ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libecb/ecb.pod
Revision: 1.106
Committed: Fri Mar 25 15:23:14 2022 UTC (2 years, 3 months ago) by root
Branch: MAIN
Changes since 1.105: +3 -7 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 root 1.85 It mainly provides a number of wrappers around many compiler built-ins,
16     together with replacement functions for other compilers. In addition
17 root 1.104 to this, it provides a number of other low-level C utilities, such as
18 root 1.85 endianness detection, byte swapping or bit rotations.
19    
20     Or in other words, things that should be built into any standard C
21     system, but aren't, implemented as efficient as possible with GCC (clang,
22 root 1.104 MSVC...), and still 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 root 1.104 F<ecb.h> makes sure that the following types are defined (in the expected way):
62 root 1.40
63 root 1.76 int8_t uint8_
64     int16_t uint16_t
65     int32_t uint32_
66     int64_t uint64_t
67     int_fast8_t uint_fast8_t
68     int_fast16_t uint_fast16_t
69     int_fast32_t uint_fast32_t
70     int_fast64_t uint_fast64_t
71     intptr_t uintptr_t
72 root 1.40
73     The macro C<ECB_PTRSIZE> is defined to the size of a pointer on this
74 root 1.45 platform (currently C<4> or C<8>) and can be used in preprocessor
75     expressions.
76 root 1.40
77 root 1.74 For C<ptrdiff_t> and C<size_t> use C<stddef.h>/C<cstddef>.
78 root 1.49
79 root 1.62 =head2 LANGUAGE/ENVIRONMENT/COMPILER VERSIONS
80 root 1.43
81 sf-exg 1.46 All the following symbols expand to an expression that can be tested in
82 root 1.44 preprocessor instructions as well as treated as a boolean (use C<!!> to
83     ensure it's either C<0> or C<1> if you need that).
84    
85 root 1.88 =over
86 root 1.43
87 root 1.44 =item ECB_C
88    
89     True if the implementation defines the C<__STDC__> macro to a true value,
90 root 1.82 while not claiming to be C++, i..e C, but not C++.
91 root 1.44
92 root 1.43 =item ECB_C99
93    
94 root 1.47 True if the implementation claims to be compliant to C99 (ISO/IEC
95 root 1.55 9899:1999) or any later version, while not claiming to be C++.
96 root 1.47
97     Note that later versions (ECB_C11) remove core features again (for
98     example, variable length arrays).
99 root 1.43
100 root 1.74 =item ECB_C11, ECB_C17
101 root 1.43
102 root 1.74 True if the implementation claims to be compliant to C11/C17 (ISO/IEC
103     9899:2011, :20187) or any later version, while not claiming to be C++.
104 root 1.44
105     =item ECB_CPP
106    
107     True if the implementation defines the C<__cplusplus__> macro to a true
108     value, which is typically true for C++ compilers.
109    
110 root 1.74 =item ECB_CPP11, ECB_CPP14, ECB_CPP17
111 root 1.44
112 root 1.74 True if the implementation claims to be compliant to C++11/C++14/C++17
113     (ISO/IEC 14882:2011, :2014, :2017) or any later version.
114 root 1.43
115 root 1.83 Note that many C++20 features will likely have their own feature test
116     macros (see e.g. L<http://eel.is/c++draft/cpp.predefined#1.8>).
117    
118 root 1.81 =item ECB_OPTIMIZE_SIZE
119    
120     Is C<1> when the compiler optimizes for size, C<0> otherwise. This symbol
121     can also be defined before including F<ecb.h>, in which case it will be
122     unchanged.
123    
124 root 1.57 =item ECB_GCC_VERSION (major, minor)
125 root 1.43
126 root 1.84 Expands to a true value (suitable for testing by the preprocessor) if the
127     compiler used is GNU C and the version is the given version, or higher.
128 root 1.43
129     This macro tries to return false on compilers that claim to be GCC
130     compatible but aren't.
131    
132 root 1.50 =item ECB_EXTERN_C
133    
134     Expands to C<extern "C"> in C++, and a simple C<extern> in C.
135    
136     This can be used to declare a single external C function:
137    
138     ECB_EXTERN_C int printf (const char *format, ...);
139    
140     =item ECB_EXTERN_C_BEG / ECB_EXTERN_C_END
141    
142     These two macros can be used to wrap multiple C<extern "C"> definitions -
143     they expand to nothing in C.
144    
145     They are most useful in header files:
146    
147     ECB_EXTERN_C_BEG
148    
149     int mycfun1 (int x);
150     int mycfun2 (int x);
151    
152     ECB_EXTERN_C_END
153    
154     =item ECB_STDFP
155    
156 root 1.84 If this evaluates to a true value (suitable for testing by the
157 root 1.50 preprocessor), then C<float> and C<double> use IEEE 754 single/binary32
158     and double/binary64 representations internally I<and> the endianness of
159     both types match the endianness of C<uint32_t> and C<uint64_t>.
160    
161     This means you can just copy the bits of a C<float> (or C<double>) to an
162     C<uint32_t> (or C<uint64_t>) and get the raw IEEE 754 bit representation
163     without having to think about format or endianness.
164    
165     This is true for basically all modern platforms, although F<ecb.h> might
166     not be able to deduce this correctly everywhere and might err on the safe
167     side.
168    
169 root 1.87 =item ECB_64BIT_NATIVE
170    
171     Evaluates to a true value (suitable for both preprocessor and C code
172     testing) if 64 bit integer types on this architecture are evaluated
173 sf-exg 1.92 "natively", that is, with similar speeds as 32 bit integers. While 64 bit
174     integer support is very common (and in fact required by libecb), 32 bit
175 root 1.104 CPUs have to emulate operations on them, so you might want to avoid them.
176 root 1.87
177 root 1.54 =item ECB_AMD64, ECB_AMD64_X32
178    
179     These two macros are defined to C<1> on the x86_64/amd64 ABI and the X32
180     ABI, respectively, and undefined elsewhere.
181    
182     The designers of the new X32 ABI for some inexplicable reason decided to
183     make it look exactly like amd64, even though it's completely incompatible
184     to that ABI, breaking about every piece of software that assumed that
185     C<__x86_64> stands for, well, the x86-64 ABI, making these macros
186     necessary.
187    
188 root 1.43 =back
189    
190 root 1.62 =head2 MACRO TRICKERY
191    
192 root 1.88 =over
193 root 1.62
194     =item ECB_CONCAT (a, b)
195    
196     Expands any macros in C<a> and C<b>, then concatenates the result to form
197     a single token. This is mainly useful to form identifiers from components,
198     e.g.:
199    
200     #define S1 str
201     #define S2 cpy
202    
203     ECB_CONCAT (S1, S2)(dst, src); // == strcpy (dst, src);
204    
205     =item ECB_STRINGIFY (arg)
206    
207     Expands any macros in C<arg> and returns the stringified version of
208     it. This is mainly useful to get the contents of a macro in string form,
209     e.g.:
210    
211     #define SQL_LIMIT 100
212     sql_exec ("select * from table limit " ECB_STRINGIFY (SQL_LIMIT));
213    
214 root 1.64 =item ECB_STRINGIFY_EXPR (expr)
215    
216     Like C<ECB_STRINGIFY>, but additionally evaluates C<expr> to make sure it
217     is a valid expression. This is useful to catch typos or cases where the
218     macro isn't available:
219    
220     #include <errno.h>
221    
222     ECB_STRINGIFY (EDOM); // "33" (on my system at least)
223     ECB_STRINGIFY_EXPR (EDOM); // "33"
224    
225     // now imagine we had a typo:
226    
227     ECB_STRINGIFY (EDAM); // "EDAM"
228     ECB_STRINGIFY_EXPR (EDAM); // error: EDAM undefined
229    
230 root 1.62 =back
231    
232 sf-exg 1.60 =head2 ATTRIBUTES
233 root 1.1
234 sf-exg 1.60 A major part of libecb deals with additional attributes that can be
235     assigned to functions, variables and sometimes even types - much like
236     C<const> or C<volatile> in C. They are implemented using either GCC
237     attributes or other compiler/language specific features. Attributes
238     declarations must be put before the whole declaration:
239 root 1.20
240     ecb_const int mysqrt (int a);
241     ecb_unused int i;
242    
243 root 1.88 =over
244 root 1.1
245 root 1.3 =item ecb_unused
246    
247     Marks a function or a variable as "unused", which simply suppresses a
248 root 1.85 warning by the compiler when it detects it as unused. This is useful when
249     you e.g. declare a variable but do not always use it:
250 root 1.3
251 root 1.15 {
252 sf-exg 1.61 ecb_unused int var;
253 root 1.3
254 root 1.15 #ifdef SOMECONDITION
255     var = ...;
256     return var;
257     #else
258     return 0;
259     #endif
260     }
261 root 1.3
262 root 1.56 =item ecb_deprecated
263    
264     Similar to C<ecb_unused>, but marks a function, variable or type as
265     deprecated. This makes some compilers warn when the type is used.
