server/server/dynbuf.C

#include "global.h"

#include <cstdio>

dynbuf::dynbuf (int initial, int extend)
{
  ext = extend;
  _size = 0;

  first = last = (chunk *)salloc<char> (sizeof (chunk) + initial);
  first->alloc = sizeof (chunk) + initial;
  first->next = 0;

  ptr = first->data;
  end = ptr + initial;
}

dynbuf::~dynbuf ()
{
  _clear ();
}

void
dynbuf::_clear ()
{
  while (first)
    {
      chunk *next = first->next;

      sfree<char> ((char *)first, first->alloc);
      first = next;
    }
}

void
dynbuf::clear ()
{
  _clear ();
  _size = 0;

  first = last = (chunk *)salloc<char> (sizeof (chunk) + ext);
  first->alloc = sizeof (chunk) + ext;
  first->next = 0;

  ptr = first->data;
  end = ptr + ext;
}

void
dynbuf::finish ()
{
  // finalise current chunk
  _size += last->size = ptr - last->data;
}

void
dynbuf::_reserve (int size)
{
  finish ();

  do
    {
      ext += ext >> 1;
      ext = (ext + 15) & ~15;
    }
  while (ext < size);

  chunk *add = (chunk *) salloc<char> (sizeof (chunk) + ext);
  add->alloc = sizeof (chunk) + ext;
  add->next = 0;

  last->next = add;
  last = add;

  ptr = last->data;
  end = ptr + ext;
}

void
dynbuf::linearise (void *data)
{
  last->size = ptr - last->data;

  for (chunk *c = first; c; c = c->next)
    {
      memcpy (data, c->data, c->size);
      data = (void *)(((char *)data) + c->size);
    }
}

char *
dynbuf::linearise ()
{
  if (first->next)
    {
      finish ();

      chunk *add = (chunk *) salloc<char> (sizeof (chunk) + _size);
      add->alloc = sizeof (chunk) + _size;
      add->next = 0;

      linearise ((void *)add->data);
      _clear ();

      first = last = add;
      ptr = last->data + _size;
      end = ptr;
      _size = 0;
    }

  return first->data;
}

dynbuf::operator std::string ()
{
  // could optimise
  return std::string (linearise (), size ());
}

void
dynbuf_text::printf (const char *format, ...)
{
  int len;

  {
    force (128);

    va_list ap;
    va_start (ap, format);
    len = vsnprintf (ptr, end - ptr, format, ap);
    va_end (ap);

    assert (len >= 0); // shield against broken vsnprintf's

    // was enough room available
    if (ptr + len < end)
      {
        ptr += len;
        return;
      }
  }

  // longer, try harder
  va_list ap;
  va_start (ap, format);
  vsnprintf (force (len + 1), len + 1, format, ap);
  va_end (ap);

  ptr += len;
}

// simply return a mask with "bits" bits set
inline uint64
m (int b)
{
  return (uint64 (1) << b) - 1;
}

// convert 9 digits to ascii, using only a single multiplication
// (depending on cpu and compiler).
// will generate a single 0 as output when v=lz=0
inline char *
i2a_9 (char *ptr, uint32 v, bool lz)
{
  // convert to 4.56 fixed-point representation
  // this should be optimal on 64 bit cpus, and rather
  // slow on 32 bit cpus. go figure :)
  const int bits = 7*8; // 7 bits per post-comma digit

  uint64 u = v * ((m (bits) + 100000000) / 100000000); // 10**8

  if (lz)
    {
      // output leading zeros
      // good compilers will compile this into only shifts, masks and adds
      *ptr++ = char (u >> (bits - 0)) + '0'; u = (u & m (bits - 0)) * 5;
      *ptr++ = char (u >> (bits - 1)) + '0'; u = (u & m (bits - 1)) * 5;
      *ptr++ = char (u >> (bits - 2)) + '0'; u = (u & m (bits - 2)) * 5;
      *ptr++ = char (u >> (bits - 3)) + '0'; u = (u & m (bits - 3)) * 5;
      *ptr++ = char (u >> (bits - 4)) + '0'; u = (u & m (bits - 4)) * 5;
      *ptr++ = char (u >> (bits - 5)) + '0'; u = (u & m (bits - 5)) * 5;
      *ptr++ = char (u >> (bits - 6)) + '0'; u = (u & m (bits - 6)) * 5;
      *ptr++ = char (u >> (bits - 7)) + '0'; u = (u & m (bits - 7)) * 5;
      *ptr++ = char (u >> (bits - 8)) + '0';
    }
  else
    {
      // do not output leading zeroes (except if v == 0)
      // good compilers will compile this into completely branchless code
      char digit, nz = 0;

