tests: Port `split-debuginfo` to rmake.rs #135572

jieyouxu · 2025-01-16T08:30:55Z

This PR supersedes #128754 and is co-authored with @Oneirical.

Known limitations

In general, like the Makefile version, this test in its present form is also somewhat funny because for the most part it merely checks for existence/absence of output artifacts but makes no attempt to actually check if the debuginfo is at all usable.

Changes

This PR ports tests/run-make/split-debuginfo to rmake.rs. This is an initial port, and certainly could be cleaned up and/or enhanced.

The original Makefile version had several functional problems. I fixed some of them, but also left some existing issues as-is.

The linux/non-linux final branch had a conditional interpolation of UNSTABLE_OPTIONS := -Zunstable-options. However, one of the use sites was -C $(UNSTABLE_OPTIONS) split-debuginfo. This indicates to me that this run-make test is not run in CI under a non-linux + non-windows + non-darwin environment, because that would've failed as this would expand to -C -Zunstable-options split-debuginfo. I fixed this in the rmake.rs version, but I'm not sure if this distinction is worth keeping at all if it's not tested in CI.
There are several comments that were discovered to be wrong. I tried to fix them in the rmake.rs version as well.
The check for path remapping / lack of path remapping through
```
objdump -Wi $(TMPDIR)/foo | grep DW_AT_GNU_dwo_name | (! grep $(TMPDIR)) || exit 1
```
is incorrect, because that looks at the single line of that contains DW_AT_GNU_dwo_name. This is unfortunately wrong because empirical evidence shows that with objdump¹, the check actually needs to look at the attribute value of DW_AT_comp_dir on the previous line not DW_AT_GNU_dwo_name². Example output of objdump:
```
<10>   DW_AT_comp_dir    : (indirect string, offset: 0xafb48): /home/joe/repos/rust
<14>   DW_AT_GNU_dwo_name: (indirect string, offset: 0x5d1b0): foo.foo.fc848df41df7a00d-cgu.0.rcgu.dwo
```
In the rmake.rs version I used a 2-line sliding window to check for DW_AT_comp_dir and DW_AT_GNU_dwo_name, but to look at DW_AT_comp_dir specifically.
I included a bunch of FIXMEs and ENHANCEMENTs I noticed regarding the test because I didn't want to fix them in this initial port³.
The Makefile version didn't test anything on Windows (both windows-msvc and windows-gnu). I added some very basic and very sparse checks for windows-msvc, but I am not willing to spend the effort to expand test coverage to windows-gnu in this initial port.
This run-make test is way too big. But I didn't want to expend the effort of breaking this up in this initial port.

Review advice

I'm sorry for how long the rmake.rs test ended up, but a lot of it is comments and just vertical space due to formatting. If there's any ways to make this test less long / convoluted, advice would be appreciated.
This PR intentionally introduces several intermediate commits for the Makefile, mostly to illustrate the problems I discovered when looking at the original Makefile version. This is intended to highlight the existing problems in the Makefile version for the reviewer⁴.
- There are several intentional non-functional commits:
  1. Reindent the Makefile to make the platform conditional gating more obvious.
  2. Collapse nested if-else branches into an else if construct, which is not supported by GNU Make 3.80.
  3. Remove all redundant -C debuginfo=2 when -g is already specified.
This PR is best reviewed commit-by-commit.

try-job: x86_64-msvc
try-job: i686-msvc
try-job: i686-mingw
try-job: x86_64-mingw-1
try-job: x86_64-apple-1
try-job: aarch64-apple
try-job: test-various

the output format differs between objdump and llvm-objdump, but the same is true for llvm-objdump that this is looking at the wrong line. ↩
AFAICT that is a GNU DWARF attribute extension, since it isn't mentioned in DWARFv5 spec ↩
For instance, the previous path remapping check could in theory be precisely inspected by inspecting .debug_info section to look for attribute value of DW_AT_comp_dir. But that involves resolving the value of the indirect string, which means you have to: (1) look for offset into string offset table and (2) use that offset to find the string itself in the string table. The split part of "split-debuginfo" makes this murky for me, so I wasn't able to replace llvm-objdump textual output substring matches with more precise object + gimli inspections. ↩
I intend to squash these intermediate commits away after the reviewer concludes that the current form of the rmake.rs test is acceptable for merge. Before then, I'll keep them to help with review. ↩

jieyouxu · 2025-01-16T08:33:22Z