266    
267 root 1.62 =item ecb_deprecated_message (message)
268    
269 root 1.67 Same as C<ecb_deprecated>, but if possible, the specified diagnostic is
270 root 1.62 used instead of a generic depreciation message when the object is being
271     used.
272    
273 root 1.31 =item ecb_inline
274 root 1.29
275 root 1.73 Expands either to (a compiler-specific equivalent of) C<static inline> or
276     to just C<static>, if inline isn't supported. It should be used to declare
277     functions that should be inlined, for code size or speed reasons.
278 root 1.29
279 root 1.104 Example: inline this function, it surely will reduce code size.
280 root 1.29
281 root 1.31 ecb_inline int
282 root 1.29 negmul (int a, int b)
283     {
284     return - (a * b);
285     }
286    
287 root 1.2 =item ecb_noinline
288    
289 sf-exg 1.66 Prevents a function from being inlined - it might be optimised away, but
290 root 1.3 not inlined into other functions. This is useful if you know your function
291     is rarely called and large enough for inlining not to be helpful.
292    
293 root 1.2 =item ecb_noreturn
294    
295 root 1.17 Marks a function as "not returning, ever". Some typical functions that
296     don't return are C<exit> or C<abort> (which really works hard to not
297     return), and now you can make your own:
298    
299     ecb_noreturn void
300     my_abort (const char *errline)
301     {
302     puts (errline);
303     abort ();
304     }
305    
306 sf-exg 1.19 In this case, the compiler would probably be smart enough to deduce it on
307     its own, so this is mainly useful for declarations.
308 root 1.17
309 root 1.53 =item ecb_restrict
310    
311     Expands to the C<restrict> keyword or equivalent on compilers that support
312     them, and to nothing on others. Must be specified on a pointer type or
313     an array index to indicate that the memory doesn't alias with any other
314     restricted pointer in the same scope.
315    
316     Example: multiply a vector, and allow the compiler to parallelise the
317     loop, because it knows it doesn't overwrite input values.
318    
319     void
320 sf-exg 1.61 multiply (ecb_restrict float *src,
321     ecb_restrict float *dst,
322 root 1.53 int len, float factor)
323     {
324     int i;
325    
326     for (i = 0; i < len; ++i)
327     dst [i] = src [i] * factor;
328     }
329    
330 root 1.2 =item ecb_const
331    
332 sf-exg 1.19 Declares that the function only depends on the values of its arguments,
333 root 1.17 much like a mathematical function. It specifically does not read or write
334     any memory any arguments might point to, global variables, or call any
335     non-const functions. It also must not have any side effects.
336    
337     Such a function can be optimised much more aggressively by the compiler -
338     for example, multiple calls with the same arguments can be optimised into
339     a single call, which wouldn't be possible if the compiler would have to
340     expect any side effects.
341    
342     It is best suited for functions in the sense of mathematical functions,
343 sf-exg 1.19 such as a function returning the square root of its input argument.
344 root 1.17
345     Not suited would be a function that calculates the hash of some memory
346     area you pass in, prints some messages or looks at a global variable to
347     decide on rounding.
348    
349     See C<ecb_pure> for a slightly less restrictive class of functions.
350    
351 root 1.2 =item ecb_pure
352    
353 root 1.17 Similar to C<ecb_const>, declares a function that has no side
354     effects. Unlike C<ecb_const>, the function is allowed to examine global
355     variables and any other memory areas (such as the ones passed to it via
356     pointers).
357    
358     While these functions cannot be optimised as aggressively as C<ecb_const>
359     functions, they can still be optimised away in many occasions, and the
360     compiler has more freedom in moving calls to them around.
361    
362     Typical examples for such functions would be C<strlen> or C<memcmp>. A
363     function that calculates the MD5 sum of some input and updates some MD5
364     state passed as argument would I<NOT> be pure, however, as it would modify
365     some memory area that is not the return value.
366    
367 root 1.2 =item ecb_hot
368    
369 root 1.17 This declares a function as "hot" with regards to the cache - the function
370     is used so often, that it is very beneficial to keep it in the cache if
371     possible.
372    
373     The compiler reacts by trying to place hot functions near to each other in
374     memory.
375    
376 sf-exg 1.19 Whether a function is hot or not often depends on the whole program,
377 root 1.17 and less on the function itself. C<ecb_cold> is likely more useful in
378     practise.
379    
380 root 1.2 =item ecb_cold
381    
382 root 1.17 The opposite of C<ecb_hot> - declares a function as "cold" with regards to
383     the cache, or in other words, this function is not called often, or not at
384     speed-critical times, and keeping it in the cache might be a waste of said
385     cache.
386    
387     In addition to placing cold functions together (or at least away from hot
388     functions), this knowledge can be used in other ways, for example, the
389 root 1.104 function will be optimised for size, as opposed to speed, and code paths
390 root 1.17 leading to calls to those functions can automatically be marked as if
391 root 1.27 C<ecb_expect_false> had been used to reach them.
392 root 1.17
393     Good examples for such functions would be error reporting functions, or
394     functions only called in exceptional or rare cases.
395    
396 root 1.2 =item ecb_artificial
397    
398 root 1.17 Declares the function as "artificial", in this case meaning that this
399 root 1.52 function is not really meant to be a function, but more like an accessor
400 root 1.17 - many methods in C++ classes are mere accessor functions, and having a
401     crash reported in such a method, or single-stepping through them, is not
402     usually so helpful, especially when it's inlined to just a few instructions.
403    
404     Marking them as artificial will instruct the debugger about just this,
405     leading to happier debugging and thus happier lives.
406    
407     Example: in some kind of smart-pointer class, mark the pointer accessor as
408     artificial, so that the whole class acts more like a pointer and less like
409     some C++ abstraction monster.
410    
411     template<typename T>
412     struct my_smart_ptr
413     {
414     T *value;
415    
416     ecb_artificial
417     operator T *()
418     {
419     return value;
420     }
421     };
422    
423 root 1.2 =back
424 root 1.1
425     =head2 OPTIMISATION HINTS
426    
427 root 1.88 =over
428 root 1.1
429 root 1.58 =item bool ecb_is_constant (expr)
430 root 1.1
431 root 1.3 Returns true iff the expression can be deduced to be a compile-time
432     constant, and false otherwise.
