CPU precompiled headers are rejected with LLVM 22

[shi-eric]

Bug Description

When Warp is built against LLVM 22, Clang rejects the CPU precompiled header with an error similar to:

error: OptimizationLevel differs in precompiled file ... vs. current file
Warp: Compilation with PCH failed, retrying without precompiled headers

The PCH is generated at -O3, while CPU kernel modules are compiled at -O2. LLVM 22 checks the code-generation options stored in the PCH and rejects it when the optimization levels differ.

Warp then deletes the rejected PCH and recompiles the module without it. The module still compiles, but the next module generates another incompatible PCH and repeats the same fallback. As a result:

• CPU module builds do not benefit from the PCH cache.
• Each module pays for PCH generation followed by a full non-PCH compilation.
• The PCH cache directory is left without a reusable PCH.
• The existing CPU PCH tests fail.

Reproduction

1. Build Warp against LLVM 22.1.8.
2. Set WARP_CACHE_PATH to a new, empty directory.
3. Save and run the following script:

import warp as wp

wp.config.verbose = True
wp.config.use_precompiled_headers = True


@wp.kernel(module="unique")
def kernel(values: wp.array[float]):
    tid = wp.tid()
    values[tid] = float(tid)


values = wp.zeros(8, dtype=float, device="cpu")
wp.launch(kernel, dim=8, inputs=[values], device="cpu")
print(values.numpy())

The kernel runs successfully, but the compilation log reports the PCH optimization-level mismatch and the fallback compilation.

Expected behavior

CPU modules should compile using the generated PCH without Clang rejecting it. The PCH should remain in the cache and be reusable by subsequent compatible module builds.

The stricter PCH validation was introduced by llvm/llvm-project#146422. LLVM 21 and earlier accept the mismatched PCH because they do not perform this validation.

System Information

Reproduced with:

• Warp built from main
• LLVM 22.1.8
• Windows x86-64 and Windows Arm64
• CPU-only Warp build

[shi-eric]

This is addressed by ea157f2.

The change generates CPU precompiled headers at the same effective optimization level as the modules that consume them and includes that level in the PCH cache key. Default -O2 modules can therefore reuse a compatible PCH, while modules using distinct optimization levels receive separate cache entries. This prevents LLVM 22 from rejecting the PCH and triggering repeated fallback compilation.

Public API behavior:

• No new API is added. Existing wp.config.use_precompiled_headers behavior now retains and reuses compatible CPU PCHs when Warp is built against LLVM 22.
• Existing CPU optimization_level options are reflected in PCH compatibility; debug modules use their effective -O0 level.

Test coverage:

• Modified warp/tests/test_cpu_precompiled_headers.py::TestCpuPrecompiledHeaders::test_compiler_flags_cache_key to validate the optimization-level-aware cache filename while preserving distinct compiler-flag keys.
• Added warp/tests/test_cpu_precompiled_headers.py::TestCpuPrecompiledHeaders::test_debug_optimization_level_cache_key, verifying that debug modules requesting different optimization levels execute correctly and share a single effective -O0 PCH.