issues with Int64 `atomic_cmpxchg!`

[simeonschaub]

The original issue uncovered in #347:

julia> using OpenCL, pocl_jll

julia> function atomic_arrayset(A::AbstractArray{T}, I::Integer, op::Function, val) where {T}
           ptr = pointer(A, I)
           old = Base.unsafe_load(ptr, 1)
           while true
               cmp = old
               new = convert(T, op(old, val))
               old = OpenCL.SPIRVIntrinsics.atomic_cmpxchg!(ptr, cmp, new)
               (old == cmp) && return new
           end
       end
atomic_arrayset (generic function with 1 method)

julia> function foo(a)
           atomic_arrayset(a, 1, +, 1)
           return
       end
foo (generic function with 1 method)

julia> a = OpenCL.zeros(Int64)
0-dimensional CLArray{Int64, 0, OpenCL.cl.UnifiedDeviceMemory}:
0

julia> @opencl global_size=1000 foo(a)
error: line 53: OpPhi's result type <id> '9[%ulong]' does not match incoming value <id> '35[%35]' type <id> '12[%uint]'.
  %value_phi = OpPhi %ulong %33 %top %35 %L8

ERROR: Failed to validate generated SPIR-V.
If you think this is a bug, please file an issue and attach /tmp/jl_87gxwPQ5yh.bc and /tmp/jl_cdmHQxoFiF.spv.
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/spirv.jl:163 [inlined]
  [3] mcgen(job::GPUCompiler.CompilerJob{…}, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:134
  [4] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [5] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:406 [inlined]
  [6] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:403 [inlined]
  [8] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:116
  [9] compile_unhooked(output::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:115
 [10] compile_unhooked
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:80 [inlined]
 [11] compile(target::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:67
 [12] compile
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:55 [inlined]
 [13] #36
    @ ~/.julia/dev/OpenCL/src/compiler/compilation.jl:62 [inlined]
 [14] JuliaContext(f::OpenCL.var"#36#37"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:34
 [15] JuliaContext(f::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:25
 [16] compile(job::GPUCompiler.CompilerJob)
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/compilation.jl:61
 [17] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(OpenCL.compile), linker::typeof(OpenCL.link))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:245
 [18] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:159
 [19] macro expansion
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:186 [inlined]
 [20] macro expansion
    @ ./lock.jl:273 [inlined]
 [21] clfunction(f::typeof(foo), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}}; kwargs::@Kwargs{})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:181
 [22] clfunction(f::typeof(foo), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:177
 [23] top-level scope
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:66

caused by: failed process: Process(setenv(`/home/simeon/.julia/artifacts/954f1d0245f593658ca964009b1b0c17834a59a7/bin/spirv-val /tmp/jl_cdmHQxoFiF.spv`,["P9K_SSH=0", "PATH=/home/simeon/.julia/artifacts/954f1d0245f593658ca964009b1b0c17834a59a7/bin:/home/simeon/.local/bin:/home/simeon/.config/Code/User/globalStorage/ms-vscode-remote.remote-containers/cli-bin:/home/simeon/.juliaup/bin:/home/simeon/.cargo/bin:/home/simeon/.juliaup/bin:/usr/share/Modules/bin:/usr/lib64/ccache:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/home/simeon/.config/Code/User/globalStorage/github.copilot-chat/debugCommand", "EDITOR=vim", "NO_AT_BRIDGE=1", "POCL_ARGS_CLANG=-fuse-ld=lld;--ld-path=/home/simeon/.julia/scratchspaces/627d6b7a-bbe6-5189-83e7-98cc0a5aeadd/bin/lld;-L;/home/simeon/.julia/artifacts/335e91adc50bc361682926789360ef6150603b44/share/lib;-L;/home/simeon/.julia/scratchspaces/627d6b7a-bbe6-5189-83e7-98cc0a5aeadd/lib", "LD_LIBRARY_PATH=/home/simeon/.julia/artifacts/954f1d0245f593658ca964009b1b0c17834a59a7/lib:/home/simeon/.julia/juliaup/julia-1.11.6+0.x64.linux.gnu/bin/../lib/julia:/home/simeon/.julia/juliaup/julia-1.11.6+0.x64.linux.gnu/bin/../lib", "DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/1000/bus", "MANPATH=/usr/share/man:", "MODULEPATH=/usr/share/Modules/modulefiles:/etc/modulefiles:/usr/share/modulefiles", "DEBUGINFOD_IMA_CERT_PATH=/etc/keys/ima:"  …  "HISTCONTROL=ignoredups", "ORIGINAL_XDG_CURRENT_DESKTOP=GNOME", "MAIL=/var/spool/mail/simeon", "JOURNAL_STREAM=9:22034", "KDEDIRS=/usr", "JULIA_NUM_PRECOMPILE_TASKS=4", "POCL_PATH_LLVM_SPIRV=/home/simeon/.julia/scratchspaces/627d6b7a-bbe6-5189-83e7-98cc0a5aeadd/bin/llvm-spirv", "GIO_LAUNCHED_DESKTOP_FILE_PID=24479", "LS_COLORS=", "OPENBLAS_MAIN_FREE=1"]), ProcessExited(1)) [1]