      digit = (u >> (bits - 0)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 0)) * 5;
      digit = (u >> (bits - 1)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 1)) * 5;
      digit = (u >> (bits - 2)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 2)) * 5;
      digit = (u >> (bits - 3)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 3)) * 5;
      digit = (u >> (bits - 4)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 4)) * 5;
      digit = (u >> (bits - 5)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 5)) * 5;
      digit = (u >> (bits - 6)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 6)) * 5;
      digit = (u >> (bits - 7)); *ptr = digit + '0'; nz |= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 7)) * 5;
      digit = (u >> (bits - 8)); *ptr = digit + '0'; nz |= digit; ptr += 1;
    }

  return ptr;
}

void
dynbuf_text::add (sint32 i)
{
  force (11); // 10 digits + '-'

  *ptr = '-'; ptr += i < 0 ? 1 : 0;
  uint32 u = i < 0 ? -i : i;

  if (expect_true (u < 10)) // we have a lot of single-digit numbers, so optimise
    fadd (char (u + '0'));
  else if (expect_true (u < 1000000000)) // 9 0's
    ptr = i2a_9 (ptr, u, false);
  else
    {
      sint32 div = u / 1000000000;
      uint32 rem = u % 1000000000;

      ptr = i2a_9 (ptr, div, false);
      ptr = i2a_9 (ptr, rem, true);
    }
}

void
dynbuf_text::add (sint64 i)
{
  force (20); // 19 digits + '-'

  *ptr = '-'; ptr += i < 0 ? 1 : 0;
  uint64 u = i < 0 ? -i : i;

  // split the number into a 1-digit part
  // (#19) and two 9 digit parts (9..18 and 0..8)

  // good compilers will only use multiplications here

  if (u < 10) // we have a lot of single-digit numbers, so optimise
    fadd (char (u + '0'));
  else if (expect_true (u < 1000000000)) // 9 0's
    ptr = i2a_9 (ptr, u, false);
  else if (expect_true (u < UINT64_C (1000000000000000000))) // 18 0's
    {
      sint32 div = u / 1000000000;
      uint32 rem = u % 1000000000;

      ptr = i2a_9 (ptr, div, false);
      ptr = i2a_9 (ptr, rem, true);
    }
  else
    {
      // a biggy
      sint32 div = u / UINT64_C (1000000000000000000);
      uint64 rem = u % UINT64_C (1000000000000000000);

      fadd (char (div + '0'));
      u = rem;

      {
        sint32 div = u / 1000000000;
        uint32 rem = u % 1000000000;

        ptr = i2a_9 (ptr, div, true);
        ptr = i2a_9 (ptr, rem, true);
      }
    }
}

#if 0
struct dynbuf_test_class {
  dynbuf_test_class ()
  {
    sint64 s = 0;
    for (int i = 0; i < 10000000; ++i)
      {
        char b1[256], b2[256];

        dynbuf_text db;
        db.add (s);
        db.add (char (0));

        db.linearise (b1);
        sprintf (b2, "%ld", s);

        if (strcmp (b1, b2))
          printf ("<%s,%s>\n", b1, b2);

        if (i < 20)
          s = (sint64) pow (10., i);
        else
          s = (sint64) exp (random () * (43.6682723752766 / RAND_MAX));
      }

