2 |
- |
1 |
/* Copyright (C) 1989-2004 artofcode LLC. All rights reserved.
|
|
|
2 |
|
|
|
3 |
This software is provided AS-IS with no warranty, either express or
|
|
|
4 |
implied.
|
|
|
5 |
|
|
|
6 |
This software is distributed under license and may not be copied,
|
|
|
7 |
modified or distributed except as expressly authorized under the terms
|
|
|
8 |
of the license contained in the file LICENSE in this distribution.
|
|
|
9 |
|
|
|
10 |
For more information about licensing, please refer to
|
|
|
11 |
http://www.ghostscript.com/licensing/. For information on
|
|
|
12 |
commercial licensing, go to http://www.artifex.com/licensing/ or
|
|
|
13 |
contact Artifex Software, Inc., 101 Lucas Valley Road #110,
|
|
|
14 |
San Rafael, CA 94903, U.S.A., +1(415)492-9861.
|
|
|
15 |
*/
|
|
|
16 |
|
|
|
17 |
/* $Id: genarch.c,v 1.11 2004/06/17 21:42:53 giles Exp $ */
|
|
|
18 |
/*
|
|
|
19 |
* Generate a header file (arch.h) with parameters
|
|
|
20 |
* reflecting the machine architecture and compiler characteristics.
|
|
|
21 |
*/
|
|
|
22 |
|
|
|
23 |
#include "stdpre.h"
|
|
|
24 |
#include <ctype.h>
|
|
|
25 |
#include <stdio.h>
|
|
|
26 |
/*
|
|
|
27 |
* In theory, not all systems provide <string.h> or <setjmp.h>, or declare
|
|
|
28 |
* memset in <string.h>, but at this point I don't think we care about any
|
|
|
29 |
* that don't.
|
|
|
30 |
*/
|
|
|
31 |
#include <string.h>
|
|
|
32 |
#include <time.h>
|
|
|
33 |
#include <setjmp.h>
|
|
|
34 |
|
|
|
35 |
/* We should write the result on stdout, but the original Turbo C 'make' */
|
|
|
36 |
/* can't handle output redirection (sigh). */
|
|
|
37 |
|
|
|
38 |
private void
|
|
|
39 |
section(FILE * f, const char *str)
|
|
|
40 |
{
|
|
|
41 |
fprintf(f, "\n\t /* ---------------- %s ---------------- */\n\n", str);
|
|
|
42 |
}
|
|
|
43 |
|
|
|
44 |
private clock_t
|
|
|
45 |
time_clear(char *buf, int bsize, int nreps)
|
|
|
46 |
{
|
|
|
47 |
clock_t t = clock();
|
|
|
48 |
int i;
|
|
|
49 |
|
|
|
50 |
for (i = 0; i < nreps; ++i)
|
|
|
51 |
memset(buf, 0, bsize);
|
|
|
52 |
return clock() - t;
|
|
|
53 |
}
|
|
|
54 |
|
|
|
55 |
private void
|
|
|
56 |
define(FILE *f, const char *str)
|
|
|
57 |
{
|
|
|
58 |
fprintf(f, "#define %s ", str);
|
|
|
59 |
}
|
|
|
60 |
|
|
|
61 |
private void
|
|
|
62 |
define_int(FILE *f, const char *str, int value)
|
|
|
63 |
{
|
|
|
64 |
fprintf(f, "#define %s %d\n", str, value);
|
|
|
65 |
}
|
|
|
66 |
|
|
|
67 |
private void
|
|
|
68 |
print_ffs(FILE *f, int nbytes)
|
|
|
69 |
{
|
|
|
70 |
int i;
|
|
|
71 |
|
|
|
72 |
for (i = 0; i < nbytes; ++i)
|
|
|
73 |
fprintf(f, "ff");
|
|
|
74 |
}
|
|
|
75 |
|
|
|
76 |
private int
|
|
|
77 |
ilog2(int n)
|
|
|
78 |
{
|
|
|
79 |
int i = 0, m = n;
|
|
|
80 |
|
|
|
81 |
while (m > 1)
|
|
|
82 |
++i, m = (m + 1) >> 1;
|
|
|
83 |
return i;
|
|
|
84 |
}
|
|
|
85 |
|
|
|
86 |
int
|
|
|
87 |
main(int argc, char *argv[])
|
|
|
88 |
{
|
|
|
89 |
