ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libecb/ecb.pod
Revision: 1.43
Committed: Tue May 29 14:09:49 2012 UTC (12 years ago) by root
Branch: MAIN
Changes since 1.42: +49 -0 lines
Log Message:
*** empty log message ***

File Contents

# User Rev Content
1 root 1.14 =head1 LIBECB - e-C-Builtins
2 root 1.3
3 root 1.14 =head2 ABOUT LIBECB
4    
5     Libecb is currently a simple header file that doesn't require any
6     configuration to use or include in your project.
7    
8 sf-exg 1.16 It's part of the e-suite of libraries, other members of which include
9 root 1.14 libev and libeio.
10    
11     Its homepage can be found here:
12    
13     http://software.schmorp.de/pkg/libecb
14    
15     It mainly provides a number of wrappers around GCC built-ins, together
16     with replacement functions for other compilers. In addition to this,
17 sf-exg 1.16 it provides a number of other lowlevel C utilities, such as endianness
18 root 1.14 detection, byte swapping or bit rotations.
19    
20 root 1.24 Or in other words, things that should be built into any standard C system,
21     but aren't, implemented as efficient as possible with GCC, and still
22     correct with other compilers.
23 root 1.17
24 root 1.14 More might come.
25 root 1.3
26     =head2 ABOUT THE HEADER
27    
28 root 1.14 At the moment, all you have to do is copy F<ecb.h> somewhere where your
29     compiler can find it and include it:
30    
31     #include <ecb.h>
32    
33     The header should work fine for both C and C++ compilation, and gives you
34     all of F<inttypes.h> in addition to the ECB symbols.
35    
36 sf-exg 1.16 There are currently no object files to link to - future versions might
37 root 1.14 come with an (optional) object code library to link against, to reduce
38     code size or gain access to additional features.
39    
40     It also currently includes everything from F<inttypes.h>.
41    
42     =head2 ABOUT THIS MANUAL / CONVENTIONS
43    
44     This manual mainly describes each (public) function available after
45     including the F<ecb.h> header. The header might define other symbols than
46     these, but these are not part of the public API, and not supported in any
47     way.
48    
49     When the manual mentions a "function" then this could be defined either as
50     as inline function, a macro, or an external symbol.
51    
52     When functions use a concrete standard type, such as C<int> or
53     C<uint32_t>, then the corresponding function works only with that type. If
54     only a generic name is used (C<expr>, C<cond>, C<value> and so on), then
55     the corresponding function relies on C to implement the correct types, and
56     is usually implemented as a macro. Specifically, a "bool" in this manual
57     refers to any kind of boolean value, not a specific type.
58 root 1.1
59 root 1.40 =head2 TYPES / TYPE SUPPORT
60    
61     ecb.h makes sure that the following types are defined (in the expected way):
62    
63 root 1.42 int8_t uint8_t int16_t uint16_t
64     int32_t uint32_t int64_t uint64_t
65     intptr_t uintptr_t ptrdiff_t
66 root 1.40
67     The macro C<ECB_PTRSIZE> is defined to the size of a pointer on this
68     platform (currently C<4> or C<8>).
69    
70 root 1.43 =head2 LANGUAGE/COMPILER VERSIONS
71    
72     =over 4
73    
74     =item ECB_C99
75    
76     Expands to a true value (suitable for testing in by the preprocessor)
77     if the environment claims to be C99 compliant.
78    
79     =item ECB_C11
80    
81     Expands to a true value (suitable for testing in by the preprocessor)
82     if the environment claims to be C11 compliant.
83    
84     =item ECB_GCC_VERSION(major,minor)
85    
86     Expands to a true value (suitable for testing in by the preprocessor)
87     if the compiler used is GNU C and the version is the givne version, or
88     higher.
89    
90     This macro tries to return false on compilers that claim to be GCC
91     compatible but aren't.
92    
93     =back
94    
95 root 1.1 =head2 GCC ATTRIBUTES
96    
97 root 1.20 A major part of libecb deals with GCC attributes. These are additional
98 sf-exg 1.26 attributes that you can assign to functions, variables and sometimes even
99 root 1.20 types - much like C<const> or C<volatile> in C.
