flags & BF_MARKED)) { 

The first one is in rts/sm/Compact.c, so it’s not relevant to the use with -w. The second one, in rts/sm/Evac.c, is a bit indirect.

  if ((bd->flags & (BF_LARGE | BF_MARKED | BF_EVACUATED)) != 0) {

      // pointer into to-space: just return it.  It might be a pointer
      // into a generation that we aren't collecting (> N), or it
      // might just be a pointer into to-space.  The latter doesn't
      // happen often, but allowing it makes certain things a bit
      // easier; e.g. scavenging an object is idempotent, so it's OK to
      // have an object on the mutable list multiple times.
      if (bd->flags & BF_EVACUATED) {
          // We aren't copying this object, so we have to check
          // whether it is already in the target generation.  (this is
          // the write barrier).
	  if (bd->gen < gct->evac_gen) {
	      gct->failed_to_evac = rtsTrue;

      /* evacuate large objects by re-linking them onto a different list.
      if (bd->flags & BF_LARGE) {
	  info = get_itbl(q);
	  if (info->type == TSO && 
	      ((StgTSO *)q)->what_next == ThreadRelocated) {
	      q = (StgClosure *)((StgTSO *)q)->_link;
              *p = q;
	      goto loop;
      /* If the object is in a gen that we're compacting, then we
       * need to use an alternative evacuate procedure.
      if (!is_marked((P_)q,bd)) {

The first if, in line 466, is executed if any of the three flags is present: BF_LARGE, BF_MARKED or BF_EVACUATED. The second if, in line 474, checks for BF_EVACUATED, and returns. The third if, in line 487, checks for BF_LARGE and returns. The code in lines 502-505 is only executed if BF_MARKED is present, and not the other ones. I tried commenting this code, and got an assertion fail in the user code, so I think this is not a good path to follow.

The second occurence of BF_MARKED is in the same file.

        if (bd->flags & BF_MARKED) {
            // must call evacuate() to mark this closure if evac==rtsTrue
            *q = (StgClosure *)p;
            if (evac) evacuate(q);
            unchain_thunk_selectors(prev_thunk_selector, (StgClosure *)p);

Commenting it, with or without the call to sweep() commented, causes the same segfault. So, I’m tending to think this part of the code is unrelated to the issue.

The third occurence is in rts/sm/GCAux.c. I tried commenting it with all the four combinations of the two others commented and not commented, and all resulted in the segfault in the same place.

There’s another place to check in rts/sm/GC.c. Again, commenting it made no difference. The last one is part of sweep(), so it’s avoided anyway when the call to this function is commented.

Pensamentos sobre política (1)

28/6/2010 (segunda-feira)

Já faz algum tempo que pretendo escrever sobre minhas idéias políticas. Hoje, após ler o link do post abaixo, essa vontade veio mais forte. Minha idéia era colocar isso num wiki ou numa página estática, mas isso não é possível por enquanto, já que estou trocando de servidores. Portanto, vou escrevendo aqui no blog, e depois sintetizo tudo em um wiki.

Todas as pessoas devem ter a liberdade de fazer tudo o que quiserem, desde que esses atos não impessam outras pessoas de exercerem suas liberdades. A função do estado é evitar que a vontade de uma pessoa se sobreponha ao direito das outras pessoas de exercerem suas liberdades. Não estou certo que ele deva fazer apenas isso, mas acho que não existem muitas outras coisas que ele deva fazer. Assim, a atuação do estado deve ser mínima.

Um exemplo desse minimalismo é a ação da polícia. A polícia é a manifestação do estado mais concreta na vida dos cidadãos. É ela quem garante que as pessoas não limitem a liberdade de outras pessoas, e deve agir apenas com o propósito de punir quem fizer isso e evitar que as pessoas o façam. Ao agir preventivamente, ela deve agir o mínimo possível, com a maior cautela possível e de forma a perturbar os cidadãos o mínimo possível.

