================ @@ -0,0 +1,21 @@ +#ifndef __CLC_INTEGER_CLC_CLZ_H__ +#define __CLC_INTEGER_CLC_CLZ_H__ + +#if defined(CLC_CLSPV) || defined(CLC_SPIRV) +// clspv and spir-v targets provide their own OpenCL-compatible clz +#define __clc_clz clz ---------------- arsenm wrote:
> The clspv/spirv targets don't provide implementations of all OpenCL builtins What do you mean by this exactly? I think the system today is backwards and it results in confusing terminology. There are 2 possible interpretation of "target provided" library functions. The backend notion and the libclc notion. > we need to replace that with a redirection to clz somehow By `clz` do you mean OpenCL `clz` (Itanium mangled function defined in this bitcode lib) or system/target provided (libc) `clz`? Either way, it shouldn't matter. `__clc_clz` should not be defined in terms of either. The goal here is to just get a call to `llvm.ctlz` (which should be the universal implementation) in the IR. __clc_clz should always just be a wrapper around the intrinsic call. What you have in the impl file should already produce this after instcombine in the scalar case. I see some of the target overrides are defined directly on the opencl entrypoints. These should be overriding the __clc wrappers, rather than directly overriding the entries. rocm device libs never tries to use the opencl entrypoints internally, and (libclc is originally a fork of the same lib from ~2014). I think the targets should only provide overrides of __clc wrappers, and not directly override the public entrypoints. Those wrappers may end up being implemented with the libm/libc system names (preferably by codegen) https://github.com/llvm/llvm-project/pull/116786 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
