allow TASK_RUNNING and other #defines

[brendangregg]

sched.bt:

#include <linux/sched.h>

kprobe:do_nanosleep
{
	printf("%d\n", TASK_RUNNING);
}

# bpftrace sched.bt 
5.29: syntax error, unexpected ), expecting (

Looks like it's treating TASK_RUNNING as a bpftrace call, and so is expecting it to be followed by a '('.

I can workaround it by just including the TASK_RUNNING value (0x0) for now.

[brendangregg]

keep running into this

[mmarchini]

Playing with libclang I was able to get the #define names, but not their expanded values :/

diff --git a/src/clang_parser.cpp b/src/clang_parser.cpp
index b1db8ff..f0a1f6e 100644
--- a/src/clang_parser.cpp
+++ b/src/clang_parser.cpp
@@ -245,7 +245,7 @@ void ClangParser::parse(ast::Program *program, StructMap &structs)
       "definitions.h",
       &args[0], args.size(),
       unsaved_files, sizeof(unsaved_files)/sizeof(CXUnsavedFile),
-      CXTranslationUnit_None,
+      CXTranslationUnit_DetailedPreprocessingRecord,
       &translation_unit);
   if (error)
   {
@@ -261,6 +261,11 @@ void ClangParser::parse(ast::Program *program, StructMap &structs)
       {
         auto &structs = *static_cast<StructMap*>(client_data);
 
+        if (clang_getCursorKind(c) == CXCursor_MacroDefinition)
+        {
+          std::cout << get_clang_string(clang_getCursorSpelling(c)) << std::endl;
+        }
+
         if (clang_getCursorKind(parent) != CXCursor_StructDecl &&
             clang_getCursorKind(parent) != CXCursor_UnionDecl)
           return CXChildVisit_Recurse;

I found others on the internet asking how to do this (here and here), but the workarounds are very hacky.

[mmarchini]

From what I'm reading, the Preprocessor information might be accessible through clang's C++ API.

[mmarchini]

Ok, I have a possible implementation based on this code. Basically I implemented a "preprocessor" in bpftrace: first we run our Bison parser stage to collect preprocessor stuff, structs, enums, etc., then we run our clang parser stage, and then we run our parser stage again, replacing all occurrences of macros with their literal values. This will work for multi-level macro definitions, because Bison will keep parsing and replacing until there's no macro left in its cursor. The downside (maybe it is not a downside) is that macros defining things that bpftrace can't parse will throw.

Will be opening a draft PR to discuss this further soon.