Re: RFC: Representing vector lane load/store operations

[Richard Guenther <richard.guenther@gmail.com>]

On Tue, Mar 29, 2011 at 2:44 PM, Richard Sandiford
<richard.sandiford@linaro.org> wrote:
> Richard Sandiford <richard.sandiford@linaro.org> writes:
>> Richard Guenther <richard.guenther@gmail.com> writes:
>>> Well, I meant if the user compiles with -msse, declares such a
>>> global var (which means it gets V4SFmode and not BLKmode)
>>> and then uses it in a function where he explicitly disables SSE
>>> then something is wrong.  If he declares a BLKmode global
>>> then generic vector support will happily trigger and make it work.
>>
>> Ah, OK.  I'm just not sure whether, to take a MIPS example,
>> MIPS16 functions in a "-mno-mips16" compile should behave
>> differently from unannotated functions in a "-mips16" compile.
>>
>>> If it's just three element array-of-vector types you need why not expose
>>> it via attribute((mode(xyz))) only?  You could alias that mode to BLKmode
>>> if neon is not enabled ...
>>
>> I don't think that really changes anything.  Getting the non-BLK mode
>> on the array type seems like the easy part.  The difficult part is
>> dealing with the fallout when the array is defined in a Neon context
>> and used in a non-Neon context.
>
> As a follow-up to this, I think the current definition of TYPE_MODE
> is too restrictive even for the vector case.  Single-element structures
> get the modes of their fields, and similarly for arrays.  So if we modify
> the original 38240 testcase a bit, we still get a difference:
>
> -------------------------------------------------------------------------
> #if STRUCT
> typedef struct {
>  float x __attribute__ ((__vector_size__ (16), __may_alias__));
> } V;
> #else
> typedef float V __attribute__ ((__vector_size__ (16), __may_alias__));
> #endif
>
> V __attribute__((target("sse"))) f(const V *ptr) { return *ptr; }
> -------------------------------------------------------------------------
>
> Without -DSTRUCT, this generates the same code regardless of whether
> you compile with -msse.  But with -DSTRUCT, you get:
>
>        movaps  (%rdi), %xmm0
>        ret
>
> with -msse and:
>
>        movq    (%rdi), %rax
>        movq    %rax, -24(%rsp)
>        movq    8(%rdi), %rax
>        movq    %rax, -16(%rsp)
>        movdqa  -24(%rsp), %xmm0
>        ret
>
> with -mno-sse.
>
> I think your argument is that most/all uses of TYPE_MODE are a mistake.
> But I still think it makes sense to say that types have a natural mode
> _in a given context_, just not globally.  So how about replacing it with
> a current_mode_of_type function?  That makes it obvious that TYPE_MODE is
> not a global property, and that it isn't really a simple accessor any more.
> We could then make it recompute the mode for all types, possibly with a
> cache if that's necessary for performance reasons.

Well, ok.  That current_mode_of_type wouldn't make sense when for
example expanding global initializers (neither would looking at TYPE_MODE).
But - what's the natural mode to choose for global entities?  After all
we have to stick something into TYPE_MODE and DECL_MODE.

But yes, changing the TYPE_MODE users over to current_mode_of_type
(or rather mode_of_type_in_fn (struct function *, tree)) would be nice
(and then get rid of the TYPE_MODE hack).

Richard.

> Richard
>