433    
434     For example, when you have a C<rndm16> function that returns a 16 bit
435     random number, and you have a function that maps this to a range from
436 root 1.5 0..n-1, then you could use this inline function in a header file:
437 root 1.3
438     ecb_inline uint32_t
439     rndm (uint32_t n)
440     {
441 root 1.6 return (n * (uint32_t)rndm16 ()) >> 16;
442 root 1.3 }
443    
444     However, for powers of two, you could use a normal mask, but that is only
445     worth it if, at compile time, you can detect this case. This is the case
446     when the passed number is a constant and also a power of two (C<n & (n -
447     1) == 0>):
448    
449     ecb_inline uint32_t
450     rndm (uint32_t n)
451     {
452     return is_constant (n) && !(n & (n - 1))
453     ? rndm16 () & (num - 1)
454 root 1.6 : (n * (uint32_t)rndm16 ()) >> 16;
455 root 1.3 }
456    
457 root 1.62 =item ecb_expect (expr, value)
458 root 1.1
459 root 1.7 Evaluates C<expr> and returns it. In addition, it tells the compiler that
460     the C<expr> evaluates to C<value> a lot, which can be used for static
461     branch optimisations.
462 root 1.1
463 root 1.27 Usually, you want to use the more intuitive C<ecb_expect_true> and
464     C<ecb_expect_false> functions instead.
465 root 1.1
466 root 1.27 =item bool ecb_expect_true (cond)
467 root 1.1
468 root 1.27 =item bool ecb_expect_false (cond)
469 root 1.1
470 root 1.7 These two functions expect a expression that is true or false and return
471     C<1> or C<0>, respectively, so when used in the condition of an C<if> or
472     other conditional statement, it will not change the program:
473    
474     /* these two do the same thing */
475     if (some_condition) ...;
476 root 1.27 if (ecb_expect_true (some_condition)) ...;
477 root 1.7
478 root 1.27 However, by using C<ecb_expect_true>, you tell the compiler that the
479     condition is likely to be true (and for C<ecb_expect_false>, that it is
480     unlikely to be true).
481 root 1.7
482 root 1.9 For example, when you check for a null pointer and expect this to be a
483 root 1.27 rare, exceptional, case, then use C<ecb_expect_false>:
484 root 1.7
485     void my_free (void *ptr)
486     {
487 root 1.27 if (ecb_expect_false (ptr == 0))
488 root 1.7 return;
489     }
490    
491     Consequent use of these functions to mark away exceptional cases or to
492     tell the compiler what the hot path through a function is can increase
493     performance considerably.
494    
495 root 1.27 You might know these functions under the name C<likely> and C<unlikely>
496     - while these are common aliases, we find that the expect name is easier
497     to understand when quickly skimming code. If you wish, you can use
498     C<ecb_likely> instead of C<ecb_expect_true> and C<ecb_unlikely> instead of
499     C<ecb_expect_false> - these are simply aliases.
500    
501 root 1.7 A very good example is in a function that reserves more space for some
502     memory block (for example, inside an implementation of a string stream) -
503 root 1.9 each time something is added, you have to check for a buffer overrun, but
504 root 1.7 you expect that most checks will turn out to be false:
505    
506     /* make sure we have "size" extra room in our buffer */
507     ecb_inline void
508     reserve (int size)
509     {
510 root 1.27 if (ecb_expect_false (current + size > end))
511 root 1.7 real_reserve_method (size); /* presumably noinline */
512     }
513    
514 root 1.62 =item ecb_assume (cond)
515 root 1.7
516 sf-exg 1.66 Tries to tell the compiler that some condition is true, even if it's not
517 root 1.65 obvious. This is not a function, but a statement: it cannot be used in
518     another expression.
519 root 1.7
520     This can be used to teach the compiler about invariants or other
521     conditions that might improve code generation, but which are impossible to
522     deduce form the code itself.
523    
524 root 1.27 For example, the example reservation function from the C<ecb_expect_false>
525 root 1.7 description could be written thus (only C<ecb_assume> was added):
526    
527     ecb_inline void
528     reserve (int size)
529     {
530 root 1.27 if (ecb_expect_false (current + size > end))
531 root 1.7 real_reserve_method (size); /* presumably noinline */
532    
533     ecb_assume (current + size <= end);
534     }
535    
536     If you then call this function twice, like this:
537    
538     reserve (10);
539     reserve (1);
540    
541     Then the compiler I<might> be able to optimise out the second call
542     completely, as it knows that C<< current + 1 > end >> is false and the
543     call will never be executed.
544    
545 root 1.62 =item ecb_unreachable ()
546 root 1.7
547     This function does nothing itself, except tell the compiler that it will
548 root 1.9 never be executed. Apart from suppressing a warning in some cases, this
549 root 1.65 function can be used to implement C<ecb_assume> or similar functionality.
550 root 1.7
551 root 1.62 =item ecb_prefetch (addr, rw, locality)
552 root 1.7
553 root 1.104 Tells the compiler to try to prefetch memory at the given I<addr>ess
554     for either reading (I<rw> = 0) or writing (I<rw> = 1). A I<locality> of
555 root 1.7 C<0> means that there will only be one access later, C<3> means that
556     the data will likely be accessed very often, and values in between mean
557     something... in between. The memory pointed to by the address does not
558     need to be accessible (it could be a null pointer for example), but C<rw>
559     and C<locality> must be compile-time constants.
560    
561 root 1.65 This is a statement, not a function: you cannot use it as part of an
562     expression.
563    
564 root 1.7 An obvious way to use this is to prefetch some data far away, in a big
565 root 1.9 array you loop over. This prefetches memory some 128 array elements later,
566 root 1.7 in the hope that it will be ready when the CPU arrives at that location.
567    
568     int sum = 0;
569    
570     for (i = 0; i < N; ++i)
571     {
572     sum += arr [i]
573     ecb_prefetch (arr + i + 128, 0, 0);
574     }
575    
576     It's hard to predict how far to prefetch, and most CPUs that can prefetch
577     are often good enough to predict this kind of behaviour themselves. It
578     gets more interesting with linked lists, especially when you do some fair
579     processing on each list element:
580    
581     for (node *n = start; n; n = n->next)
582     {
583     ecb_prefetch (n->next, 0, 0);
584     ... do medium amount of work with *n
585     }
586    
587     After processing the node, (part of) the next node might already be in
588     cache.
589 root 1.1
590 root 1.2 =back
591 root 1.1
592 root 1.36 =head2 BIT FIDDLING / BIT WIZARDRY
593 root 1.1
594 root 1.88 =over
595 root 1.4
596 root 1.3 =item bool ecb_big_endian ()
597    
598     =item bool ecb_little_endian ()
599    
600 sf-exg 1.11 These two functions return true if the byte order is big endian
601     (most-significant byte first) or little endian (least-significant byte
602     first) respectively.
603    
604 root 1.24 On systems that are neither, their return values are unspecified.
605    
606 root 1.3 =item int ecb_ctz32 (uint32_t x)
607    
608 root 1.35 =item int ecb_ctz64 (uint64_t x)
609    
610 root 1.77 =item int ecb_ctz (T x) [C++]
611    
612 sf-exg 1.11 Returns the index of the least significant bit set in C<x> (or
613 root 1.24 equivalently the number of bits set to 0 before the least significant bit
614 root 1.35 set), starting from 0. If C<x> is 0 the result is undefined.
615    
616 root 1.36 For smaller types than C<uint32_t> you can safely use C<ecb_ctz32>.
617    
618 root 1.77 The overloaded C++ C<ecb_ctz> function supports C<uint8_t>, C<uint16_t>,
619     C<uint32_t> and C<uint64_t> types.
620    
621 root 1.35 For example:
622 sf-exg 1.11
623 root 1.15 ecb_ctz32 (3) = 0
624     ecb_ctz32 (6) = 1
625 sf-exg 1.11
626 root 1.105 =item int ecb_clz32 (uint32_t x)
627    
628     =item int ecb_clz64 (uint64_t x)
629    
630     Counts the number of leading zero bits in C<x>. If C<x> is 0 the result is
631     undefined.
632    
633 root 1.106 It is often simpler to use one of the C<ecb_ld*> functions instead, whose
634     result only depends on the value and not the size of the type. This is
635     also the reason why there is no C++ overload.