100    
101     While GCC allows declarations to show up in many surprising places,
102 sf-exg 1.26 but not in many expected places, the safest way is to put attribute
103 root 1.20 declarations before the whole declaration:
104    
105     ecb_const int mysqrt (int a);
106     ecb_unused int i;
107    
108     For variables, it is often nicer to put the attribute after the name, and
109     avoid multiple declarations using commas:
110    
111     int i ecb_unused;
112 root 1.3
113 root 1.1 =over 4
114    
115 root 1.2 =item ecb_attribute ((attrs...))
116 root 1.1
117 root 1.15 A simple wrapper that expands to C<__attribute__((attrs))> on GCC, and to
118     nothing on other compilers, so the effect is that only GCC sees these.
119    
120     Example: use the C<deprecated> attribute on a function.
121    
122     ecb_attribute((__deprecated__)) void
123     do_not_use_me_anymore (void);
124 root 1.2
125 root 1.3 =item ecb_unused
126    
127     Marks a function or a variable as "unused", which simply suppresses a
128     warning by GCC when it detects it as unused. This is useful when you e.g.
129     declare a variable but do not always use it:
130    
131 root 1.15 {
132     int var ecb_unused;
133 root 1.3
134 root 1.15 #ifdef SOMECONDITION
135     var = ...;
136     return var;
137     #else
138     return 0;
139     #endif
140     }
141 root 1.3
142 root 1.31 =item ecb_inline
143 root 1.29
144     This is not actually an attribute, but you use it like one. It expands
145     either to C<static inline> or to just C<static>, if inline isn't
146     supported. It should be used to declare functions that should be inlined,
147     for code size or speed reasons.
148    
149     Example: inline this function, it surely will reduce codesize.
150    
151 root 1.31 ecb_inline int
152 root 1.29 negmul (int a, int b)
153     {
154     return - (a * b);
155     }
156    
157 root 1.2 =item ecb_noinline
158    
159 root 1.9 Prevent a function from being inlined - it might be optimised away, but
160 root 1.3 not inlined into other functions. This is useful if you know your function
161     is rarely called and large enough for inlining not to be helpful.
162    
163 root 1.2 =item ecb_noreturn
164    
165 root 1.17 Marks a function as "not returning, ever". Some typical functions that
166     don't return are C<exit> or C<abort> (which really works hard to not
167     return), and now you can make your own:
168    
169     ecb_noreturn void
170     my_abort (const char *errline)
171     {
172     puts (errline);
173     abort ();
174     }
175    
176 sf-exg 1.19 In this case, the compiler would probably be smart enough to deduce it on
177     its own, so this is mainly useful for declarations.
178 root 1.17
179 root 1.2 =item ecb_const
180    
181 sf-exg 1.19 Declares that the function only depends on the values of its arguments,
182 root 1.17 much like a mathematical function. It specifically does not read or write
183     any memory any arguments might point to, global variables, or call any
184     non-const functions. It also must not have any side effects.
185    
186     Such a function can be optimised much more aggressively by the compiler -
187     for example, multiple calls with the same arguments can be optimised into
188     a single call, which wouldn't be possible if the compiler would have to
189     expect any side effects.
190    
191     It is best suited for functions in the sense of mathematical functions,
192 sf-exg 1.19 such as a function returning the square root of its input argument.
193 root 1.17
194     Not suited would be a function that calculates the hash of some memory
195     area you pass in, prints some messages or looks at a global variable to
196     decide on rounding.
197    
198     See C<ecb_pure> for a slightly less restrictive class of functions.
199    
200 root 1.2 =item ecb_pure
201    
202 root 1.17 Similar to C<ecb_const>, declares a function that has no side
203     effects. Unlike C<ecb_const>, the function is allowed to examine global
204     variables and any other memory areas (such as the ones passed to it via
205     pointers).
206    
207     While these functions cannot be optimised as aggressively as C<ecb_const>
208     functions, they can still be optimised away in many occasions, and the
209     compiler has more freedom in moving calls to them around.
210    
211     Typical examples for such functions would be C<strlen> or C<memcmp>. A
212     function that calculates the MD5 sum of some input and updates some MD5
213     state passed as argument would I<NOT> be pure, however, as it would modify
214     some memory area that is not the return value.
215    
216 root 1.2 =item ecb_hot
217    
218 root 1.17 This declares a function as "hot" with regards to the cache - the function
219     is used so often, that it is very beneficial to keep it in the cache if
220     possible.
