2013/11/11 Richard Biener <richard.guent...@gmail.com>: > On Mon, Nov 11, 2013 at 3:00 PM, Ilya Enkovich <enkovich....@gmail.com> wrote: >> 2013/11/11 Richard Biener <richard.guent...@gmail.com>: >>> On Fri, Nov 8, 2013 at 11:03 AM, Ilya Enkovich <enkovich....@gmail.com> >>> wrote: >>>> 2013/11/8 Richard Biener <richard.guent...@gmail.com>: >>>>> On Thu, Nov 7, 2013 at 7:55 PM, Jeff Law <l...@redhat.com> wrote: >>>>>> On 11/07/13 04:50, Ilya Enkovich wrote: >>>>>>> >>>>>>> Hi, >>>>>>> >>>>>>> Here is an updated patch version. >>>>>> >>>>>> I think this needs to hold until we have a consensus on what the >>>>>> parameter >>>>>> passing looks like for bounded pointers. >>>>> >>>>> I still think the best thing to do on GIMPLE is >>>>> >>>>> arg_2 = __builtin_ia32_bnd_arg (arg_1(D)); >>>>> foo (arg_2); >>>>> >>>>> that is, make the parameter an implicit pair of {value, bound} where >>>>> the bound is determined by the value going through a bound association >>>>> builtin. No extra explicit argument to the calls so arguments match >>>>> the fndecl and fntype. All the complexity is defered to the expander >>>>> which can trivially lookup bound arguments via the SSA def (I suppose >>>>> it does that anyway now for getting at the explicit bound argument now). >>>>> >>>>> As far as I can see (well, think), all currently passed bound arguments >>>>> are the return value of such builtin already. >>>> >>>> All bounds are result of different builtin calls. Small example: >>>> >>>> int *global_p; >>>> void foo (int *p) >>>> { >>>> int buf[10]; >>>> bar (p, buf, global_p); >>>> } >>>> >>>> >>>> It is translated into: >>>> >>>> __bound_tmp.1_7 = __builtin_ia32_bndmk (&buf, 40); >>>> __bound_tmp.1_6 = __builtin_ia32_arg_bnd (p_3(D)(ab)); >>>> global_p.0_2 = global_p; >>>> __bound_tmp.1_8 = __builtin_ia32_bndldx (&global_p, global_p.0_2); >>>> bar (p_3(D)(ab), __bound_tmp.1_6, &buf, __bound_tmp.1_7, >>>> global_p.0_2, __bound_tmp.1_8); >>>> >>>> Bounds binding via calls as you suggest may be done as following: >>>> >>>> __bound_tmp.1_7 = __builtin_ia32_bndmk (&buf, 40); >>>> __bound_tmp.1_6 = __builtin_ia32_arg_bnd (p_3(D)(ab)); >>>> global_p.0_2 = global_p; >>>> __bound_tmp.1_8 = __builtin_ia32_bndldx (&global_p, global_p.0_2); >>>> _9 = __builtin_bind_bounds (p_3(D)(ab), __bound_tmp.1_6); >>>> _10 = __builtin_bind_bounds (&buf, __bound_tmp.1_7); >>>> _11 = __builtin_bind_bounds (global_p.0_2, __bound_tmp.1_8); >>>> bar (_9, _10, _11); >>>> >>>> Is it close to what you propose? >>> >>> Yes. >> >> Is there a way to bind bounds with structures in a similar way? > > Not to have them easy to lookup in the SSA web. A long time ago > I proposed to make SSA aggregates possible, so you could do > > tem_2 = __internal_bind_bounds (aggr(D), __bound_tmp.1_3, > __bound_tmp.1_4, ...); > bar (tem_2); > > (originally the SSA aggregates were supposed to make copy-propgagation > possible using the SSA copy propagator, and of course I needed it for > the middle-end array work) > > Not sure if that will give other people the creeps (expand would > never expand the "load" from tem_2 but instead handle aggr as > parameter). A slight complication is due to alias analysis > which would be need to taught that bar performs a load of aggr.
It would require bounds loading for aggr before __internal_bind_bounds anyway. So, why not to do it in expand? I just need to mark calls with a flag (which you've proposed few times already) to let expand know when it should load bounds. Having SSA aggregates would be nice but I suspect it has much higher impact then loading bounds in expand. I want to try simpler variant first. Thanks, Ilya > > Richard. > >> For >> SSA names I may easily find definition and check if it is a binding >> builtin call. But for structures it is not possible. The way I see it >> to mark all such args as addressable and load required bounds on >> expand pass. >> >> Ilya >>> >>> Richard. >>> >>>> Ilya >>>>> >>>>> Richard. >>>>> >>>>> >>>>> >>>>>> Jeff