--check-bounds=no results in invalid IR with AMDGPU

[pxl-th]

On Julia 1.9-beta3 setting --check-bounds=no results in invalid IR.
Note, that @inbounds works.

Launch below MWE with --check-bounds=no:

using AMDGPU

function memset_kerr!(x, v::Float32)
    i = workitemIdx().x + (workgroupIdx().x - 0x1) * workgroupDim().x
    x[i] = v
    return nothing
end

function main()
    x = ROCArray{Float32}(undef, 1024 * 1024)
    wait(@roc groupsize=256 gridsize=length(x) memset_kerr!(x, 1f0))
    return nothing
end
main()

Error:

ERROR: LoadError: InvalidIRError: compiling kernel #memset_kerr!(AMDGPU.Device.ROCDeviceVector{Float32, 1}, Float32) resulted in invalid LLVM IR
Reason: unsupported call to an unknown function (call to jl_f_apply_type)
Stacktrace:
 [1] Val
   @ ./essentials.jl:801
 [2] #setindex!
   @ ~/.julia/dev/AMDGPU/src/device/gcn/array.jl:85
 [3] memset_kerr!
   @ ~/.julia/dev/gcn_test.jl:5
Reason: unsupported call to an unknown function (call to ijl_new_structv)
Stacktrace:
 [1] Val
   @ ./essentials.jl:799
 [2] Val
   @ ./essentials.jl:801
 [3] #setindex!
   @ ~/.julia/dev/AMDGPU/src/device/gcn/array.jl:85
 [4] memset_kerr!
   @ ~/.julia/dev/gcn_test.jl:5
Reason: unsupported dynamic function invocation (call to unsafe_store!)
Stacktrace:
 [1] #setindex!
   @ ~/.julia/dev/AMDGPU/src/device/gcn/array.jl:85
 [2] memset_kerr!
   @ ~/.julia/dev/gcn_test.jl:5
Hint: catch this exception as `err` and call `code_typed(err; interactive = true)` to introspect the erronous code with Cthulhu.jl
Stacktrace:
  [1] check_ir(job::GPUCompiler.CompilerJob{GPUCompiler.GCNCompilerTarget, AMDGPU.Compiler.ROCCompilerParams, GPUCompiler.FunctionSpec{typeof(memset_kerr!), Tuple{AMDGPU.Device.ROCDeviceVector{Float32, 1}, Float32}}}, args::LLVM.Module)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/qdoh1/src/validation.jl:141
  [2] macro expansion
    @ ~/.julia/packages/GPUCompiler/qdoh1/src/driver.jl:418 [inlined]
  [3] macro expansion
    @ ~/.julia/packages/TimerOutputs/LHjFw/src/TimerOutput.jl:253 [inlined]
  [4] macro expansion
    @ ~/.julia/packages/GPUCompiler/qdoh1/src/driver.jl:416 [inlined]
  [5] emit_asm(job::GPUCompiler.CompilerJob, ir::LLVM.Module; strip::Bool, validate::Bool, format::LLVM.API.LLVMCodeGenFileType)
    @ GPUCompiler ~/.julia/packages/GPUCompiler/qdoh1/src/utils.jl:83
  [6] (::AMDGPU.Compiler.var"#33#36"{GPUCompiler.CompilerJob{GPUCompiler.GCNCompilerTarget, AMDGPU.Compiler.ROCCompilerParams, GPUCompiler.FunctionSpec{typeof(memset_kerr!), Tuple{AMDGPU.Device.ROCDeviceVector{Float32, 1}, Float32}}}, Core.MethodInstance})(ctx::LLVM.ThreadSafeContext)
    @ AMDGPU.Compiler ~/.julia/dev/AMDGPU/src/compiler.jl:182
  [7] LLVM.ThreadSafeContext(f::AMDGPU.Compiler.var"#33#36"{GPUCompiler.CompilerJob{GPUCompiler.GCNCompilerTarget, AMDGPU.Compiler.ROCCompilerParams, GPUCompiler.FunctionSpec{typeof(memset_kerr!), Tuple{AMDGPU.Device.ROCDeviceVector{Float32, 1}, Float32}}}, Core.MethodInstance})
    @ LLVM ~/.julia/packages/LLVM/qc3sa/src/executionengine/ts_module.jl:14
  [8] JuliaContext(f::AMDGPU.Compiler.var"#33#36"{GPUCompiler.CompilerJob{GPUCompiler.GCNCompilerTarget, AMDGPU.Compiler.ROCCompilerParams, GPUCompiler.FunctionSpec{typeof(memset_kerr!), Tuple{AMDGPU.Device.ROCDeviceVector{Float32, 1}, Float32}}}, Core.MethodInstance})
    @ GPUCompiler ~/.julia/packages/GPUCompiler/qdoh1/src/driver.jl:74
  [9] rocfunction_compile(job::GPUCompiler.CompilerJob)
    @ AMDGPU.Compiler ~/.julia/dev/AMDGPU/src/compiler.jl:178
 [10] cached_compilation(cache::Dict{UInt64, Any}, job::GPUCompiler.CompilerJob, compiler::typeof(AMDGPU.Compiler.rocfunction_compile), linker::typeof(AMDGPU.Compiler.rocfunction_link))
    @ GPUCompiler ~/.julia/packages/GPUCompiler/qdoh1/src/cache.jl:90
 [11] rocfunction(f::typeof(memset_kerr!), tt::Type; name::Nothing, device::ROCDevice, global_hooks::NamedTuple{(), Tuple{}})
    @ AMDGPU.Compiler ~/.julia/dev/AMDGPU/src/compiler.jl:164
 [12] rocfunction
    @ ~/.julia/dev/AMDGPU/src/compiler.jl:154 [inlined]
 [13] macro expansion
    @ ~/.julia/dev/AMDGPU/src/highlevel.jl:430 [inlined]
 [14] memset!(x::ROCVector{Float32}, v::Float32)
    @ Main ~/.julia/dev/gcn_test.jl:10
 [15] main()
    @ Main ~/.julia/dev/gcn_test.jl:25
 [16] top-level scope
    @ ~/.julia/dev/gcn_test.jl:35

[pxl-th]

Turns out, this is probably AMDGPU's issue, as it passes alignment argument to Base.unsafe_store!:
https://github.com/JuliaGPU/AMDGPU.jl/blob/09cf06ca16012887e7f6b19635936189994fdadc/src/device/gcn/array.jl#L85

I'll close it here.

[maleadt]

LLVM.jl has versions of these functions taking a Val argument: https://github.com/maleadt/LLVM.jl/blob/9f9d2faee81f907343ddf2d27eab12fa836e5cd4/src/interop/pointer.jl#L84-L85

That said, an in valid IR error is unlikely to be a compiler issue, or at least not straight away. It's more likely that the const-prop of align failed, for whatever reason.

[maleadt]

It's more likely that the const-prop of align failed, for whatever reason.

The reason is of course --check-bounds=no, see JuliaLang/julia#48245. TL;DR: don't use --check-bounds=no, as it pessimizes codegen leading issues to the const-prop failure here.