Looking for a possibility to reduce MIR memory consumption.

[qwe661234]

Based on my observations of the analysis conducted using the memory sanitization tool valgrind, it's evident that MIR_gen consumes approximately 1.3 GiB of memory, while c2mir_compile consumes around 0.45 GiB. Consequently, I am inquiring about potential strategies to reduce the memory consumption of MIR_gen. For example, we could lower memory consumption by removing certain modules that demand substantial memory but are dispensable for these two functions or reusing memory within some functions rather than repeatedly allocating and freeing it.

Issue reproducing

• test_program

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>

#include "mir.h"
#include "mir-gen.h"
#include "c2mir/c2mir.h"

/* Provide for c2mir to retrieve the generated code string. */
typedef struct {
  char *code;
  size_t code_size;
  size_t curr; /**< the current pointer to code string */
} code_string_t;

/* mir module */
typedef struct {
  MIR_context_t ctx;
  struct c2mir_options *options;
  uint8_t debug_level;
  uint8_t optimize_level;
} riscv_jit_t;

typedef struct {
  char *name;
  void *func;
} func_obj_t;

static int get_string_func (void *data) {
  code_string_t *codestr = data;
  return codestr->curr >= codestr->code_size ? EOF : codestr->code[codestr->curr++];
}

#define DLIST_ITEM_FOREACH(modules, item) \
  for (item = DLIST_HEAD (MIR_item_t, modules->items); item; item = DLIST_NEXT (MIR_item_t, item))

/* parse the funcitons are not defined in mir */
static void *import_resolver (const char *name) {
  func_obj_t func_list[] = {
    {NULL, NULL},
  };
  for (int i = 0; func_list[i].name; i++) {
    if (!strcmp (name, func_list[i].name)) return func_list[i].func;
  }
  return NULL;
}

static riscv_jit_t *jit = NULL;
static code_string_t *jit_code_string = NULL;

/* TODO: fix the segmentation fault error that occurs when invoking a function
 * compiled by mir while running on Apple M1 MacOS.
 */
static uint8_t *compile () {
  char func_name[25];
  snprintf (func_name, 25, "jit_func_");
  c2mir_init (jit->ctx);
  size_t gen_num = 0;
  MIR_gen_init (jit->ctx, gen_num);
  MIR_gen_set_optimize_level (jit->ctx, gen_num, jit->optimize_level);
  if (!c2mir_compile (jit->ctx, jit->options, get_string_func, jit_code_string, func_name, NULL)) {
    perror ("Compile failure");
    exit (EXIT_FAILURE);
  }
  FILE *fp = fopen ("add.bmir", "w");
  MIR_write (jit->ctx, fp);
  MIR_load_external (jit->ctx, "printf", printf);
  MIR_module_t module = DLIST_TAIL (MIR_module_t, *MIR_get_module_list (jit->ctx));
  MIR_load_module (jit->ctx, module);
  MIR_link (jit->ctx, MIR_set_gen_interface, import_resolver);
  MIR_item_t func = DLIST_HEAD (MIR_item_t, module->items);
  uint8_t *code = NULL;
  size_t func_len = DLIST_LENGTH (MIR_item_t, module->items);
  for (size_t i = 0; i < func_len; i++, func = DLIST_NEXT (MIR_item_t, func)) {
    if (func->item_type == MIR_func_item) code = func->addr;
  }

  MIR_gen_finish (jit->ctx);
  c2mir_finish (jit->ctx);
  return code;
}

uint8_t *block_compile () {
  if (!jit) {
    jit = calloc (1, sizeof (riscv_jit_t));
    jit->options = calloc (1, sizeof (struct c2mir_options));
    jit->ctx = MIR_init ();
    jit->optimize_level = 1;
  }

  if (!jit_code_string) {
    jit_code_string = malloc (sizeof (code_string_t));
    jit_code_string->code = malloc (64 * 1024);
  }
  jit_code_string->curr = 0;
  jit_code_string->code
    = "int add(int a, int b) { "
      " return a + b; "
      "} ";
  jit_code_string->code_size = strlen (jit_code_string->code);
    return compile ();
}

