[Clang] Do not perform integral promotion if operands of expression x ? : y have the same type

[zwuis]

According to [expr.cond]/5 and [expr.cond]/7.1, no conversion should be performed to compute result type of conditional operator if middle operand and right operand have the same type regardless of their value catagories.

Fixes #15998.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Yanzuo Liu (zwuis)

Changes

According to [[expr.cond]/5](https://eel.is/c++draft/expr.cond#5) and [[expr.cond]/7.1](https://eel.is/c++draft/expr.cond#7.1), no conversion should be performed to compute result type if middle operand and right operand have the same type regardless of their value catagories.

Fixes #15998.

---

Full diff: https://github.com/llvm/llvm-project/pull/108837.diff

3 Files Affected:

• (modified) clang/docs/ReleaseNotes.rst (+2)
• (modified) clang/lib/Sema/SemaExpr.cpp (+1-3)
• (added) clang/test/SemaCXX/conditional-gnu-ext.cpp (+19)

diff --git a/clang/docs/ReleaseNotes.rst b/clang/docs/ReleaseNotes.rst
index 17ec1fe0b946de..db8a7568a12114 100644
--- a/clang/docs/ReleaseNotes.rst
+++ b/clang/docs/ReleaseNotes.rst
@@ -389,6 +389,8 @@ Bug Fixes to C++ Support
 - Fixed a crash when clang tries to subtitute parameter pack while retaining the parameter
   pack. #GH63819, #GH107560
 - Fix a crash when a static assert declaration has an invalid close location. (#GH108687)
+- Fixed a bug in computing result type of conditional operator with omitted middle operand
+  (a GNU extension). (#GH15998)
 
 Bug Fixes to AST Handling
 ^^^^^^^^^^^^^^^^^^^^^^^^^
diff --git a/clang/lib/Sema/SemaExpr.cpp b/clang/lib/Sema/SemaExpr.cpp
index 8f3e15cc9a9bb7..8414a55e4868b9 100644
--- a/clang/lib/Sema/SemaExpr.cpp
+++ b/clang/lib/Sema/SemaExpr.cpp
@@ -8785,11 +8785,9 @@ ExprResult Sema::ActOnConditionalOp(SourceLocation QuestionLoc,
       commonExpr = result.get();
     }
     // We usually want to apply unary conversions *before* saving, except
-    // in the special case of a C++ l-value conditional.
+    // in the special case in C++ that operands have the same type.
     if (!(getLangOpts().CPlusPlus
           && !commonExpr->isTypeDependent()
-          && commonExpr->getValueKind() == RHSExpr->getValueKind()
-          && commonExpr->isGLValue()
           && commonExpr->isOrdinaryOrBitFieldObject()
           && RHSExpr->isOrdinaryOrBitFieldObject()
           && Context.hasSameType(commonExpr->getType(), RHSExpr->getType()))) {
diff --git a/clang/test/SemaCXX/conditional-gnu-ext.cpp b/clang/test/SemaCXX/conditional-gnu-ext.cpp
new file mode 100644
index 00000000000000..83a6fff8467863
--- /dev/null
+++ b/clang/test/SemaCXX/conditional-gnu-ext.cpp
@@ -0,0 +1,19 @@
+// RUN: %clang_cc1 -fsyntax-only -verify -std=c++11 %s
+// RUN: %clang_cc1 -fsyntax-only -verify -std=c++11 %s -fexperimental-new-constant-interpreter
+// expected-no-diagnostics
+
+/*
+FIXME: Support constexpr
+
+constexpr int evaluate_once(int x) {
+  return (++x) ? : 10;
+}
+static_assert(evaluate_once(0) == 1, "");
+*/
+
+namespace GH15998 {
+  enum E { Zero, One };
+  E test(E e) {
+    return e ? : One;
+  }
+}

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[shafik]

I think this test should go in clang/test/Sema/conditional-expr.c and we should add a section of the GNU extension there.

I am also a bit concerned that we don't have a specific set of sema tests for the GNU form of the conditional expression. So it makes me much left confident that we will catch any breaks with this change.

[erichkeane]

I too am very concerned about the lack of tests here, there are some disappearing conditions I cannot help but think are there for a good reason, and would like to see some study as to why those where there and when.

I'm also concerned about the constexpr support being 'FIXME', it seems that this should just 'work' unless something odd is happening.

[erichkeane]

So I don't see how losing the above two value kind exceptions fix this? Can you explain this better?

[erichkeane]

Should run with BOTH constexpr interpreters.