636 root 1.105
637     For example:
638    
639     ecb_clz32 (3) = 30
640     ecb_clz32 (6) = 29
641    
642 root 1.41 =item bool ecb_is_pot32 (uint32_t x)
643    
644     =item bool ecb_is_pot64 (uint32_t x)
645    
646 root 1.77 =item bool ecb_is_pot (T x) [C++]
647    
648 sf-exg 1.66 Returns true iff C<x> is a power of two or C<x == 0>.
649 root 1.41
650 sf-exg 1.66 For smaller types than C<uint32_t> you can safely use C<ecb_is_pot32>.
651 root 1.41
652 root 1.77 The overloaded C++ C<ecb_is_pot> function supports C<uint8_t>, C<uint16_t>,
653     C<uint32_t> and C<uint64_t> types.
654    
655 root 1.35 =item int ecb_ld32 (uint32_t x)
656    
657     =item int ecb_ld64 (uint64_t x)
658    
659 root 1.77 =item int ecb_ld64 (T x) [C++]
660    
661 root 1.35 Returns the index of the most significant bit set in C<x>, or the number
662     of digits the number requires in binary (so that C<< 2**ld <= x <
663     2**(ld+1) >>). If C<x> is 0 the result is undefined. A common use case is
664     to compute the integer binary logarithm, i.e. C<floor (log2 (n))>, for
665     example to see how many bits a certain number requires to be encoded.
666    
667     This function is similar to the "count leading zero bits" function, except
668     that that one returns how many zero bits are "in front" of the number (in
669     the given data type), while C<ecb_ld> returns how many bits the number
670     itself requires.
671    
672 root 1.36 For smaller types than C<uint32_t> you can safely use C<ecb_ld32>.
673    
674 root 1.77 The overloaded C++ C<ecb_ld> function supports C<uint8_t>, C<uint16_t>,
675     C<uint32_t> and C<uint64_t> types.
676    
677 root 1.3 =item int ecb_popcount32 (uint32_t x)
678    
679 root 1.35 =item int ecb_popcount64 (uint64_t x)
680    
681 root 1.77 =item int ecb_popcount (T x) [C++]
682    
683 root 1.36 Returns the number of bits set to 1 in C<x>.
684    
685     For smaller types than C<uint32_t> you can safely use C<ecb_popcount32>.
686    
687 root 1.77 The overloaded C++ C<ecb_popcount> function supports C<uint8_t>, C<uint16_t>,
688     C<uint32_t> and C<uint64_t> types.
689    
690 root 1.36 For example:
691 sf-exg 1.11
692 root 1.15 ecb_popcount32 (7) = 3
693     ecb_popcount32 (255) = 8
694 sf-exg 1.11
695 root 1.39 =item uint8_t ecb_bitrev8 (uint8_t x)
696    
697     =item uint16_t ecb_bitrev16 (uint16_t x)
698    
699     =item uint32_t ecb_bitrev32 (uint32_t x)
700    
701 root 1.77 =item T ecb_bitrev (T x) [C++]
702    
703 root 1.39 Reverses the bits in x, i.e. the MSB becomes the LSB, MSB-1 becomes LSB+1
704     and so on.
705    
706 root 1.77 The overloaded C++ C<ecb_bitrev> function supports C<uint8_t>, C<uint16_t> and C<uint32_t> types.
707    
708 root 1.39 Example:
709    
710     ecb_bitrev8 (0xa7) = 0xea
711     ecb_bitrev32 (0xffcc4411) = 0x882233ff
712    
713 root 1.77 =item T ecb_bitrev (T x) [C++]
714    
715     Overloaded C++ bitrev function.
716    
717     C<T> must be one of C<uint8_t>, C<uint16_t> or C<uint32_t>.
718    
719 root 1.8 =item uint32_t ecb_bswap16 (uint32_t x)
720    
721 root 1.3 =item uint32_t ecb_bswap32 (uint32_t x)
722    
723 root 1.34 =item uint64_t ecb_bswap64 (uint64_t x)
724 sf-exg 1.13
725 root 1.78 =item T ecb_bswap (T x)
726    
727 root 1.34 These functions return the value of the 16-bit (32-bit, 64-bit) value
728     C<x> after reversing the order of bytes (0x11223344 becomes 0x44332211 in
729     C<ecb_bswap32>).
730    
731 root 1.77 The overloaded C++ C<ecb_bswap> function supports C<uint8_t>, C<uint16_t>,
732     C<uint32_t> and C<uint64_t> types.
733 root 1.76
734 root 1.34 =item uint8_t ecb_rotl8 (uint8_t x, unsigned int count)
735    
736     =item uint16_t ecb_rotl16 (uint16_t x, unsigned int count)
737 root 1.3
738     =item uint32_t ecb_rotl32 (uint32_t x, unsigned int count)
739    
740 root 1.34 =item uint64_t ecb_rotl64 (uint64_t x, unsigned int count)
741    
742     =item uint8_t ecb_rotr8 (uint8_t x, unsigned int count)
743    
744     =item uint16_t ecb_rotr16 (uint16_t x, unsigned int count)
745    
746     =item uint32_t ecb_rotr32 (uint32_t x, unsigned int count)
747    
748 root 1.33 =item uint64_t ecb_rotr64 (uint64_t x, unsigned int count)
749    
750 root 1.34 These two families of functions return the value of C<x> after rotating
751     all the bits by C<count> positions to the right (C<ecb_rotr>) or left
752 root 1.94 (C<ecb_rotl>). There are no restrictions on the value C<count>, i.e. both
753 root 1.95 zero and values equal or larger than the word width work correctly. Also,
754     notwithstanding C<count> being unsigned, negative numbers work and shift
755     to the opposite direction.
756 root 1.93
757 root 1.85 Current GCC/clang versions understand these functions and usually compile
758     them to "optimal" code (e.g. a single C<rol> or a combination of C<shld>
759     on x86).
760 root 1.20
761 root 1.77 =item T ecb_rotl (T x, unsigned int count) [C++]
762    
763     =item T ecb_rotr (T x, unsigned int count) [C++]
764    
765     Overloaded C++ rotl/rotr functions.
766    
767     C<T> must be one of C<uint8_t>, C<uint16_t>, C<uint32_t> or C<uint64_t>.
768    
769 root 1.102 =item uint_fast8_t ecb_gray8_encode (uint_fast8_t b)
770    
771     =item uint_fast16_t ecb_gray16_encode (uint_fast16_t b)
772    
773     =item uint_fast32_t ecb_gray32_encode (uint_fast32_t b)
774    
775     =item uint_fast64_t ecb_gray64_encode (uint_fast64_t b)
776    
777     Encode an unsigned into its corresponding (reflective) gray code - the
778     kind of gray code meant when just talking about "gray code". These
779     functions are very fast and all have identical implementation, so there is
780 root 1.103 no need to use a smaller type, as long as your CPU can handle it natively.
781 root 1.102
782     =item T ecb_gray_encode (T b) [C++]
783    
784     Overloaded C++ version of the above, for C<uint{8,16,32,64}_t>.
785    
786     =item uint_fast8_t ecb_gray8_decode (uint_fast8_t b)
787    
788     =item uint_fast16_t ecb_gray16_decode (uint_fast16_t b)
789    
790     =item uint_fast32_t ecb_gray32_decode (uint_fast32_t b)
791    
792     =item uint_fast64_t ecb_gray64_decode (uint_fast64_t b)
793    
794 root 1.103 Decode a gray code back into linear index form (the reverse of
795     C<ecb_gray*_encode>. Unlike the encode functions, the decode functions
796     have higher time complexity for larger types, so it can pay off to use a
797 root 1.102 smaller type here.
798    
799     =item T ecb_gray_decode (T b) [C++]
800    
801     Overloaded C++ version of the above, for C<uint{8,16,32,64}_t>.
802    
803 root 1.3 =back
804 root 1.1
805 root 1.104 =head2 HILBERT CURVES
806    
807     These functions deal with (square, pseudo) Hilbert curves. The parameter
808     I<order> indicates the size of the square and is specified in bits, that
809     means for order C<8>, the coordinates range from C<0>..C<255>, and the
810     curve index ranges from C<0>..C<65535>.
