Assertion failed: ((!Result.isArray() || Result.getArrayInitializedElts() == 0) && "zero-initialized array shouldn't have any initialized elts"

[ahatanak]

$ cat test.cpp

struct ty { int *ptr; };
const struct ty arr[] = {(int *)("")};
void f(void) {
  int *x;
  x = arr[0].ptr;
}

$ clang++ -std=c++98 -c test.cpp
Assertion failed: ((!Result.isArray() || Result.getArrayInitializedElts() == 0) && "zero-initialized array shouldn't have any initialized elts"), function VisitCXXParenListOrInitListExpr, file ExprConstant.cpp, line 11315

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[ahatanak]

clang doesn't crash if I revert 610ec95

@nickdesaulniers

[nickdesaulniers]

cc @efriedma-quic presumably this is related to the !Ctx.getLangOpts().CPlusPlus11 from the removed guard.

[nickdesaulniers]

Slightly reduced test case:

struct {
  int *ptr;
} const arr[] = {(int *)""};
void f() { *arr; }

[nickdesaulniers]

The ASTs look the same between language modes.

ArrayExprEvaluator::VisitCXXParenListOrInitListExpr gets called twice; once where Result.isArray is false for the top most InitListExpr. The second time it's true for c++98 but still false for c++11. I wonder why that is? (digging)

[nickdesaulniers]

Not really sure where Result is getting reset.

gdb watch points seem to think that performLifetimeExtension is somehow resetting these objects, but if I add print statements to that code or breakpoints to that function, it seems they are never run.

@efriedma-quic can we just delete the assertion? Removing it doesn't seem to cause any test failures for clang.

[efriedma-quic]

The assertion was added in 1b9f2eb; I think the idea is that zero-initialization is supposed to happen at a particular point, and not later.

If Result.getArrayInitializedElts() is non-zero, that means there's something in the array already; what are the elements?

[nickdesaulniers]

Result.dump():

Array size=1
`-element: Struct
  `-field: LValue <todo>