[clang] [Clang] Implement P2809: Trivial infinite loops are not Undefined Behavior (PR #90066)

[Eli Friedman via cfe-commits]

================
@@ -908,6 +908,73 @@ void CodeGenFunction::EmitIfStmt(const IfStmt &S) {
     incrementProfileCounter(&S);
 }
 
+bool CodeGenFunction::checkIfLoopMustProgress(const Expr *ControllingExpression,
+                                              bool IsTrivialCXXLoop) {
+  if (CGM.getCodeGenOpts().getFiniteLoops() ==
+      CodeGenOptions::FiniteLoopsKind::Always)
+    return true;
+  if (CGM.getCodeGenOpts().getFiniteLoops() ==
+      CodeGenOptions::FiniteLoopsKind::Never)
+    return false;
+
+  // Now apply rules for plain C (see  6.8.5.6 in C11).
+  // Loops with constant conditions do not have to make progress in any C
+  // version.
+  // As an extension, we consisider loops whose constant expression
+  // can be constant-folded.
+  Expr::EvalResult Result;
+  bool CondIsConstInt =
+      !ControllingExpression ||
+      (ControllingExpression->EvaluateAsInt(Result, getContext()) &&
+       Result.Val.isInt());
+  bool IsTrue = CondIsConstInt &&
+                (!ControllingExpression || Result.Val.getInt().getBoolValue());
+
+  if (getLangOpts().C99 && CondIsConstInt)
+    return false;
+
+  // Loops with non-constant conditions must make progress in C11 and later.
+  if (getLangOpts().C11)
+    return true;
+
+  // [C++26][intro.progress] (DR)
+  // The implementation may assume that any thread will eventually do one of the
+  // following:
+  // [...]
+  // - continue execution of a trivial infinite loop ([stmt.iter.general]).
+  if (getLangOpts().CPlusPlus11) {
+    if (IsTrivialCXXLoop && IsTrue) {
+      CurFn->removeFnAttr(llvm::Attribute::MustProgress);
----------------
efriedma-quic wrote:

I... have to apologize here.  I was pretty sure I remembered how mustprogress worked, but then I went back to read the patches, and it's actually slightly different than what I remembered. mustprogress applies to control flow and calls within a function, but it doesn't propagate: an infinite loop in a non-mustprogress counts as an "effect".  I'll propose a patch to clarify LangRef.

So your original patch was essentially right.  If we drop the mustprogress attribute from functions that contain an infinite loop, we end up with essentially the right behavior: if a function doesn't contain a trivially-infinite loop, we can mark it mustprogress.  And we mark individual loops with metadata either way.

https://github.com/llvm/llvm-project/pull/90066