Compiler related: Intel's code slower on AMD-CPUs?!

[Vacon]

Hello everyone,

since every now and then a thread deals with the question which compiler is more practical than the other, maybe here is an interesting link:
http://www.swallowtail.org/naughty-intel.html
Since I have no experience in programming / compiling I give to the audience to decide or test.
In short: Intel compilers check on which CPU the code runs and slow down on AMD
Edit: ok, nothing brand-new, but found on 7-zip's forum:
http://sourceforge.net/forum/forum.p...forum_id=45797

Best regards!

[encode]

Intel's code slower on AMD-CPUs?!
In short: Intel compilers check on which CPU the code runs and slow down on AMD

I doubt it. Intel Compiler can generate native code - without CPU detection/code selection. Furthermore, in Intel's Optimization Guide you may find which options to use with each processor or processor group, including AMD CPUs. Concluding, Intel has optimizations for Generic/AMD processors as well as it's own Intel-oriented...

[Shelwien]

That's kinda expected... Probably, the whole compiler is made to improve CPU selling by making use of their new instruction sets which other compilers don't.
Instead you should say thanks that they didn't insert some less identifiable brakes.

1. Read section 12.1 from http://agner.org/optimize/optimizing_cpp.pdf
2. I just compile my executables with /QxN without any patches,
because its better to use separate binaries for different CPUs anyway... Imho /Qax switches cause too much code caching problems.
3. Intel's math library is bloated, and anyway its a bad practice to directly use the math.h functions in internal loops.

[SvenBent]

I've seen it before. people using hacks on the compiler to get the propper code executed on AMD cpu's

[roytam1]

/me always use iccpatch for icc builds.

[Black_Fox]

I doubt it. Intel Compiler can generate native code - without CPU detection/code selection.

Well, yes, ICC _can_ do it, but usually doesn't. This article is mainly about Intel Fortran Compiler, but author adds ICC behaves completely the same way...