Broken code produced by Intel 11.1 compiler

[GoogleCodeExporter]

Chao Liu reported the following issue. He is using ADDA on a cluster with Intel 
Xeon processors. Using gcc, everything is fine. But when using icc with version 
1.0 and later, code execution results in zero cross sections (although 
compilation proceeds fine without warnings). Moreover, ADDA 0.79 compiles and 
works fine with icc.

I have reproduced this issue on another (similar) cluster in Novosibirsk, using 
recent (1.1b6) version of ADDA and icc 11.1. However, when I tried it on a 
desktop Linux with icc 12.0, everything was fine.

Original issue reported on code.google.com by yurkin on 11 May 2012 at 5:48

[GoogleCodeExporter]

The code is broken only for three iterative solvers: QMR, QMR2, and CSYM. The 
problem occurs only when optimization is performed by compiler, probably due to 
incorrect handling of static (inside function) variables, which carry 
information between function calls. Also built-in intrinsics can be involved.

In particular, compiling with options '-g' or '-O0' produces correct code, 
while '-g -fbuiltin' breaks the code. Also, '-O1' and larger breaks code in 
combination with any other options, even '-O1 -fno-builtin' (which is strange).

I have localized the problematic variables and added several dumb statements to 
iterative.c in r1136. This resolves the issue for icc 11.1 on cluster in 
Novosibirsk. I have even tested the suite in tests/2exec on intel vs. gnu 
compiler - and everything is fine.

This fix should be verified on other hardware/compiler versions. 

Overall, I believe that this is a compiler bug, which is troubling that it may 
affect other parts of code (in the future). Hopefully, this bug is fixed in 
later versions of the compiler (at least one test with icc 12.0 suggests so).

Original comment by yurkin on 13 May 2012 at 12:25

• Changed state: Fixed

[GoogleCodeExporter]

Chao Liu verified that the fix works for his cluster (with icc 11.1.073).

I have run a few extra tests on LISA cluster (in Amsterdam). For icc 11.0.074 
the issue is irrelevant - code compiles and works fine both before and after 
the fix. For icc 11.1.072 the issue is relevant (code is broken before the fix, 
but fine afterwards).

Original comment by yurkin on 14 May 2012 at 8:33

[GoogleCodeExporter]

There is another problem with icc 11.1 using -O3 (but not -O2). It is 
highlighted by r1309. Unfortunately, this time I couldn't find any simple fix, 
so I just recommend not to use this version of icc. 

Original comment by yurkin on 17 Feb 2014 at 11:02

• Changed state: WontFix
• Removed labels: Milestone-1.1