221    
222     The compiler reacts by trying to place hot functions near to each other in
223     memory.
224    
225 sf-exg 1.19 Whether a function is hot or not often depends on the whole program,
226 root 1.17 and less on the function itself. C<ecb_cold> is likely more useful in
227     practise.
228    
229 root 1.2 =item ecb_cold
230    
231 root 1.17 The opposite of C<ecb_hot> - declares a function as "cold" with regards to
232     the cache, or in other words, this function is not called often, or not at
233     speed-critical times, and keeping it in the cache might be a waste of said
234     cache.
235    
236     In addition to placing cold functions together (or at least away from hot
237     functions), this knowledge can be used in other ways, for example, the
238     function will be optimised for size, as opposed to speed, and codepaths
239     leading to calls to those functions can automatically be marked as if
240 root 1.27 C<ecb_expect_false> had been used to reach them.
241 root 1.17
242     Good examples for such functions would be error reporting functions, or
243     functions only called in exceptional or rare cases.
244    
245 root 1.2 =item ecb_artificial
246    
247 root 1.17 Declares the function as "artificial", in this case meaning that this
248     function is not really mean to be a function, but more like an accessor
249     - many methods in C++ classes are mere accessor functions, and having a
250     crash reported in such a method, or single-stepping through them, is not
251     usually so helpful, especially when it's inlined to just a few instructions.
252    
253     Marking them as artificial will instruct the debugger about just this,
254     leading to happier debugging and thus happier lives.
255    
256     Example: in some kind of smart-pointer class, mark the pointer accessor as
257     artificial, so that the whole class acts more like a pointer and less like
258     some C++ abstraction monster.
259    
260     template<typename T>
261     struct my_smart_ptr
262     {
263     T *value;
264    
265     ecb_artificial
266     operator T *()
267     {
268     return value;
269     }
270     };
271    
272 root 1.2 =back
273 root 1.1
274     =head2 OPTIMISATION HINTS
275    
276     =over 4
277    
278 root 1.14 =item bool ecb_is_constant(expr)
279 root 1.1
280 root 1.3 Returns true iff the expression can be deduced to be a compile-time
281     constant, and false otherwise.
282    
283     For example, when you have a C<rndm16> function that returns a 16 bit
284     random number, and you have a function that maps this to a range from
285 root 1.5 0..n-1, then you could use this inline function in a header file:
286 root 1.3
287     ecb_inline uint32_t
288     rndm (uint32_t n)
289     {
290 root 1.6 return (n * (uint32_t)rndm16 ()) >> 16;
291 root 1.3 }
292    
293     However, for powers of two, you could use a normal mask, but that is only
294     worth it if, at compile time, you can detect this case. This is the case
295     when the passed number is a constant and also a power of two (C<n & (n -
296     1) == 0>):
297    
298     ecb_inline uint32_t
299     rndm (uint32_t n)
300     {
301     return is_constant (n) && !(n & (n - 1))
302     ? rndm16 () & (num - 1)
303 root 1.6 : (n * (uint32_t)rndm16 ()) >> 16;
304 root 1.3 }
305    
306 root 1.14 =item bool ecb_expect (expr, value)
307 root 1.1
308 root 1.7 Evaluates C<expr> and returns it. In addition, it tells the compiler that
309     the C<expr> evaluates to C<value> a lot, which can be used for static
310     branch optimisations.
311 root 1.1
312 root 1.27 Usually, you want to use the more intuitive C<ecb_expect_true> and
313     C<ecb_expect_false> functions instead.
314 root 1.1
315 root 1.27 =item bool ecb_expect_true (cond)
316 root 1.1
317 root 1.27 =item bool ecb_expect_false (cond)
318 root 1.1
319 root 1.7 These two functions expect a expression that is true or false and return
320     C<1> or C<0>, respectively, so when used in the condition of an C<if> or
321     other conditional statement, it will not change the program:
322    
323     /* these two do the same thing */
324     if (some_condition) ...;
325 root 1.27 if (ecb_expect_true (some_condition)) ...;
326 root 1.7
327 root 1.27 However, by using C<ecb_expect_true>, you tell the compiler that the
328     condition is likely to be true (and for C<ecb_expect_false>, that it is
329     unlikely to be true).
