[PATCH] Don't copy objects with trivial, deleted copy ctors

[Reid Kleckner]

OK.  I added a two-copy ctor test case, but it's not enough of a corner case.  We appear to eagerly declare the implicit special members for it.

PTAL

================
Comment at: lib/CodeGen/CGCXXABI.cpp:49-50
@@ +48,4 @@
+
+  // If all trivial copy and move constructors are deleted, we cannot copy the
+  // argument.
+  return !(AnyDeleted && AllDeleted);
----------------
Richard Smith wrote:
> Please add a FIXME saying that we assume that all lazily-declared trivial copy/move constructors are not deleted, and that this assumption might not actually be true in some corner cases.
done

================
Comment at: lib/CodeGen/ItaniumCXXABI.cpp:65-66
@@ -65,1 +64,4 @@
+      return RAA_Indirect;
+    // If the class has a destructor, the ABI requires us to pass by address.
+    if (RD->hasNonTrivialDestructor())
       return RAA_Indirect;
----------------
Richard Smith wrote:
> This should probably be part of the 'canCopy' check; the proposed fix for DR1590 also requires a trivial destructor.
I wasn't thinking about C++11 braced, copyless initialization when I named this.  I intended to share it with win64, but those rules are more complicated.  I'll fold this in.

================
Comment at: lib/CodeGen/MicrosoftCXXABI.cpp:431-435
@@ +430,7 @@
+
+    // MSVC passes objects with a trivial, non-deleted copy ctor directly.
+    for (const CXXConstructorDecl *CD : RD->ctors()) {
+      if (CD->isCopyConstructor() && CD->isTrivial() && !CD->isDeleted())
+        return RAA_Default;
+    }
+
----------------
Richard Smith wrote:
> As an optimization, you could skip this if `!hasTrivialCopyConstructor()`.
done

http://reviews.llvm.org/D3660