Slow ICC...

[Apopatos]

I wanted to do some benchmarks comparing lame encoding with different versions of gcc and various CFLAGS.
The best time was with GCC-4.4.2 and CFLAGS="-march=k8 -O3 -pipe" in my 64bit system.
The time was 1m5.916s.
Then thought to try a Window$ binary via Wine. I found a binary of lame that had bin compiled with ICC-11.1.046. The results were quite amazing since this 32bit binary via emulation gave me a score of 1m2.794s!
So I decided to try ICC natively. Followed the instructions and installed ICC-11.1.056 via portage but whatever ICCFLAGS I use the times are always over 1.10s. I even patched with the AuthenticAMD patch without better results.
Now I have these flags in my /etc/make.conf:

[Akkara]

These are differences of 10% or less. This is not enough to draw conclusions from because many things can cause a 10% speed difference.

What could be happening:

• The code is placed differently in memory, leading to different caching effects.
  
• The 64-bit code has pointers that are twice as long as the 32-bit code. This results in more memory traffic and higher cache pressure on the 64-bit versions.
  
• While 64-bit arithmetic can be expected to help, the code might not be written in a way that allows the compiler to take advantage of that, at least not enough to offset the larger pointers.

...just a few guesses.

[Apopatos]

But it says that lame works pretty well with ICC. Also always lame64 is faster than lame32, despite the CFLAGS

[Apopatos]

Here it says I must have the libelf library for AMD patch to work http://server01.iiiii.info/patch-AuthenticAMD.html while I have elfutils. Could it be the problem? I suppose not, since both do the same job, right?