Why is the Linux compilers are wrong?

[just a guest]

Why not recommend to make anything with a flag of a max performance? I just can't understand - how the compiler can compile a unworking program? After all, this is not the correct compiler, right? So he miscalculated somewhere, perhaps the address is not correctly written, it's for programs fatal.
Or here's another: I has fumbled in net about building a system with the ICC and found phrases like "kernel can not even try to make with it" and "It is going to compile archiver, and archiver even works, but when you try to open the archive says that he is corrupt. Although the archive is completely good ...". In my opinion in the source code is pretty clear describes - what, where and why. So how then possible to compile archiver, that it will not open archives?

[aCOSwt]

Well, Pardon me but I had troubles understanding your post.
Do you mean that a compiler should refuse to build a program with features incompatible with the application run-time environment, with functions that might bug at run-time ?

[Goverp]

[aCOSwt]

Thanks to Goverp's post, I understand better your post !

About the question why gentoo's guide does not recommend -O3, in addition with what Goverp rightly points out, one should mention that -O3 does not necessarily improve performances :

Quickly said, -O3 = -O2 + finline-functions + some others

1/ Many of the specific -O3 optimizations lead to ignoring developer's explicit will.

finline-functions for instance, tells the compiler to ignore specific declarations in the code and decide by itself if some function will or will not be inlined.
As it is impossible for the compiler to actually benchmark the result... the compiler makes its decision from some so-called heuristical analysis which... is not always right and might lead to unexpected results !
Activating such an optimization, you suppose that the compiler will be wiser than the dev ! Honestly ! If it happens it is not, whose fault is it ?

Well, as a consequence of this, you can understand that dev's would refuse supporting builds ignoring what they specifically took time and care to code.

2/ Many of the specific -O3 optimizations lead to results highly depending on the arch, the amount of cache...
If you decide to duplicate the code of a loop (what -funswitch-loop actually does) you will necessarily need more cache and can understand that the result will depend on the speed of your memory transfers.
-O3 specific optimizations, contrarily to -O2, are actually trade-offs !

As a conclusion, I would understand that you put this on the compiler's fault but, the only bug imho is in its documentation :

[just a guest]

Yeah, thanks for understanding, seems me is only one, who thinking that I've no any syntax errors Sorry for bad English.
I'm grateful for all answers, and will be glad, if anyone decides to add one more about it. Thanks.

[just a guest]

[Yamakuzure]

It might be worth reading the documentation of something as aggressive as O3 before posting what is more imagination than fact. However, the compiler uses the "inline" keyword of the programmer, and then decides whether it is really inlined by complexity of the function. In short: You do not want overly complex functions to be inlined. But as a programmer you can force inlining with a special function attribute. Basically it is a bad idea, because the compiler actually is smarter than you. Or in other words: If you try to outsmart your compiler, it will get it's revenge very soon and very hard.

But here we go with the documentation:

[Jaglover]

[aCOSwt]

[Goverp]

[Goverp]