Para isso, naturalmente, precisamos de uma polícia extremamente bem preparada. Por constituirem um setor chave do funcionamento da sociedade, os policiais deveriam ser muito bem preparados e remunerados. Curso superior em direito deveria ser exigência mínima.

A prevenção de crimes com o mínimo de ação pode ser feita apenas com a presença policial nas ruas, em pontos estratégicos e fazendo ronda. Também considero válida a abordagem de carros e até mesmo a utilização de bafômetros. Não considero válida uma pergunta sem justificativa para um cidadão, do tipo: “Aonde você está indo?” Não considero válida, de forma alguma, a proibição de aglomerações, nem o bloqueamento das vias de acesso a uma área da cidade.

Tenho dúvidas sobre se vale a pena preservar o corpo policial atual ou se é melhor simplesmente construir uma nova polícia, com pessoas mais bem preparadas. Tendo a achar que a instituição já está corrompida, e que seria melhor a extinção dessa instituição e a criação de uma nova, do zero. Mas não estou completamente certo disso. Penso em algo parecido para a educação de base, ou seja, ensino fundamental e médio.

São Bento sitiado

28/6/2010 (segunda-feira)

Immix on GHC Summer of Code weekly report #6

18/6/2010 (sexta-feira)

My project.

This post assumes that the reader has read my last post.

I started the change in the line representation. Instead of a single linked list of free lines, it’s better to work with line groups, to avoid fragmentation inside lines in allocation. I’ve create a struct to represent this line group, with a pointer to the next group, and the size, in lines, of the group. It is stored in includes/rts/storage/GC.h currently, but I’m not sure if this is the best place to put it, and I’m thinking about changing it latter.

#include "rts/OSThreads.h"

typedef struct line_ {
    struct line_ *next;
    StgWord size;
} line;

/* -----------------------------------------------------------------------------

Then we need to change the code of liberation of lines in rts/sm/Sweep.c.

            else if(!(bd->flags & BF_MEDIUM)) {
                StgBool sequence;
                sequence = 0;

                for(i = 1; i < BLOCK_SIZE_W / BITS_IN(W_); i++) {
                    StgPtr start;
                    start = bd->start + BITS_IN(W_) * i;
                    if(bd->u.bitmap[i] == 0 && bd->u.bitmap[i - 1] == 0 &&
                       start + BITS_IN(W_) <= bd->free) {
                        printf("DEBUG: line_found(%p)\n", start); fflush(stdout);
                        if(gen->first_line == NULL) {
                            gen->first_line = (line *) start;
                        if(sequence) {
                        else {
                            if(last_line != NULL) {
                                last_line->next = (line *) start;
                            last_line = (line *) start;
                            last_line->next = NULL;
                            last_line->size = 1;
                        sequence = 1;
                    else {
                        sequence = 0;

I’ve tested this code using the same technique as before, checking the produced list trasversing it after it’s done. The list obtained showed the correct lines.

Now I’m back to the allocation code, which never worked. The changes in rts/sm/Evac.c are straightfoward.

        if (gen->first_line != NULL &&
            size <= BITS_IN(W_) * gen->first_line->size) {
            ws->todo_free = (StgPtr) gen->first_line;
            ws->todo_lim = ws->todo_free + BITS_IN(W_) * gen->first_line->size;
            gen->first_line = gen->first_line->next;
            to = ws->todo_free;

This cause the same kind of errors I was getting before. I should go back to debugging. My co-supervisor in my Oriented Project in Computer Science suggested me using valgrind. I’ll try it.

Last week I forgot to mention I’ve presented my final presentation about my Oriented Project in Computer Science. This week I finished writing my monograph.

Summer of Code weekly report #5

15/6/2010 (terça-feira)

My project.

This post assumes that the reader has read my last post.

