Crash in LLVM

[maleadt]

using CUDA

A = CUDA.rand(4, 4, 3)
B = CUDA.rand(4, 4)
C = similar(A)

function kernel!(C, A, B)
	i = threadIdx().x

	if (i <= size(A, 3))
		@inbounds C[:, :, i] = A[:, :, i] * B
	end

	return nothing
end

@cuda threads = size(A, 3) kernel!(C, A, B)

[1202] signal (11.1): Segmentation fault
in expression starting at REPL[6]:1
_ZN4llvm27DeadArgumentEliminationPass27RemoveDeadStuffFromFunctionEPNS_8FunctionE.part.424 at /home/tim/Julia/depot/juliaup/julia-1.9.1+0.x64.linux.gnu/bin/../lib/julia/libLLVM-14jl.so (unknown line)
_ZN4llvm27DeadArgumentEliminationPass3runERNS_6ModuleERNS_15AnalysisManagerIS1_JEEE at /home/tim/Julia/depot/juliaup/julia-1.9.1+0.x64.linux.gnu/bin/../lib/julia/libLLVM-14jl.so (unknown line)
_ZN12_GLOBAL__N_13DAE11runOnModuleERN4llvm6ModuleE.part.425 at /home/tim/Julia/depot/juliaup/julia-1.9.1+0.x64.linux.gnu/bin/../lib/julia/libLLVM-14jl.so (unknown line)
_ZN4llvm6legacy15PassManagerImpl3runERNS_6ModuleE at /home/tim/Julia/depot/juliaup/julia-1.9.1+0.x64.linux.gnu/bin/../lib/julia/libLLVM-14jl.so (unknown line)
LLVMRunPassManager at /home/tim/Julia/depot/juliaup/julia-1.9.1+0.x64.linux.gnu/bin/../lib/julia/libLLVM-14jl.so (unknown line)
LLVMRunPassManager at /home/tim/Julia/depot/packages/LLVM/bsdku/lib/13/libLLVM_h.jl:4898 [inlined]
run! at /home/tim/Julia/depot/packages/LLVM/bsdku/src/passmanager.jl:39 [inlined]
macro expansion at /home/tim/Julia/pkg/GPUCompiler/src/optim.jl:292 [inlined]
macro expansion at /home/tim/Julia/depot/packages/LLVM/bsdku/src/base.jl:98 [inlined]
optimize! at /home/tim/Julia/pkg/GPUCompiler/src/optim.jl:286

Should check with an assertions build, the CPU code here probably triggers something.

[maleadt]

Reduced to:

target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64-ni:10:11:12:13"
target triple = "nvptx64-nvidia-cuda"

define internal fastcc void @outer() {
top:
  call void (void ({} *, {} **, i32)*) null(void ({} *, {} **, i32)* @inner)
  ret void
}

define internal void @inner({} *, {} **, i32) {
fail7:
  %state = call [1 x i64] @julia.gpu.state_getter()
  ret void
}

declare [1 x i64] @julia.gpu.state_getter()

LLVM error: Call parameter type does not match function signature!
void ([1 x i64], {}*, {}**, i32)* @inner
 void ({}*, {}**, i32)*  call void null(void ([1 x i64], {}*, {}**, i32)* @inner)

IR corruption due to kernel rewrite pass.