[clang] [llvm] [clang] Implement lifetime analysis for lifetime_capture_by(X) (PR #115921)

[Utkarsh Saxena via cfe-commits]

================
@@ -1412,17 +1438,34 @@ static void checkExprLifetimeImpl(Sema &SemaRef,
     return false;
   };
 
+  bool HasReferenceBinding = Init->isGLValue();
   llvm::SmallVector<IndirectLocalPathEntry, 8> Path;
-  if (LK == LK_Assignment &&
-      shouldRunGSLAssignmentAnalysis(SemaRef, *AEntity)) {
-    Path.push_back(
-        {isAssignmentOperatorLifetimeBound(AEntity->AssignmentOperator)
-             ? IndirectLocalPathEntry::LifetimeBoundCall
-             : IndirectLocalPathEntry::GslPointerAssignment,
-         Init});
+  switch (LK) {
+  case LK_Assignment: {
+    if (shouldRunGSLAssignmentAnalysis(SemaRef, *AEntity))
+      Path.push_back(
+          {isAssignmentOperatorLifetimeBound(AEntity->AssignmentOperator)
+               ? IndirectLocalPathEntry::LifetimeBoundCall
+               : IndirectLocalPathEntry::GslPointerAssignment,
+           Init});
+    break;
+  }
+  case LK_LifetimeCapture: {
+    Path.push_back({IndirectLocalPathEntry::LifetimeCapture, Init});
+    if (isRecordWithAttr<PointerAttr>(Init->getType()))
+      HasReferenceBinding = false;
+    // Skip the top MTE if it is a temporary object of the pointer-like type
+    // itself.
+    if (auto *MTE = dyn_cast<MaterializeTemporaryExpr>(Init);
+        MTE && isPointerLikeType(Init->getType()))
+      Init = MTE->getSubExpr();
----------------
usx95 wrote:

Hmm. I see this method being used only for `LK_Extended` so this seems safe. I always disliked dealing with lifetime extension as part of this analysis as it is the only AST/codegen-side effect of this analysis.

But this feels safe now. Thanks for digging into this.

https://github.com/llvm/llvm-project/pull/115921