This was a week full of segfaults, failed assertions, changes in the user data and other crazy stuff. I started working on the allocation of memory on the freed lines. In the first glance, it seemed to be much harder than the small changes I’ve done to free memory in lines, because of the a never-seen-before-by-me datatypes being dealed with in alloc_for_copy(), from rts/sm/Evac.c: gen_workspace. But in a short time the code of alloc_for_copy() has shown itself to be simple, and it was easy to do an initial implementation. I went in the same path I was going on before, trying the simplest solution that works. For instance, the line representation used is not ideal, because groups of lines are not considered, but it was the simplest at that time, and I plan to improve it latter.

The gen_workspace data type contains a pointer to an area of a block that is not being used (todo_free) and a pointer to the end of that area (todo_lim). When a space for an object is requested, it’s allocated in this area, and todo_free is adjusted. If the area is full, a new block is requested. My intended change was to return the first free line of the generation when the object is smaller or equal to the size of a line, and use the current approach otherwise. This is the simplest way I could think, but has the problem that the only one object is allocated per line. This was a known issue.

    if(size <= BITS_IN(W_) && gen->first_line != NULL) {
        to = gen->first_line;
        gen->first_line = (StgPtr) *gen->first_line;
        return to;

    ws = &gct->gens[gen->no];

After that I got the first round of segfaults. One problem I could spot so far was in the code in rts/sm/Sweep.c, related to the liberation of the lines. The block is allocated by need, and there’s a pointer to the first free byte, free. This free space is used by alloc_for_copy(). So, we should only think of lines in the region already allocated in the block, that is, where the address is smaller than free.

                    if(bd->u.bitmap[i] == 0 && bd->u.bitmap[i - 1] == 0 &&
                       start + BITS_IN(W_) <= bd->free)

But the segfaults, memory corruptions, etc, kept going on. I tried some restrictions, like allocating in lines only objects with the exact size of a line, or very small objects (size == 2), or allocating just one object for generation, or only one object at all. I also tried checking for the type of the allocated object, to see a correlation with the problems I was having. Nothing helped by now.

So I thought I should work on other things, so that maybe my mind gets clearer and I can spot the problem. I followed the suggestion of using todo_free and todo_lim to make it possible to allocate more than one object in a line. I liked this change, since I had the impression that it fits better with the rest of the code than the initial implementation, and it will be easier to addapt it when I improve the representation of the free lines. As I said, I wanted something to work while I think in how I’ll solve the segfault problems. The bad side of this choice is that I’ll not be able to test it, since it’s just a reimplementation of the same idea in the allocation code, and is not expected to work. The good size is that it may work, and solve the problems I was having by accident.

    if (ws->todo_free > ws->todo_lim) {
        if (size <= BITS_IN(W_) && gen->first_line != NULL) {
            ws->todo_free = gen->first_line;
            gen->first_line = (StgPtr) *gen->first_line;
            to = ws->todo_free;
            ws->todo_lim = ws->todo_free + BITS_IN(W_);
        } else {
            to = todo_block_full(size, ws);

I got the same results as with the older implementation, and no ideas about how to solve it (yet). So I thought about another thing to work on, which I’m talking about all through this post: improve the free line representation. This has the advantage of being testable, since it’s unrelated to the allocation code, and may give me an idea about how I can fix the segfaults. The bad side is that it probably won’t fix my problems directly. I’m working on it right now.

Following a suggestion, I’ve started using gdb to see what was happening in the GC instead of including printfs everywhere, and it’s been useful. I’ve noticed that sometimes printf is more effective, but sometimes gdb is much better.

Licença maternidade na Prefeitura de Belo Horizonte

11/6/2010 (sexta-feira)

Meu nome é Marco Túlio e passei por uma situação com a Prefeitura de Belo Horizonte que gostaria de relatar. Minha filha nasceu no dia 12 de janeiro de 2010 e minha esposa, professora da rede municipal, teria direito a 4 meses de licença. Tendo em vista que a Organização Mundial de Saúde (OMS) aconselha que a criança se alimente apenas do leite materno até os 6 meses, minha esposa entrou com um pedido na Prefeitura, para que sua licença fosse extendida para 6 meses. Naquele momento, o empregador poderia legalmente escolher se daria à funcionária 4 ou 6 meses de licença. A Prefeitura negou o pedido. Minha esposa, então, entrou com um processo junto ao sindicato contra a Prefeitura, demandando a licença de 6 meses. O juiz concedeu a liminar, e ficamos tranquilos com a notícia. Quando soubemos que, no dia 8 de março, Dia Internacional da Mulher, o Prefeito Márcio Lacerda assinou um projeto de lei que ampliava a licença para 6 meses, ficamos ainda mais tranquilos.

