Does compiling on your computer really improve performance?

Question

Any programmer knows that when building a C/C++, the compiler can optimize the code to generate faster executables. But, it is also said that there is the compiler optimization for your processor.

This means that if I created a C program and compiled it on a desktop with an Intel Core i3, the program is optimized to run at peak performance on the Intel Core i3 but will normally run on other x86_64 processors. I don’t know if that’s true, just saying.

Based on this, I see a lot to say around that the best way to have a program running at maximum efficiency on your computer is to install it directly from the source code. So much so that some Gentoo Linux enthusiasts make use of this idea to say that Gentoo Linux is the fastest distribution on the planet (Gentoo is a distribution where everything is installed directly from source code, from kernel to most basic applications).

But I’m suspicious of all this. Is this talk that something compiled on your computer will run at peak performance on it real? And, adding: The fact that Gentoo Linux is a distribution where everything is installed directly from the source code actually makes it a system that has a remarkable performance?

Answer 1

The fact is that when you compile a code with your compiler, by default it assumes that you want to distribute the executable to others. So as much as it optimizes, the code needs to continue running on any popular processor on the market. So the compiler cannot simply use very advanced instructions like vectorization and the like.

To improve the performance in this case there are two very interesting GCC options:

• -march=cpu-type

  Sets the minimum that the compiler should assume over the CPU. If you specify one that supports AVX instructions for example, the compiler will use them freely without worrying about the code not working on previous processors. The cool thing is you can write like this: -march=native and the optimization will be based on the processor of your machine. If the code is to be executed only by you, use this flag. Not forgetting of course the other optimization options like -O3 and -fomit-frame-pointer.

• -mtune=cpu-type

  That option is similar. The compiler will generate code optimized for that particular model, but it also generates conditionals to check whether the functionality is actually present in the processor and includes alternative implementations. In practice the code gets bigger, but runs on any processor and has performance comparable to -march on the target processor.

  if (cpu suporta AVX)
      ComputeComAVX();
  else if (cpu suporta SSE2)
      ComputeComSSE2();
  else
      ComputeGenerico();
  

For GCC 4.9.0, you can find here the list of supported processors and other relevant flags, such as the -mavx to indicate that you can use AVX.

3.17.17 Intel 386 and AMD x86-64 Options

In my machine the architecture used by default is the i686 (Pentium Pro) of 1995. So without a doubt there is a significant improvement in performance, especially in code that can be vectored.