[PATCH] Fix call expansion ICE (PR c++/79085)

[Jason Merrill]

LGTM.

On Thu, Mar 15, 2018 at 3:25 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> The following testcase ICEs on arm (or other strict alignment targets).
> The problem is we have a call that returns a TREE_ADDRESSABLE type,
> and the lhs of the call is some memory that isn't known to be sufficiently
> aligned.  expand_call in that case allocates a temporary, lets the call
> return into that object and copies it; this can't work for TREE_ADDRESSABLE
> types, assign_temp ICEs for those, we don't want to create temporaries in
> that case.
> This patch just avoids the temporary for TREE_ADDRESSABLE types and returns
> directly into the provided target memory.
> Either the compiler just doesn't know whether the target is properly aligned
> or not and it is properly aligned at runtime, then it will work properly,
> or it isn't properly aligned at runtime (perhaps compiler can even prove
> that) and it will be UB, which is user's fault.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2018-03-15  Jakub Jelinek  <jakub@redhat.com>
>
>         PR c++/79085
>         * calls.c (expand_call): For TREE_ADDRESSABLE rettype ignore alignment
>         check and use address of target always.
>
>         * g++.dg/opt/pr79085.C: New test.
>
> --- 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
>           {
> --- gcc/testsuite/g++.dg/opt/pr79085.C.jj       2018-03-15 15:42:05.382927611 +0100
> +++ gcc/testsuite/g++.dg/opt/pr79085.C  2018-03-15 15:32:56.626552182 +0100
> @@ -0,0 +1,24 @@
> +// PR c++/79085
> +// { dg-do compile }
> +// { dg-options "-Os" }
> +// { dg-additional-options "-mstrict-align" { target { aarch64*-*-* powerpc*-*-linux* powerpc*-*-elf* } } }
> +
> +void *operator new (__SIZE_TYPE__, void *p) { return p; }
> +
> +struct S
> +{
> +  S ();
> +  S (const S &);
> +  ~S (void);
> +  int i;
> +};
> +
> +S foo ();
> +
> +static char buf [sizeof (S) + 1];
> +
> +S *
> +bar ()
> +{
> +  return new (buf + 1) S (foo ());
> +}
>
>         Jakub