Add an experimental opaque closure type.

[maleadt]

Opaque closures don't make that much sense in the CUDA environment (where we don't have a linker and need to recompile when the OC changes, also because we require specsig), but it can be useful to inline typed IR into a kernel and call it.

Current state of this PR is 1.12+ only (because of JuliaLang/julia#53219). It could be made more general if and when we decide to merge this.

Depends on JuliaGPU/GPUCompiler.jl#572

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 94 lines in your changes are missing coverage. Please review.

Project coverage is 59.97%. Comparing base (a011e73) to head (3f9cde3).
Report is 2 commits behind head on master.

❗ Current head 3f9cde3 differs from pull request most recent head 7782216. Consider uploading reports for the commit 7782216 to get more accurate results

Files	Patch %	Lines
src/compiler/compilation.jl	0.00%	94 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #1853       +/-   ##
===========================================
- Coverage   71.83%   59.97%   -11.87%     
===========================================
  Files         155      155               
  Lines       15013    14979       -34     
===========================================
- Hits        10785     8984     -1801     
- Misses       4228     5995     +1767     

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

[pepijndevos]

How much of this is actually CUDA specific?

[maleadt]

Not much. Only the fact that we need to be able to create a compiler invocation, which depends on backend-specific state (which device is active, etc): https://github.com/JuliaGPU/CUDA.jl/pull/1853/files#diff-ecbfaf5b99ab10dcafff5717c7cc5f856768e4313446fa3fb58b839a25b17cfcR317

[maleadt]

Seems to cause some IR verification errors:

ERROR: LLVM error: Called function is not the same type as the call!
  call void bitcast (
      void ([1 x i64],
            { { i8 addrspace(1)*, i64, [2 x i64], i64 }, { [1 x [1 x i64]], i64 }, i64, i64 }*,
            { { i8 addrspace(1)*, i64, [2 x i64], i64 }, { [1 x [1 x i64]], i64 }, i64, i64 }*,
            [3 x float]*,
            double)* @julia_opaque_gpu_closure_9593
    to
      void ([1 x i64],
            { { i8 addrspace(1)*, i64, [2 x i64], i64 }, { [1 x [1 x i64]], i64 }, i64, i64 }*,
            { { i8 addrspace(1)*, i64, [2 x i64], i64 }, { [1 x [1 x i64]], i64 }, i64, i64 }*,
            { double, float },
            double)*)(
      [1 x i64] %state,
      { { i8 addrspace(1)*, i64, [2 x i64], i64 }, { [1 x [1 x i64]], i64 }, i64, i64 }* nonnull %5,
      { { i8 addrspace(1)*, i64, [2 x i64], i64 }, { [1 x [1 x i64]], i64 }, i64, i64 }* nonnull %6,
      { double, float } %.fca.1.insert,
      double %25), !dbg !300

[vchuravy]

After #582 you should be able to just emit a gpuc.lookup(mi, oc, args...) or maybe gpuc.deferred(oc, args...)?