[llvm] [IR][FunctionAttrs] Clarify memory effects of atomics (PR #193768)

[Nikita Popov via llvm-commits]

================
@@ -1,10 +1,10 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-attributes
 ; RUN: opt -passes=function-attrs -S < %s | FileCheck %s
 
-; Atomic load/store to local doesn't affect whether a function is
-; readnone/readonly.
+; Even though the load/store is on alloca, we can't mark the function as
+; readnone due to the synchronization effect.
----------------
nikic wrote:

Yes, this specific case is fine, but ignoring atomic accesses on allocas in general is not. I tried to make the comment clearer.

In most cases where we can optimize this, it will happen as part of SROA (which runs before FunctionAttrs). I guess there is a small optimization gap where we have an alloca that does not escape, but is accessed atomically with variable GEPs. In that case we could relax the accesses to not-atomic. Not sure whether that case is common enough to handle, but it should be a separate opt in any case.

https://github.com/llvm/llvm-project/pull/193768