clang plugin: modify AST during compilation

Question

I'm developing a clang plugin for externalising log-strings (in order to reduce the binary size for an embedded project), which needs to modify the AST (so after preprocessor, before compilation).

Currently, I'm successfully able to match the nodes that I want to replace (using RecursiveASTVisitor<T>), and correctly extract the data which is needed in order to build the replacement.

In the VisitStmt method of my visitor, I'm iterating over the child nodes, and when finding a match, then I'm attempting to replace the matched expression nodes (CallExpr, returning int) with a new node (IntegerLiteral) by assigning via the iterator:

*it = clang::IntegerLiteral::Create(context, { 32, value, false }, context.IntTy, old->getLocStart()).

However when I try to compile an example program using my plugin, clang returns a success status code but no object file is produced:

$ clang -std=c11 -Xclang -load -Xclang /tmp/plugin.so -Xclang -plugin -Xclang string-externalise -Xclang -plugin-arg-string-externalise -Xclang -lut-file=lut.txt -c -o example.o example.c

$ echo $?
0

$ stat example.o
stat: cannot stat 'example.o': No such file or directory

If I try to dump the AST in order to verify that I've modified it correctly, then clang also gives no output. But if I disable my plugin (remove all arguments to clang relating to it) then I can dump the (original, unmodified) AST.

The specific AST modifications I'm trying to make are replacing calls to a placeholder function (returning int) with integer literals (value determined by the plugin).

For example, replacing this CallExpr subtree with an IntegerLiteral:

| |-CallExpr 0x556d460b3ab8 <line:7:42, col:78> 'int'
| | |-ImplicitCastExpr 0x556d460b3aa0 <col:42> 'int (*)(const char *, int)' <FunctionToPointerDecay>
| | | `-DeclRefExpr 0x556d460b3a20 <col:42> 'int (const char *, int)' Function 0x556d460b22d8 '__externalise_location' 'int (const char *, int)'
| | |-ImplicitCastExpr 0x556d460b3b08 <<scratch space>:26:1> 'const char *' <BitCast>
| | | `-ImplicitCastExpr 0x556d460b3af0 <col:1> 'char *' <ArrayToPointerDecay>
| | |   `-StringLiteral 0x556d460b3a48 <col:1> 'char [10]' lvalue "example.c"
| | `-IntegerLiteral 0x556d460b3a80 <line:27:1> 'int' 22

Or replacing this CallExpr subtree with an IntegerLiteral:

| |-CallExpr 0x556d460b3b98 <example.c:6:26, col:46> 'int'
| | |-ImplicitCastExpr 0x556d460b3b80 <col:26> 'int (*)(const char *)' <FunctionToPointerDecay>
| | | `-DeclRefExpr 0x556d460b3b20 <col:26> 'int (const char *)' Function 0x556d460b20c0 '__externalise' 'int (const char *)'
| | `-ImplicitCastExpr 0x556d460b3be0 <line:22:6> 'const char *' <BitCast>
| |   `-ImplicitCastExpr 0x556d460b3bc8 <col:6> 'char *' <ArrayToPointerDecay>
| |     `-StringLiteral 0x556d460b3b48 <col:6> 'char [10]' lvalue "hello %s\n"

Answer 1

While it's been years and I don't have access to the original codebase any more, I was looking into this for someone else recently.

Based on what I could remember, plus the Clang docs, and another SO answer that I'd upvoted years ago, the solution was basically:

• to use -add-plugin instead of -plugin (original solution from 2018), or
• to expose getActionType as described in the linked Clang docs.