Stacktrace:
  [1] pipeline_error
    @ ./process.jl:598 [inlined]
  [2] run(::Cmd; wait::Bool)
    @ Base ./process.jl:513
  [3] run
    @ ./process.jl:510 [inlined]
  [4] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/spirv.jl:161 [inlined]
  [5] mcgen(job::GPUCompiler.CompilerJob{…}, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:134
  [6] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:406 [inlined]
  [8] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [9] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:403 [inlined]
 [10] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:116
 [11] compile_unhooked(output::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:115
 [12] compile_unhooked
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:80 [inlined]
 [13] compile(target::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:67
 [14] compile
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:55 [inlined]
 [15] #36
    @ ~/.julia/dev/OpenCL/src/compiler/compilation.jl:62 [inlined]
 [16] JuliaContext(f::OpenCL.var"#36#37"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:34
 [17] JuliaContext(f::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:25
 [18] compile(job::GPUCompiler.CompilerJob)
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/compilation.jl:61
 [19] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(OpenCL.compile), linker::typeof(OpenCL.link))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:245
 [20] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:159
 [21] macro expansion
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:186 [inlined]
 [22] macro expansion
    @ ./lock.jl:273 [inlined]
 [23] clfunction(f::typeof(foo), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}}; kwargs::@Kwargs{})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:181
 [24] clfunction(f::typeof(foo), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:177
 [25] top-level scope
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:66
Some type information was truncated. Use `show(err)` to see complete types.

Trying to reduce this further I stumbled upon another error, not sure it's the same root cause:

julia> function bar(a)
           OpenCL.SPIRVIntrinsics.atomic_cmpxchg!(pointer(a), 0, 1)
           return
       end
bar (generic function with 1 method)

julia> @opencl global_size=1000 bar(a)
error: line 35: AtomicCompareExchange: 64-bit atomics require the Int64Atomics capability
  %24 = OpAtomicCompareExchange %uint %23 %uint_2 %uint_0 %uint_0 %ulong_1 %14

ERROR: Failed to validate generated SPIR-V.
If you think this is a bug, please file an issue and attach /tmp/jl_YrsN2NK3yp.bc and /tmp/jl_4PwzOnCRyH.spv.
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/spirv.jl:163 [inlined]
  [3] mcgen(job::GPUCompiler.CompilerJob{…}, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:134
  [4] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [5] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:406 [inlined]
  [6] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:403 [inlined]
  [8] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:116
  [9] compile_unhooked(output::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:115
 [10] compile_unhooked
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:80 [inlined]
 [11] compile(target::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:67
 [12] compile
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:55 [inlined]
 [13] #36
    @ ~/.julia/dev/OpenCL/src/compiler/compilation.jl:62 [inlined]
 [14] JuliaContext(f::OpenCL.var"#36#37"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:34
 [15] JuliaContext(f::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:25
 [16] compile(job::GPUCompiler.CompilerJob)
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/compilation.jl:61
 [17] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(OpenCL.compile), linker::typeof(OpenCL.link))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:245
 [18] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:159
 [19] macro expansion
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:186 [inlined]
 [20] macro expansion
    @ ./lock.jl:273 [inlined]
 [21] clfunction(f::typeof(bar), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}}; kwargs::@Kwargs{})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:181
 [22] clfunction(f::typeof(bar), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:177
 [23] top-level scope
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:66

