LLVM+Clang 3.3: clang: for the -vectorize-loops option: may only occur zero or one times!

Clang Frontend
1

Hi,

My application generates some C code which I then compile using LLVM+Clang before running it. For the compilation side of things, I have a class which one of the methods (the CompilerEngine::compileCode() method in https://raw.github.com/opencor/opencor/llvm_3.3/src/plugins/misc/Compiler/src/compilerengine.cpp) is based on the code from the clang-interpreter example (http://llvm.org/viewvc/llvm-project/cfe/trunk/examples/clang-interpreter/main.cpp?view=co). The idea is therefore to create an instance of the ‘compiler’ class and, among other things, use its compileCode() whenever needed.

All of this works perfectly fine using LLVM+Clang 3.2, but not with LLVM+Clang 3.3. At least, not out of the box. Basically, the first time I try to compile some C code, everything is fine. However, if I try to compile some other C code, then I will get the following error message:

clang: for the -vectorize-loops option: may only occur zero or one times!

FWIW, I use the -O3 option which I understand makes implicit use of the loop vectorizer, and that seems to be the problem. Indeed, if I disable the loop vectorizer (using the -fno-vectorize option), then everything is fine but I clearly don’t want to have to do that. So, right now, my solution consists of commenting out the test for the number of option occurrences in the Option::addOccurrence() method (http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Support/CommandLine.cpp?view=co), but it is clearly not satisfactory either (even though it’s good enough for my particular purpose).

So, would anyone have any idea about what I might be missing, if anything?..

Cheers, Alan.

2

Hi,

It seems to me that the CompilerEngine is not reentrant, either by design or by mistake (the latter seems more likely).

I think that a reset of the count is missing from the initialization when you call compileCode(), and thus the second time it (incorrectly) diagnose that you are trying to pass the option a second time, not realizing it is a different run.

– Matthieu

3

Hi Matthieu,

You are correct in that CompilerEngine is not re-entrant, and it’s actually by design. The idea is to create an instance of CompilerEngine, compile some code with it, and then use the compiled code. From there, once you are done with that code and don’t need it anymore, you can compile some other code. So, in the current use of CompilerEngine, there cannot be a case where a call to compileCode() is made while another call to compileCode() is still being performed. If you ever wanted to do that, then you would create two instances of CompilerEngine.

Anyway, I don’t think the problem lies with CompilerEngine. As I mentioned in my original message, everything works fine with LLVM+Clang 3.2. It seems like LLVM keeps track of command line options (through llvm::cl::RegisteredOptionList) and this is what is causing me problems. So, I am now wondering whether there is a way to ‘reset’ LLVM so that I can ‘properly’ call clang more than once?

Alan