[llvm] [EarlyCSE] Remove void return restriction for call CSE (PR #145320)

[Nikita Popov via llvm-commits]

================
@@ -146,6 +146,28 @@ define i32 @test5(ptr%P) {
   ret i32 %Diff
 }
 
+declare void @void_func()
+
+define void @void_func_cse_readonly(ptr %P) {
----------------
nikic wrote:

> Once upon a time, we had a baked in assumption that in order to throw, a routine must write memory. Have we changed that at some point?

I think this holds as a practical matter, because the implementation of the unwinding mechanism is probably going to involved writes in some form. But I don't think we ever make IR level assumptions that readonly calls cannot unwind.

> I'm asking because otherwise, a void readonly call would seem to have no observable effect and thus be trivially dead?

Even if we ignore the unwind case, you can have a non-willreturn readnone function, e.g. a function that only contains an infinite loop.

https://github.com/llvm/llvm-project/pull/145320