]> git.pld-linux.org Git - packages/crossavr-libc.git/blob - 506-avr-libc-optimize_dox.patch
- synchronized patches with official AVR toolchain 3.4.1.830
[packages/crossavr-libc.git] / 506-avr-libc-optimize_dox.patch
1 diff -Naurp doc/api/optimize.dox doc/api/optimize.dox
2 --- doc/api/optimize.dox        1970-01-01 05:30:00.000000000 +0530
3 +++ doc/api/optimize.dox        2012-07-25 14:29:02.000000000 +0530
4 @@ -0,0 +1,137 @@
5 +/* Copyright (c) 2010 Jan Waclawek
6 +   Copyright (c) 2010 Joerg Wunsch
7 +   All rights reserved.
8 +
9 +   Redistribution and use in source and binary forms, with or without
10 +   modification, are permitted provided that the following conditions are met:
11 +
12 +   * Redistributions of source code must retain the above copyright
13 +     notice, this list of conditions and the following disclaimer.
14 +   * Redistributions in binary form must reproduce the above copyright
15 +     notice, this list of conditions and the following disclaimer in
16 +     the documentation and/or other materials provided with the
17 +     distribution.
18 +
19 +  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20 +  AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 +  IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 +  ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
23 +  LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 +  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 +  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 +  INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 +  CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 +  ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 +  POSSIBILITY OF SUCH DAMAGE. */
30 +
31 +/* $Id$ */
32 +
33 +/** \page optimization Compiler optimization
34 +
35 +\section optim_code_reorder Problems with reordering code
36 +\author Jan Waclawek
37 +
38 +Programs contain sequences of statements, and a naive compiler would
39 +execute them exactly in the order as they are written. But an
40 +optimizing compiler is free to \e reorder the statements - or even
41 +parts of them - if the resulting "net effect" is the same. The
42 +"measure" of the "net effect" is what the standard calls "side
43 +effects", and is accomplished exclusively through accesses (reads and
44 +writes) to variables qualified as \c volatile. So, as long as all
45 +volatile reads and writes are to the same addresses and in the same
46 +order (and writes write the same values), the program is correct,
47 +regardless of other operations in it. (One important point to note
48 +here is, that time duration between consecutive volatile accesses is
49 +not considered at all.)
50 +
51 +Unfortunately, there are also operations which are not covered by
52 +volatile accesses. An example of this in avr-gcc/avr-libc are the
53 +cli() and sei() macros defined in <avr/interrupt.h>, which convert
54 +directly to the respective assembler mnemonics through the __asm__()
55 +statement. These don't constitute a variable access at all, not even
56 +volatile, so the compiler is free to move them around. Although there
57 +is a "volatile" qualifier which can be attached to the __asm__()
58 +statement, its effect on (re)ordering is not clear from the
59 +documentation (and is more likely only to prevent complete removal by
60 +the optimiser), as it (among other) states:
61 +
62 +<em>Note that even a volatile asm instruction can be moved
63 +relative to other code, including across jump instructions. [...]
64 +Similarly, you can't expect a sequence of volatile asm instructions to
65 +remain perfectly consecutive.</em>
66 +
67 +\sa http://gcc.gnu.org/onlinedocs/gcc-4.3.4/gcc/Extended-Asm.html
68 +
69 +There is another mechanism which can be used to achieve something
70 +similar: <em>memory barriers</em>. This is accomplished through adding a
71 +special "memory" clobber to the inline \c asm statement, and ensures that
72 +all variables are flushed from registers to memory before the
73 +statement, and then re-read after the statement. The purpose of memory
74 +barriers is slightly different than to enforce code ordering: it is
75 +supposed to ensure that there are no variables "cached" in registers,
76 +so that it is safe to change the content of registers e.g. when
77 +switching context in a multitasking OS (on "big" processors with
78 +out-of-order execution they also imply usage of special instructions
79 +which force the processor into "in-order" state (this is not the case
80 +of AVRs)).
81 +
82 +However, memory barrier works well in ensuring that all volatile
83 +accesses before and after the barrier occur in the given order with
84 +respect to the barrier. However, it does not ensure the compiler
85 +moving non-volatile-related statements across the barrier. Peter
86 +Dannegger provided a nice example of this effect:
87 +
88 +\code
89 +#define cli() __asm volatile( "cli" ::: "memory" )
90 +#define sei() __asm volatile( "sei" ::: "memory" )
91 +
92 +unsigned int ivar;
93 +
94 +void test2( unsigned int val )
95 +{
96 +  val = 65535U / val;
97 +
98 +  cli();
99 +
100 +  ivar = val;
101 +
102 +  sei();
103 +}
104 +\endcode
105 +
106 +compiles with optimisations switched on (-Os) to
107 +
108 +\verbatim
109 +00000112 <test2>:
110 + 112:  bc 01           movw    r22, r24
111 + 114:  f8 94           cli
112 + 116:  8f ef           ldi     r24, 0xFF       ; 255
113 + 118:  9f ef           ldi     r25, 0xFF       ; 255
114 + 11a:  0e 94 96 00     call    0x12c   ; 0x12c <__udivmodhi4>
115 + 11e:  70 93 01 02     sts     0x0201, r23
116 + 122:  60 93 00 02     sts     0x0200, r22
117 + 126:  78 94           sei
118 + 128:  08 95           ret
119 +\endverbatim
120 +
121 +where the potentially slow division is moved across cli(),
122 +resulting in interrupts to be disabled longer than intended. Note,
123 +that the volatile access occurs in order with respect to cli() or
124 +sei(); so the "net effect" required by the standard is achieved as
125 +intended, it is "only" the timing which is off. However, for most of
126 +embedded applications, timing is an important, sometimes critical
127 +factor.
128 +
129 +\sa https://www.mikrocontroller.net/topic/65923
130 +
131 +Unfortunately, at the moment, in avr-gcc (nor in the C standard),
132 +there is no mechanism to enforce complete match of written and
133 +executed code ordering - except maybe of switching the optimization
134 +completely off (-O0), or writing all the critical code in assembly.
135 +
136 +To sum it up:
137 +
138 +\li memory barriers ensure proper ordering of volatile accesses
139 +\li memory barriers don't ensure statements with no volatile accesses to be reordered across the barrier
140 +
141 +*/
This page took 0.106277 seconds and 4 git commands to generate.