char *fname = argv[1];
|
|
|
90 |
long one = 1;
|
|
|
91 |
struct {
|
|
|
92 |
char c;
|
|
|
93 |
short s;
|
|
|
94 |
} ss;
|
|
|
95 |
struct {
|
|
|
96 |
char c;
|
|
|
97 |
int i;
|
|
|
98 |
} si;
|
|
|
99 |
struct {
|
|
|
100 |
char c;
|
|
|
101 |
long l;
|
|
|
102 |
} sl;
|
|
|
103 |
struct {
|
|
|
104 |
char c;
|
|
|
105 |
char *p;
|
|
|
106 |
} sp;
|
|
|
107 |
struct {
|
|
|
108 |
char c;
|
|
|
109 |
float f;
|
|
|
110 |
} sf;
|
|
|
111 |
struct {
|
|
|
112 |
char c;
|
|
|
113 |
double d;
|
|
|
114 |
} sd;
|
|
|
115 |
/* Some architectures have special alignment requirements for jmpbuf. */
|
|
|
116 |
struct {
|
|
|
117 |
char c;
|
|
|
118 |
jmp_buf j;
|
|
|
119 |
} sj;
|
|
|
120 |
long lm1 = -1;
|
|
|
121 |
long lr1 = lm1 >> 1, lr2 = lm1 >> 2;
|
|
|
122 |
unsigned long um1 = ~(unsigned long)0;
|
|
|
123 |
int im1 = -1;
|
|
|
124 |
int ir1 = im1 >> 1, ir2 = im1 >> 2;
|
|
|
125 |
union {
|
|
|
126 |
long l;
|
|
|
127 |
char *p;
|
|
|
128 |
} pl0, pl1;
|
|
|
129 |
int ars;
|
|
|
130 |
int lwidth = size_of(long) * 8;
|
|
|
131 |
union {
|
|
|
132 |
float f;
|
|
|
133 |
int i;
|
|
|
134 |
long l;
|
|
|
135 |
} f0, f1, fm1;
|
|
|
136 |
int floats_are_IEEE;
|
|
|
137 |
FILE *f = fopen(fname, "w");
|
|
|
138 |
|
|
|
139 |
if (f == NULL) {
|
|
|
140 |
fprintf(stderr, "genarch.c: can't open %s for writing\n", fname);
|
|
|
141 |
return exit_FAILED;
|
|
|
142 |
}
|
|
|
143 |
fprintf(f, "/* Parameters derived from machine and compiler architecture. */\n");
|
|
|
144 |
fprintf(f, "/* This file is generated mechanically by genarch.c. */\n");
|
|
|
145 |
|
|
|
146 |
/* We have to test the size dynamically here, */
|
|
|
147 |
/* because the preprocessor can't evaluate sizeof. */
|
|
|
148 |
f0.f = 0.0, f1.f = 1.0, fm1.f = -1.0;
|
|
|
149 |
floats_are_IEEE =
|
|
|
150 |
(size_of(float) == size_of(int) ?
|
|
|
151 |
f0.i == 0 && f1.i == (int)0x3f800000 && fm1.i == (int)0xbf800000 :
|
|
|
152 |
f0.l == 0 && f1.l == 0x3f800000L && fm1.l == 0xbf800000L);
|
|
|
153 |
|
|
|
154 |
section(f, "Scalar alignments");
|
|
|
155 |
|
|
|
156 |
#define OFFSET_IN(s, e) (int)((char *)&s.e - (char *)&s)
|
|
|
157 |
define_int(f, "ARCH_ALIGN_SHORT_MOD", OFFSET_IN(ss, s));
|
|
|
158 |
define_int(f, "ARCH_ALIGN_INT_MOD", OFFSET_IN(si, i));
|
|
|
159 |
define_int(f, "ARCH_ALIGN_LONG_MOD", OFFSET_IN(sl, l));
|
|
|
160 |
define_int(f, "ARCH_ALIGN_PTR_MOD", OFFSET_IN(sp, p));
|
|
|
161 |
define_int(f, "ARCH_ALIGN_FLOAT_MOD", OFFSET_IN(sf, f));
|
|
|
162 |
define_int(f, "ARCH_ALIGN_DOUBLE_MOD", OFFSET_IN(sd, d));
|
|
|
163 |
define_int(f, "ARCH_ALIGN_STRUCT_MOD", OFFSET_IN(sj, j));
|
|
|
164 |
#undef OFFSET_IN
|
|
|
165 |
|
|
|
166 |
section(f, "Scalar sizes");
|
|
|
167 |
|
|
|
168 |
define_int(f, "ARCH_LOG2_SIZEOF_CHAR", ilog2(size_of(char)));
|
|
|
169 |
define_int(f, "ARCH_LOG2_SIZEOF_SHORT", ilog2(size_of(short)));
|
|
|
170 |
define_int(f, "ARCH_LOG2_SIZEOF_INT", ilog2(size_of(int)));
|
|
|