Apesar disso, no dia 9 de junho de 2010, recebemos um telefonema de uma funcionária da Regional, dizendo que minha esposa deveria retornar às suas atividades imediatamente. Como achávamos que ela voltaria ao trabalho somente a partir do dia 12 de julho, não começamos ainda a dar outros alimentos à nossa filha. Estamos alimentando-a apenas com o leite materno, como recomendado pela OMS.

A adaptação para que a criança comece a ingerir outros alimentos leva um certo tempo, e é muito difícil fazer isso de um dia para o outro. Levando isso em consideração e com um atestado médico que pedia 30 dias de licença para que minha esposa pudesse fazer essa adaptação, fomos à junta médica da Prefeitura, pedindo uma licença para acompanhamento. Os funcionários da Prefeitura tem direito a 30 dias dessa licença por ano, para acompanhar familiares doentes. O pedido foi negado, com a justificativa de que essa licença só se aplica em casos de doença. Nessa ida à junta médica, me impressionou quantas cópias haviam da notícia do projeto de lei assinado pelo Prefeito Márcio Lacerda, pregadas nas paredes em todos os lugares.

Além do problema da amamentação, temos que providenciar um lugar para deixar nossa filha, de 5 meses, de uma hora para outra. Estamos procurando alternativas, mas a creche com berçário mais próxima, a Unidade Municipal de Educação Infantil Alaíde Lisboa, localizada na UFMG, não tem vagas. Uma funcionária disse que, para atender toda a fila de espera, seriam necessárias 3 escolas.

O que mais me incomodou nessa história foi a decisão de recorrer da liminar que extendeu a licença para 6 meses. Se o próprio Prefeito Márcio Lacerda já tinha ampliado a licença, por que recorrer contra quem conseguiu essa ampliação na justiça? Tentando tirar essa dúvida, liguei para o gabinete do Promotor Saulo Converso Lara (3277-4029), da Procuradoria Geral do Município, mas não consegui falar com ele. Expliquei o que havia acontecido para uma servidora municipal que atendeu o telefone, e ela disse que o Promotor Saulo Converso Lara age em interesse do município, independentemente do contexto. Quando perguntei qual era o interesse do município nesse caso específico, a servidora disse que precisava desligar o telefone. Tentei perguntar novamente, dizendo que ela só estava desligando porque não conseguia responder essa pergunta. Ela me disse que não era obrigada a dar satisfações e desligou na minha cara. Infelizmente, não perguntei o nome da servidora. Tentei ligar novamente, para falar diretamente com o Promotor Saulo Converso Lara mas, ao ouvir a minha voz, a servidora desligou na minha cara novamente.

Summer of Code weekly report #4

4/6/2010 (sexta-feira)

My project.

I’m publishing my report earlier this week, because there was a lot to talk about. This week I started to make changes in the code going in the direction of what I want to do. I haven’t started a final implementation, but I’m studying about how what I want to do will affect the rest of the Garbage Collection (GC). I noticed the code in rts/sm/Sweep.c was simple and similar to what I’m planning to do, so I started changing how it works.