static uint8_t *gen () {
  if (!jit) {
    jit = calloc (1, sizeof (riscv_jit_t));
    jit->options = calloc (1, sizeof (struct c2mir_options));
    jit->ctx = MIR_init ();
    jit->optimize_level = 1;
  }
  size_t gen_num = 0;
  MIR_gen_init (jit->ctx, gen_num);
  MIR_gen_set_optimize_level (jit->ctx, gen_num, jit->optimize_level);
  FILE *fp = fopen ("add.bmir", "r");
  MIR_read (jit->ctx, fp);
  MIR_load_external (jit->ctx, "printf", printf);
  MIR_module_t module = DLIST_TAIL (MIR_module_t, *MIR_get_module_list (jit->ctx));
  MIR_load_module (jit->ctx, module);
  MIR_link (jit->ctx, MIR_set_gen_interface, import_resolver);
  MIR_item_t func = DLIST_HEAD (MIR_item_t, module->items);
  MIR_gen(jit->ctx, gen_num, func);
  uint8_t *code = NULL;
  size_t func_len = DLIST_LENGTH (MIR_item_t, module->items);
  for (size_t i = 0; i < func_len; i++, func = DLIST_NEXT (MIR_item_t, func)) {
    if (func->item_type == MIR_func_item) code = func->addr;
  }
  MIR_gen_finish (jit->ctx);
  return code;
}

typedef int (*func) (int a, int b);
int main () {
  int a = 10, b = 5;
  a = ((func) (block_compile ())) (a, b); /* test c2mir_compile */
//   a = ((func) (gen ())) (a, b);  /* test MIR_gen */
  printf ("a = %d\n", a);
  return 0;
}

test

gcc test.c ./libmir.a 
valgrind --tool=massif ./a.out 
ms_print massif.out.{number}

• MIR_gen memory analysis

->78.08% (1,312,816B) 0x130517: reduce_decode_start (mir-reduce.h:346)
| ->78.08% (1,312,816B) 0x130517: MIR_read_with_func (mir.c:4892)
|   ->78.08% (1,312,816B) 0x11217F: gen (in /home/qwe661234/mir/a.out)
|     ->78.08% (1,312,816B) 0x1122CA: main (in /home/qwe661234/mir/a.out)
|       
->11.47% (192,794B) in 242 places, all below massif's threshold (1.00%)
| 
->04.38% (73,728B) 0x13AE58: VARR_HTAB_EL_val_tcreate (mir.c:2736)
| ->04.38% (73,728B) 0x13AE58: HTAB_val_t_create (mir.c:2736)
|   ->04.38% (73,728B) 0x13AE58: simplify_init (mir.c:2774)
|     ->04.38% (73,728B) 0x13AE58: _MIR_init (mir.c:628)
|       ->04.38% (73,728B) 0x111C6C: MIR_init (in /home/qwe661234/mir/a.out)
|         ->04.38% (73,728B) 0x1120F1: gen (in /home/qwe661234/mir/a.out)
|           ->04.38% (73,728B) 0x1122CA: main (in /home/qwe661234/mir/a.out)
|             
->01.95% (32,768B) 0x115616: VARR_HTAB_EL_string_tcreate (mir.c:335)
| ->01.95% (32,768B) 0x115616: HTAB_string_t_create (mir.c:335)
|   ->01.95% (32,768B) 0x115616: string_init (mir.c:357)
|     ->01.95% (32,768B) 0x13ACDC: _MIR_init (mir.c:624)
|       ->01.95% (32,768B) 0x111C6C: MIR_init (in /home/qwe661234/mir/a.out)
|         ->01.95% (32,768B) 0x1120F1: gen (in /home/qwe661234/mir/a.out)
|           ->01.95% (32,768B) 0x1122CA: main (in /home/qwe661234/mir/a.out)
|             
->01.95% (32,768B) 0x165AA9: VARR_HTAB_EL_def_tab_el_tcreate (mir-gen.c:1500)
| ->01.95% (32,768B) 0x165AA9: HTAB_def_tab_el_t_create (mir-gen.c:1500)
|   ->01.95% (32,768B) 0x165AA9: init_ssa (mir-gen.c:2011)
|     ->01.95% (32,768B) 0x165AA9: MIR_gen_init (mir-gen.c:5684)
|       ->01.95% (32,768B) 0x112121: gen (in /home/qwe661234/mir/a.out)
|         ->01.95% (32,768B) 0x1122CA: main (in /home/qwe661234/mir/a.out)
|           
->01.95% (32,768B) 0x165C04: VARR_size_tcreate (mir-gen.c:1505)
  ->01.95% (32,768B) 0x165C04: init_ssa (mir-gen.c:2013)
    ->01.95% (32,768B) 0x165C04: MIR_gen_init (mir-gen.c:5684)
      ->01.95% (32,768B) 0x112121: gen (in /home/qwe661234/mir/a.out)
        ->01.95% (32,768B) 0x1122CA: main (in /home/qwe661234/mir/a.out)

