Sync to LLVM 18.1.3 #14

dpaoliello · 2024-04-11T19:59:25Z

Syncs to LLVM 18.1.3:

Adds support for COFF.
Adds support for Arm64EC.
Adds support for AIX symbol tables.

Other changes:

Move LLVM C++ files into reference directory, to make it clear that they aren't built.
Moved all object parsing into a new object_reader module, and replaced the single get_symbol callback with an ObjectReader struct that has all the callbacks we require for reading object files.
Bumped object to latest.
Synced round-trip tests with object.
Requires beta Rust toolchain (1.78) since this has LLVM 18.
Got clippy clean and added clippy to CI.

Fixes #9

src/lib.rs

src/archive_writer.rs

bjorn3 · 2024-04-12T17:32:15Z

src/archive_writer.rs

+
+struct SymbolicFile<'a> {
+    data: &'a [u8],
+    parsed: File<'a>,


ar_archive_writer intentionally doesn't parse object files. When used as part of rustc symbols we need to get symbols from LLVM bitcode files, which only LLVM itself can. This is why the get_symbols field is used for getting symbols instead. Rustc overrides it with a function which calls into LLVM to get all symbols.

I should be able to avoid parsing in most cases, but I'm not sure what to do about AIX symbol tables: we need to know if the obj is 32 or 64 bit to place the symbols in the correct table, and we need to get the alignment. Should I add new callbacks for both of those?

Ok, I've moved everything that was using the object crate into a new object_reader module and replaced the get_symbol function pointer with an ObjectReader struct that has function pointers for each of the functions that need to parse object files.

src/archive_writer.rs

bjorn3 · 2024-04-12T17:56:32Z

src/archive_writer.rs

+    match &obj.parsed {
+        File::Coff(coff_obj) => {
+            coff_obj.architecture() != Architecture::Aarch64
+                || coff_obj.sub_architecture() == Some(SubArchitecture::Arm64EC)


Does this cover import files?

And it doesn't handle llvm bitcode.

Added in support for import files, but not for bitcode - we can either try to add this later, or have the LLVM backend provide an implementation that does check for bitcode.

Support for reading bitcode is currently done by the LLVM backend of rustc by overriding get_symbols (now object_reader).

src/archive_writer.rs

bjorn3 · 2024-04-12T18:17:05Z

I'm done with this review round.

bjorn3 · 2024-04-16T13:03:26Z

Thanks a lot for this!

bjorn3 · 2024-04-16T13:04:41Z

Anything else to do before I publish it to crates.io?

dpaoliello · 2024-04-16T15:18:34Z

Anything else to do before I publish it to crates.io?

Nope, I think this is good to go!

bjorn3 · 2024-04-16T15:24:04Z

Done. The rustc update to 0.2.0 got approved earlier today.

Move reference files into their own directory

1185cf6

dpaoliello force-pushed the sync branch from fd9a13a to d50a6d3 Compare April 11, 2024 20:05