[llvm] [LangRef] Specify that dereferenceable implies a spurious read (PR #218413)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-ir

Author: Nikita Popov (nikic)

<details>
<summary>Changes</summary>

Specify `dereferenceable` in terms of performing a spurious read of `n` bytes at function entry/exit. This is necessary because dereferenceability is used to introduce spurious reads through speculation. For this it is not sufficient to know that such a read cannot trap: Introducing such a read may introduce a noalias violation. The annotation itself needs to perform a spurious read to justify introduction of further such reads.

It should be noted that this places a standardization constraint on restrict in C++ (or alternatively, may ultimately render the use of dereferenceable for C++ references non-viable). I don't believe there should be any issues in pure C code, because we don't mark C pointers dereferenceable.

That is, something like this needs to be UB in C++:
```
void foo(int &a, int *__restrict b) {
    *b = 0;
}
foo(x, &x);
```

As far as I know, defining dereferenceable in this way is the only way to make the model coherent, unless we can come up with some magic way to have noalias with poison on load instead of IUB, which appears to be impossible.

Related discussion: https://discourse.llvm.org/t/interaction-of-noalias-and-dereferenceable/66979

---
Full diff: https://github.com/llvm/llvm-project/pull/218413.diff


1 Files Affected:

- (modified) llvm/docs/LangRef.md (+4) 


``````````diff
diff --git a/llvm/docs/LangRef.md b/llvm/docs/LangRef.md
index e77076bb68158..537d82007863a 100644
--- a/llvm/docs/LangRef.md
+++ b/llvm/docs/LangRef.md
@@ -1669,6 +1669,10 @@ Currently, only the following parameter attributes are defined:
     freed after that point. Other attributes such as `nofree` can be used
     to exclude frees.
 
+    `dereferenceable` implies a spurious read of `n` bytes (on function entry
+    for arguments and exit for return values). In particular, the bytes are
+    considered read for the purpose of `noalias`.
+
 (attr_dereferenceable_or_null)=
 
 `dereferenceable_or_null(<n>)`

``````````

</details>


https://github.com/llvm/llvm-project/pull/218413