330 root 1.7
331 root 1.9 For example, when you check for a null pointer and expect this to be a
332 root 1.27 rare, exceptional, case, then use C<ecb_expect_false>:
333 root 1.7
334     void my_free (void *ptr)
335     {
336 root 1.27 if (ecb_expect_false (ptr == 0))
337 root 1.7 return;
338     }
339    
340     Consequent use of these functions to mark away exceptional cases or to
341     tell the compiler what the hot path through a function is can increase
342     performance considerably.
343    
344 root 1.27 You might know these functions under the name C<likely> and C<unlikely>
345     - while these are common aliases, we find that the expect name is easier
346     to understand when quickly skimming code. If you wish, you can use
347     C<ecb_likely> instead of C<ecb_expect_true> and C<ecb_unlikely> instead of
348     C<ecb_expect_false> - these are simply aliases.
349    
350 root 1.7 A very good example is in a function that reserves more space for some
351     memory block (for example, inside an implementation of a string stream) -
352 root 1.9 each time something is added, you have to check for a buffer overrun, but
353 root 1.7 you expect that most checks will turn out to be false:
354    
355     /* make sure we have "size" extra room in our buffer */
356     ecb_inline void
357     reserve (int size)
358     {
359 root 1.27 if (ecb_expect_false (current + size > end))
360 root 1.7 real_reserve_method (size); /* presumably noinline */
361     }
362    
363 root 1.14 =item bool ecb_assume (cond)
364 root 1.7
365     Try to tell the compiler that some condition is true, even if it's not
366     obvious.
367    
368     This can be used to teach the compiler about invariants or other
369     conditions that might improve code generation, but which are impossible to
370     deduce form the code itself.
371    
372 root 1.27 For example, the example reservation function from the C<ecb_expect_false>
373 root 1.7 description could be written thus (only C<ecb_assume> was added):
374    
375     ecb_inline void
376     reserve (int size)
377     {
378 root 1.27 if (ecb_expect_false (current + size > end))
379 root 1.7 real_reserve_method (size); /* presumably noinline */
380    
381     ecb_assume (current + size <= end);
382     }
383    
384     If you then call this function twice, like this:
385    
386     reserve (10);
387     reserve (1);
388    
389     Then the compiler I<might> be able to optimise out the second call
390     completely, as it knows that C<< current + 1 > end >> is false and the
391     call will never be executed.
392    
393     =item bool ecb_unreachable ()
394    
395     This function does nothing itself, except tell the compiler that it will
396 root 1.9 never be executed. Apart from suppressing a warning in some cases, this
397 root 1.7 function can be used to implement C<ecb_assume> or similar functions.
398    
399 root 1.14 =item bool ecb_prefetch (addr, rw, locality)
400 root 1.7
401     Tells the compiler to try to prefetch memory at the given C<addr>ess
402 root 1.10 for either reading (C<rw> = 0) or writing (C<rw> = 1). A C<locality> of
403 root 1.7 C<0> means that there will only be one access later, C<3> means that
404     the data will likely be accessed very often, and values in between mean
405     something... in between. The memory pointed to by the address does not
406     need to be accessible (it could be a null pointer for example), but C<rw>
407     and C<locality> must be compile-time constants.
408    
409     An obvious way to use this is to prefetch some data far away, in a big
410 root 1.9 array you loop over. This prefetches memory some 128 array elements later,
411 root 1.7 in the hope that it will be ready when the CPU arrives at that location.
412    
413     int sum = 0;
414    
415     for (i = 0; i < N; ++i)
416     {
417     sum += arr [i]
418     ecb_prefetch (arr + i + 128, 0, 0);
419     }
420    
421     It's hard to predict how far to prefetch, and most CPUs that can prefetch
422     are often good enough to predict this kind of behaviour themselves. It
423     gets more interesting with linked lists, especially when you do some fair
424     processing on each list element:
425    
426     for (node *n = start; n; n = n->next)
427     {
428     ecb_prefetch (n->next, 0, 0);
429     ... do medium amount of work with *n
430     }
431    
432     After processing the node, (part of) the next node might already be in
433     cache.
