pgoodman
commented
1 year ago <pasta-input> is a synthetic input file that clang's parser operates on. PASTA first runs an instrumented preprocessor on the input, and in the process of this, forms <pasta-input>, which is just a very large std::string in memory. Each unique token is rendered onto a different line of this input file, so that a source location's line number acts as a kind of side-channel for a token index. Previously, I had attempted to patch Clang to embed more information into SourceLocations, but this became problematic as the underlying integral representation of SourceLocations is often stored directly in AST nodes. In retrospect, I regret the choice to use line numbers as a side-channel, as SourceLocations are basically byte offsets, and so I could have just used the byte offset itself with a list of "end of token bytes offsets", and then some binary searching. Oh well.
Anyway, probably you aren't holding it wrong. The reality is just that PASTA does some sketchy things to achieve its goals. Here's how I would proceed:
Patch the following code into lib/Compile/Run.cpp, around line 353:
auto fd = open("/tmp/source.cpp", O_TRUNC | O_CREAT | O_WRONLY, 0666);
write(fd, ast->preprocessed_code.data(), ast->preprocessed_code.size());
close(fd);Then build and run, and attach /tmp/source.cpp here. Another thing that could help is seeing the backtrace.
seelabs
I read about pasta on your blog post and decide to play with it. I am not familiar with the internals of libclang or llvm, so this issue may be user error.
I successfully built pasta and was able to run the demo programs. When I tried the
dump-astexample on a toy example it worked as expected. However, when I tried it against my real project, it hits an assert. In particular, when I run it against this project: https://github.com/XRPLF/rippled using the following command line:It hits the following assert:
I attempted to debug this. I built the debug version of the libraries and went into gdb and attempted to try to find the offending code so I could create a minimal repro, but unfortunately my unfamiliarity with llvm is an obstacle. Two things we may be able to use as staring points.
In SemaType.cpp:8299
So the offending attribute is called "access".
In ParseDeclCXX.cpp:2107
What I'm trying to do here is find the name of the file that contains the offending code. Some searching pointed me to
getPresumedLoc, but clearly<pasta-input>isn't useful to find the offending code. Perhaps you have a suggestion to find the code?Also, I should mention I was only playing with pasta. Pasta obviously works and it's very possible I'm "holding it wrong". I'm not expecting help here, but thought I'd report the issue in case it was a possible bug in pasta (and will help trying to debug it after regular work hours if it's helpful to you).