[llvm] [AMDGPU] Fix simplification of VM_VSRC based on VMEM waits (PR #178711)

[Ryan Mitchell via llvm-commits]

================
@@ -625,9 +625,11 @@ class WaitcntBrackets {
   }
 
   void setStateOnFunctionEntryOrReturn() {
-    setScoreUB(AMDGPU::STORE_CNT,
-               getScoreUB(AMDGPU::STORE_CNT) + getLimit(AMDGPU::STORE_CNT));
-    PendingEvents |= Context->getWaitEvents(AMDGPU::STORE_CNT);
+    if (Context->ST.hasVscnt()) {
+      setScoreUB(AMDGPU::STORE_CNT,
+                 getScoreUB(AMDGPU::STORE_CNT) + getLimit(AMDGPU::STORE_CNT));
+      PendingEvents |= Context->getWaitEvents(AMDGPU::STORE_CNT);
+    }
----------------
RyanRio wrote:

Explanation on this change? Might just be archaic and I'm forgetting something..

https://github.com/llvm/llvm-project/pull/178711