434 root 1.1
435 root 1.2 =back
436 root 1.1
437 root 1.36 =head2 BIT FIDDLING / BIT WIZARDRY
438 root 1.1
439 root 1.4 =over 4
440    
441 root 1.3 =item bool ecb_big_endian ()
442    
443     =item bool ecb_little_endian ()
444    
445 sf-exg 1.11 These two functions return true if the byte order is big endian
446     (most-significant byte first) or little endian (least-significant byte
447     first) respectively.
448    
449 root 1.24 On systems that are neither, their return values are unspecified.
450    
451 root 1.3 =item int ecb_ctz32 (uint32_t x)
452    
453 root 1.35 =item int ecb_ctz64 (uint64_t x)
454    
455 sf-exg 1.11 Returns the index of the least significant bit set in C<x> (or
456 root 1.24 equivalently the number of bits set to 0 before the least significant bit
457 root 1.35 set), starting from 0. If C<x> is 0 the result is undefined.
458    
459 root 1.36 For smaller types than C<uint32_t> you can safely use C<ecb_ctz32>.
460    
461 root 1.35 For example:
462 sf-exg 1.11
463 root 1.15 ecb_ctz32 (3) = 0
464     ecb_ctz32 (6) = 1
465 sf-exg 1.11
466 root 1.41 =item bool ecb_is_pot32 (uint32_t x)
467    
468     =item bool ecb_is_pot64 (uint32_t x)
469    
470     Return true iff C<x> is a power of two or C<x == 0>.
471    
472     For smaller types then C<uint32_t> you can safely use C<ecb_is_pot32>.
473    
474 root 1.35 =item int ecb_ld32 (uint32_t x)
475    
476     =item int ecb_ld64 (uint64_t x)
477    
478     Returns the index of the most significant bit set in C<x>, or the number
479     of digits the number requires in binary (so that C<< 2**ld <= x <
480     2**(ld+1) >>). If C<x> is 0 the result is undefined. A common use case is
481     to compute the integer binary logarithm, i.e. C<floor (log2 (n))>, for
482     example to see how many bits a certain number requires to be encoded.
483    
484     This function is similar to the "count leading zero bits" function, except
485     that that one returns how many zero bits are "in front" of the number (in
486     the given data type), while C<ecb_ld> returns how many bits the number
487     itself requires.
488    
489 root 1.36 For smaller types than C<uint32_t> you can safely use C<ecb_ld32>.
490    
491 root 1.3 =item int ecb_popcount32 (uint32_t x)
492    
493 root 1.35 =item int ecb_popcount64 (uint64_t x)
494    
495 root 1.36 Returns the number of bits set to 1 in C<x>.
496    
497     For smaller types than C<uint32_t> you can safely use C<ecb_popcount32>.
498    
499     For example:
500 sf-exg 1.11
501 root 1.15 ecb_popcount32 (7) = 3
502     ecb_popcount32 (255) = 8
503 sf-exg 1.11
504 root 1.39 =item uint8_t ecb_bitrev8 (uint8_t x)
505    
506     =item uint16_t ecb_bitrev16 (uint16_t x)
507    
508     =item uint32_t ecb_bitrev32 (uint32_t x)
509    
510     Reverses the bits in x, i.e. the MSB becomes the LSB, MSB-1 becomes LSB+1
511     and so on.
512    
513     Example:
514    
515     ecb_bitrev8 (0xa7) = 0xea
516     ecb_bitrev32 (0xffcc4411) = 0x882233ff
517    
518 root 1.8 =item uint32_t ecb_bswap16 (uint32_t x)
519    
520 root 1.3 =item uint32_t ecb_bswap32 (uint32_t x)
521    
522 root 1.34 =item uint64_t ecb_bswap64 (uint64_t x)
523 sf-exg 1.13
524 root 1.34 These functions return the value of the 16-bit (32-bit, 64-bit) value
525     C<x> after reversing the order of bytes (0x11223344 becomes 0x44332211 in
526     C<ecb_bswap32>).
527    
528     =item uint8_t ecb_rotl8 (uint8_t x, unsigned int count)
529    
530     =item uint16_t ecb_rotl16 (uint16_t x, unsigned int count)
531 root 1.3
532     =item uint32_t ecb_rotl32 (uint32_t x, unsigned int count)
533    
534 root 1.34 =item uint64_t ecb_rotl64 (uint64_t x, unsigned int count)
535    
536     =item uint8_t ecb_rotr8 (uint8_t x, unsigned int count)
537    
538     =item uint16_t ecb_rotr16 (uint16_t x, unsigned int count)
539    
540     =item uint32_t ecb_rotr32 (uint32_t x, unsigned int count)
541    
542 root 1.33 =item uint64_t ecb_rotr64 (uint64_t x, unsigned int count)
543    
544 root 1.34 These two families of functions return the value of C<x> after rotating
545     all the bits by C<count> positions to the right (C<ecb_rotr>) or left
546     (C<ecb_rotl>).
