| @@ -1,5 +1,5 @@ |
| 1 | 1 | /* |
| 2 | | −** $Id: lgc.c,v 2.128 2012/05/23 15:43:14 roberto Exp $ |
| 2 | +** $Id: lgc.c,v 2.129 2012/05/28 20:41:00 roberto Exp $ |
| 3 | 3 | ** Garbage Collector |
| 4 | 4 | ** See Copyright Notice in lua.h |
| 5 | 5 | */ |
| @@ -24,8 +24,11 @@ · 18 unmodified lines |
| 24 | 24 | |
| 25 | 25 | |
| 26 | 26 | |
| 27 | | −/* cost of sweeping one element (half the size of a small object) */ |
| 28 | | −#define GCSWEEPCOST ((sizeof(TString) + 4) / 2) |
| 27 | +/* |
| 28 | +** cost of sweeping one element (the size of a small object divided |
| 29 | +** by some adjust for the sweep speed) |
| 30 | +*/ |
| 31 | +#define GCSWEEPCOST ((sizeof(TString) + 4) / 4) |
| 29 | 32 | |
| 30 | 33 | /* maximum number of elements to sweep in each single step */ |
| 31 | 34 | #define GCSWEEPMAX (cast_int((GCSTEPSIZE / GCSWEEPCOST) / 4)) |
| @@ -33,25 +36,27 @@ · 1 unmodified line |
| 33 | 36 | /* maximum number of finalizers to call in each GC step */ |
| 34 | 37 | #define GCFINALIZENUM 4 |
| 35 | 38 | |
| 36 | | −/* (arbitrary) cost of atomic step */ |
| 37 | | −#define GCATOMICCOST GCSTEPSIZE |
| 38 | 39 | |
| 40 | +/* |
| 41 | +** macro to adjust 'stepmul': 'stepmul' is actually used like |
| 42 | +** 'stepmul / STEPMULADJ' (value chosen by tests) |
| 43 | +*/ |
| 44 | +#define STEPMULADJ 200 |
| 39 | 45 | |
| 40 | 46 | /* |
| 41 | | −** macro to apply the "speed" of the garbage collector: the constant |
| 42 | | −** 80 makes the standard 'stepmul' of 200 results in the GC handling |
| 43 | | −** 80/200 = 1/2.5 = 0.4Kbytes for every 1Kb allocated. |
| 44 | | −** (The computation tries to avoid overflows or underflows.) |
| 47 | +** macro to adjust 'pause': 'pause' is actually used like |
| 48 | +** 'pause / PAUSEADJ' (value chosen by tests) |
| 45 | 49 | */ |
| 46 | | −#define workrate(x,mul) \ |
| 47 | | − ((x) < MAX_INT/80 ? ((x) * 80) / mul : ((x) / mul) * 80) |
| 50 | +#define PAUSEADJ 200 |
| 51 | + |
| 52 | + |
| 48 | 53 | |
| 49 | 54 | |
| 50 | 55 | /* |
| 51 | 56 | ** standard negative debt for GC; a reasonable "time" to wait before |
| 52 | 57 | ** starting a new cycle |
| 53 | 58 | */ |
| 54 | | −#define stddebtest(g,e) (-cast(l_mem, (e)/100) * g->gcpause) |
| 59 | +#define stddebtest(g,e) (-cast(l_mem, (e)/PAUSEADJ) * g->gcpause) |
| 55 | 60 | #define stddebt(g) stddebtest(g, gettotalbytes(g)) |
| 56 | 61 | |
| 57 | 62 | |
| @@ -672,7 +677,7 @@ static void freeobj (lua_State *L, GCObject *o) { · 614 unmodified lines |
| 672 | 677 | case LUA_TTABLE: luaH_free(L, gco2t(o)); break; |
| 673 | 678 | case LUA_TTHREAD: luaE_freethread(L, gco2th(o)); break; |
| 674 | 679 | case LUA_TUSERDATA: luaM_freemem(L, o, sizeudata(gco2u(o))); break; |
| 675 | | − case LUA_TSHRSTR: |
| 680 | + case LUA_TSHRSTR: |
| 676 | 681 | G(L)->strt.nuse--; |
| 677 | 682 | /* go through */ |
| 678 | 683 | case LUA_TLNGSTR: { |
| @@ -955,8 +960,9 @@ void luaC_freeallobjects (lua_State *L) { · 276 unmodified lines |
