AlexVlx wrote: > > `__constant__` may not necessarily be `const` for IR purposes. I.e. IR may > > not rely on the 'known' values, as seen in IR, as the data may actually be > > populated by the host via CUDA API calls `cudaMemcpyToSymbol` before the > > GPU kernel launch. > > But since this is marked `externally_initialised` on the device side, one > would assume that those semantics still hold i.e. there's no known value to > assume?
Oh, I guess the concern might be around doing a `cuda/hipMemcpyToSymbol` during a kernel's execution (i.e. `__global__ foo()` is executing, host sets a __constant__, notifies foo, and expects foo to observe the updated constant)? I admit that I'm not certain if that is allowed, if it is then yes, this is problematic / not viable. https://github.com/llvm/llvm-project/pull/110182 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