547 sf-exg 1.11
548 root 1.20 Current GCC versions understand these functions and usually compile them
549 root 1.34 to "optimal" code (e.g. a single C<rol> or a combination of C<shld> on
550     x86).
551 root 1.20
552 root 1.3 =back
553 root 1.1
554     =head2 ARITHMETIC
555    
556 root 1.3 =over 4
557    
558 root 1.14 =item x = ecb_mod (m, n)
559 root 1.3
560 root 1.25 Returns C<m> modulo C<n>, which is the same as the positive remainder
561     of the division operation between C<m> and C<n>, using floored
562     division. Unlike the C remainder operator C<%>, this function ensures that
563     the return value is always positive and that the two numbers I<m> and
564     I<m' = m + i * n> result in the same value modulo I<n> - in other words,
565     C<ecb_mod> implements the mathematical modulo operation, which is missing
566     in the language.
567 root 1.14
568 sf-exg 1.23 C<n> must be strictly positive (i.e. C<< >= 1 >>), while C<m> must be
569 root 1.14 negatable, that is, both C<m> and C<-m> must be representable in its
570 root 1.30 type (this typically excludes the minimum signed integer value, the same
571 root 1.25 limitation as for C</> and C<%> in C).
572 sf-exg 1.11
573 root 1.24 Current GCC versions compile this into an efficient branchless sequence on
574 root 1.28 almost all CPUs.
575 root 1.24
576     For example, when you want to rotate forward through the members of an
577     array for increasing C<m> (which might be negative), then you should use
578     C<ecb_mod>, as the C<%> operator might give either negative results, or
579     change direction for negative values:
580    
581     for (m = -100; m <= 100; ++m)
582     int elem = myarray [ecb_mod (m, ecb_array_length (myarray))];
583    
584 sf-exg 1.37 =item x = ecb_div_rd (val, div)
585    
586     =item x = ecb_div_ru (val, div)
587    
588     Returns C<val> divided by C<div> rounded down or up, respectively.
589     C<val> and C<div> must have integer types and C<div> must be strictly
590 sf-exg 1.38 positive. Note that these functions are implemented with macros in C
591     and with function templates in C++.
592 sf-exg 1.37
593 root 1.3 =back
594 root 1.1
595     =head2 UTILITY
596    
597 root 1.3 =over 4
598    
599 sf-exg 1.23 =item element_count = ecb_array_length (name)
600 root 1.3
601 sf-exg 1.13 Returns the number of elements in the array C<name>. For example:
602    
603     int primes[] = { 2, 3, 5, 7, 11 };
604     int sum = 0;
605    
606     for (i = 0; i < ecb_array_length (primes); i++)
607     sum += primes [i];
608    
609 root 1.3 =back
610 root 1.1
611 root 1.43 =head2 SYMBOLS GOVERNING COMPILATION OF ECB.H ITSELF
612    
613     These symbols need to be defined before including F<ecb.h> the first time.
614    
615     =over 4
616    
617     =item ECB_NO_THRADS
618    
619     If F<ecb.h> is never used from multiple threads, then this symbol can
620     be defined, in which case memory fences (and similar constructs) are
621     completely removed, leading to more efficient code and fewer dependencies.
622    
623     Setting this symbol to a true value implies C<ECB_NO_SMP>.
624    
625     =item ECB_NO_SMP
626    
627     The weaker version of C<ECB_NO_THREADS> - if F<ecb.h> is used from
628     multiple threads, but never concurrently (e.g. if the system the program
629     runs on has only a single CPU with a single core, no hyperthreading and so
630     on), then this symbol can be defined, leading to more efficient code and
631     fewer dependencies.
632    
633     =back
634    
635 root 1.1