[Bug target/71460] Copying structs can trap (on x86-32) due to SNaN to QNaN

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71460

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jamborm at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Uroš Bizjak from comment #5)
> Following patch fixes the failure:
> 
> --cut here--
> diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
> index b807a9a..4526c7d 100644
> --- a/gcc/config/i386/i386.c
> +++ b/gcc/config/i386/i386.c
> @@ -10310,10 +10310,8 @@ ix86_promote_function_mode (const_tree type,
> machine_mode mode,
>  static bool
>  ix86_member_type_forces_blk (const_tree field, machine_mode mode)
>  {
> -  /* Union with XFmode must be in BLKmode.  */
> -  return (mode == XFmode
> -         && (TREE_CODE (DECL_FIELD_CONTEXT (field)) == UNION_TYPE
> -             || TREE_CODE (DECL_FIELD_CONTEXT (field)) == QUAL_UNION_TYPE));
> +  return (IS_STACK_MODE (mode) && flag_signaling_nans
> +         && RECORD_OR_UNION_TYPE_P (DECL_FIELD_CONTEXT (field)));
>  }
>  
>  rtx
> --cut here--
> 
> Patch was tested with various -mfpmath=... compile flags. IMO, we don't need
> BLKmode when we copy the value using SSE registers (and the test indeed does
> not fail with -mfpmath=sse), and we still can copy XFmode using x87 regs
> without -fsignalling-nans.
> 
> Richi, Joseph - does the new condition look OK to you?

Looks good to me.

Btw, the mentioned SRA issue probably still exists for full scalarization
and re-mat for a call.  SRA should probably check member_type_forces_blk
as well before creating full scalarization accesses.  Would be nice to
have a testcase for the SRA case as well.