811    
812     The 32 bit variants of these functions map a 32 bit index to two 16 bit
813     coordinates, stored in a 32 bit variable, where the high order bits are
814     the x-coordinate, and the low order bits are the y-coordinate, thus,
815     these functions map 32 bit linear index on the curve to a 32 bit packed
816     coordinate pair, and vice versa.
817    
818     The 64 bit variants work similarly.
819    
820     The I<order> can go from C<1> to C<16> for the 32 bit curve, and C<1> to
821     C<32> for the 64 bit curve.
822    
823     When going from one order to the next higher order, these functions
824     replace the curve segments by smaller versions of the generating shape,
825     while doubling the size (since they use integer coordinates), which is
826     what you would expect mathematically. This means that the curve will be
827     mirrored at the diagonal. If your goal is to simply cover more area while
828     retaining existing point coordinates you should increase or decrease the
829     I<order> by C<2> or, in the case of C<ecb_hilbert2d_index_to_coord>,
830     simply specify the maximum I<order> of C<16> or C<32>, respectively, as
831     these are constant-time.
832    
833     =over
834    
835     =item uint32_t ecb_hilbert2d_index_to_coord32 (int order, uint32_t index)
836    
837     =item uint64_t ecb_hilbert2d_index_to_coord64 (int order, uint64_t index)
838    
839     Map a point on a pseudo Hilbert curve from its linear distance from the
840     origin on the curve to a x|y coordinate pair. The result is a packed
841     coordinate pair, to get the actual x and < coordinates, you could do
842     something like this:
843    
844     uint32_t xy = ecb_hilbert2d_index_to_coord32 (16, 255);
845     uint16_t x = xy >> 16;
846     uint16_t y = xy & 0xffffU;
847    
848     uint64_t xy = ecb_hilbert2d_index_to_coord64 (32, 255);
849     uint32_t x = xy >> 32;
850     uint32_t y = xy & 0xffffffffU;
851    
852     These functions work in constant time, so for many applications it is
853     preferable to simply hard-code the order to the maximum (C<16> or C<32>).
854    
855     This (production-ready, i.e. never run) example generates an SVG image of
856     an order 8 pseudo Hilbert curve:
857    
858     printf ("<svg xmlns='http://www.w3.org/2000/svg' width='%d' height='%d'>\n", 64 * 8, 64 * 8);
859     printf ("<g transform='translate(4) scale(8)' stroke-width='0.25' stroke='black'>\n");
860     for (uint32_t i = 0; i < 64*64 - 1; ++i)
861     {
862     uint32_t p1 = ecb_hilbert2d_index_to_coord32 (6, i );
863     uint32_t p2 = ecb_hilbert2d_index_to_coord32 (6, i + 1);
864     printf ("<line x1='%d' y1='%d' x2='%d' y2='%d'/>\n",
865     p1 >> 16, p1 & 0xffff,
866     p2 >> 16, p2 & 0xffff);
867     }
868     printf ("</g>\n");
869     printf ("</svg>\n");
870    
871     =item uint32_t ecb_hilbert2d_coord_to_index32 (int order, uint32_t xy)
872    
873     =item uint64_t ecb_hilbert2d_coord_to_index64 (int order, uint64_t xy)
874    
875     The reverse of C<ecb_hilbert2d_index_to_coord> - map a packed pair of
876     coordinates to their linear index on the pseudo Hilbert curve of order
877     I<order>.
878    
879     They are an exact inverse of the C<ecb_hilbert2d_coord_to_index> functions
880     for the same I<order>:
881    
882     assert (
883     u == ecb_hilbert2d_coord_to_index (32,
884     ecb_hilbert2d_index_to_coord32 (32,
885     u)));
886    
887     Packing coordinates is done the same way, as well, from I<x> and I<y>:
888    
889     uint32_t xy = ((uint32_t)x << 16) | y; // for ecb_hilbert2d_coord_to_index32
890     uint64_t xy = ((uint64_t)x << 32) | y; // for ecb_hilbert2d_coord_to_index64
891    
892     Unlike C<ecb_hilbert2d_coord_to_index>, these functions are O(I<order>),
893     so it is preferable to use the lowest possible order.
894    
895     =back
896    
897 root 1.96 =head2 BIT MIXING, HASHING
898    
899     Sometimes you have an integer and want to distribute its bits well, for
900 root 1.104 example, to use it as a hash in a hash table. A common example is pointer
901 root 1.96 values, which often only have a limited range (e.g. low and high bits are
902     often zero).
903    
904     The following functions try to mix the bits to get a good bias-free
905     distribution. They were mainly made for pointers, but the underlying
906     integer functions are exposed as well.
907    
908     As an added benefit, the functions are reversible, so if you find it
909     convenient to store only the hash value, you can recover the original
910 root 1.104 pointer from the hash ("unmix"), as long as your pointers are 32 or 64 bit
911 root 1.96 (if this isn't the case on your platform, drop us a note and we will add
912     functions for other bit widths).
913    
914     The unmix functions are very slightly slower than the mix functions, so
915     it is equally very slightly preferable to store the original values wehen
916     convenient.
917    
918     The underlying algorithm if subject to change, so currently these
919     functions are not suitable for persistent hash tables, as their result
920 root 1.104 value can change between different versions of libecb.
921 root 1.96
922     =over
923    
924     =item uintptr_t ecb_ptrmix (void *ptr)
925    
926     Mixes the bits of a pointer so the result is suitable for hash table
927     lookups. In other words, this hashes the pointer value.
928    
929     =item uintptr_t ecb_ptrmix (T *ptr) [C++]
930    
931     Overload the C<ecb_ptrmix> function to work for any pointer in C++.
932    
933     =item void *ecb_ptrunmix (uintptr_t v)
934    
935     Unmix the hash value into the original pointer. This only works as long
936     as the hash value is not truncated, i.e. you used C<uintptr_t> (or
937     equivalent) throughout to store it.
938    
939     =item T *ecb_ptrunmix<T> (uintptr_t v) [C++]
940    
941     The somewhat less useful template version of C<ecb_ptrunmix> for
942     C++. Example:
943    
944     sometype *myptr;
945     uintptr_t hash = ecb_ptrmix (myptr);
946     sometype *orig = ecb_ptrunmix<sometype> (hash);
947    
948     =item uint32_t ecb_mix32 (uint32_t v)
949    
950     =item uint64_t ecb_mix64 (uint64_t v)
951    
952     Sometimes you don't have a pointer but an integer whose values are very
953 root 1.104 badly distributed. In this case you can use these integer versions of the
954 root 1.96 mixing function. No C++ template is provided currently.
955    
956     =item uint32_t ecb_unmix32 (uint32_t v)
957    
958     =item uint64_t ecb_unmix64 (uint64_t v)
959    
960     The reverse of the C<ecb_mix> functions - they take a mixed/hashed value
961     and recover the original value.
962    
963     =back
964    
965 root 1.76 =head2 HOST ENDIANNESS CONVERSION
966    
967 root 1.88 =over
968 root 1.76
969     =item uint_fast16_t ecb_be_u16_to_host (uint_fast16_t v)
970    
971     =item uint_fast32_t ecb_be_u32_to_host (uint_fast32_t v)
972    
973     =item uint_fast64_t ecb_be_u64_to_host (uint_fast64_t v)
974    
975     =item uint_fast16_t ecb_le_u16_to_host (uint_fast16_t v)
976    
977     =item uint_fast32_t ecb_le_u32_to_host (uint_fast32_t v)
978    
979     =item uint_fast64_t ecb_le_u64_to_host (uint_fast64_t v)
980    
981     Convert an unsigned 16, 32 or 64 bit value from big or little endian to host byte order.
982    
983     The naming convention is C<ecb_>(C<be>|C<le>)C<_u>C<16|32|64>C<_to_host>,
984 root 1.79 where C<be> and C<le> stand for big endian and little endian, respectively.
