@llvm/pr-subscribers-clang

Author: Timm Baeder (tbaederr)

diff --git a/clang/lib/AST/ExprConstant.cpp b/clang/lib/AST/ExprConstant.cpp
index e5950f461e4b2..46a87fd6c7e1c 100644
--- a/clang/lib/AST/ExprConstant.cpp
+++ b/clang/lib/AST/ExprConstant.cpp
@@ -11788,6 +11788,11 @@ bool ArrayExprEvaluator::VisitCXXParenListOrInitListExpr(
   LLVM_DEBUG(llvm::dbgs() << "The number of elements to initialize: "
                           << NumEltsToInit << ".\n");
 
+  if (!Info.CheckArraySize(ExprToVisit->getExprLoc(),
+                           CAT->getNumAddressingBits(Info.Ctx), NumEltsToInit,
+                           /*Diag=*/true))
+    return false;
+
   Result = APValue(APValue::UninitArray(), NumEltsToInit, NumElts);
 
   // If the array was previously zero-initialized, preserve the

That fixes compile-time computation, but this example still breaks in codegen:

#include <array>
using namespace std;
constexpr size_t kMemoryChunk = 1024 * 8;
constexpr size_t kNumberOfIterations = 2000000;
constexpr size_t kThreadsNumber = 2 * kMemoryChunk;
struct S {};

int main() {
    array<S, kThreadsNumber * kNumberOfIterations> futures{};
}

here:

because NumElements is 32768000000.

Edit: Hmm, but that's only if S is empty.

I have not added the test since I thought it might take too long, but it's probably not a problem since we do the check before even touching the array...

If I could come up with a reproducer that doesn't require <future>...

I have not added the test since I thought it might take too long, but it's probably not a problem since we do the check before even touching the array...

Can’t you set the eval steps to like 100 or sth to avoid it taking too long or am I missing something here?

That comment is obsolete, I've added two test cases that work