[pocl] Drop dead deferred-codegen bodies before SPIR-V translation

[vchuravy]

Companion to EnzymeAD/Enzyme.jl#3405. Needed to make Enzyme work over the POCL backend (x-ref #583, EnzymeAD/Enzyme.jl#3309).

Problem

Deferred-codegen entrypoints — notably the wrappers Enzyme generates — are held externally live across GPUCompiler's InternalizePass so linking can resolve them. Once linked and alwaysinlined they are dead, but external linkage keeps GlobalDCEPass from dropping them, so the SPIR-V backend still has to translate a function it cannot express.

These wrappers take first-class aggregates containing addrspace(1) pointers. Extracting one back out miscompiles — the backend types the struct member as pointer-to-uchar but the extract result as pointer-to-double:

error: Result type (OpTypePointer) does not match the type that results from indexing into the composite (OpTypePointer).
  %105 = OpCompositeExtract %_ptr_CrossWorkgroup_double %91 0

This reproduces with no Enzyme involved:

target datalayout = "e-i64:64-...-G1"
target triple = "spirv64-unknown-unknown"

define spir_func void @extract_from_byval_struct({ ptr addrspace(1), i64 } %arg) {
  %p = extractvalue { ptr addrspace(1), i64 } %arg, 0
  store double 1.0, ptr addrspace(1) %p, align 8
  ret void
}

so the real fix belongs upstream in the LLVM SPIR-V backend. Setting spir_func on the wrapper is not sufficient — I checked.

This change

Empty the bodies of unreferenced non-kernel definitions in finish_ir!, so only declarations reach the backend. finish_ir! is the right hook because it runs after AlwaysInlinerPass (which is what makes these functions dead) and after the cleanup GlobalDCEPass that would otherwise have handled them. We empty rather than erase so the finish_ir! loop over the remaining deferred jobs can still look them up by name.

Guarded on job.config.kernel so it runs once for the entry job, not again for each deferred job.

Result

With this plus EnzymeAD/Enzyme.jl#3405, forward-mode Enzyme over a POCL kernel compiles, runs on device and produces correct derivatives:

KA POCL FORWARD-MODE ENZYME: PASS  (dA[1:4] = [2.0, 4.0, 6.0, 8.0])

Reverse mode remains blocked on a separate LLVM SPIR-V backend segfault in SPIRVInstructionSelector::selectExtractVal; see the Enzyme PR for the minimal reproducer.

🤖 Generated with Claude Code

[vchuravy]

Superseded — branch accidentally included #583's commits. Reopened as a standalone PR off main.

[github-actions]