Sweep in the Glasgow Haskell Compiler (GHC) is done by a bitmap, which contains a bit for each word in a memory block, and is set to 1 when there’s an object starting in the mapped area and 0 otherwise. When there’s a block with no objects starting at it, that is, all bits of the bitmap are set to 0, the block is freed.

        if (resid == 0)
            if (prev == NULL) {
                gen->old_blocks = next;
            } else {
                prev->link = next;

The bits are analyzed in a group of BITS_IN(W_), where BITS_IN(W_) is the number of bits in a word.

        for (i = 0; i < BLOCK_SIZE_W / BITS_IN(W_); i++)
            if (bd->u.bitmap[i] != 0) resid++;

If more than ¼ of the groups are completely set to 0, the block is considered fragmented.

            if (resid < (BLOCK_SIZE_W * 3) / (BITS_IN(W_) * 4)) {
                bd->flags |= BF_FRAGMENTED;

Immix, the GC algorithm I plan to implement in GHC, divides the blocks of memory in lines. My initial plan was to identify free lines. I decided to consider a the size of a line fixed in BITS_IN(W_) words, because it will map to a word in the bitmap, and the code was already analyzing in groups of BITS_IN(W_) words. This was very easy with the current code.

            if (bd->u.bitmap[i] != 0) resid++;
            else printf("DEBUG: line_found(%p)\n", bd->start + BITS_IN(W_) * i);

This worked, and showed some free lines. I’m sure there are other ways of logging in GHC, but printf was the simplest way I could thought of. I measured the occurrence of free lines using the bernouilli program from the NoFib benchmark suite, calling it with 500 +RTS -w, to make it uses sweep. In 782 calls to GarbageCollect(), sweep() was called 171 times, for 41704 blocks to be swept and found 230461 free lines. This gives us about 5.5 free lines per block, from the 8 lines in each block (on 64 bits systems).

The problem is that the bitmap is marked only in the start of the objects allocated, so even in a line that all bits are marked with 0 we can’t assume that it’s completely free, because there may be an object that starts in the previous line that is using the space of the line. Checking only for the previous line doesn’t work either, because a big object can span several lines. What we can do here is a variation of conservative marking, as proposed in the Immix paper, checking only the previous line and working only with objects smaller than a line.

To make sure I was working only with objects smaller than a line, I had to mark the blocks that contains medium objects and avoid them when seeking free lines. The block flags are defined in includes/rts/storage/Block.h, so I included another flag in this file, BF_MEDIUM.

/* Block contains objects evacuated during this GC */
#define BF_EVACUATED 1
/* Block is a large object */
#define BF_LARGE     2
/* Block is pinned */
#define BF_PINNED    4
/* Block is to be marked, not copied */
#define BF_MARKED    8
/* Block is free, and on the free list  (TODO: is this used?) */
#define BF_FREE      16
/* Block is executable */
#define BF_EXEC	     32
/* Block contains only a small amount of live data */
#define BF_FRAGMENTED 64
/* we know about this block (for finding leaks) */
#define BF_KNOWN     128
/* Block contains objects larger than a line */
#define BF_MEDIUM    256

The GHC GC is generational, there is, objects are allocated in a generation and, after a time, the ones that are still being used are moved the next generation. This idea assumes the death probability of younger objects is higher, so few objects are moved to the next generation. Sweep and Immix work only in the last generation so, to mark blocks with medium objects we have to check the size of the objects that are moved to the next generation.

This is done in the copy_tag function of rts/sm/Evac.c. I inserted a code that checks for the object size and marks the block when it’s bigger than BITS_IN(W_).

copy_tag(StgClosure **p, const StgInfoTable *info, 
         StgClosure *src, nat size, generation *gen, StgWord tag)
    StgPtr to, from;
    nat i;

    to = alloc_for_copy(size,gen);

    if(size > 8) {
        Bdescr(to)->flags |= BF_MEDIUM;

So I updated the code in rts/sm/Sweep.c to only inspect for free lines in blocks without BF_MEDIUM mark.

            if (bd->u.bitmap[i] != 0) resid++;
            else if(!(bd->flags & BF_MEDIUM)) {
                printf("DEBUG: line_found(%p)\n", bd->start + BITS_IN(W_) * i);