985 root 1.76
986     =item uint_fast16_t ecb_host_to_be_u16 (uint_fast16_t v)
987    
988     =item uint_fast32_t ecb_host_to_be_u32 (uint_fast32_t v)
989    
990     =item uint_fast64_t ecb_host_to_be_u64 (uint_fast64_t v)
991    
992     =item uint_fast16_t ecb_host_to_le_u16 (uint_fast16_t v)
993    
994     =item uint_fast32_t ecb_host_to_le_u32 (uint_fast32_t v)
995    
996     =item uint_fast64_t ecb_host_to_le_u64 (uint_fast64_t v)
997    
998     Like above, but converts I<from> host byte order to the specified
999     endianness.
1000    
1001     =back
1002    
1003 root 1.77 In C++ the following additional template functions are supported:
1004 root 1.76
1005 root 1.88 =over
1006 root 1.76
1007     =item T ecb_be_to_host (T v)
1008    
1009     =item T ecb_le_to_host (T v)
1010    
1011     =item T ecb_host_to_be (T v)
1012    
1013     =item T ecb_host_to_le (T v)
1014    
1015 root 1.86 =back
1016    
1017 root 1.77 These functions work like their C counterparts, above, but use templates,
1018     which make them useful in generic code.
1019 root 1.76
1020     C<T> must be one of C<uint8_t>, C<uint16_t>, C<uint32_t> or C<uint64_t>
1021     (so unlike their C counterparts, there is a version for C<uint8_t>, which
1022     again can be useful in generic code).
1023    
1024     =head2 UNALIGNED LOAD/STORE
1025    
1026     These function load or store unaligned multi-byte values.
1027    
1028 root 1.88 =over
1029 root 1.76
1030     =item uint_fast16_t ecb_peek_u16_u (const void *ptr)
1031    
1032     =item uint_fast32_t ecb_peek_u32_u (const void *ptr)
1033    
1034     =item uint_fast64_t ecb_peek_u64_u (const void *ptr)
1035    
1036     These functions load an unaligned, unsigned 16, 32 or 64 bit value from
1037     memory.
1038    
1039     =item uint_fast16_t ecb_peek_be_u16_u (const void *ptr)
1040    
1041     =item uint_fast32_t ecb_peek_be_u32_u (const void *ptr)
1042    
1043     =item uint_fast64_t ecb_peek_be_u64_u (const void *ptr)
1044    
1045     =item uint_fast16_t ecb_peek_le_u16_u (const void *ptr)
1046    
1047     =item uint_fast32_t ecb_peek_le_u32_u (const void *ptr)
1048    
1049     =item uint_fast64_t ecb_peek_le_u64_u (const void *ptr)
1050    
1051     Like above, but additionally convert from big endian (C<be>) or little
1052     endian (C<le>) byte order to host byte order while doing so.
1053    
1054     =item ecb_poke_u16_u (void *ptr, uint16_t v)
1055    
1056     =item ecb_poke_u32_u (void *ptr, uint32_t v)
1057    
1058     =item ecb_poke_u64_u (void *ptr, uint64_t v)
1059    
1060     These functions store an unaligned, unsigned 16, 32 or 64 bit value to
1061     memory.
1062    
1063     =item ecb_poke_be_u16_u (void *ptr, uint_fast16_t v)
1064    
1065     =item ecb_poke_be_u32_u (void *ptr, uint_fast32_t v)
1066    
1067     =item ecb_poke_be_u64_u (void *ptr, uint_fast64_t v)
1068    
1069     =item ecb_poke_le_u16_u (void *ptr, uint_fast16_t v)
1070    
1071     =item ecb_poke_le_u32_u (void *ptr, uint_fast32_t v)
1072    
1073     =item ecb_poke_le_u64_u (void *ptr, uint_fast64_t v)
1074    
1075     Like above, but additionally convert from host byte order to big endian
1076     (C<be>) or little endian (C<le>) byte order while doing so.
1077    
1078     =back
1079    
1080 root 1.77 In C++ the following additional template functions are supported:
1081 root 1.76
1082 root 1.88 =over
1083 root 1.76
1084 root 1.80 =item T ecb_peek<T> (const void *ptr)
1085 root 1.76
1086 root 1.80 =item T ecb_peek_be<T> (const void *ptr)
1087 root 1.76
1088 root 1.80 =item T ecb_peek_le<T> (const void *ptr)
1089 root 1.76
1090 root 1.80 =item T ecb_peek_u<T> (const void *ptr)
1091 root 1.76
1092 root 1.80 =item T ecb_peek_be_u<T> (const void *ptr)
1093 root 1.76
1094 root 1.80 =item T ecb_peek_le_u<T> (const void *ptr)
1095 root 1.76
1096     Similarly to their C counterparts, these functions load an unsigned 8, 16,
1097     32 or 64 bit value from memory, with optional conversion from big/little
1098     endian.
1099    
1100 root 1.80 Since the type cannot be deduced, it has to be specified explicitly, e.g.
1101 root 1.76
1102     uint_fast16_t v = ecb_peek<uint16_t> (ptr);
1103    
1104     C<T> must be one of C<uint8_t>, C<uint16_t>, C<uint32_t> or C<uint64_t>.
1105    
1106     Unlike their C counterparts, these functions support 8 bit quantities
1107     (C<uint8_t>) and also have an aligned version (without the C<_u> prefix),
1108     all of which hopefully makes them more useful in generic code.
1109    
1110     =item ecb_poke (void *ptr, T v)
1111    
1112     =item ecb_poke_be (void *ptr, T v)
1113    
1114     =item ecb_poke_le (void *ptr, T v)
1115    
1116     =item ecb_poke_u (void *ptr, T v)
1117    
1118     =item ecb_poke_be_u (void *ptr, T v)
1119    
1120     =item ecb_poke_le_u (void *ptr, T v)
1121    
1122     Again, similarly to their C counterparts, these functions store an
1123 root 1.104 unsigned 8, 16, 32 or 64 bit value to memory, with optional conversion to
1124 root 1.76 big/little endian.
1125    
1126     C<T> must be one of C<uint8_t>, C<uint16_t>, C<uint32_t> or C<uint64_t>.
1127    
1128     Unlike their C counterparts, these functions support 8 bit quantities
1129     (C<uint8_t>) and also have an aligned version (without the C<_u> prefix),
1130     all of which hopefully makes them more useful in generic code.
1131    
1132     =back
1133    
1134 root 1.89 =head2 FAST INTEGER TO STRING
1135    
1136     Libecb defines a set of very fast integer to decimal string (or integer
1137 root 1.104 to ASCII, short C<i2a>) functions. These work by converting the integer
1138 root 1.89 to a fixed point representation and then successively multiplying out
1139     the topmost digits. Unlike some other, also very fast, libraries, ecb's
1140     algorithm should be completely branchless per digit, and does not rely on
1141 root 1.104 the presence of special CPU functions (such as C<clz>).
1142 root 1.89
1143     There is a high level API that takes an C<int32_t>, C<uint32_t>,
1144     C<int64_t> or C<uint64_t> as argument, and a low-level API, which is
1145     harder to use but supports slightly more formatting options.
1146    
1147     =head3 HIGH LEVEL API
1148    
1149     The high level API consists of four functions, one each for C<int32_t>,
1150     C<uint32_t>, C<int64_t> and C<uint64_t>:
1151    
1152 root 1.91 Example:
1153    
1154     char buf[ECB_I2A_MAX_DIGITS + 1];
1155     char *end = ecb_i2a_i32 (buf, 17262);
1156     *end = 0;
1157     // buf now contains "17262"
1158    
1159 root 1.89 =over
1160    
1161     =item ECB_I2A_I32_DIGITS (=11)
1162    
1163     =item char *ecb_i2a_u32 (char *ptr, uint32_t value)
1164    
1165     Takes an C<uint32_t> I<value> and formats it as a decimal number starting
1166     at I<ptr>, using at most C<ECB_I2A_I32_DIGITS> characters. Returns a
1167     pointer to just after the generated string, where you would normally put
1168 sf-exg 1.92 the terminating C<0> character. This function outputs the minimum number
1169 root 1.89 of digits.
