Fix a crash in constant evaluation of ExtVectorElementExprs

[ahatanak]

Handle the case where the base expression is a pointer to a vector type.

rdar://149223362

[llvmbot]

@llvm/pr-subscribers-clang

Author: Akira Hatanaka (ahatanak)

Changes

Handle the case where the base expression is a pointer to a vector type.

rdar://149223362

---

Full diff: https://github.com/llvm/llvm-project/pull/136771.diff

2 Files Affected:

• (modified) clang/lib/AST/ExprConstant.cpp (+4-1)
• (modified) clang/test/SemaCXX/constexpr-vectors-access-elements.cpp (+2)

diff --git a/clang/lib/AST/ExprConstant.cpp b/clang/lib/AST/ExprConstant.cpp
index f598ef5929aa4..fd870c84796a5 100644
--- a/clang/lib/AST/ExprConstant.cpp
+++ b/clang/lib/AST/ExprConstant.cpp
@@ -9197,7 +9197,10 @@ bool LValueExprEvaluator::VisitExtVectorElementExpr(
 
   if (Success) {
     Result.setFrom(Info.Ctx, Val);
-    const auto *VT = E->getBase()->getType()->castAs<VectorType>();
+    QualType BaseType = E->getBase()->getType();
+    if (E->isArrow())
+      BaseType = BaseType->getPointeeType();
+    const auto *VT = BaseType->castAs<VectorType>();
     HandleLValueVectorElement(Info, E, Result, VT->getElementType(),
                               VT->getNumElements(), Indices[0]);
   }
diff --git a/clang/test/SemaCXX/constexpr-vectors-access-elements.cpp b/clang/test/SemaCXX/constexpr-vectors-access-elements.cpp
index 08223e15feb72..58efcde414af2 100644
--- a/clang/test/SemaCXX/constexpr-vectors-access-elements.cpp
+++ b/clang/test/SemaCXX/constexpr-vectors-access-elements.cpp
@@ -43,4 +43,6 @@ static_assert(b.lo.lo == 1); // expected-error {{not an integral constant expres
 // make sure clang rejects taking address of a vector element
 static_assert(&b[1]); // expected-error {{address of vector element requested}}
 
+constexpr const FourIntsExtVec *p = &b;
+static_assert(p->x == 1);
 }

[erichkeane]

I'm a little surprised this isn't already handled in the AST. I would expect the sub-expression to be a dereference of the vector type in the AST instead to having to do this (for the same reason that I would expect *vecTyPtr to have one).

[ahatanak]

Is sub-expression the base expression in this case? For ExtVectorElementExpr, the base expression is whatever the LHS of -> or . is and that decision was made a long time ago. I think MemberExpr's base expression can be a pointer type too.

[erichkeane]

Ah, huh, I didn't realize that about MemberExpr, I guess this IS consistent. Odd choice that we made...