| … | |
… | |
| 6 | #include <cstdlib> |
6 | #include <cstdlib> |
| 7 | |
7 | |
| 8 | #include <tr1/unordered_set> |
8 | #include <tr1/unordered_set> |
| 9 | |
9 | |
| 10 | #include "shstr.h" |
10 | #include "shstr.h" |
|
|
11 | #include "util.h" |
| 11 | |
12 | |
| 12 | struct hash |
|
|
| 13 | { |
|
|
| 14 | std::size_t operator ()(const char *s) const |
|
|
| 15 | { |
|
|
| 16 | unsigned long hash = 0; |
|
|
| 17 | unsigned int i = 0; |
|
|
| 18 | |
|
|
| 19 | /* use the one-at-a-time hash function, which supposedly is |
|
|
| 20 | * better than the djb2-like one used by perl5.005, but |
|
|
| 21 | * certainly is better then the bug used here before. |
|
|
| 22 | * see http://burtleburtle.net/bob/hash/doobs.html |
|
|
| 23 | */ |
|
|
| 24 | while (*s) |
|
|
| 25 | { |
|
|
| 26 | hash += *s++; |
|
|
| 27 | hash += hash << 10; |
|
|
| 28 | hash ^= hash >> 6; |
|
|
| 29 | } |
|
|
| 30 | |
|
|
| 31 | hash += hash << 3; |
|
|
| 32 | hash ^= hash >> 11; |
|
|
| 33 | hash += hash << 15; |
|
|
| 34 | |
|
|
| 35 | return hash; |
|
|
| 36 | } |
|
|
| 37 | }; |
|
|
| 38 | |
|
|
| 39 | struct equal |
|
|
| 40 | { |
|
|
| 41 | bool operator ()(const char *a, const char *b) const |
|
|
| 42 | { |
|
|
| 43 | return !strcmp (a, b); |
|
|
| 44 | } |
|
|
| 45 | }; |
|
|
| 46 | |
|
|
| 47 | typedef std::tr1::unordered_set<const char *, hash, equal> HT; |
13 | typedef std::tr1::unordered_set<const char *, str_hash, str_equal> HT; |
| 48 | |
14 | |
| 49 | static HT ht; |
15 | static HT ht; |
|
|
16 | |
|
|
17 | static const char *makevec (const char *s) |
|
|
18 | { |
|
|
19 | int len = strlen (s); |
|
|
20 | |
|
|
21 | const char *v = (const char *)(2 + (int *)malloc (sizeof (int) * 2 + len + 1)); |
|
|
22 | |
|
|
23 | shstr::length (v) = len; |
|
|
24 | shstr::refcnt (v) = 1; |
|
|
25 | |
|
|
26 | memcpy ((char *)v, s, len + 1); |
|
|
27 | |
|
|
28 | return v; |
|
|
29 | } |
|
|
30 | |
|
|
31 | const char *shstr::null = makevec ("<nil>"); |
|
|
32 | |
|
|
33 | // what weird misoptimisation is this again? |
|
|
34 | const shstr undead_name ("undead"); |
| 50 | |
35 | |
| 51 | const char * |
36 | const char * |
| 52 | shstr::find (const char *s) |
37 | shstr::find (const char *s) |
| 53 | { |
38 | { |
| 54 | if (!s) |
39 | if (!s) |
| 55 | return s; |
40 | return s; |
| 56 | |
41 | |
| 57 | HT::iterator i = ht.find (s); |
42 | HT::iterator i = ht.find (s); |
| 58 | |
43 | |
| 59 | return i != ht.end () |
44 | return i != ht.end () |
| 60 | ? (char *)*i |
45 | ? *i |
| 61 | : 0; |
46 | : 0; |
| 62 | } |
47 | } |
| 63 | |
48 | |
| 64 | const char * |
49 | const char * |
| 65 | shstr::intern (const char *s) |
50 | shstr::intern (const char *s) |
| 66 | { |
51 | { |
| 67 | if (!s) |
52 | if (!s) |
| 68 | return s; |
53 | return null; |
| 69 | |
54 | |
| 70 | if (const char *found = find (s)) |
55 | if (const char *found = find (s)) |
|
|
56 | { |
|
|
57 | ++refcnt (found); |
| 71 | return found; |
58 | return found; |
|
|
59 | } |
| 72 | |
60 | |
| 73 | int len = strlen (s); |
61 | s = makevec (s); |
| 74 | |
|
|
| 75 | int *v = (int *)malloc (sizeof (int) * 2 + len + 1); |
|
|
| 76 | |
|
|
| 77 | v [0] = len; |
|
|
| 78 | v [1] = 0; |
|
|
| 79 | |
|
|
| 80 | v += 2; |
|
|
| 81 | |
|
|
| 82 | memcpy (v, s, len + 1); |
|
|
| 83 | |
|
|
| 84 | ht.insert ((char *)v); |
62 | ht.insert (s); |
| 85 | |
63 | return s; |
| 86 | return (char *)v; |
|
|
| 87 | } |
64 | } |
| 88 | |
65 | |
| 89 | // TODO: periodically test refcounts == 0 for a few strings (e.g. one hash bucket, |
66 | // periodically test refcounts == 0 for a few strings |
| 90 | // exploiting the fatc that iterators stay valid for unordered_set). |
67 | // this is the ONLY thing that erases stuff from ht. keep it that way. |
| 91 | void |
68 | void |
| 92 | shstr::gc () |
69 | shstr::gc () |
| 93 | { |
70 | { |
|
|
71 | static const char *curpos; |
|
|
72 | |
|
|
73 | HT::iterator i = curpos ? ht.find (curpos) : ht.begin (); |
|
|
74 | |
|
|
75 | if (i == ht.end ()) |
|
|
76 | i = ht.begin (); |
|
|
77 | |
|
|
78 | // go through all strings roughly once every 4 minutes |
|
|
79 | int n = ht.size () / 256 + 16; |
|
|
80 | |
|
|
81 | for (;;) |
|
|
82 | { |
|
|
83 | if (i == ht.end ()) |
|
|
84 | { |
|
|
85 | curpos = 0; |
|
|
86 | return; |
|
|
87 | } |
|
|
88 | else if (!--n) |
|
|
89 | break; |
|
|
90 | else if (!refcnt (*i)) |
|
|
91 | { |
|
|
92 | HT::iterator o = i++; |
|
|
93 | const char *s = *o; |
|
|
94 | ht.erase (o); |
|
|
95 | |
|
|
96 | //printf ("GC %4d %3d %d >%s<%d\n", (int)ht.size (), n, shstr::refcnt (s), s, shstr::length (s)); |
|
|
97 | free (-2 + (int *)s); |
|
|
98 | } |
|
|
99 | else |
|
|
100 | ++i; |
|
|
101 | } |
|
|
102 | |
|
|
103 | curpos = *i; |
| 94 | } |
104 | } |
| 95 | |
105 | |
|
|
106 | //TODO: this should of course not be here |
| 96 | /* buf_overflow() - we don't want to exceed the buffer size of |
107 | /* buf_overflow() - we don't want to exceed the buffer size of |
| 97 | * buf1 by adding on buf2! Returns true if overflow will occur. |
108 | * buf1 by adding on buf2! Returns true if overflow will occur. |
| 98 | */ |
109 | */ |
| 99 | |
110 | |
| 100 | int |
111 | int |