caused by: failed process: Process(setenv(`/home/simeon/.julia/artifacts/954f1d0245f593658ca964009b1b0c17834a59a7/bin/spirv-val /tmp/jl_4PwzOnCRyH.spv`,["P9K_SSH=0", "PATH=/home/simeon/.julia/artifacts/954f1d0245f593658ca964009b1b0c17834a59a7/bin:/home/simeon/.local/bin:/home/simeon/.config/Code/User/globalStorage/ms-vscode-remote.remote-containers/cli-bin:/home/simeon/.juliaup/bin:/home/simeon/.cargo/bin:/home/simeon/.juliaup/bin:/usr/share/Modules/bin:/usr/lib64/ccache:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/home/simeon/.config/Code/User/globalStorage/github.copilot-chat/debugCommand", "EDITOR=vim", "NO_AT_BRIDGE=1", "POCL_ARGS_CLANG=-fuse-ld=lld;--ld-path=/home/simeon/.julia/scratchspaces/627d6b7a-bbe6-5189-83e7-98cc0a5aeadd/bin/lld;-L;/home/simeon/.julia/artifacts/335e91adc50bc361682926789360ef6150603b44/share/lib;-L;/home/simeon/.julia/scratchspaces/627d6b7a-bbe6-5189-83e7-98cc0a5aeadd/lib", "LD_LIBRARY_PATH=/home/simeon/.julia/artifacts/954f1d0245f593658ca964009b1b0c17834a59a7/lib:/home/simeon/.julia/juliaup/julia-1.11.6+0.x64.linux.gnu/bin/../lib/julia:/home/simeon/.julia/juliaup/julia-1.11.6+0.x64.linux.gnu/bin/../lib", "DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/1000/bus", "MANPATH=/usr/share/man:", "MODULEPATH=/usr/share/Modules/modulefiles:/etc/modulefiles:/usr/share/modulefiles", "DEBUGINFOD_IMA_CERT_PATH=/etc/keys/ima:"  …  "HISTCONTROL=ignoredups", "ORIGINAL_XDG_CURRENT_DESKTOP=GNOME", "MAIL=/var/spool/mail/simeon", "JOURNAL_STREAM=9:22034", "KDEDIRS=/usr", "JULIA_NUM_PRECOMPILE_TASKS=4", "POCL_PATH_LLVM_SPIRV=/home/simeon/.julia/scratchspaces/627d6b7a-bbe6-5189-83e7-98cc0a5aeadd/bin/llvm-spirv", "GIO_LAUNCHED_DESKTOP_FILE_PID=24479", "LS_COLORS=", "OPENBLAS_MAIN_FREE=1"]), ProcessExited(1)) [1]

