Re: RFC: VTA alias set discrepancy

[Richard Guenther <richard dot guenther at gmail dot com>]

On Thu, Mar 18, 2010 at 4:33 PM, Aldy Hernandez <aldyh@redhat.com> wrote:
> On Thu, Mar 18, 2010 at 04:08:13PM +0100, Jakub Jelinek wrote:
>> Hi!
>>
>> 1) as cfgexpand.c will need to call the same function, the new routine
>> shouldn't live in dwarf2out.c, instead varasm.c or cfgexpand.c
>> 2) I wouldn't do this conditionally on flag_compare_debug, make_decl_rtl
>> doesn't need to be called at all in -g0 case and I think we should better
>> avoid having any aliasing changes.
>
> Fixed.
>
> Like this?
>
> ? ? ? ?* output.h: Declare make_decl_rtl_for_debug.
> ? ? ? ?* varasm.c (make_decl_rtl_for_debug): New.
> ? ? ? ?* dwarf2out.c (rtl_for_decl_location): Call it.
>
> Index: testsuite/gfortran.dg/assign-debug.f90
> ===================================================================
> --- testsuite/gfortran.dg/assign-debug.f90 ? ? ?(revision 0)
> +++ testsuite/gfortran.dg/assign-debug.f90 ? ? ?(revision 0)
> @@ -0,0 +1,8 @@
> +! { dg-do compile }
> +! { dg-options "-fcompare-debug -O2" }
> + ? ? ?program test
> + ? ? ?integer i
> + ? ? ?common i
> + ? ? ?assign 2000 to i ?! { dg-warning "Deleted feature: ASSIGN statement" }
> +2000 ?continue
> + ? ? ?end
> Index: dwarf2out.c
> ===================================================================
> --- dwarf2out.c (revision 157481)
> +++ dwarf2out.c (working copy)
> @@ -15770,10 +15770,7 @@ rtl_for_decl_location (tree decl)
> ? ? ? && !DECL_HARD_REGISTER (decl)
> ? ? ? && DECL_MODE (decl) != VOIDmode)
> ? ? {
> - ? ? ?rtl = DECL_RTL (decl);
> - ? ? ?/* Reset DECL_RTL back, as various parts of the compiler expects
> - ? ? ? ?DECL_RTL set meaning it is actually going to be output. ?*/
> - ? ? ?SET_DECL_RTL (decl, NULL);
> + ? ? ?rtl = make_decl_rtl_for_debug (decl);
> ? ? ? if (!MEM_P (rtl)
> ? ? ? ? ?|| GET_CODE (XEXP (rtl, 0)) != SYMBOL_REF
> ? ? ? ? ?|| SYMBOL_REF_DECL (XEXP (rtl, 0)) != decl)
> Index: varasm.c
> ===================================================================
> --- varasm.c ? ?(revision 157481)
> +++ varasm.c ? ?(working copy)
> @@ -1476,6 +1476,35 @@ make_decl_rtl (tree decl)
> ? if (flag_mudflap && TREE_CODE (decl) == VAR_DECL)
> ? ? mudflap_enqueue_decl (decl);
> ?}
> +
> +/* Like make_decl_rtl, but inhibit creation of new alias sets when
> + ? calling make_decl_rtl. ?Also, reset DECL_RTL before returning the
> + ? rtl. ?*/
> +
> +rtx
> +make_decl_rtl_for_debug (tree decl)
> +{
> + ?unsigned int save_aliasing_flag;
> + ?rtx rtl;

While existing callers are under !DECL_RTL_SET_P it is not obvious
that that is true here, so please add

  if (DECL_RTL_SET_P (decl))
    return DECL_RTL (decl);

here.

Please also adjust cfgexpand.c:expand_debug_expr (just see
where we do SET_DECL_RTL (.., NULL)).

Ok with that changes.

Thanks,
Richard.

> + ?/* Kludge alert! ?Somewhere down the call chain, make_decl_rtl will
> + ? ? call new_alias_set. ?If running with -fcompare-debug, sometimes
> + ? ? we do not want to create alias sets that will throw the alias
> + ? ? numbers off in the comparison dumps. ?So... clearing
> + ? ? flag_strict_aliasing will keep new_alias_set() from creating a
> + ? ? new set. ?*/
> + ?save_aliasing_flag = flag_strict_aliasing;
> + ?flag_strict_aliasing = 0;
> +
> + ?rtl = DECL_RTL (decl);
> + ?/* Reset DECL_RTL back, as various parts of the compiler expects
> + ? ? DECL_RTL set meaning it is actually going to be output. ?*/
> + ?SET_DECL_RTL (decl, NULL);
> +
> + ?flag_strict_aliasing = save_aliasing_flag;
> +
> + ?return rtl;
> +}
>
> ?/* Output a string of literal assembler code
> ? ?for an `asm' keyword used between functions. ?*/
> Index: output.h
> ===================================================================
> --- output.h ? ?(revision 157481)
> +++ output.h ? ?(working copy)
> @@ -318,6 +318,10 @@ extern bool initializer_constant_valid_f
> ? ?ALIGN is the alignment in bits that may be assumed for the data. ?*/
> ?extern void output_constant (tree, unsigned HOST_WIDE_INT, unsigned int);
>
> +extern rtx make_decl_rtl_for_debug (tree);
> +
> +/* End of functions in varasm.c. ?*/
> +
> ?/* When outputting delayed branch sequences, this rtx holds the
> ? ?sequence being output. ?It is null when no delayed branch
> ? ?sequence is being output, so it can be used as a test in the
>