I can trigger 3 bugs in the OpenCL compiler for which I can supply code:
- The compiler crashes in case of infinite loop and "-cl-opt-disable". Just put a goto and it segfaults.
- The compiler from ~2011:
$ md5sum libamdocl64.so
d55c853a7b8c892576eb2355298ca59f libamdocl64.so
is 30% faster on a piece of code (when "-cl-opt-disable" is given) than the new compiler (with or without the "-cl-opt-disable"):
$ md5sum libamdocl64.so
7d2a234f375b693d9d2e71bf93f28edf libamdocl64.so
- The OpenCL compiler miscompiles a piece of code. I know people at this point stop reading, but I know what I am talking about. I have wrestled with a bug in gcc before (and had it reported&confirmed&fixed). It is a miscompiled, it's easy to check and verify.
I have separate codes that triggers every one of these. Unfortunately, it's not public code, but I would be happy to share them with any AMD developer. I have been using the old compiler from 2011 for some while, but it crashes when compiling for 7970's, so I am forced to use 7d2a234f375 (13.4), but that miscompiles, so I am in a deadlock. I would really appreciate if the miscompilation issue would be fixed. I think it's very little to ask, to be honest.