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

[Florian Hahn via llvm-commits]

================
@@ -2260,6 +2271,48 @@ bool CallAnalyzer::simplifyCallSite(Function *F, CallBase &Call) {
   return false;
 }
 
+bool CallAnalyzer::isLoweredToCall(Function *F, CallBase &Call) {
+  // Calls to memcpy with a known constant size 1/2/4/8 should not incur a call
+  // penalty, as the calls will be folded away by InstCombine. 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.
+  if (GetTLI) {
+    auto TLI = GetTLI(*F);
+    LibFunc LF;
+    if (TLI.getLibFunc(*F, LF) && TLI.has(LF)) {
+      switch (LF) {
+      case LibFunc_memcpy_chk:
+      case LibFunc_memmove_chk:
+      case LibFunc_mempcpy_chk:
+      case LibFunc_memset_chk: {
+        auto LenOp = dyn_cast_or_null<ConstantInt>(Call.getOperand(2));
+        if (!LenOp)
+          LenOp = dyn_cast_or_null<ConstantInt>(
+              SimplifiedValues.lookup(Call.getOperand(2)));
----------------
fhahn wrote:

As a follow-up refactoring, it would be nice to introduce a helper that returns a constant if either the value itself is a constant or if it has an entry in SimplifiedValues that is constant. I think there are other places that could also use such a helper.

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