incr: cache dwarf objects in work products

[davidtwco]

Cache DWARF objects alongside object files in work products when those exist so that DWARF object files are available for thorin in packed mode in incremental scenarios.

r? @michaelwoerister

[rustbot]

Some changes occurred in compiler/rustc_codegen_cranelift

cc @bjorn3

[michaelwoerister]

Looks good, @davidtwco! Now we just need a regression test. Compiling twice with -Zsplit-dwarf-kind=split -Csplit-debuginfo=packed and no changes in between triggered bug for me.

[davidtwco]

Looks good, @davidtwco! Now we just need a regression test. Compiling twice with -Zsplit-dwarf-kind=split -Csplit-debuginfo=packed and no changes in between triggered bug for me.

I'm struggling to reproduce the issue with these instructions - using a simple case and compiling twice with those flags doesn't cause an issue for me. I've been trying with the following test case:

// Check that compiling with packed Split DWARF twice succeeds. This should confirm that DWARF
// objects are cached as work products and available to the incremental compilation for `thorin` to
// pack into a DWARF package.

// ignore-tidy-linelength
// only-x86_64-unknown-linux-gnu
// revisions:rpass1 rpass2

// [rpass1]compile-flags: -Zquery-dep-graph -Zunstable-options -Csplit-debuginfo=packed -Zsplit-dwarf-kind=split
// [rpass2]compile-flags: -Zquery-dep-graph -Zunstable-options -Csplit-debuginfo=packed -Zsplit-dwarf-kind=split

#![feature(rustc_attrs)]
// For `rpass2`, nothing has changed so everything should re-used.
#![rustc_partition_reused(module = "split_debuginfo_cached", cfg = "rpass2")]
#![rustc_partition_reused(module = "split_debuginfo_cached-another_module", cfg = "rpass2")]

mod another_module {
    pub fn foo() -> &'static str {
        "hello world"
    }
}

pub fn main() {
    println!("{}", another_module::foo());
}

Can't reproduce outside of the test suite either (without this patch obviously).

Nevermind, figured it out, just being stupid - will have a commit up shortly.

[davidtwco]

Looks good, @davidtwco! Now we just need a regression test. Compiling twice with -Zsplit-dwarf-kind=split -Csplit-debuginfo=packed and no changes in between triggered bug for me.

Okay, added a test now. I lost some time to a missing -g flag in the test I wrote, which was a little frustrating, and then it turned out that link_dwarf_object wasn't calling abort_if_errors so tests wouldn't fail when thorin had.

I also noticed that we weren't deleting all the DWARF object temporaries in some cases, so I fixed that - we weren't deleting module.dwarf_object for allocator or metadata modules.

[michaelwoerister]

Great, thanks a lot, @davidtwco!

@bors r+

[bors]

📌 Commit e106523 has been approved by michaelwoerister