[Bug tree-optimization/87325] determine_base_object does not consider consider address space of base expr

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87325

--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 18 Sep 2018, amker at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87325
> 
> --- Comment #2 from bin cheng <amker at gcc dot gnu.org> ---
> Seems alloc_iv doesn't infer base object correctly:
>   /* Lower address expression in base except ones with DECL_P as operand.
>      By doing this:
>        1) More accurate cost can be computed for address expressions;
>        2) Duplicate candidates won't be created for bases in different
>           forms, like &a[0] and &a.  */
>   STRIP_NOPS (expr);
>   if ((TREE_CODE (expr) == ADDR_EXPR && !DECL_P (TREE_OPERAND (expr, 0)))
>       || contain_complex_addr_expr (expr))
>     {
>       aff_tree comb;
>       tree_to_aff_combination (expr, TREE_TYPE (expr), &comb);
>       base = fold_convert (TREE_TYPE (base), aff_combination_to_tree (&comb));
>     }
> 
>   iv->base = base;
>   iv->base_object = determine_base_object (base);
> 
> Given:
> (gdb) call debug_generic_expr(expr)
> &ts[1].y
> 
> in which "ts" is a VAR_DECL in specific address space:
> (gdb) call debug_generic_expr(*addr_p)
> ts
> (gdb) p (*addr_p)->base.code
> $8 = VAR_DECL
> (gdb) p (*addr_p)->typed.type->base.u.bits.address_space
> $10 = 7
> 
> Question is, should be build appropriate pointer type for "ts[1].y", for now, 
>       base = build_fold_addr_expr (base);
> returns:
> (gdb) call debug_generic_expr(base)
> &ts[1].y
> (gdb) call debug_generic_expr(base->typed.type)
> float *

I think build_fold_addr_expr handles things correctly since it uses
built_pointer_type which should handle address-spaces correctly.

The question is probably whether ts[1].y has the proper address-space
qualified type (which IIRC it doesn't ...).  So we'd need sth
like ref_alias_ptr_type - a ref_addr_space () we can use to extract
the address-space from an object and transfer that properly in
build_fold_addr_expr (and elsewhere we run into this issue).  Whereever
RTL expansion takes the address-space from is probably the code to
mimic in said new API.