On VIEW_CONVERT_EXPRs and TBAA

On VIEW_CONVERT_EXPRs and TBAA

[Richard Guenther]

While looking at PR38747 again I was diving into the alias.c code
and tried to confirm that it does what I think it does for 
VIEW_CONVERT_EXPR.  And in fact it basically does.

In general there are odds between what get_alias_set does if you
pass it a reference tree compared to what it does if you pass it
a type.  In particular whether it returns the alias-set of the
innermost or the outermost aliased component is not really consistent.

But - onto that VIEW_CONVERT_EXPR.  VIEW_CONVERT_EXPR is basically
ignored / skipped, which means that if you have C++ code that does
*(type2 *)&X and translate it to VIEW_CONVERT_EXPR <type2> (X) then
you have generated wrong code (this is PR38747, forwprop does that).

With VIEW_CONVERT_EXPR you can also easily create the situation
where for a reference tree, let it be VIEW_CONVERT_EXPR <T2> (X.a).b
like commonly seen in Ada, the alias-set of the outermost component
is not a subset of that of the innermost one (the relationship that
is usually assured to be true by the record_component_aliases
machinery).  You are probably safe here if only your frontend generates
such conversions and it is very consistent on how it balances
V_C_Es with accesses through pointers (because in the above case
if .b is an addressable component an access via a pointer to type
of b wouldn't necessarily alias the cited reference tree).

The alias-oracle tries to use both the outermost and innermost
alias-sets for disambiguations, but relies on the outermost alias-set
being a subset of the innermost one - a relationship that V_C_E can
break.

So, what I'd like to know is in which circumstances the Ada
frontend uses VIEW_CONVERT_EXPRs and what counter-measures it
applies to ensure consistent TBAA.

Thanks,
Richard.

Re: On VIEW_CONVERT_EXPRs and TBAA

[Eric Botcazou]

> With VIEW_CONVERT_EXPR you can also easily create the situation
> where for a reference tree, let it be VIEW_CONVERT_EXPR <T2> (X.a).b
> like commonly seen in Ada, the alias-set of the outermost component
> is not a subset of that of the innermost one (the relationship that
> is usually assured to be true by the record_component_aliases
> machinery).  You are probably safe here if only your frontend generates
> such conversions and it is very consistent on how it balances
> V_C_Es with accesses through pointers (because in the above case
> if .b is an addressable component an access via a pointer to type
> of b wouldn't necessarily alias the cited reference tree).

I presume you're talking exclusively about the addressable case here, because
in Ada the alias set of (the type of) the outermost component is not a subset 
of (that of) the innermost one by default (at least for scalar types, it is 
for aggregate types because of implementation limitations), there being a VCE 
or not; the field must explicitly be declared addressable.

> So, what I'd like to know is in which circumstances the Ada
> frontend uses VIEW_CONVERT_EXPRs and what counter-measures it
> applies to ensure consistent TBAA.

The Ada compiler generates a lot of VIEW_CONVERT_EXPRs for conversions between 
aggregate (sub)types because it generates a lot of aggregate subtypes from a 
given aggregate type.  These conversions are purely formal and their purpose 
is to ensure type consistency.  The Ada compiler also uses VIEW_CONVERT_EXPR
to implement up-casting

  type Base is tagged record
    I : Integer;
  end record;

  type Derived is new Base with record
    J : Integer;
  end record;

  D : Derived;
  Base (D).I 

is translated into VIEW_CONVERT_EXPR<base>(d).i (a questionable use of VCE, no 
doubt about that).  There are a few other, more obscur uses, but we get rid 
of the most controversial one (optimization barrier on scalars for VRP).

The main countermeasure to ensure consistent TBAA is decl.c:relate_alias_sets 
which is aimed at doing the right thing when types are VCEd to each other.

-- 
Eric Botcazou

Re: On VIEW_CONVERT_EXPRs and TBAA

[Richard Guenther]

On Sat, 26 Sep 2009, Eric Botcazou wrote:

> > With VIEW_CONVERT_EXPR you can also easily create the situation
> > where for a reference tree, let it be VIEW_CONVERT_EXPR <T2> (X.a).b
> > like commonly seen in Ada, the alias-set of the outermost component
> > is not a subset of that of the innermost one (the relationship that
> > is usually assured to be true by the record_component_aliases
> > machinery).  You are probably safe here if only your frontend generates
> > such conversions and it is very consistent on how it balances
> > V_C_Es with accesses through pointers (because in the above case
> > if .b is an addressable component an access via a pointer to type
> > of b wouldn't necessarily alias the cited reference tree).
> 
> I presume you're talking exclusively about the addressable case here, because
> in Ada the alias set of (the type of) the outermost component is not a subset 
> of (that of) the innermost one by default (at least for scalar types, it is 
> for aggregate types because of implementation limitations), there being a VCE 
> or not; the field must explicitly be declared addressable.

Yes, addressable in the way that get_alias_set would not skip the
component as component_uses_parent_alias_set.

> > So, what I'd like to know is in which circumstances the Ada
> > frontend uses VIEW_CONVERT_EXPRs and what counter-measures it
> > applies to ensure consistent TBAA.
> 
> The Ada compiler generates a lot of VIEW_CONVERT_EXPRs for conversions between 
> aggregate (sub)types because it generates a lot of aggregate subtypes from a 
> given aggregate type.  These conversions are purely formal and their purpose 
> is to ensure type consistency.  The Ada compiler also uses VIEW_CONVERT_EXPR
> to implement up-casting
> 
>   type Base is tagged record
>     I : Integer;
>   end record;
> 
>   type Derived is new Base with record
>     J : Integer;
>   end record;
> 
>   D : Derived;
>   Base (D).I 
> 
> is translated into VIEW_CONVERT_EXPR<base>(d).i (a questionable use of VCE, no 
> doubt about that).

C++ for this case has a FIELD_DECL in Derived of type Base, so the
middle-end sees d.<anonymous>.i here (and thus also automatically
gets the TBAA hierarchy correct by means of recording component aliases).

The Ada way doesn't look too obscure to me.

> There are a few other, more obscur uses, but we get rid 
> of the most controversial one (optimization barrier on scalars for VRP).
> 
> The main countermeasure to ensure consistent TBAA is decl.c:relate_alias_sets 
> which is aimed at doing the right thing when types are VCEd to each other.

Ok, thanks.

Richard.

Re: On VIEW_CONVERT_EXPRs and TBAA

[Eric Botcazou]

> C++ for this case has a FIELD_DECL in Derived of type Base, so the
> middle-end sees d.<anonymous>.i here (and thus also automatically
> gets the TBAA hierarchy correct by means of recording component aliases).

Ada has a _Parent FIELD_DECL in Derived (of type a subtype of Base) so it 
should be able to do d._Parent.i and this would be a lot nicer IMO.

-- 
Eric Botcazou