On Tue, 14 Jun 2011, Mike Stump wrote:
> On Jun 13, 2011, at 3:57 AM, Richard Guenther wrote:
> > That's not exactly an example - I can't think of how you want or need
> > to use VIEW_CONVERT_EXPRs to implement said divmod instruction or why
> > you would need anything special for the _argument_ of said instruction.
>
> Oh, I completely misunderstood your question. In my case, as I previously
> stated, was with a vector type that was identical, save the name of the type:
>
> mod = a%b
>
> where mod didn't have the type of the expression (a%b), so someone created
> the VIEW_CONVERT_EXPR on the mod. The person creating it _thought_ it would
> be a rvalue context, but ultimately, it was an lvalue context. We discover
> the lvalue/rvalue state of the expression at target_fold_builtin time. The
> actual code looks more like:
>
> __builtin_divmod (div, mod, a, b);
>
> In fold_builtin, we do all the processing to handle the semantics.
>
> > An
> > instruction or call with multiple outputs would simply be something
> > like
> >
> > { div_1, mod_2 } = __builtin_divmod (arg_3);
> >
> > with two SSA defs. A nice representation for the tree for { div_1,
> > mod_2 } remains to be found (if it should be a single tree at all, or
> > possibly multiple ones).
>
> At target_fold_builtin time we regenerate it as:
>
> s = builtin_divmod_final (a, b);
> div_1 = s.div
> mod_2 = s.mod
>
> and generate a type { div, mod } on the fly. We expect the optimizer to
> handle extra moves reasonably, and we want to keep the one instruction as one
> unit.
>
> > We already play tricks for sincos for example via
> >
> > tem_1 = __builtin_cexpi (arg_2);
> > sin_3 = REALPART_EXPR <tem_1>;
> > cos_4 = IMAGPART_EXPR <tem_1>;
> >
> > which avoids the two defs by using a single def which is then decomposed.
> >
> > So, can you elaborate a bit more on what you want to do with special
> > argument kinds? Elaborate with an actual example, not words.
>
> We support tagging any parameter to a builtin as define_outputs,
> define_inputs or define_in_outs in a part of the .md file that describes the
> builtins for the machine, the actual divmod builtin for example is:
>
> (define_builtin "divmod<T_ALL_DI:sign_u>" "divmod<T_ALL_DI:sign_u>_<type>"
> [
> (define_outputs [(var_operand:T_ALL_DI 0) ;;dividend
>
> (var_operand:T_ALL_DI 1)]) ;;mod
>
> (define_inputs [(var_operand:T_ALL_DI 2)
> (var_operand:T_ALL_DI 3)])
> (define_rtl_pattern "<T_ALL_DI:sign_u>divmod<m_mode>4" [0 1 2 3])
> (attributes [pure])
> ]
> )
>
> that's the actual code. The testcase looks like:
>
> t_v4udi_0 = divmodu_t_v4udi (t_v4udi_1, t_v4udi_2, t_v4udi_3);
I see. So it's the C side of the representation that needs the
special operands.
> The VIEW_CONVERT_EXPR looks like:
>
> <view_convert_expr 0x7ffff5a872d0
> type <vector_type 0x7ffff7f4b930 __attribute__((vector_size(32)))
> unsigned long
> type <integer_type 0x7ffff7e8c690 long unsigned int public unsigned DI
> size <integer_cst 0x7ffff7e76730 constant 64>
> unit size <integer_cst 0x7ffff7e76758 constant 8>
> align 64 symtab 0 alias set -1 canonical type 0x7ffff7e8c690
> precision 64 min <integer_cst 0x7ffff7e76780 0> max\
> <integer_cst 0x7ffff7e76708 18446744073709551615>
> pointer_to_this <pointer_type 0x7ffff7f50738> reference_to_this
> <reference_type 0x7ffff7f513f0>>
> unsigned V4DI
> size <integer_cst 0x7ffff7f43de8 constant 256>
> unit size <integer_cst 0x7ffff7e76348 constant 32>
> align 256 symtab 0 alias set -1 canonical type 0x7ffff7f4b930 nunits
> 4 reference_to_this <reference_type 0x7ffff7f51\
> 540>>
>
> arg 0 <var_decl 0x7ffff59d9640 t_v4udi_1
> type <vector_type 0x7ffff5ac3888 type <integer_type 0x7ffff7e8c690
> long unsigned int>
> unsigned V4DI size <integer_cst 0x7ffff7f43de8 256> unit size
> <integer_cst 0x7ffff7e76348 32>
> align 256 symtab 0 alias set -1 canonical type 0x7ffff5ac3888
> nunits 4>
> used public static unsigned V4DI defer-output file t22.c line 262 col
> 48 size <integer_cst 0x7ffff7f43de8 256> unit \
> size <integer_cst 0x7ffff7e76348 32>
> align 256>>
This VIEW_CONVERT_EXPR looks useless - in fact useless_type_conversion_p
will tell you that, so you can omit it.
Richard.
