Support caching CUBINs generated with name_expressions - #9801

- Add cuModuleEnumerateFunctions and cuFuncGetName bindings (CUDA 11.6+)
- Implement C++ demangling using __cxa_demangle in function.pyx
- Add _enumerate_and_build_mapping to Module class
- Update compiler.py to use enumeration on cache hit for both CUDA and HIP
- HIP/ROCm: Raise NotImplementedError if enumeration not supported

Co-authored-by: leofang <[email protected]>

- Test cache hit with name_expressions
- Test C++ demangling and name matching with whitespace variations
- Verify mapping is built from cached CUBIN via enumeration APIs
- Skip on HIP/ROCm as it doesn't support function enumeration yet

Co-authored-by: leofang <[email protected]>

…evert HIP changes

- Replace __cxa_demangle with __cu_demangle from nv_decode.h and link libcufilt
- Change name_expressions parameter from list to tuple throughout
- Revert HIP enumeration changes, add TODO comment instead
- Functions are available in cuda.bindings (CUDA 11.6+ standard APIs)

Co-authored-by: leofang <[email protected]>

- Include name_expressions in cache hash key (sorted, comma-separated)
- Explicitly unload module if enumeration fails to prevent CUBIN leak
- Move libcufilt from libraries to static_libraries (it's a static lib)
- Clarify comment: return mangled name on demangle failure for exact matching

Co-authored-by: leofang <[email protected]>

- Change moduleEnumerateFunctions to cdef returning std::vector<Function>
- Change funcGetName to cdef returning char*
- Remove match_name_expression, normalize_name, and demangle_cxx_name functions
- Simplify _enumerate_and_build_mapping to use identity mapping
- Remove unused imports (re, libc.stdlib.free)

Co-authored-by: leofang <[email protected]>

- Restore demangling code and name matching logic in function.pyx
- Swap HIP check before mapping check in _enumerate_and_build_mapping
- Change except * to except +* in moduleEnumerateFunctions
- Add back re import and libc.stdlib.free import

Co-authored-by: leofang <[email protected]>

- Add cdef declarations for moduleEnumerateFunctions and funcGetName in driver.pxd
- Simplify _enumerate_and_build_mapping to build demangled->mangled mapping directly
- Remove match_name_expression and normalize_name functions (no longer needed)
- Remove re import (no longer used)

Co-authored-by: leofang <[email protected]>

- Wrap nv_decode.h extern declaration with IF CUPY_HIP_VERSION == 0
- Guard demangle_cxx_name function body for CUDA vs HIP
- On HIP, return mangled name as-is since libcufilt is not available

Co-authored-by: leofang <[email protected]>

- Change demangle_cxx_name to take const char* instead of str
- Avoid encode→decode round-trip by passing char* directly
- Decode only once when building mapping
- Remove TestNameExpressionsCaching test class (tests unreliable due to multiple caching levels)

Co-authored-by: leofang <[email protected]>

- Change IF CUPY_HIP_VERSION == 0 to IF CUPY_CUDA_VERSION > 0
- Move cdef declarations outside try block to fix Cython compilation error
- "cdef statement not allowed here" error resolved

Co-authored-by: leofang <[email protected]>

- Change except +* to except + in moduleEnumerateFunctions (Cython syntax)
- Add _enumerate_and_build_mapping declaration to function.pxd
- Fixes compiler crash and missing method declaration errors

Co-authored-by: leofang <[email protected]>

…ibcufilt API

Co-authored-by: leofang <[email protected]>

…ibcufilt API

Co-authored-by: leofang <[email protected]>

…location

Co-authored-by: leofang <[email protected]>

… reuse decoded string

Co-authored-by: leofang <[email protected]>

…ames

…ames