    exit (0);
  }
} dynbuf_test;
#endif
Revision:	1.13
Committed:	Sun May 27 23:22:29 2007 UTC (17 years ago) by root
Content type:	text/plain
Branch:	MAIN
Changes since 1.12:	+2 -2 lines
Log Message:	minor cleanups/optimisations
#	User	Rev	Content
1	root	1.1	#include "global.h"
2
3			#include <cstdio>
4
5			dynbuf::dynbuf (int initial, int extend)
6			{
7	root	1.10	ext = extend;
8	root	1.1	_size = 0;
9	root	1.5
10	root	1.8	first = last = (chunk *)salloc<char> (sizeof (chunk) + initial);
11			first->alloc = sizeof (chunk) + initial;
12	root	1.1	first->next = 0;
13	root	1.8
14	root	1.5	ptr = first->data;
15	root	1.9	end = ptr + initial;
16	root	1.1	}
17
18			dynbuf::~dynbuf ()
19			{
20	root	1.10	_clear ();
21	root	1.1	}
22
23	root	1.5	void
24	root	1.10	dynbuf::_clear ()
25	root	1.1	{
26			while (first)
27			{
28			chunk *next = first->next;
29	root	1.5
30	root	1.8	sfree<char> ((char *)first, first->alloc);
31	root	1.1	first = next;
32			}
33	root	1.5	}
34	root	1.1
35	root	1.5	void
36	root	1.10	dynbuf::clear ()
37			{
38			_clear ();
39			_size = 0;
40
41			first = last = (chunk *)salloc<char> (sizeof (chunk) + ext);
42			first->alloc = sizeof (chunk) + ext;
43			first->next = 0;
44
45			ptr = first->data;
46			end = ptr + ext;
47			}
48
49			void
50	root	1.5	dynbuf::finish ()
51	root	1.1	{
52			// finalise current chunk
53			_size += last->size = ptr - last->data;
54			}
55
56	root	1.5	void
57			dynbuf::_reserve (int size)
58	root	1.1	{
59			finish ();
60
61			do
62			{
63			ext += ext >> 1;
64			ext = (ext + 15) & ~15;
65			}
66			while (ext < size);
67
68	root	1.8	chunk add = (chunk ) salloc<char> (sizeof (chunk) + ext);
69			add->alloc = sizeof (chunk) + ext;
70	root	1.1	add->next = 0;
71
72			last->next = add;
73			last = add;
74
75	root	1.5	ptr = last->data;
76	root	1.9	end = ptr + ext;
77	root	1.1	}
78
79	root	1.5	void
80			dynbuf::linearise (void *data)
81	root	1.1	{
82			last->size = ptr - last->data;
83
84	root	1.10	for (chunk *c = first; c; c = c->next)
85	root	1.1	{
86	root	1.10	memcpy (data, c->data, c->size);
87			data = (void )(((char )data) + c->size);
88	root	1.1	}
89			}
90
91	root	1.5	char *
92			dynbuf::linearise ()
93	root	1.1	{
94			if (first->next)
95			{
96			finish ();
97
98	root	1.8	chunk add = (chunk ) salloc<char> (sizeof (chunk) + _size);
99			add->alloc = sizeof (chunk) + _size;
100			add->next = 0;
101	root	1.1
102	root	1.8	linearise ((void *)add->data);
103	root	1.10	_clear ();
104	root	1.1
105			first = last = add;
106	root	1.5	ptr = last->data + _size;
107	root	1.9	end = ptr;
108	root	1.1	_size = 0;
109			}
110
111			return first->data;
112			}
113
114	root	1.7	dynbuf::operator std::string ()
115			{
116			// could optimise
117			return std::string (linearise (), size ());
118			}
119
120			void
121			dynbuf_text::printf (const char *format, ...)
122			{
123			int len;
124
125			{
126			force (128);
127
128			va_list ap;
129			va_start (ap, format);
130	root	1.9	len = vsnprintf (ptr, end - ptr, format, ap);
131	root	1.7	va_end (ap);
132
133			assert (len >= 0); // shield against broken vsnprintf's
134
135			// was enough room available
136	root	1.9	if (ptr + len < end)
137	root	1.7	{
138	root	1.13	ptr += len;
139	root	1.7	return;
140			}
141			}
142
143			// longer, try harder
144			va_list ap;
145			va_start (ap, format);
146			vsnprintf (force (len + 1), len + 1, format, ap);
147			va_end (ap);
148
149	root	1.13	ptr += len;
150	root	1.7	}
151
152	root	1.11	// simply return a mask with "bits" bits set
153			inline uint64
154			m (int b)
155	root	1.1	{
156	root	1.11	return (uint64 (1) << b) - 1;
157			}
158	root	1.1
159	root	1.11	// convert 9 digits to ascii, using only a single multiplication
160			// (depending on cpu and compiler).
161			// will generate a single 0 as output when v=lz=0
162			inline char *
163			i2a_9 (char *ptr, uint32 v, bool lz)
164			{
165			// convert to 4.56 fixed-point representation
166			// this should be optimal on 64 bit cpus, and rather
167			// slow on 32 bit cpus. go figure :)
168	root	1.12	const int bits = 7*8; // 7 bits per post-comma digit
169	root	1.11
170			uint64 u = v * ((m (bits) + 100000000) / 100000000); // 10**8
171
172			if (lz)
173			{
174			// output leading zeros
