The real problem of GCC is 64 bits code.

It has big problems with integer code and most importantly the PGO pass even in the latest GCC is very buggy, which simply is not explainable other than that some amateurs worked at it.

The "commercial compilers" really get 20% (roughly) out of pgo pass whereas GCC maybe wins 1.5% there or so, at least for Diep.

Other software that lays more stress onto 64 bits datastructure and especially C++ code, the difference gcc versus commercial compilers is even bigger. Not seldom factor 2.

All this for integer codes.

It is at its worst in integer codes at bit manipulation level where deep layered down classes get called to manipulate data.

Now my code lucky doesn't suffer from all this, as it is simple C code with a bunch of pointers.

If you look to the code produced by GCC, the evil is not in just 1 small tiny code fragment. It's not like 1 massive bug. It's a number of total idiocies and total inefficiency.

From metaviewpoint seen it never will get ok. The commercial interests are too big to not have it outperform commercial compilers and big money can in free tiny projects always outmanipulate the volunteers. Note for Diep GCC does actually DOES outperform open64 and commercial compilers based upon it, which doesn't surprise me. I have very little respect for the mipspro compiler.

It was a piece of junk at the time and still is.

Tested open64 at 16 cores barcelona hardware.

Yet the real compare of GCC is of course versus visual c++ and intel c++. There GCC really gets annihilated in a manner that's just not normal and shouldn't be the case at simple C code.

Mark Roulo (nothanks@xxx.com) on 6/23/10 wrote:
---------------------------
>anon (anon@anon.com) on 6/23/10 wrote:
>---------------------------
>>Rohit (@.) on 6/23/10 wrote:
>>---------------------------
>>>David Kanter (dkanter@realworldtech.com) on 6/23/10 wrote:
>>>---------------------------
>>>>Rohit (@.) on 6/22/10 wrote:
>>>>---------------------------
>>>>>Why did you not test it with GCC. On linux it shouldn't be >too bad either.
>>>>
>>>>Carlie didn't really recommend using GCC and frankly I expect ICC to be quite a bit better for this sort of code.
>>>>
>>>>David
>>>
>>
>>IMO Sun has done its fair share of misleading benchmarking, however this entry is relatively interesting.
>>
>>http://blogs.sun.com/BestPerf/entry/free_compiler_wins_nehalem_race
>>
>>Without autopar, and with base flags only, it handily outperforms gcc on specfp2006 on Nehalem.
>>
>>http://blogs.sun.com/jhenning/entry/sun_studio_trounces_intel_compiler
>>
>>>
>>
>>Showed it beating icc in a similar configuration, although both using peak flags in that case.
>>
>>I have a theory that on a rather large and popular (commercial) enterprise software
>>package (so primarily integer), sun cc does significantly better than gcc as well
>>(though not as large gap as these FP tests). I'm handwaving a bit based on reported
>>benchmark numbers from various sources, so my anecdote doesn't carry much weight unfortunately.
>
>I'm not trying to make excuses for GCC, but does anyone have an idea how much focus/emphasis
>GCC puts on optimizing for floating point loads versus optimizing for integer loads?
>My *guess* is that GCC is most competitive as you get closer to integer code and
>least competitive as you get closer to vectorized floating point code.
>
>-Mark Roulo