I’m experiencing a weird problem with llvm 3.7(rc2/rc3) that did not occur in llvm 3.6.2
I created a bug for it: https://llvm.org/bugs/show_bug.cgi?id=24521
I’m building a app where multiple code gen can happen in parallel. The documentation state that I need to use separate context. Each thread has it’s own context.
When code generating a constant number I use the *InContext() calls to create the types. The assert fails since the optimizer replaces some numbers with a global type. The values are equal but the types are not. I did patch some calls to compare on getTypeID() but got another assert a couple of days later in another source file.
I needed 3.7 because of the LLVMAddGlobalMapping C api fixes.
Could someone help me out ? Or point me in the right direction? I’m using the C api.
It sounds like the two values (or the types of the values) are in different contexts. If you are creating the Types, make sure each type is created in the correct context. If some LLVM API function is creating the type for you, then it may be creating it in the wrong context (i.e., you have found a bug) or the LLVM API function has to assume a context (i.e., there is a limitation in the API for your application). If you determine that it is a bug, you should file a bug report. If it’s a limitation, you should file an enhancement request. Sorry I can’t be more helpful than that; I haven’t used LLVM 3.7 yet, and I haven’t used multiple contexts within a multi-threaded application. These are educated guesses on my part. Regards, John Criswell
Thanks for your reply. I did verify that all my calls are using the correct *InContext api calls.
I discovered that the types were in different context by testing with:
auto t1 = LLVMInt1TypeInContext(context);
auto t2 = LLVMInt1Type();
The substituted value type in GVN.cpp equals to t2 instead of t1. So It somehow is using a global context instead of my own during the GVN pass. So I think it's a bug. It could be C api related.
I willing to spend time on it and patch it if someone can point me to the part where this might happen. I'm quite new to LLVM and unfamiliar with the source tree.
I don’t know where the C bindings are located in the source tree. However, if you think the GVN pass is causing the issue, its source code is in the lib/Transforms/Scalar/GVN.cpp file. Regards, John Criswell