You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
A workaround: create an "alias" pointer of float4 array while allocating the register file (which is of int4 array), then we can load/store from/to the alias pointer without problem, and seems llvm can still perfectly promote mem2reg optimization. (which makes sense, since llvm has migrated to opaque pointer type (but we are using an older version of llvm, since the AIR generated by metal still contains typed pointers.
This
is equivalent to
Although bitcast shouldn't impact the runtime behavior... just curious why llvm -O2 optimization doesn't cover this
The text was updated successfully, but these errors were encountered: