[Bug c++/79085] [6/7/8 Regression] ICE with placement new to unaligned location

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79085

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2
   Target Milestone|---                         |6.5

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started to ICE with r232167, before that it has been just silent wrong-code,
calling the function/method with a temporary and then copying it over (and not
destructing) despite perhaps having non-trivial copy ctor and/or dtor.

Untested fix:
--- gcc/calls.c.jj      2018-02-22 12:37:02.641387687 +0100
+++ gcc/calls.c 2018-03-15 15:17:42.596835316 +0100
@@ -3422,9 +3422,14 @@ expand_call (tree exp, rtx target, int i
        if (CALL_EXPR_RETURN_SLOT_OPT (exp)
            && target
            && MEM_P (target)
-           && !(MEM_ALIGN (target) < TYPE_ALIGN (rettype)
-                && targetm.slow_unaligned_access (TYPE_MODE (rettype),
-                                                  MEM_ALIGN (target))))
+           /* If rettype is addressable, we may not create a temporary.
+              If target is properly aligned at runtime and the compiler
+              just doesn't know about it, it will work fine, otherwise it
+              will be UB.  */
+           && (TREE_ADDRESSABLE (rettype)
+               || !(MEM_ALIGN (target) < TYPE_ALIGN (rettype)
+                    && targetm.slow_unaligned_access (TYPE_MODE (rettype),
+                                                      MEM_ALIGN (target)))))
          structure_value_addr = XEXP (target, 0);
        else
          {