[llvm] [AMDGPU] Avoid resource propagation for recursion through multiple functions (PR #111004)

[Joseph Huber via llvm-commits]

================
@@ -91,13 +91,75 @@ MCSymbol *MCResourceInfo::getMaxSGPRSymbol(MCContext &OutContext) {
   return OutContext.getOrCreateSymbol("amdgpu.max_num_sgpr");
 }
 
+// The (partially complete) expression should have no recursion in it. After
+// all, we're trying to avoid recursion using this codepath.
+static bool findSymbolInExpr(MCSymbol *Sym, const MCExpr *Expr,
+                             SmallVectorImpl<const MCExpr *> &Exprs,
+                             SmallPtrSetImpl<const MCExpr *> &Visited) {
+  // Assert if any of the expressions is already visited (i.e., there is
+  // existing recursion).
+  if (!Visited.insert(Expr).second)
----------------
jhuber6 wrote:

I guess the compiler will optimize out the unused variable if `llvm_unreachable` is a no-op. Guess it's
```c++
if (!Visited.insert(Expr).second)
  llvm_unreachable("already visited expression");
```
vs
```c++
[[maybe_unused]] bool AlreadyVisited = Visited.insert(Expr).second;
assert(!AlreadyVisited && "already visited expression");
```
Roughly equivalent so I guess it's up to taste.

https://github.com/llvm/llvm-project/pull/111004