[llvm] InlineFunction: Split inlining into predicate and apply functions (PR #134213)

[Nick Desaulniers via llvm-commits]

================
@@ -271,9 +274,34 @@ class InlineFunctionInfo {
     StaticAllocas.clear();
     InlinedCalls.clear();
     InlinedCallSites.clear();
+    ConvergenceControlToken = nullptr;
+    CallSiteEHPad = nullptr;
   }
 };
 
+/// Check if it is legal to perform inlining of the function called by \p CB
+/// into the caller at this particular use, and sets fields in \p IFI.
+///
+/// This does not consider whether it is possible for the function callee itself
+/// to be inlined; for that see isInlineViable.
+InlineResult CanInlineCallSite(const CallBase &CB, InlineFunctionInfo &IFI);
+
+/// This should generally not be used, use InlineFunction instead.
+///
+/// Perform mechanical inlining of \p CB into the caller.
+///
+/// This does not perform any legality or profitability checks for the
+/// inlining. This assumes that CanInlineCallSite was already called, populated
+/// \p IFI, and returned InlineResult::success.
+///
+/// Also assumes that isInlineViable returned InlineResult::success for the
+/// called function.
----------------
nickdesaulniers wrote:

Perhaps worth an assert then?

https://github.com/llvm/llvm-project/pull/134213