[Clang] [Sema] Support matrix types in pseudo-destructor expressions

[Sirraide]

We already support vector types, and since matrix element types have to be scalar types, there should be no problem w/ just enabling this.

This now also allows matrix types to be stored in STL containers.

[llvmbot]

@llvm/pr-subscribers-clang

Author: None (Sirraide)

Changes

We already support vector types, and since matrix element types have to be scalar types, there should be no problem w/ just enabling this.

This now also allows matrix types to be stored in STL containers.

---

Full diff: https://github.com/llvm/llvm-project/pull/117483.diff

3 Files Affected:

• (modified) clang/docs/ReleaseNotes.rst (+3)
• (modified) clang/lib/Sema/SemaExprCXX.cpp (+1-1)
• (added) clang/test/SemaCXX/matrix-types-pseudo-destructor.cpp (+19)

diff --git a/clang/docs/ReleaseNotes.rst b/clang/docs/ReleaseNotes.rst
index 8bd06fadfdc984..ad3c0d3e67f031 100644
--- a/clang/docs/ReleaseNotes.rst
+++ b/clang/docs/ReleaseNotes.rst
@@ -371,6 +371,9 @@ Non-comprehensive list of changes in this release
 - ``__builtin_reduce_and`` function can now be used in constant expressions.
 - ``__builtin_reduce_or`` and ``__builtin_reduce_xor`` functions can now be used in constant expressions.
 
+- Matrix types (a Clang extension) can now be used in pseudo-destructor expressions,
+  which allows them to be stored in STL containers.
+
 New Compiler Flags
 ------------------
 
diff --git a/clang/lib/Sema/SemaExprCXX.cpp b/clang/lib/Sema/SemaExprCXX.cpp
index d85819b21c8265..d440755ee70665 100644
--- a/clang/lib/Sema/SemaExprCXX.cpp
+++ b/clang/lib/Sema/SemaExprCXX.cpp
@@ -8189,7 +8189,7 @@ ExprResult Sema::BuildPseudoDestructorExpr(Expr *Base,
     return ExprError();
 
   if (!ObjectType->isDependentType() && !ObjectType->isScalarType() &&
-      !ObjectType->isVectorType()) {
+      !ObjectType->isVectorType() && !ObjectType->isMatrixType()) {
     if (getLangOpts().MSVCCompat && ObjectType->isVoidType())
       Diag(OpLoc, diag::ext_pseudo_dtor_on_void) << Base->getSourceRange();
     else {
diff --git a/clang/test/SemaCXX/matrix-types-pseudo-destructor.cpp b/clang/test/SemaCXX/matrix-types-pseudo-destructor.cpp
new file mode 100644
index 00000000000000..f0ee953ad2557c
--- /dev/null
+++ b/clang/test/SemaCXX/matrix-types-pseudo-destructor.cpp
@@ -0,0 +1,19 @@
+// RUN: %clang_cc1 -fsyntax-only -fenable-matrix -std=c++11 -verify %s
+// expected-no-diagnostics
+
+template <typename T>
+void f() {
+  T().~T();
+}
+
+using mat4 = float __attribute__((matrix_type(4, 4)));
+using mat4i = int __attribute__((matrix_type(4, 4)));
+
+template <typename T>
+using mat4_t = T __attribute__((matrix_type(4, 4)));
+
+void g() {
+  f<mat4>();
+  f<mat4i>();
+  f<mat4_t<double>>();
+}

[Fznamznon]

This looks reasonable.
Does it make sense to add a sanity test for codegen? I.e. so there is no crashes/assertions and no attempts to generate anything

[Sirraide]

Does it make sense to add a sanity test for codegen?

Probably; good idea

[erichkeane]

I've got no problem with this, but I definitely agree with a need to test this/validate codegen to make sure the destruciton 'looks' reasonable.

[Sirraide]

I definitely agree with a need to test this/validate codegen to make sure the destruciton 'looks' reasonable.

Done.

(And while adding the tests, I also discovered an unrelated bug: #118604)

[AaronBallman]

Should we have a constexpr test to ensure that this noop works fine in a constant expression too?

[Sirraide]

Probably a good idea

[Sirraide]

Well, it turns out we currently don’t support matrix types in constant expressions, so, er, I’ll open a separate issue about that.

[Sirraide]

#120167