171 |
define_int(f, "ARCH_LOG2_SIZEOF_LONG", ilog2(size_of(long)));
|
|
|
172 |
#ifdef HAVE_LONG_LONG
|
|
|
173 |
define_int(f, "ARCH_LOG2_SIZEOF_LONG_LONG", ilog2(size_of(long long)));
|
|
|
174 |
#endif
|
|
|
175 |
define_int(f, "ARCH_SIZEOF_PTR", size_of(char *));
|
|
|
176 |
define_int(f, "ARCH_SIZEOF_FLOAT", size_of(float));
|
|
|
177 |
define_int(f, "ARCH_SIZEOF_DOUBLE", size_of(double));
|
|
|
178 |
if (floats_are_IEEE) {
|
|
|
179 |
define_int(f, "ARCH_FLOAT_MANTISSA_BITS", 24);
|
|
|
180 |
define_int(f, "ARCH_DOUBLE_MANTISSA_BITS", 53);
|
|
|
181 |
} else {
|
|
|
182 |
/*
|
|
|
183 |
* There isn't any general way to compute the number of mantissa
|
|
|
184 |
* bits accurately, especially if the machine uses hex rather
|
|
|
185 |
* than binary exponents. Use conservative values, assuming
|
|
|
186 |
* the exponent is stored in a 16-bit word of its own.
|
|
|
187 |
*/
|
|
|
188 |
define_int(f, "ARCH_FLOAT_MANTISSA_BITS", sizeof(float) * 8 - 17);
|
|
|
189 |
define_int(f, "ARCH_DOUBLE_MANTISSA_BITS", sizeof(double) * 8 - 17);
|
|
|
190 |
}
|
|
|
191 |
|
|
|
192 |
section(f, "Unsigned max values");
|
|
|
193 |
|
|
|
194 |
/*
|
|
|
195 |
* We can't use fprintf with a numeric value for PRINT_MAX, because
|
|
|
196 |
* too many compilers produce warnings or do the wrong thing for
|
|
|
197 |
* complementing or widening unsigned types.
|
|
|
198 |
*/
|
|
|
199 |
#define PRINT_MAX(str, typ, tstr, l)\
|
|
|
200 |
BEGIN\
|
|
|
201 |
define(f, str);\
|
|
|
202 |
fprintf(f, "((%s)0x", tstr);\
|
|
|
203 |
print_ffs(f, sizeof(typ));\
|
|
|
204 |
fprintf(f, "%s + (%s)0)\n", l, tstr);\
|
|
|
205 |
END
|
|
|
206 |
PRINT_MAX("ARCH_MAX_UCHAR", unsigned char, "unsigned char", "");
|
|
|
207 |
PRINT_MAX("ARCH_MAX_USHORT", unsigned short, "unsigned short", "");
|
|
|
208 |
/*
|
|
|
209 |
* For uint and ulong, a different approach is required to keep gcc
|
|
|
210 |
* with -Wtraditional from spewing out pointless warnings.
|
|
|
211 |
*/
|
|
|
212 |
define(f, "ARCH_MAX_UINT");
|
|
|
213 |
fprintf(f, "((unsigned int)~0 + (unsigned int)0)\n");
|
|
|
214 |
define(f, "ARCH_MAX_ULONG");
|
|
|
215 |
fprintf(f, "((unsigned long)~0L + (unsigned long)0)\n");
|
|
|
216 |
#undef PRINT_MAX
|
|
|
217 |
|
|
|
218 |
section(f, "Cache sizes");
|
|
|
219 |
|
|
|
220 |
/*
|
|
|
221 |
* Determine the primary and secondary cache sizes by looking for a
|
|
|
222 |
* non-linearity in the time required to fill blocks with memset.
|
|
|
223 |
*/
|
|
|
224 |
{
|
|
|
225 |
#define MAX_BLOCK (1 << 22) /* max 4M cache */
|
|
|
226 |
#define MAX_NREPS (1 << 10) /* limit the number of reps we try */
|
|
|
227 |
static char buf[MAX_BLOCK];
|
|
|
228 |
int bsize = 1 << 10;
|
|
|
229 |
int nreps = 1;
|
|
|
230 |
clock_t t = 0;
|
|
|
231 |
clock_t t_eps;
|
|
|
232 |
|
|
|
233 |
/*
|
|
|
234 |
* Increase the number of repetitions until the time is
|
|
|
235 |
* long enough to exceed the likely uncertainty.
