yxsamliu wrote: > It has nothing to do with writing to those arrays while the kernel is > running. That would indeed be UB. > > > both would still work just the same even with this change, > > No, they will not. Here's the demonstration of the behavior change that > `const` brings to the table: https://cuda.godbolt.org/z/h6e5Wb4PT > > It shows that adding `const` allows compiler to use the initializer values as > written (it does not even bother instantiating `cxxx`), while without `const` > it reads the array values, assuming that they may be changed externally.
EarlyCSE pass propagates the constant value https://cuda.godbolt.org/z/Ta8rKjMTo However, ` @xxx = dso_local addrspace(4) externally_initialized global [3 x i32] [i32 1, i32 2, i32 3], align 4 @_ZL4cxxx = internal addrspace(4) constant [3 x i32] [i32 1, i32 2, i32 3], align 4 ` In this case, _ZL4cxxx does not have externally_initialized . If this patch does not remove externally_initialized, probably this constant propagation won't happen. Maybe Alex can add a lit test to prove that. 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