175			// good compilers will compile this into only shifts, masks and adds
176			ptr++ = char (u >> (bits - 0)) + '0'; u = (u & m (bits - 0)) 5;
177			ptr++ = char (u >> (bits - 1)) + '0'; u = (u & m (bits - 1)) 5;
178			ptr++ = char (u >> (bits - 2)) + '0'; u = (u & m (bits - 2)) 5;
179			ptr++ = char (u >> (bits - 3)) + '0'; u = (u & m (bits - 3)) 5;
180			ptr++ = char (u >> (bits - 4)) + '0'; u = (u & m (bits - 4)) 5;
181			ptr++ = char (u >> (bits - 5)) + '0'; u = (u & m (bits - 5)) 5;
182			ptr++ = char (u >> (bits - 6)) + '0'; u = (u & m (bits - 6)) 5;
183			ptr++ = char (u >> (bits - 7)) + '0'; u = (u & m (bits - 7)) 5;
184			*ptr++ = char (u >> (bits - 8)) + '0';
185	root	1.1	}
186			else
187	root	1.4	{
188	root	1.11	// do not output leading zeroes (except if v == 0)
189			// good compilers will compile this into completely branchless code
190			char digit, nz = 0;
191
192			digit = (u >> (bits - 0)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 0)) 5;
193			digit = (u >> (bits - 1)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 1)) 5;
194			digit = (u >> (bits - 2)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 2)) 5;
195			digit = (u >> (bits - 3)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 3)) 5;
196			digit = (u >> (bits - 4)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 4)) 5;
197			digit = (u >> (bits - 5)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 5)) 5;
198			digit = (u >> (bits - 6)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 6)) 5;
199			digit = (u >> (bits - 7)); ptr = digit + '0'; nz \|= digit; ptr += nz ? 1 : 0; u = (u & m (bits - 7)) 5;
200			digit = (u >> (bits - 8)); *ptr = digit + '0'; nz \|= digit; ptr += 1;
201	root	1.4	}
202	root	1.1
203	root	1.11	return ptr;
204			}
205
206			void
207			dynbuf_text::add (sint32 i)
208			{
209			force (11); // 10 digits + '-'
210
211			*ptr = '-'; ptr += i < 0 ? 1 : 0;
212			uint32 u = i < 0 ? -i : i;
213
214			if (expect_true (u < 10)) // we have a lot of single-digit numbers, so optimise
215			fadd (char (u + '0'));
216			else if (expect_true (u < 1000000000)) // 9 0's
217			ptr = i2a_9 (ptr, u, false);
218			else
219	root	1.1	{
220	root	1.11	sint32 div = u / 1000000000;
221			uint32 rem = u % 1000000000;
222	root	1.5
223	root	1.11	ptr = i2a_9 (ptr, div, false);
224			ptr = i2a_9 (ptr, rem, true);
225	root	1.1	}
226			}
227
228	root	1.5	void
229	root	1.7	dynbuf_text::add (sint64 i)
230	root	1.1	{
231	root	1.11	force (20); // 19 digits + '-'
232
233			*ptr = '-'; ptr += i < 0 ? 1 : 0;
234			uint64 u = i < 0 ? -i : i;
235
236			// split the number into a 1-digit part
237			// (#19) and two 9 digit parts (9..18 and 0..8)
238	root	1.4
239	root	1.11	// good compilers will only use multiplications here
240	root	1.2
241	root	1.11	if (u < 10) // we have a lot of single-digit numbers, so optimise
242			fadd (char (u + '0'));
243			else if (expect_true (u < 1000000000)) // 9 0's
244			ptr = i2a_9 (ptr, u, false);
245			else if (expect_true (u < UINT64_C (1000000000000000000))) // 18 0's
246	root	1.1	{
247	root	1.11	sint32 div = u / 1000000000;
248			uint32 rem = u % 1000000000;
249
250			ptr = i2a_9 (ptr, div, false);
251			ptr = i2a_9 (ptr, rem, true);
252	root	1.1	}
253	root	1.2	else
254	root	1.1	{
255	root	1.11	// a biggy
256			sint32 div = u / UINT64_C (1000000000000000000);
257			uint64 rem = u % UINT64_C (1000000000000000000);
258
259			fadd (char (div + '0'));
260			u = rem;
261
262			{
263			sint32 div = u / 1000000000;
264			uint32 rem = u % 1000000000;
265
266			ptr = i2a_9 (ptr, div, true);
267			ptr = i2a_9 (ptr, rem, true);
268			}
269	root	1.1	}
270	root	1.11	}
271	root	1.1
272	root	1.11	#if 0
273			struct dynbuf_test_class {
274			dynbuf_test_class ()
275			{
276			sint64 s = 0;
277			for (int i = 0; i < 10000000; ++i)
278			{
279			char b1[256], b2[256];
280
281			dynbuf_text db;
282			db.add (s);
283			db.add (char (0));
284
285			db.linearise (b1);
286			sprintf (b2, "%ld", s);
287
288			if (strcmp (b1, b2))
289			printf ("<%s,%s>\n", b1, b2);
290
291			if (i < 20)
292			s = (sint64) pow (10., i);
293			else
294			s = (sint64) exp (random () * (43.6682723752766 / RAND_MAX));
295			}
296
297			exit (0);
298			}
299			} dynbuf_test;
300			#endif