|
|
|
236 |
*/
|
|
|
237 |
|
|
|
238 |
while (nreps < MAX_NREPS && (t = time_clear(buf, bsize, nreps)) == 0)
|
|
|
239 |
nreps <<= 1;
|
|
|
240 |
t_eps = t;
|
|
|
241 |
while (nreps < MAX_NREPS && (t = time_clear(buf, bsize, nreps)) < t_eps * 10)
|
|
|
242 |
nreps <<= 1;
|
|
|
243 |
|
|
|
244 |
/*
|
|
|
245 |
* Increase the block size until the time jumps non-linearly.
|
|
|
246 |
*/
|
|
|
247 |
for (; bsize <= MAX_BLOCK;) {
|
|
|
248 |
clock_t dt = time_clear(buf, bsize, nreps);
|
|
|
249 |
|
|
|
250 |
if (dt > t + (t >> 1)) {
|
|
|
251 |
t = dt;
|
|
|
252 |
break;
|
|
|
253 |
}
|
|
|
254 |
bsize <<= 1;
|
|
|
255 |
nreps >>= 1;
|
|
|
256 |
if (nreps == 0)
|
|
|
257 |
nreps = 1, t <<= 1;
|
|
|
258 |
}
|
|
|
259 |
define_int(f, "ARCH_CACHE1_SIZE", bsize >> 1);
|
|
|
260 |
/*
|
|
|
261 |
* Do the same thing a second time for the secondary cache.
|
|
|
262 |
*/
|
|
|
263 |
if (nreps > 1)
|
|
|
264 |
nreps >>= 1, t >>= 1;
|
|
|
265 |
for (; bsize <= MAX_BLOCK;) {
|
|
|
266 |
clock_t dt = time_clear(buf, bsize, nreps);
|
|
|
267 |
|
|
|
268 |
if (dt > t * 1.25) {
|
|
|
269 |
t = dt;
|
|
|
270 |
break;
|
|
|
271 |
}
|
|
|
272 |
bsize <<= 1;
|
|
|
273 |
nreps >>= 1;
|
|
|
274 |
if (nreps == 0)
|
|
|
275 |
nreps = 1, t <<= 1;
|
|
|
276 |
}
|
|
|
277 |
define_int(f, "ARCH_CACHE2_SIZE", bsize >> 1);
|
|
|
278 |
}
|
|
|
279 |
|
|
|
280 |
section(f, "Miscellaneous");
|
|
|
281 |
|
|
|
282 |
define_int(f, "ARCH_IS_BIG_ENDIAN", 1 - *(char *)&one);
|
|
|
283 |
pl0.l = 0;
|
|
|
284 |
pl1.l = -1;
|
|
|
285 |
define_int(f, "ARCH_PTRS_ARE_SIGNED", (pl1.p < pl0.p));
|
|
|
286 |
define_int(f, "ARCH_FLOATS_ARE_IEEE", (floats_are_IEEE ? 1 : 0));
|
|
|
287 |
|
|
|
288 |
/*
|
|
|
289 |
* There are three cases for arithmetic right shift:
|
|
|
290 |
* always correct, correct except for right-shifting a long by 1
|
|
|
291 |
* (a bug in some versions of the Turbo C compiler), and
|
|
|
292 |
* never correct.
|
|
|
293 |
*/
|
|
|
294 |
ars = (lr2 != -1 || ir1 != -1 || ir2 != -1 ? 0 :
|
|
|
295 |
lr1 != -1 ? 1 : /* Turbo C problem */
|
|
|
296 |
2);
|
|
|
297 |
define_int(f, "ARCH_ARITH_RSHIFT", ars);
|
|
|
298 |
/*
|
|
|
299 |
* Some machines can't handle a variable shift by
|
|
|
300 |
* the full width of a long.
|
|
|
301 |
*/
|
|
|
302 |
define_int(f, "ARCH_CAN_SHIFT_FULL_LONG", um1 >> lwidth == 0);
|
|
|
303 |
/*
|
|
|
304 |
* Determine whether dividing a negative integer by a positive one
|
|
|
305 |
* takes the floor or truncates toward zero.
|
|
|
306 |
*/
|
|
|
307 |
define_int(f, "ARCH_DIV_NEG_POS_TRUNCATES", im1 / 2 == 0);
|
|
|
308 |
|
|
|
309 |
/* ---------------- Done. ---------------- */
|
|
|
310 |
|
|
|
311 |
fclose(f);
|
|
|
312 |
return exit_OK;
|
|
|
313 |
}
|