[Bug tree-optimization/66768] address space gets lost on literal pointer

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66768

--- Comment #9 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 7 Jul 2015, amker at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66768
> 
> --- Comment #8 from amker at gcc dot gnu.org ---
> So address space info is kept and checked in base object's type of MEM_REF.  As
> in function expand_expr_real_1:
> 
>     case TARGET_MEM_REF:
>       {
>         addr_space_t as
>           = TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (TREE_OPERAND (exp, 0))));
>         enum insn_code icode;
>         unsigned int align;
> 
> The AVR ICE happens when base of MEM_REF[base:0, index:(sizetype)ivtmp] has
> __memx address space attribute.  Since pointer to memory object in __memx
> address space (PSImode) has 24 bits type length, while the type of index is
> sizetype(Pmode == HImode in this case), which has 16 bits type length.  The
> expression "(sizetype)ivtmp" is expanded into (subreg:HI (reg:PSI 40) 0).
> 
> So I still think IVO should distribute ivtmp as base part of MEM_REF since it
> stands for a memory object.  Otherwise, we have below IVOed code:
> 
>   <bb 3>:
>   # total_10 = PHI <total_5(4), 0(2)>
>   # ivtmp.7_8 = PHI <ivtmp.7_7(4), 4660(2)>
>   _12 = (sizetype) ivtmp.7_8;
>   _4 = MEM[base: 0B, index: _12, offset: 0B];
>   total_5 = _4 + total_10;
>   ivtmp.7_7 = ivtmp.7_8 + 2;
>   if (ivtmp.7_7 != 4700)
>     goto <bb 4>;
>   else
>     goto <bb 5>;
> 
>   <bb 4>:
>   goto <bb 3>;
> 
> 
> This is wrong since truncation of ivtmp.7_8 to sizetype could result in wrong
> address.

Indeed truncating it to sizetype just to make it fit into INDEX
looks wrong (extension is ok, truncation is not).