[llvm] [Inliner] Don't count a call penalty for foldable __memcpy_chk and similar (PR #117876)

[Amara Emerson via llvm-commits]

================
@@ -2260,6 +2271,45 @@ bool CallAnalyzer::simplifyCallSite(Function *F, CallBase &Call) {
   return false;
 }
 
+bool CallAnalyzer::isLoweredToCall(Function *F, CallBase &Call) {
+  const TargetLibraryInfo *TLI = GetTLI ? &GetTLI(*F) : nullptr;
+  LibFunc LF;
+  if (!TLI || !TLI->getLibFunc(*F, LF) || !TLI->has(LF))
+    return TTI.isLoweredToCall(F);
+
+  switch (LF) {
+  case LibFunc_memcpy_chk:
+  case LibFunc_memmove_chk:
+  case LibFunc_mempcpy_chk:
+  case LibFunc_memset_chk:
+    // Calls to  __memcpy_chk whose length is known to fit within the object
+    // size will eventually be replaced by inline stores. Therefore, these
+    // should not incur a call penalty. This is only really relevant on
+    // platforms whose headers redirect memcpy to __memcpy_chk (e.g. Mac), as
+    // other platforms use memcpy intrinsics, which are already exempt from the
+    // call penalty.
+    {
+      auto LenOp = dyn_cast_or_null<ConstantInt>(Call.getOperand(2));
----------------
aemerson wrote:

`auto *` for pointers. But also it shouldn't be possible for the call operand to be null so a vanilla `dyn_cast` should suffice?

https://github.com/llvm/llvm-project/pull/117876