Benchmark Results

Show table

	main	e3e4799...	main / e3e4799...
saxpy/default/Float32/1024	0.0734 ± 0.03 ms	0.0746 ± 0.031 ms	0.983 ± 0.58
saxpy/default/Float32/1048576	0.494 ± 0.022 ms	0.493 ± 0.023 ms	1 ± 0.064
saxpy/default/Float32/16384	0.061 ± 0.03 ms	0.063 ± 0.03 ms	0.968 ± 0.67
saxpy/default/Float32/2048	0.0607 ± 0.029 ms	0.0589 ± 0.029 ms	1.03 ± 0.7
saxpy/default/Float32/256	0.0474 ± 0.031 ms	0.0515 ± 0.032 ms	0.921 ± 0.83
saxpy/default/Float32/262144	0.17 ± 0.029 ms	0.174 ± 0.029 ms	0.975 ± 0.23
saxpy/default/Float32/32768	0.0687 ± 0.029 ms	0.0691 ± 0.029 ms	0.993 ± 0.59
saxpy/default/Float32/4096	0.0614 ± 0.027 ms	0.0682 ± 0.027 ms	0.901 ± 0.53
saxpy/default/Float32/512	0.0496 ± 0.033 ms	0.069 ± 0.031 ms	0.719 ± 0.58
saxpy/default/Float32/64	0.0453 ± 0.028 ms	0.0491 ± 0.031 ms	0.922 ± 0.82
saxpy/default/Float32/65536	0.0854 ± 0.03 ms	0.0839 ± 0.03 ms	1.02 ± 0.51
saxpy/default/Float64/1024	0.0684 ± 0.031 ms	0.0717 ± 0.031 ms	0.954 ± 0.6
saxpy/default/Float64/1048576	0.578 ± 0.093 ms	0.564 ± 0.075 ms	1.02 ± 0.21
saxpy/default/Float64/16384	0.0599 ± 0.025 ms	0.0605 ± 0.026 ms	0.989 ± 0.6
saxpy/default/Float64/2048	0.0545 ± 0.031 ms	0.0564 ± 0.03 ms	0.967 ± 0.75
saxpy/default/Float64/256	0.0465 ± 0.031 ms	0.0556 ± 0.032 ms	0.835 ± 0.74
saxpy/default/Float64/262144	0.19 ± 0.036 ms	0.195 ± 0.033 ms	0.977 ± 0.25
saxpy/default/Float64/32768	0.0724 ± 0.028 ms	0.0713 ± 0.027 ms	1.02 ± 0.56
saxpy/default/Float64/4096	0.0711 ± 0.025 ms	0.0712 ± 0.024 ms	0.999 ± 0.49
saxpy/default/Float64/512	0.0503 ± 0.034 ms	0.0684 ± 0.03 ms	0.735 ± 0.59
saxpy/default/Float64/64	0.0455 ± 0.029 ms	0.0481 ± 0.031 ms	0.946 ± 0.85
saxpy/default/Float64/65536	0.089 ± 0.029 ms	0.091 ± 0.029 ms	0.978 ± 0.44
saxpy/static workgroup=(1024,)/Float32/1024	0.0677 ± 0.031 ms	0.0703 ± 0.031 ms	0.964 ± 0.61
saxpy/static workgroup=(1024,)/Float32/1048576	0.422 ± 0.029 ms	0.424 ± 0.028 ms	0.994 ± 0.094
saxpy/static workgroup=(1024,)/Float32/16384	0.056 ± 0.028 ms	0.0582 ± 0.03 ms	0.963 ± 0.68
saxpy/static workgroup=(1024,)/Float32/2048	0.0541 ± 0.029 ms	0.0553 ± 0.029 ms	0.979 ± 0.73
saxpy/static workgroup=(1024,)/Float32/256	0.0467 ± 0.03 ms	0.0492 ± 0.029 ms	0.95 ± 0.82
saxpy/static workgroup=(1024,)/Float32/262144	0.145 ± 0.029 ms	0.151 ± 0.029 ms	0.961 ± 0.27
saxpy/static workgroup=(1024,)/Float32/32768	0.0609 ± 0.025 ms	0.0631 ± 0.026 ms	0.964 ± 0.57
saxpy/static workgroup=(1024,)/Float32/4096	0.0561 ± 0.027 ms	0.0662 ± 0.026 ms	0.847 ± 0.53
saxpy/static workgroup=(1024,)/Float32/512	0.0486 ± 0.03 ms	0.065 ± 0.029 ms	0.748 ± 0.57
saxpy/static workgroup=(1024,)/Float32/64	0.046 ± 0.03 ms	0.0473 ± 0.029 ms	0.972 ± 0.86
saxpy/static workgroup=(1024,)/Float32/65536	0.0733 ± 0.029 ms	0.0754 ± 0.03 ms	0.972 ± 0.54
saxpy/static workgroup=(1024,)/Float64/1024	0.0548 ± 0.031 ms	0.0656 ± 0.031 ms	0.836 ± 0.62
saxpy/static workgroup=(1024,)/Float64/1048576	0.54 ± 0.097 ms	0.553 ± 0.093 ms	0.976 ± 0.24
saxpy/static workgroup=(1024,)/Float64/16384	0.0557 ± 0.024 ms	0.0595 ± 0.027 ms	0.937 ± 0.58
saxpy/static workgroup=(1024,)/Float64/2048	0.0519 ± 0.031 ms	0.0542 ± 0.029 ms	0.958 ± 0.77
saxpy/static workgroup=(1024,)/Float64/256	0.0472 ± 0.03 ms	0.0537 ± 0.029 ms	0.878 ± 0.73
saxpy/static workgroup=(1024,)/Float64/262144	0.185 ± 0.032 ms	0.193 ± 0.033 ms	0.959 ± 0.23
saxpy/static workgroup=(1024,)/Float64/32768	0.0668 ± 0.025 ms	0.069 ± 0.027 ms	0.967 ± 0.52
saxpy/static workgroup=(1024,)/Float64/4096	0.0582 ± 0.026 ms	0.0664 ± 0.026 ms	0.876 ± 0.52
saxpy/static workgroup=(1024,)/Float64/512	0.0483 ± 0.031 ms	0.0643 ± 0.029 ms	0.751 ± 0.59
saxpy/static workgroup=(1024,)/Float64/64	0.0458 ± 0.029 ms	0.0473 ± 0.028 ms	0.969 ± 0.84
saxpy/static workgroup=(1024,)/Float64/65536	0.0854 ± 0.025 ms	0.088 ± 0.027 ms	0.97 ± 0.41
time_to_load	0.928 ± 0.0079 s	0.927 ± 0.01 s	1 ± 0.014

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[codecov]

Codecov Report

❌ Patch coverage is 0% with 33 lines in your changes missing coverage. Please review.
✅ Project coverage is 0.83%. Comparing base (65690fd) to head (e3e4799).

Files with missing lines	Patch %	Lines
src/pocl/compiler/compilation.jl	0.00%	12 Missing ⚠️
ext/EnzymeCore07Ext.jl	0.00%	9 Missing ⚠️
ext/EnzymeCore08Ext.jl	0.00%	9 Missing ⚠️
ext/EnzymeExt.jl	0.00%	3 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (65690fd) and HEAD (e3e4799). Click for more details.

HEAD has 28 uploads less than BASE

Flag	BASE (65690fd)	HEAD (e3e4799)
	52	24

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #735       +/-   ##
==========================================
- Coverage   64.77%   0.83%   -63.94%     
==========================================
  Files          24      23        -1     
  Lines        2024    1788      -236     
==========================================
- Hits         1311      15     -1296     
- Misses        713    1773     +1060     

☔ 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.