Stacktrace:
  [1] pipeline_error
    @ ./process.jl:598 [inlined]
  [2] run(::Cmd; wait::Bool)
    @ Base ./process.jl:513
  [3] run
    @ ./process.jl:510 [inlined]
  [4] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/spirv.jl:161 [inlined]
  [5] mcgen(job::GPUCompiler.CompilerJob{…}, mod::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:134
  [6] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [7] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:406 [inlined]
  [8] macro expansion
    @ ~/.julia/packages/Tracy/slmNc/src/tracepoint.jl:163 [inlined]
  [9] macro expansion
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:403 [inlined]
 [10] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/utils.jl:116
 [11] compile_unhooked(output::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:115
 [12] compile_unhooked
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:80 [inlined]
 [13] compile(target::Symbol, job::GPUCompiler.CompilerJob; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:67
 [14] compile
    @ ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:55 [inlined]
 [15] #36
    @ ~/.julia/dev/OpenCL/src/compiler/compilation.jl:62 [inlined]
 [16] JuliaContext(f::OpenCL.var"#36#37"{GPUCompiler.CompilerJob{…}}; kwargs::@Kwargs{})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:34
 [17] JuliaContext(f::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/driver.jl:25
 [18] compile(job::GPUCompiler.CompilerJob)
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/compilation.jl:61
 [19] actual_compilation(cache::Dict{…}, src::Core.MethodInstance, world::UInt64, cfg::GPUCompiler.CompilerConfig{…}, compiler::typeof(OpenCL.compile), linker::typeof(OpenCL.link))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:245
 [20] cached_compilation(cache::Dict{…}, src::Core.MethodInstance, cfg::GPUCompiler.CompilerConfig{…}, compiler::Function, linker::Function)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/Ecaql/src/execution.jl:159
 [21] macro expansion
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:186 [inlined]
 [22] macro expansion
    @ ./lock.jl:273 [inlined]
 [23] clfunction(f::typeof(bar), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}}; kwargs::@Kwargs{})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:181
 [24] clfunction(f::typeof(bar), tt::Type{Tuple{CLDeviceArray{Int64, 0, 1}}})
    @ OpenCL ~/.julia/dev/OpenCL/src/compiler/execution.jl:177
 [25] top-level scope
    @ ~/.julia/dev/OpenCL/src/compiler/execution.jl:66
Some type information was truncated. Use `show(err)` to see complete types.

This is even though my device definitely supports extended Int64 atomics:

julia> cl.device().extensions
26-element Vector{String}:
 "cl_khr_byte_addressable_store"
 "cl_khr_global_int32_base_atomics"
 "cl_khr_global_int32_extended_atomics"
 "cl_khr_local_int32_base_atomics"
 "cl_khr_local_int32_extended_atomics"
 "cl_khr_3d_image_writes"
 "cl_ext_float_atomics"
 "cl_intel_unified_shared_memory"
 "cl_ext_buffer_device_address"
 "cl_pocl_svm_rect"
 "cl_pocl_command_buffer_svm"
 "cl_khr_command_buffer"
 "cl_khr_command_buffer_multi_device"
 "cl_khr_command_buffer_mutable_dispatch"
 "cl_pocl_command_buffer_host_buffer"
 "cl_khr_subgroups"
 "cl_khr_subgroup_ballot"
 "cl_khr_subgroup_shuffle"
 "cl_intel_subgroups"
 "cl_intel_subgroups_short"
 "cl_intel_subgroups_char"
 "cl_intel_required_subgroup_size"
 "cl_khr_il_program"
 "cl_khr_fp64"
 "cl_khr_int64_base_atomics"
 "cl_khr_int64_extended_atomics"

Are the capabilities perhaps not passed through to the validation pass correctly?

[simeonschaub]

I just ran into this again. Thinking about this some more, I think the atomics might just be a red herring and the actual issue could be in the way we lower control flow here. @maleadt Any tips for reducing this further?

[maleadt]

I'd take the generated LLVM module, create a reproducer script that feeds it to llc (the one from LLVM 20, which we use through the LLVM_SPIRV_Backend_jll) and spirv-val (which should reproduce the failure here), and feed both to llvm-reduce to get a MWE. That is probably then something to file on the LLVM bug tracker.

If this actually turns out to be yet another bug in the LLVM SPIR-V back-end, I'd vote for temporarily going back to the Khronos' translator, see #349. There's a handful of issues to look into for that, but the move back should be reasonable (while unblocking JuliaGPU/GPUCompiler.jl#709 and thus allowing progress on #328).

Let me know if you're having issues, happy to take a look when I have a bit of time next week.

[simeonschaub]

I opened llvm/llvm-project#152863 with a reduced test case, so it seems like there is indeed another bug in the LLVM SPIR-V backend