1170    
1171     =item ECB_I2A_U32_DIGITS (=10)
1172    
1173     =item char *ecb_i2a_i32 (char *ptr, int32_t value)
1174    
1175     Same as C<ecb_i2a_u32>, but formats a C<int32_t> value, including a minus
1176     sign if needed.
1177    
1178     =item ECB_I2A_I64_DIGITS (=20)
1179    
1180     =item char *ecb_i2a_u64 (char *ptr, uint64_t value)
1181    
1182     =item ECB_I2A_U64_DIGITS (=21)
1183    
1184     =item char *ecb_i2a_i64 (char *ptr, int64_t value)
1185    
1186     Similar to their 32 bit counterparts, these take a 64 bit argument.
1187    
1188 root 1.90 =item ECB_I2A_MAX_DIGITS (=21)
1189 root 1.89
1190 root 1.97 Instead of using a type specific length macro, you can just use
1191 root 1.90 C<ECB_I2A_MAX_DIGITS>, which is good enough for any C<ecb_i2a> function.
1192 root 1.89
1193     =back
1194    
1195     =head3 LOW-LEVEL API
1196    
1197     The functions above use a number of low-level APIs which have some strict
1198 root 1.98 limitations, but can be used as building blocks (studying C<ecb_i2a_i32>
1199 sf-exg 1.92 and related functions is recommended).
1200 root 1.89
1201     There are three families of functions: functions that convert a number
1202     to a fixed number of digits with leading zeroes (C<ecb_i2a_0N>, C<0>
1203     for "leading zeroes"), functions that generate up to N digits, skipping
1204     leading zeroes (C<_N>), and functions that can generate more digits, but
1205     the leading digit has limited range (C<_xN>).
1206    
1207 sf-exg 1.92 None of the functions deal with negative numbers.
1208 root 1.89
1209 root 1.104 Example: convert an IP address in an C<uint32_t> into dotted-quad:
1210 root 1.91
1211     uint32_t ip = 0x0a000164; // 10.0.1.100
1212     char ips[3 * 4 + 3 + 1];
1213     char *ptr = ips;
1214     ptr = ecb_i2a_3 (ptr, ip >> 24 ); *ptr++ = '.';
1215     ptr = ecb_i2a_3 (ptr, (ip >> 16) & 0xff); *ptr++ = '.';
1216     ptr = ecb_i2a_3 (ptr, (ip >> 8) & 0xff); *ptr++ = '.';
1217     ptr = ecb_i2a_3 (ptr, ip & 0xff); *ptr++ = 0;
1218     printf ("ip: %s\n", ips); // prints "ip: 10.0.1.100"
1219    
1220 root 1.89 =over
1221    
1222     =item char *ecb_i2a_02 (char *ptr, uint32_t value) // 32 bit
1223    
1224     =item char *ecb_i2a_03 (char *ptr, uint32_t value) // 32 bit
1225    
1226     =item char *ecb_i2a_04 (char *ptr, uint32_t value) // 32 bit
1227    
1228     =item char *ecb_i2a_05 (char *ptr, uint32_t value) // 64 bit
1229    
1230     =item char *ecb_i2a_06 (char *ptr, uint32_t value) // 64 bit
1231    
1232     =item char *ecb_i2a_07 (char *ptr, uint32_t value) // 64 bit
1233    
1234     =item char *ecb_i2a_08 (char *ptr, uint32_t value) // 64 bit
1235    
1236     =item char *ecb_i2a_09 (char *ptr, uint32_t value) // 64 bit
1237    
1238 root 1.101 The C<< ecb_i2a_0I<N> >> functions take an unsigned I<value> and convert
1239 root 1.89 them to exactly I<N> digits, returning a pointer to the first character
1240     after the digits. The I<value> must be in range. The functions marked with
1241     I<32 bit> do their calculations internally in 32 bit, the ones marked with
1242     I<64 bit> internally use 64 bit integers, which might be slow on 32 bit
1243     architectures (the high level API decides on 32 vs. 64 bit versions using
1244     C<ECB_64BIT_NATIVE>).
1245    
1246     =item char *ecb_i2a_2 (char *ptr, uint32_t value) // 32 bit
1247    
1248     =item char *ecb_i2a_3 (char *ptr, uint32_t value) // 32 bit
1249    
1250     =item char *ecb_i2a_4 (char *ptr, uint32_t value) // 32 bit
1251    
1252     =item char *ecb_i2a_5 (char *ptr, uint32_t value) // 64 bit
1253    
1254     =item char *ecb_i2a_6 (char *ptr, uint32_t value) // 64 bit
1255    
1256     =item char *ecb_i2a_7 (char *ptr, uint32_t value) // 64 bit
1257    
1258     =item char *ecb_i2a_8 (char *ptr, uint32_t value) // 64 bit
1259    
1260     =item char *ecb_i2a_9 (char *ptr, uint32_t value) // 64 bit
1261    
1262 root 1.101 Similarly, the C<< ecb_i2a_I<N> >> functions take an unsigned I<value>
1263 root 1.89 and convert them to at most I<N> digits, suppressing leading zeroes, and
1264     returning a pointer to the first character after the digits.
1265    
1266     =item ECB_I2A_MAX_X5 (=59074)
1267    
1268     =item char *ecb_i2a_x5 (char *ptr, uint32_t value) // 32 bit
1269    
1270     =item ECB_I2A_MAX_X10 (=2932500665)
1271    
1272     =item char *ecb_i2a_x10 (char *ptr, uint32_t value) // 64 bit
1273    
1274 root 1.101 The C<< ecb_i2a_xI<N> >> functions are similar to the C<< ecb_i2a_I<N> >>
1275 root 1.89 functions, but they can generate one digit more, as long as the number
1276     is within range, which is given by the symbols C<ECB_I2A_MAX_X5> (almost
1277     16 bit range) and C<ECB_I2A_MAX_X10> (a bit more than 31 bit range),
1278     respectively.
1279    
1280 sf-exg 1.92 For example, the digit part of a 32 bit signed integer just fits into the
1281 root 1.89 C<ECB_I2A_MAX_X10> range, so while C<ecb_i2a_x10> cannot convert a 10
1282     digit number, it can convert all 32 bit signed numbers. Sadly, it's not
1283     good enough for 32 bit unsigned numbers.
1284    
1285     =back
1286    
1287 root 1.50 =head2 FLOATING POINT FIDDLING
1288    
1289 root 1.88 =over
1290 root 1.50
1291 root 1.71 =item ECB_INFINITY [-UECB_NO_LIBM]
1292 root 1.62
1293     Evaluates to positive infinity if supported by the platform, otherwise to
1294     a truly huge number.
1295    
1296 root 1.71 =item ECB_NAN [-UECB_NO_LIBM]
1297 root 1.62
1298     Evaluates to a quiet NAN if supported by the platform, otherwise to
1299     C<ECB_INFINITY>.
1300    
1301 root 1.71 =item float ecb_ldexpf (float x, int exp) [-UECB_NO_LIBM]
1302 root 1.62
1303     Same as C<ldexpf>, but always available.
1304    
1305 root 1.71 =item uint32_t ecb_float_to_binary16 (float x) [-UECB_NO_LIBM]
1306    
1307 root 1.50 =item uint32_t ecb_float_to_binary32 (float x) [-UECB_NO_LIBM]
1308    
1309     =item uint64_t ecb_double_to_binary64 (double x) [-UECB_NO_LIBM]
1310    
1311     These functions each take an argument in the native C<float> or C<double>
1312 root 1.71 type and return the IEEE 754 bit representation of it (binary16/half,
1313     binary32/single or binary64/double precision).
