Porting PathScale's EKOPath Compiler Suite

[Tak Pui Lou]

On Nov 25, 2011, at 12:23 PM, C. Bergström wrote:

> On 11/25/11 04:38 PM, Tak Pui Lou wrote:
>> Hello,
>> 
>> I have tested the port from http://people.freebsd.org/~jkim/path64-devel-20111117.tar.bz2 and http://people.freebsd.org/~jkim/path64-20111115.tar.xz but the compiler failed in the following tests:
>> 
>> 3/6 Test #3: regression_tests .................***Failed    0.81 sec
>>     Start 4: hello_c
>> 4/6 Test #4: hello_c ..........................   Passed    0.14 sec
>>     Start 5: hello_cpp
>> 5/6 Test #5: hello_cpp ........................   Passed    0.67 sec
>>     Start 6: path64_bootstrap_test
>> 6/6 Test #6: path64_bootstrap_test ............***Failed   42.28 sec
>> 
>> 67% tests passed, 2 tests failed out of 6
>> 
>> Total Test time (real) =  44.74 sec
>> 
>> The following tests FAILED:
>>           3 - regression_tests (Failed)
>>           6 - path64_bootstrap_test (Failed)
>> Errors while running CTest
>> 
>> Are these known errors for that build?
> Normally I'd bug you about using vanilla upstream, but in this case I think JK's branch is in better shape.  (Apologies about not merging it yet, but we have a QA project we'll be testing it with and open sourcing soon - compiler agnostic fwiw)
> 
I did search on the Internet to check if the upstream has got the patches merged or not. But, I did not find too much information about this. So, I tried JK's branch instead. When you feel that I should try the source on github, please let me know.
> Specifically about your question - It's probably unexpected and I'm curious what processor and version of FBSD this is.
The kernel is compiled from 9.0-RC2 (releng/9.0 r227910) with gcc 4.2 that comes with the OS. I cannot give you the 'uname -a' output now because I have just compiled and installed a kernel with clang but I remembered it was updated two days ago before I upgraded from stable/8 to releng/9.0. The CPU is an AMD Athlon II 270u x2 running at 2 GHz.
>> 
>> I also tested it on a fortran code. Here is the runtime result:
>> 
>> 0.923u /usr/local/path64/bin/pathf95 -O3 -LANG:copyinout=ON:recursive=ON -OPT:goto=ON
>> 1.283u gfortran46 -O3
>> 
>> I actually compiled gfortran with CLooG-PPL but the optimization flags from GRAPHITE does not change the run time of this code.
> Am I reading the result correctly that we're faster?  You may also want to add/test -ipa to your flags..
> 
Yes, this code compiled from pathf95 runs faster than that compiled from gfortran46.  It may be more interesting to mention that I also have OpenIndiana 151a installed on the same computer and tested the code with Solaris Studio 12.2. The runtime for the same code compiled with Solaris Studio 12.2 is ~1.0xx u. On OpenIndiana, I have only tested the optimization flags that do not require SUNWprivate_1.5 version of libmtsk.so. All results are checked in those run.

I will try -ipa later and let you know if it makes any difference in runtime. (I think I have already tried that but let me do this again.)
> Side notes :
>    1) -ipa == LTO in gcc which I don't know if it works at all on FBSD (We have some linker work that may help this situation in the future)
>    2) I don't care what others say - Graphite isn't afaik production ready so *if* you ever do see any performance gains from it - ensure that you strongly validate before using in production setup
>    3) We've added the latest User Guide online - http://www.pathscale.com/EKOPath-User-Guide
> 
> Thanks a lot for testing!
> 
> ./C
Thank you for making PathScale Compilers open source!
---L