Sporadic CPU kernel crashes in Newton examples

[c0d1f1ed]

Description

Running Newton examples sporadically crashes with an access violation when the CUDA driver is also initialized in the same process. The crash occurs during kernel execution, not compilation.

Reproduction

# Crashes ~40% of the time with cold cache + CUDA GPU present
uv run python -c "import warp as wp; wp.clear_kernel_cache()"
uv run --extra dev -m newton.tests -k test_cloth_body_collision_semi_implicit -j 1

# Does not crash without CUDA
CUDA_VISIBLE_DEVICES="-1" uv run --extra dev -m newton.tests -k test_cloth_body_collision_semi_implicit -j 1

Observed on:

• Windows 10 Pro x64
• NVIDIA GeForce RTX 5080, Driver 13.1, CUDA Toolkit 13.0
• Warp 1.13.0.dev (LLVM 15.0.7 on Windows)

Warp unit tests appear unaffected, despite the same environment. Newton kernels are larger and more complex though.

[c0d1f1ed]

Root cause analysis: 2026-04-06-crash_analysis.md.

It's another 2 GB relocation offset overflow bug in LLVM's RTDyld JIT linker. Potential fixes:

• Switch to the JITLink linker. It improves relocation handling but is recent in LLVM 15 so might have its own bugs.
• Write a custom MemoryMapper which uses a large region of contiguous reserved memory.
• Upgrade LLVM, if it fixed the RTDyld bug (it did for the similar AArch64 bug).

[c0d1f1ed]

Switching to JITLink worked suprisingly well. It loses JIT step-through debugging support, but that is rarely needed and I made it fall back to RTDyld in that case. JITLink does support debugging in recent versions of LLVM, so developers can also build with --build-llvm as a workaround until we upgrade LLVM.

Fixed in 397af7d.