This also worked. Now, in the 32012 blocks there were 189015 free lines, found in the same number of GCs, making about 5.9 free lines per block. We considered only blocks with small objects, but we didn’t ignore the first line of each group of free lines. This can be achieved by checking if the previous line was also free.

            if (bd->u.bitmap[i] != 0) resid++;
            else if(!(bd->flags & BF_MEDIUM) && i > 0 && bd->u.bitmap[i] == 0) {
                printf("DEBUG: line_found(%p)\n", bd->start + BITS_IN(W_) * i);

Now, from the 32239 blocks, 165547 free lines were found, giving 5.1 free lines per block. But there are more things to improve. If the whole block is free, we want to free it, instead of marking it’s lines as free. So it’s better to mark the lines after we know that the blocks are not completely free. So I left the code that checks the bitmap as it was, and included a line check only for blocks that are not completely free. At this point, I also associated the fragmentation test with blocks with medium objects, because in blocks of small objects we plan to allocate in free lines, so fragmentation is not a (big) issue.

            if (resid < (BLOCK_SIZE_W * 3) / (BITS_IN(W_) * 4) &&
                (bd->flags & BF_MEDIUM)) {
                printf("DEBUG: BF_FRAGMENTED\n");
                bd->flags |= BF_FRAGMENTED;
            else if(!(bd->flags & BF_MEDIUM)) {
                for(i = 1; i < BLOCK_SIZE_W / BITS_IN(W_); i++)
                    if(bd->u.bitmap[i] == 0 && bd->u.bitmap[i - 1] == 0) {
                        printf("DEBUG: line_found(%p)\n", bd->start + BITS_IN(W_) * i);

The total ammount of blocks increased dramatically: the blocks that become fragmented and were not called again in sweep made a huge difference. From the 345143 blocks, 1633268 free lines were found, or about 4.7 free lines per block. 9434 blocks were free, so, from the remaining blocks, we have about 4.9 free lines per block.

Something we’ll need then is a way to access these lines latter. The simplest way I thought to achieve it is constructing a list of lines, in each the first word of each free line is a pointer to the next free line, and the first word of the last free line is 0. It’s useful to keep reporting the lines to stdout, so that we can then follow the list and check if we went to the same lines.

                    if(bd->u.bitmap[i] == 0 && bd->u.bitmap[i - 1] == 0) {
                        StgPtr start = bd->start + BITS_IN(W_) * i;
                        printf("DEBUG: line_found(%p)\n", start);
                        if(line_first == NULL) {
                            line_first = start;
                        if(line_last != NULL) {
                            *line_last = (StgWord) start;
                        line_last = start;
                        *line_last = 0;

    for(line_last = line_first; line_last; line_last = (StgPtr) *line_last) {
        fprintf(stderr, "DEBUG: line_found(%p)\n", line_last);

I printed the inclusion of the lines on the list to stdout, and the walking on the list in stderr, so that it’d be easy to diff. There was no difference between the lists.

There’re another improvements that can be made, like using a list of groups of free lines, but I think it’ll be better to think about this after studying how the allocation in the free lines will be done. That’s where I’m going to now.

There are some minor things I learned, and thought they worth blogging. The current GHC uses three strategies for collecting the last generation: copying, mark-compact and mark-sweep. Copying is the default until the memory reaches 30% of the maximum heap size; after that, mark-compact is used. Sweep can be chosen by a Real Time System (RTS) flag, -w. To use mark-compact always, the flag is -c.

I’ve been submitting small patches to the cvs-ghc mailling list, mostly about outdated comments. Most of them were accepted, except for one which contained a lot of commentary, and that indeed was not completely correct. I corrected it and resend to the list, but the message is waiting for approval because the message header matched a filter rule. I believe this is because I replied the message generated from darcs.

There’s a very useful ghc option, specially for testing the compiler, because in this case you need to rebuild the source, even when there’re no changes in it. It’s -fforce-recomp, and it makes only sense when used with --make.