Emit deprecation warnings without switching tasks.

[maleadt]

Base.depwarn logs through the active logger, and the I/O involved can switch tasks. That is illegal while holding the typeinf lock, which our @locked compilation entry points do on Julia 1.10/1.11, and caused hangs (#830).

Replace it with safe_depwarn, which writes the warning synchronously to Core.stderr like @safe_warn does. Caller attribution, warning once per call site, and throwing under --depwarn=error still behave like Base.depwarn; on 1.12+, where compilation doesn't hold the typeinf lock, it just forwards there.

Closes #831

[codecov]

Codecov Report

❌ Patch coverage is 67.74194% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 79.36%. Comparing base (a0d1302) to head (68043ee).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/driver.jl	0.00%	4 Missing ⚠️
src/deprecated.jl	0.00%	3 Missing ⚠️
src/irgen.jl	0.00%	2 Missing ⚠️
src/utils.jl	95.45%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #831      +/-   ##
==========================================
+ Coverage   79.28%   79.36%   +0.07%     
==========================================
  Files          25       25              
  Lines        4562     4584      +22     
==========================================
+ Hits         3617     3638      +21     
- Misses        945      946       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[simeonschaub]

Thanks, that's a much better solution! Would you mind also bumping the GPUCompiler version as well as the CUDA compat, so we can fix CoolPDLP CI?

[maleadt]

as well as the CUDA compat

Releasing CUDA.jl is quite a bit of effort, and won't happen right away. Or what do you mean?

[simeonschaub]

I mean that the currently released CUDA version does not seem to be compatible with GPUCompiler 1.18. If I install the latest release of CUDA.jl inside an environment and then try to add GPUCompiler#main, I get:

    Updating git-repo `https://github.com/JuliaGPU/GPUCompiler.jl.git`
   Resolving package versions...
ERROR: Unsatisfiable requirements detected for package CUDACore [bd0ed864]:
 CUDACore [bd0ed864] log:
 ├─possible versions are: 6.0.1 - 6.1.1 or uninstalled
 ├─restricted by compatibility requirements with GPUCompiler [61eb1bfa] to versions: uninstalled
 │ └─GPUCompiler [61eb1bfa] log:
 │   ├─possible versions are: 1.18.0 or uninstalled
 │   └─GPUCompiler [61eb1bfa] is fixed to version 1.18.0
 └─required (without additional version restrictions) by CUDA [052768ef]
   └─CUDA [052768ef] log:
     ├─possible versions are: 0.1.0 - 6.1.0 or uninstalled
     ├─restricted to versions * by project [0dc50556], leaving only versions: 0.1.0 - 6.1.0
     │ └─project [0dc50556] log:
     │   ├─possible versions are: 0.0.0 or uninstalled
     │   └─project [0dc50556] is fixed to version 0.0.0
     ├─restricted by julia compatibility requirements to versions: [2.3.0, 2.5.0 - 6.1.0] or uninstalled, leaving only versions: [2.3.0, 2.5.0 - 6.1.0]
     ├─restricted by compatibility requirements with LLVM [929cbde3] to versions: 5.5.0 - 6.1.0 or uninstalled, leaving only versions: 5.5.0 - 6.1.0
     │ └─LLVM [929cbde3] log:
     │   ├─possible versions are: 0.9.0 - 9.8.2 or uninstalled
     │   └─restricted to versions 9.8.1 - 9 by GPUCompiler [61eb1bfa], leaving only versions: 9.8.1 - 9.8.2
     │     └─GPUCompiler [61eb1bfa] log: see above
     └─restricted by compatibility requirements with GPUCompiler [61eb1bfa] to versions: 6.0.0 - 6.1.0 or uninstalled, leaving only versions: 6.0.0 - 6.1.0
       └─GPUCompiler [61eb1bfa] log: see above

Is there a reason CUDA.jl declares such strict compat bounds for GPUCompiler?

[maleadt]

Yes, GPUCompiler switched to llc from LLVM 22, while older versions were using the in-process one. This prompted retroactively changing compat bounds: JuliaRegistries/General@67cd467
Normally we don't have to do this, and play by semver, but this was an exception (not worth tagging a breaking release for).

[simeonschaub]

Ah, I see! Any ETA when CUDA will be compatible with the latest GPUCompiler again?

[simeonschaub]

Since this is a bugfix, should we perhaps backport this to GPUCompiler 1.17?

[maleadt]

Yes, backporting would be fine too.