[vchuravy]

Backports the LLVM.jl 10 port from main (#3323) to the 6.2 line and releases it as 6.2.3.

Why

6.2 is the last CUDA.jl line that resolves against GPUCompiler 1.x. GPUCompiler 1.23.1 (JuliaGPU/GPUCompiler.jl#989) ports 1.x to LLVM.jl 10 and only allows LLVM = "10". Enzyme.jl is moving to LLVM.jl 10 while still supporting GPUCompiler 1.x and 2.x (EnzymeAD/Enzyme.jl#3741). Its Buildkite CI tests CUDA with GPUCompiler 1 pinned, which resolves CUDA down to 6.2. Those jobs need a 6.2 release that accepts LLVM.jl 10.

6.2.3 is the next patch after 6.2.2, and no newer 6.2.x is registered, so this release doesn't skip any fixes from that line.

Changes

• Port to LLVM.jl 10. This is the subset of Upgrade to LLVM.jl 10 #3323 that LLVM.jl 10 needs, adapted to 6.2.
  • create_function/call_function were removed, so the IR generators use @llvmgenerated/generate_llvmcall instead. That covers indexing, shared memory, malloc, atomics, @cuassert, @cuprintf, ldg, @strptr and the randn constant tables.
  • CUDACore imports LLVM.IR and LLVM.Build, and accessors become properties.
  • The RNG prologue is positioned with insertion points, and it no longer overwrites the debug location of the kernel's first instruction.
  • nvvm-reflect-ftz is created with get!, which drops the empty pass manager.
  • compile() inspects and disposes of the IR inside the JuliaContext that owns it.
  • malloc and @strptr convert pointers with pointercast!.
  • Compat: LLVM = "10" for CUDACore, CUDATools and cuBLAS, and GPUCompiler = "1.23.1" for CUDACore and CUDATools.
• Behavior otherwise stays as in 6.2.2.
  • Atomics keep their default (system) sync scope and their signatures, without main's scope argument.
  • The ldg helper keeps the caller's index type.
  • I left out main's other changes: the scoped enums instead of LLVM.API constants (those still work), the IR string templates converted to @llvmgenerated, and the device-IR intrinsic rewrites.
• Version bump. All packages go to 6.2.3.

Testing

All runs were local, on a Quadro RTX 4000 (sm_75), with LLVM.jl 10.1.1 and GPUCompiler 1.23.1 (commit 65a2831):

• Julia 1.12.7, full default suite (test/runtests.jl): 18848 pass, 14 broken, and 20 failures in gpuarrays/broadcasting "mismatched sizes" (+/- through geam with mismatched shapes). Unmodified v6.2.2 with LLVM.jl 9.13.2, GPUCompiler 1.23.0 and the same GPUArrays 11.5.16 fails the same 20, so they predate this PR. They come from a test that GPUArrays 11.5.16 added.
• core/execution crashed its worker once during the parallel run. It passes on a rerun on its own (105/105).
• libraries/cublas: 4378/4378 pass.
• Julia 1.10.12, core tests: 7740 pass, 14 broken, no failures.
• The Enzyme extension tests weren't run, because no Enzyme release supports LLVM.jl 10 yet.

Depends on

The GPUCompiler 1.23.1 registration, JuliaRegistries/General#171343. Until that merges, CI can't resolve this PR without adding GPUCompiler at commit 65a28315f8518ba87bc64d2c4617f56337984fcf (or rev="release-1.23").

🤖 Generated with Claude Code

[maleadt]

Lots of CI failures?