我认为汇编程序更快的一般情况是，当一个聪明的汇编程序员看到编译器的输出并说“这是性能的关键路径，我可以写这个更有效”，然后那个人调整汇编程序或从头重写它。

在运行时创建机器代码怎么样?

我的兄弟曾经(大约在2000年)通过在运行时生成代码实现了一个非常快速的实时光线跟踪器。我不记得细节了，但有一些主模块是通过对象循环的，然后它准备和执行一些特定于每个对象的机器代码。

然而，随着时间的推移，这种方法被新的图形硬件淘汰，变得毫无用处。

今天，我认为大数据(数百万条记录)上的一些操作，如数据透视表、钻孔、实时计算等，都可以用这种方法进行优化。问题是:这样的努力值得吗?

GCC已经成为广泛使用的编译器。它的优化通常不是很好。比编写汇编程序的普通程序员好得多，但就实际性能而言，并没有那么好。有些编译器产生的代码简直令人难以置信。所以一般来说，有很多地方你可以进入编译器的输出，调整汇编器的性能，和/或简单地从头重写例程。

在Amiga上，CPU和图形/音频芯片会为了访问特定区域的RAM(具体来说是前2MB的RAM)而争斗。因此，当你只有2MB RAM(或更少)时，显示复杂的图形加上播放声音会杀死CPU的性能。

在汇编程序中，你可以巧妙地交错你的代码，使CPU只在图形/音频芯片内部繁忙时(即当总线空闲时)才尝试访问RAM。因此，通过重新排序指令，巧妙地使用CPU缓存，总线定时，你可以实现一些使用任何高级语言都不可能实现的效果，因为你必须为每个命令定时，甚至在这里或那里插入nop，以使不同的芯片不受彼此的雷达影响。

这也是为什么CPU的NOP (No Operation -什么都不做)指令实际上可以让你的整个应用程序运行得更快的另一个原因。

当然，这种技术取决于特定的硬件设置。这就是为什么许多Amiga游戏无法适应更快的cpu的主要原因:指令的计时错误。

尽管C语言“接近”于对8位、16位、32位和64位数据的低级操作，但仍有一些C语言不支持的数学操作通常可以在某些汇编指令集中优雅地执行:

Fixed-point multiplication: The product of two 16-bit numbers is a 32-bit number. But the rules in C says that the product of two 16-bit numbers is a 16-bit number, and the product of two 32-bit numbers is a 32-bit number -- the bottom half in both cases. If you want the top half of a 16x16 multiply or a 32x32 multiply, you have to play games with the compiler. The general method is to cast to a larger-than-necessary bit width, multiply, shift down, and cast back: int16_t x, y; // int16_t is a typedef for "short" // set x and y to something int16_t prod = (int16_t)(((int32_t)x*y)>>16);` In this case the compiler may be smart enough to know that you're really just trying to get the top half of a 16x16 multiply and do the right thing with the machine's native 16x16multiply. Or it may be stupid and require a library call to do the 32x32 multiply that's way overkill because you only need 16 bits of the product -- but the C standard doesn't give you any way to express yourself. Certain bitshifting operations (rotation/carries): // 256-bit array shifted right in its entirety: uint8_t x[32]; for (int i = 32; --i > 0; ) { x[i] = (x[i] >> 1) | (x[i-1] << 7); } x[0] >>= 1; This is not too inelegant in C, but again, unless the compiler is smart enough to realize what you are doing, it's going to do a lot of "unnecessary" work. Many assembly instruction sets allow you to rotate or shift left/right with the result in the carry register, so you could accomplish the above in 34 instructions: load a pointer to the beginning of the array, clear the carry, and perform 32 8-bit right-shifts, using auto-increment on the pointer. For another example, there are linear feedback shift registers (LFSR) that are elegantly performed in assembly: Take a chunk of N bits (8, 16, 32, 64, 128, etc), shift the whole thing right by 1 (see above algorithm), then if the resulting carry is 1 then you XOR in a bit pattern that represents the polynomial.

尽管如此，除非有严重的性能限制，否则我不会求助于这些技术。正如其他人所说，汇编代码比C代码更难记录/调试/测试/维护:性能的提高伴随着一些严重的代价。

编辑:3。溢出检测在汇编中是可能的(在C中不能真正做到)，这使得一些算法更容易。

以下是我个人经历中的几个例子:

Access to instructions that are not accessible from C. For instance, many architectures (like x86-64, IA-64, DEC Alpha, and 64-bit MIPS or PowerPC) support a 64 bit by 64 bit multiplication producing a 128 bit result. GCC recently added an extension providing access to such instructions, but before that assembly was required. And access to this instruction can make a huge difference on 64-bit CPUs when implementing something like RSA - sometimes as much as a factor of 4 improvement in performance. Access to CPU-specific flags. The one that has bitten me a lot is the carry flag; when doing a multiple-precision addition, if you don't have access to the CPU carry bit one must instead compare the result to see if it overflowed, which takes 3-5 more instructions per limb; and worse, which are quite serial in terms of data accesses, which kills performance on modern superscalar processors. When processing thousands of such integers in a row, being able to use addc is a huge win (there are superscalar issues with contention on the carry bit as well, but modern CPUs deal pretty well with it). SIMD. Even autovectorizing compilers can only do relatively simple cases, so if you want good SIMD performance it's unfortunately often necessary to write the code directly. Of course you can use intrinsics instead of assembly but once you're at the intrinsics level you're basically writing assembly anyway, just using the compiler as a register allocator and (nominally) instruction scheduler. (I tend to use intrinsics for SIMD simply because the compiler can generate the function prologues and whatnot for me so I can use the same code on Linux, OS X, and Windows without having to deal with ABI issues like function calling conventions, but other than that the SSE intrinsics really aren't very nice - the Altivec ones seem better though I don't have much experience with them). As examples of things a (current day) vectorizing compiler can't figure out, read about bitslicing AES or SIMD error correction - one could imagine a compiler that could analyze algorithms and generate such code, but it feels to me like such a smart compiler is at least 30 years away from existing (at best).

On the other hand, multicore machines and distributed systems have shifted many of the biggest performance wins in the other direction - get an extra 20% speedup writing your inner loops in assembly, or 300% by running them across multiple cores, or 10000% by running them across a cluster of machines. And of course high level optimizations (things like futures, memoization, etc) are often much easier to do in a higher level language like ML or Scala than C or asm, and often can provide a much bigger performance win. So, as always, there are tradeoffs to be made.