On Fri, Sep 12, 2014 at 2:14 PM, Bernd Schmidt <ber...@codesourcery.com> wrote: > On 09/12/2014 01:45 PM, Richard Biener wrote: >> >> Fixing up the vector type in advance is ok with me but I'd like us to >> move away from address-space-on-types. > > > After thinking about it for a while, this idea makes no sense. Address > spaces must be represented in the type system somehow - consider a pointer > to an object in address space 0 vs. a pointer to an object in address space > 1. These are different types, they may even have different sizes. > > So by adding address spaces to references (_DECLs and _REFs) the only thing > we'd accomplish is duplicating existing information, with enhanced chances > of getting inconsistencies.
Well. There are two parts of adress-space suport. First is pointer types which may have different size/mode. Second is memory references to different address-spaces which may require different insns in the end (RTL). On RTL we get away with "lowering" pointers properly (the size/precision should be encoded correctly on the tree/GIMPLE level as well). And on RTL we have the address-space of a MEM in its MEM_ATTRs. On GIMPLE we weirdly use some TYPE_QUALS on some type contained in a memory reference tree. I'd like to fix the latter by placing address-space info on the reference itself (like on RTL), not on the types. Conveniently that would be on the base object we access which is either a DECL or a MEM_REF/TARGET_MEM_REF. Yes, the _frontends_ would be required to properly build memory references to objects in different address-spaces. Richard. > > Bernd >
