Support -Zembed-source in debuginfo

[kehrazy]

Fixes #1528

• honour the session DWARF version, enabling v5 when embed-source is requested and plumb the flag through the debug context
• attach embedded source bytes to line table FileInfo records alongside existing MD5 hashes
• stop excluding rustc’s embed-source-dwarf run-make test so CG_CLIF exercises the feature

[bjorn3]

Thanks for working on this!

[bjorn3]

Unrelated to this PR, but while trying to test this PR I noticed that with DWARFv5 anonymous sources get invalid file names. They are supposed to be called <anon>, but get called /path/to/working_directory/<anon> instead by gdb. By the way do you know of any debugger that already supports this? Even the gdb of Fedora Rawhide doesn't seem to support it.

[kehrazy]

Unrelated to this PR, but while trying to test this PR I noticed that with DWARFv5 anonymous sources get invalid file names. They are supposed to be called <anon>, but get called /path/to/working_directory/<anon> instead by gdb.

The fix might be FileName::prefer_remapped_unconditionally - internal files and all that.

How do you check the output? With what RUSTFLAGS do you compile a binary? Can you give me a repro?

[bjorn3]

I used dist/rustc-clif -g -Zdwarf-version=5 -Zembed-source foo.rs and then removing foo.rs for trying to test -Zembed-source.

echo 'fn main() {}' | dist/rustc-clif -g -Zdwarf-version=5 - reproduces the <anon> issue.

[kehrazy]

both GDB and LLDB seem to not have DWARFv5 embedded source support :/

When using

echo 'fn main() {}' | ./dist/rustc-clif -g -Zdwarf-version=5 - 2>&1
echo 'int main() { return 0; }' | gcc -x c -gdwarf-5 - -o gcc_test

echo "y" | gdb -batch -ex "file rust_out" -ex "b main" -ex "list" 2>&1 | tail -1`
echo "y" | gdb -batch -ex "file gcc_test" -ex "b main" -ex "list" 2>&1 | tail -1

(after the latest change) I get:

warning: 1      <anon>: No such file or directory
warning: 1      <stdin>: No such file or directory

which looks okay!

[kehrazy]

which looks okay!

I really hate this solution, though, and I think it should be reverted - I can't seem to find a better approach here.

[bjorn3]

This is what cg_llvm does for determining the file name: https://github.com/rust-lang/rust/blob/caccb4d0368bd918ef6668af8e13834d07040417/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs#L552-L648 It should be lifted into cg_ssa at some point, but for now what you have LGTM. Thanks!