[maleadt]

LLVM creates globals in the datalayout's globals address space. On AMDGPU and SPIR-V this is address space 1, but several codegen paths use their addresses where a generic ptr is required. Runtime symbol lookup passes the library-handle global to jl_load_and_lookup without a cast; union typeof bitcasts type slots across address spaces or selects a small-tag global against a generic null; and @cfunction uses globals as runtime call arguments and elements of a generic-pointer initializer. The resulting IR fails verification with Invalid bitcast or Call parameter type does not match function signature!.

This causes the verifier abort described in JuliaGPU/AMDGPU.jl#1149, for which JuliaGPU/GPUCompiler.jl#1003 adds a runtime-lookup workaround. The union typeof paths also affect ordinary device code, such as selecting between Int32 and Float32 values and comparing the selected value's type.

Add generic_global_ptr, using ConstantExpr::getPointerBitCastOrAddrSpaceCast, at uses that require a generic pointer. It also replaces the existing casts for string constants and small-type slots. Casting at each use preserves direct loads and stores in the global's address space, which permits better AMDGPU code than casting every global at creation. Globals already in address space 0 are unchanged, and emission_context.cfuncs continues to retain the raw GlobalVariable.

This follows Clang's treatment of globals whose storage address space differs from the pointer address space expected by the language: CodeGenModule::GetOrCreateLLVMGlobal returns an address-space cast in that case. Julia applies the cast at the uses that require a generic pointer, rather than returning a cast when the global is created, so direct memory accesses retain the storage address space.

Validated with GPUCompiler main (6783e66) and LLVM.jl 10.1.0, using separate depots for the baseline nightly (1.14.0-DEV.3501, f5fdd1cdc80) and the assertions-enabled fixed build (1.14.0-DEV.3587, based on 3d57df8728). These are different Julia revisions. check.jl verifies raw output from GPUCompiler.compile_method_instance on GCN and SPIR-V (backend=:llvm for SPIR-V). The fixed results are identical with and without CI=true:

Raw IR case	Before, GCN / SPIR-V	After, GCN / SPIR-V
tag: union typeof comparison	Invalid bitcast / Invalid bitcast	Verifies / Verifies
tag2: three-way union typeof	Invalid bitcast / Invalid bitcast	Verifies / Verifies
smalltag: union isa	Verifies / Verifies	Verifies / Verifies
alloc_string: runtime ccall lookup	Call parameter mismatch / Call parameter mismatch	Verifies / Verifies
cfun: closure @cfunction	Call parameter mismatch / Call parameter mismatch	Verifies / Verifies
cfun_static: static @cfunction	Call parameter mismatch / Call parameter mismatch	Verifies / Verifies

GCN end-to-end case	Before	After, with and without CI=true
Union typeof kernel	Unresolved cglobal load after relocation lowering	Same GPUCompiler relocation failure; LLVM IR verifies
alloc_string	InvalidIRError	InvalidIRError, without a verifier abort

The remaining kernel failure is a GPUCompiler follow-up.

Assisted by: Sol 6.1