| 955 | 960 | } |
| 956 | 961 | |
| 957 | 962 | |
| 958 | | −static void atomic (lua_State *L) { |
| 963 | +static l_mem atomic (lua_State *L) { |
| 959 | 964 | global_State *g = G(L); |
| 965 | + l_mem trav = g->GCmemtrav; |
| 960 | 966 | GCObject *origweak, *origall; |
| 961 | 967 | lua_assert(!iswhite(obj2gco(g->mainthread))); |
| 962 | 968 | markobject(g, L); /* mark running thread */ |
| @@ -976,6 +982,7 @@ static void atomic (lua_State *L) { · 13 unmodified lines |
| 976 | 982 | separatetobefnz(L, 0); /* separate objects to be finalized */ |
| 977 | 983 | markbeingfnz(g); /* mark userdata that will be finalized */ |
| 978 | 984 | propagateall(g); /* remark, to propagate `preserveness' */ |
| 985 | + trav = g->GCmemtrav - trav; /* avoid adding convergence twice */ |
| 979 | 986 | convergeephemerons(g); |
| 980 | 987 | /* at this point, all resurrected objects are marked. */ |
| 981 | 988 | /* remove dead objects from weak tables */ |
| @@ -987,6 +994,7 @@ static void atomic (lua_State *L) { · 5 unmodified lines |
| 987 | 994 | g->currentwhite = cast_byte(otherwhite(g)); /* flip current white */ |
| 988 | 995 | entersweep(L); /* prepare to sweep strings */ |
| 989 | 996 | /*lua_checkmemory(L);*/ |
| 997 | + return trav; /* reasonable estimate of the work done by 'atomic' */ |
| 990 | 998 | } |
| 991 | 999 | |
| 992 | 1000 | |
| @@ -1012,8 +1020,7 @@ static lu_mem singlestep (lua_State *L) { · 19 unmodified lines |
| 1012 | 1020 | else { /* no more `gray' objects */ |
| 1013 | 1021 | g->gcstate = GCSatomic; /* finish mark phase */ |
| 1014 | 1022 | g->GCestimate = g->GCmemtrav; /* save what was counted */ |
| 1015 | | − atomic(L); |
| 1016 | | − return GCATOMICCOST; |
| 1023 | + return atomic(L); |
| 1017 | 1024 | } |
| 1018 | 1025 | } |
| 1019 | 1026 | case GCSsweepstring: { |
| @@ -1086,14 +1093,18 @@ static void step (lua_State *L) { · 66 unmodified lines |
| 1086 | 1093 | global_State *g = G(L); |
| 1087 | 1094 | l_mem debt = g->GCdebt; |
| 1088 | 1095 | int stepmul = g->gcstepmul; |
| 1089 | | − if (stepmul <= 0) stepmul = 1; |
| 1096 | + if (stepmul < 40) stepmul = 40; /* avoid ridiculous low values */ |
| 1097 | + /* convert debt from Kb to 'work units' (avoid zero debt and overflows) */ |
| 1098 | + debt = (debt / STEPMULADJ) + 1; |
| 1099 | + debt = (debt < MAX_LMEM / stepmul) ? debt * stepmul : MAX_LMEM; |
| 1090 | 1100 | do { /* always perform at least one single step */ |
| 1091 | 1101 | lu_mem work = singlestep(L); /* do some work */ |
| 1092 | | − work = workrate(work, stepmul); /* apply work rate */ |
| 1093 | 1102 | debt -= work; |
| 1094 | 1103 | } while (debt > -GCSTEPSIZE && g->gcstate != GCSpause); |
| 1095 | 1104 | if (g->gcstate == GCSpause) |
| 1096 | 1105 | debt = stddebtest(g, g->GCestimate); /* pause until next cycle */ |
| 1106 | + else |
| 1107 | + debt = (debt / stepmul) * STEPMULADJ; /* convert 'work units' to Kb */ |
| 1097 | 1108 | luaE_setdebt(g, debt); |
| 1098 | 1109 | } |
| 1099 | 1110 | |