• c2mir_compile memory analysis

->33.61% (458,752B) 0x1B2EF2: VARR_HTAB_EL_symbol_tcreate (c2mir.c:119)
| ->33.61% (458,752B) 0x1B2EF2: HTAB_symbol_t_create (c2mir.c:119)
|   ->33.61% (458,752B) 0x1B2EF2: symbol_init (c2mir.c:5482)
|     ->33.61% (458,752B) 0x1B2EF2: context_init (c2mir.c:9731)
|       ->33.61% (458,752B) 0x1B2EF2: compile_init (c2mir.c:13441)
|         ->33.61% (458,752B) 0x1B2EF2: c2mir_compile (c2mir.c:13484)
|           ->33.61% (458,752B) 0x111E3D: compile (in /home/qwe661234/mir/a.out)
|             ->33.61% (458,752B) 0x112096: block_compile (in /home/qwe661234/mir/a.out)
|               ->33.61% (458,752B) 0x1122CA: main (in /home/qwe661234/mir/a.out)

[vnmakarov]

Sorry I can not reproduce these huge numbers (1.3GiB). What I have with my massif usage (gen):

    MB
1.604^                                                           #
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                                       ::@:#
     |                                     ::::::::::::::::@:::@:#:::@:::::
     |                               ::::::::              @:::@:#:::@:::::
     |                          ::::::  :  ::              @:::@:#:::@:::::@@
     |                     :::::@::  :  :  ::              @:::@:#:::@:::::@@:
   0 +----------------------------------------------------------------------->ki
     0                                                                   641.0

And when block_compile is called:

    MB
1.877^                                                       ############
     |                                                       #
     |                                                       #
     |                                                       #
     |                                                       #
     |                                                       #
     |                                                       #
     |                                      @::::::::::::::: #
     |                        :::@:::@:@::::@:::: ::: ::: :  #
     |                  :::::::::@:::@:@::::@:::: ::: ::: :  #
     |                  :     :::@:::@:@::::@:::: ::: ::: :  #
     |                  :     :::@:::@:@::::@:::: ::: ::: :  #
     |                  :     :::@:::@:@::::@:::: ::: ::: :  #
     |                  :     :::@:::@:@::::@:::: ::: ::: : :#
     |               ::@:     :::@:::@:@::::@:::: ::: ::: : :#           :::
     |          ::::@: @:     :::@:::@:@::::@:::: ::: ::: : :#           :::
     |         ::   @: @:     :::@:::@:@::::@:::: ::: ::: : :#           :::
     |        :::   @: @:     :::@:::@:@::::@:::: ::: ::: : :#           :::::
     |        :::   @: @:     :::@:::@:@::::@:::: ::: ::: : :#           :::::
     |     :@@:::   @: @:     :::@:::@:@::::@:::: ::: ::: : :#           :::::
   0 +----------------------------------------------------------------------->Mi
     0                                                                   2.434

This is consistent with max resident memory (3MB) reported by /usr/bin/time.

Also the first snapshot you posted shows 1,312,816B which is about 1.3MiB (not 1.3GiB as you wrote in your first paragraph).