[llvm] [Instcombine] Write Instcombine pass to strength reduce lock xadd to lock sub (PR #184715)

[Nikita Popov via llvm-commits]

================
@@ -118,6 +119,17 @@ Instruction *InstCombinerImpl::visitAtomicRMWInst(AtomicRMWInst &RMWI) {
          RMWI.getOrdering() != AtomicOrdering::Unordered &&
          "AtomicRMWs don't make sense with Unordered or NotAtomic");
 
+  // Canonicalize atomicrmw add(ptr, neg(X)) -> atomicrmw sub(ptr, X).
+  // old + (-X) == old - X; the returned old value is identical.
+  // This allows strength reduction on targets where atomic sub is cheaper,
+  // e.g. lock sub instead of lock xadd on x86.
----------------
nikic wrote:

```suggestion
```


https://github.com/llvm/llvm-project/pull/184715