http://www.agner.org/optimize/#manuals
阅读笔记Optimizing software in C++
7. The efficiency of different C++ constructs
栈的速度快是因为，总是反复访问同一段地址，如果没有大的数组，肯定实在L1 cahce中.
全局静态区，global,static变量，float constants, string constants, array initializer lists,switch statement jump tables, and virtual function tables.
unsigned 和signed:除以常数，unsigned快，%同样。转换到float,signed快些。
指针加整数偏移其实是加上整数乘以对象的size:
// Example
struct abc {int a; int b; int c;};
abc * p; int i;
p = p + i; ->p += i*sizeof(abc);
所以对象的size最好是2^n。而两个指针减时，还需要除法效率更低，除非对象size是2^n.
x = *(p++) is more efficient than x = *(++p),后者需要等待p自增后才能读取x.
Conversion of a signed integer to a float or double takes 4 - 16 clock cycles。Conversion of an unsigned integer takes longer time. It is faster to first convert the unsigned integer to a signed integer if there is no risk of overflow:
// Example
unsigned int u; double d;
d = (double)(signed int)u; // Faster, but risk of overflow
Conversion of a floating point number to an integer takes a very long time unless the SSE2 or later instruction set is enabled. Typically, the conversion takes 50 - 100 clock cycles. The reason is that the C/C++ standard specifies truncation so the floating point rounding mode has to be changed to truncation and back again. 所以避免浮点到整型的转换，必要时：循环内使用浮点中间结果，算完再只进行一次转换；使用rounding函数转换而不是直接截断。
分支预测，branch misprediction代价很大，需要12-25个cycles恢复，而预测正常只需要0-2个cycles.
switch case语句，case 的value越接近(递增)效率越高，因为可以处理为跳转表。values越分散效率越低，因为要处理为branch tree。
函数跳转返回takes 4+ cycles.
8 Optimizations in the compiler
理解编译器能做的优化
9 Optimizing memory access
1.cache 结构
例如,8kb的cache,line size = 64,8*1024/64 = 128，128条lines 组织成32 sets x 4 ways . 任一内存地址只能放到某一个set,计算方法为：
(set) = (memory address) / (line size) % (number of sets). 地址10000的内容会放到28组中的一条cache line 中，(set) = (10000 / 64) % 32 = 28.
当读取的两个地址映射到同一组set时，可能就会产生cache miss.
4. Variables that are used together should be stored together
12 Using vector operations
讲述simd指令,使用支持automatic vectorization的编译器，使用vector class library (VCL) make code readable
13.CPU dispatching
making the most critical parts of the code in multiple versions for different CPUs.
应根据cpu支持的指令集来适配，而不是brand, family and model number
14 Specific optimization topics
1.lookup tables
静态变量可能分散在不同的内存地址，不利于cache(破坏了局部访问).可以查找表从静态区拷贝到栈上.
如下，FactorialTableA存在静态区，而FactorialTable会在调用函数CriticalInnerFunction时，把数据从静态区拷贝到栈上。
// Example 14.1c
const int FactorialTableA[13] = {1, 1, 2, 6, 24, 120, 720,
5040, 40320, 362880, 3628800, 39916800, 479001600};
void CriticalInnerFunction () {
// Table of factorials:
const int FactorialTable[13] = {1, 1, 2, 6, 24, 120, 720,
5040, 40320, 362880, 3628800, 39916800, 479001600};
...
}
2 Bounds checking
// Example 14.4a
const int size = 16; int i;
float list[size];
...
if (i < 0 || i >= size) {
cout << "Error: Index out of range";
}
else {}
替换为只有一次比较：
// Example 14.4b
if ((unsigned int)i >= (unsigned int)size) {
cout << "Error: Index out of range";
}
else {}
// Example 14.5a
const int min = 100, max = 110; int i;
...
if (i >= min && i <= max) { ...
can be changed to:
// Example 14.5b
if ((unsigned int)(i - min) <= (unsigned int)(max - min)) { ..
3 Use bitwise operators for checking multiple values at once
使用位运算简化条件判断，比较操作.
// Example 14.7a. Testing multiple conditions
enum Weekdays {Sunday, Monday, Tuesday, Wednesday, Thursday, Friday, Saturday
};
Weekdays Day;
if (Day == Tuesday || Day == Wednesday || Day == Friday) {
DoThisThreeTimesAWeek();
}
会有三次比较，可以使用位运算替代：
// Example 14.7b. Testing multiple conditions using &
enum Weekdays {
Sunday = 1, Monday = 2, Tuesday = 4, Wednesday = 8,
Thursday = 0x10, Friday = 0x20, Saturday = 0x40
};
Weekdays Day;
if (Day & (Tuesday | Wednesday | Friday)) {
DoThisThreeTimesAWeek();
}
Boolean operators &&, ||, ! 与 bitwise operators &, |, ~ 的区别：前者结果只有两种true (1) or false (0),而位运算更快因为不产生分支。
4.整型乘法(3 - 10 clock cycles),编译器会用右移和加代替乘法
5.整型除法(27 - 80 clock cycles)
除以常数更快是因为编译器会优化a / b 为a * (2^n / b) >> n. %与/规律相同。
// Example 14.10
int a, b, c;
a = b / c; // This is slow
a = b / 10; // Division by a constant is faster
a = (unsigned int)b / 10; // Still faster if unsigned
a = b / 16; // Faster if divisor is a power of 2
a = (unsigned int)b / 16; // Still faster if unsigned
6.浮点除法(20 - 45 clock cycles),使用乘以倒数替代
7.Don't mix float and double
float与double运算同样快。但是互相转换很耗时。C/C++ 默认浮点常量是double精度的。
8.浮点转整型
C++默认浮点到整型转换都是向0截断(truncation)。而truncation(40 clock cycles)耗时是rounding的3倍，除非使用sse2指令集.
9.整型转浮点
比浮点转整型快(5-20 clock cycles),有符号比无符号的整型快。所以不溢出时先转换成有符号再转换成浮点更快：
// Example 14.22b
unsigned int u; double d;
d = (double)(signed int)u;