1314 root 1.50
1315     The bit representation is just as IEEE 754 defines it, i.e. the sign bit
1316     will be the most significant bit, followed by exponent and mantissa.
1317    
1318     This function should work even when the native floating point format isn't
1319     IEEE compliant, of course at a speed and code size penalty, and of course
1320     also within reasonable limits (it tries to convert NaNs, infinities and
1321     denormals, but will likely convert negative zero to positive zero).
1322    
1323     On all modern platforms (where C<ECB_STDFP> is true), the compiler should
1324 root 1.99 be able to completely optimise away the 32 and 64 bit functions.
1325 root 1.50
1326     These functions can be helpful when serialising floats to the network - you
1327 root 1.71 can serialise the return value like a normal uint16_t/uint32_t/uint64_t.
1328 root 1.50
1329     Another use for these functions is to manipulate floating point values
1330     directly.
1331    
1332     Silly example: toggle the sign bit of a float.
1333    
1334     /* On gcc-4.7 on amd64, */
1335     /* this results in a single add instruction to toggle the bit, and 4 extra */
1336     /* instructions to move the float value to an integer register and back. */
1337    
1338     x = ecb_binary32_to_float (ecb_float_to_binary32 (x) ^ 0x80000000U)
1339    
1340 root 1.58 =item float ecb_binary16_to_float (uint16_t x) [-UECB_NO_LIBM]
1341    
1342 root 1.50 =item float ecb_binary32_to_float (uint32_t x) [-UECB_NO_LIBM]
1343    
1344 root 1.70 =item double ecb_binary64_to_double (uint64_t x) [-UECB_NO_LIBM]
1345 root 1.50
1346 sf-exg 1.59 The reverse operation of the previous function - takes the bit
1347 root 1.71 representation of an IEEE binary16, binary32 or binary64 number (half,
1348     single or double precision) and converts it to the native C<float> or
1349     C<double> format.
1350 root 1.50
1351     This function should work even when the native floating point format isn't
1352     IEEE compliant, of course at a speed and code size penalty, and of course
1353     also within reasonable limits (it tries to convert normals and denormals,
1354     and might be lucky for infinities, and with extraordinary luck, also for
1355     negative zero).
1356    
1357     On all modern platforms (where C<ECB_STDFP> is true), the compiler should
1358     be able to optimise away this function completely.
1359    
1360 root 1.71 =item uint16_t ecb_binary32_to_binary16 (uint32_t x)
1361    
1362     =item uint32_t ecb_binary16_to_binary32 (uint16_t x)
1363    
1364     Convert a IEEE binary32/single precision to binary16/half format, and vice
1365 root 1.72 versa, handling all details (round-to-nearest-even, subnormals, infinity
1366     and NaNs) correctly.
1367 root 1.71
1368     These are functions are available under C<-DECB_NO_LIBM>, since
1369     they do not rely on the platform floating point format. The
1370     C<ecb_float_to_binary16> and C<ecb_binary16_to_float> functions are
1371     usually what you want.
1372    
1373 root 1.50 =back
1374    
1375 root 1.1 =head2 ARITHMETIC
1376    
1377 root 1.88 =over
1378 root 1.3
1379 root 1.14 =item x = ecb_mod (m, n)
1380 root 1.3
1381 root 1.25 Returns C<m> modulo C<n>, which is the same as the positive remainder
1382     of the division operation between C<m> and C<n>, using floored
1383     division. Unlike the C remainder operator C<%>, this function ensures that
1384     the return value is always positive and that the two numbers I<m> and
1385     I<m' = m + i * n> result in the same value modulo I<n> - in other words,
1386     C<ecb_mod> implements the mathematical modulo operation, which is missing
1387     in the language.
1388 root 1.14
1389 sf-exg 1.23 C<n> must be strictly positive (i.e. C<< >= 1 >>), while C<m> must be
1390 root 1.14 negatable, that is, both C<m> and C<-m> must be representable in its
1391 root 1.30 type (this typically excludes the minimum signed integer value, the same
1392 root 1.25 limitation as for C</> and C<%> in C).
1393 sf-exg 1.11
1394 root 1.85 Current GCC/clang versions compile this into an efficient branchless
1395     sequence on almost all CPUs.
1396 root 1.24
1397     For example, when you want to rotate forward through the members of an
1398     array for increasing C<m> (which might be negative), then you should use
1399     C<ecb_mod>, as the C<%> operator might give either negative results, or
1400     change direction for negative values:
1401    
1402     for (m = -100; m <= 100; ++m)
1403     int elem = myarray [ecb_mod (m, ecb_array_length (myarray))];
1404    
1405 sf-exg 1.37 =item x = ecb_div_rd (val, div)
1406    
1407     =item x = ecb_div_ru (val, div)
1408    
1409     Returns C<val> divided by C<div> rounded down or up, respectively.
1410     C<val> and C<div> must have integer types and C<div> must be strictly
1411 sf-exg 1.38 positive. Note that these functions are implemented with macros in C
1412     and with function templates in C++.
1413 sf-exg 1.37
1414 root 1.3 =back
1415 root 1.1
1416     =head2 UTILITY
1417    
1418 root 1.88 =over
1419 root 1.3
1420 sf-exg 1.23 =item element_count = ecb_array_length (name)
1421 root 1.3
1422 sf-exg 1.13 Returns the number of elements in the array C<name>. For example:
1423    
1424     int primes[] = { 2, 3, 5, 7, 11 };
1425     int sum = 0;
1426    
1427     for (i = 0; i < ecb_array_length (primes); i++)
1428     sum += primes [i];
1429    
1430 root 1.3 =back
1431 root 1.1
1432 root 1.43 =head2 SYMBOLS GOVERNING COMPILATION OF ECB.H ITSELF
1433    
1434     These symbols need to be defined before including F<ecb.h> the first time.
1435    
1436 root 1.88 =over
1437 root 1.43
1438 root 1.51 =item ECB_NO_THREADS
1439 root 1.43
1440     If F<ecb.h> is never used from multiple threads, then this symbol can
1441     be defined, in which case memory fences (and similar constructs) are
1442     completely removed, leading to more efficient code and fewer dependencies.
1443    
1444     Setting this symbol to a true value implies C<ECB_NO_SMP>.
1445    
1446     =item ECB_NO_SMP
1447    
1448     The weaker version of C<ECB_NO_THREADS> - if F<ecb.h> is used from
1449     multiple threads, but never concurrently (e.g. if the system the program
1450 root 1.104 runs on has only a single CPU with a single core, no hyper-threading and so
1451 root 1.43 on), then this symbol can be defined, leading to more efficient code and
1452     fewer dependencies.
1453    
1454 root 1.50 =item ECB_NO_LIBM
1455    
1456     When defined to C<1>, do not export any functions that might introduce
1457     dependencies on the math library (usually called F<-lm>) - these are
1458     marked with [-UECB_NO_LIBM].
1459    
1460 sf-exg 1.69 =back
1461    
1462 root 1.68 =head1 UNDOCUMENTED FUNCTIONALITY
1463    
1464     F<ecb.h> is full of undocumented functionality as well, some of which is
1465     intended to be internal-use only, some of which we forgot to document, and
1466     some of which we hide because we are not sure we will keep the interface
1467     stable.
1468    
1469     While you are welcome to rummage around and use whatever you find useful
1470 root 1.100 (we don't want to stop you), keep in mind that we will change undocumented
1471 root 1.68 functionality in incompatible ways without thinking twice, while we are
1472     considerably more conservative with documented things.
1473    
1474     =head1 AUTHORS
1475    
1476     C<libecb> is designed and maintained by:
1477    
1478     Emanuele Giaquinta <e.giaquinta@glauco.it>
1479     Marc Alexander Lehmann <schmorp@schmorp.de>