Re: [PATCH] rs6000: Use NO_EXPR to cast to MMA pointer types

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

> Am 05.09.2022 um 16:53 schrieb Segher Boessenkool <segher@kernel.crashing.org>:
> 
> On Mon, Sep 05, 2022 at 11:25:21AM +0200, Richard Biener wrote:
>>> On Fri, Sep 2, 2022 at 7:24 PM Segher Boessenkool
>>> <segher@kernel.crashing.org> wrote:
>>> On Fri, Sep 02, 2022 at 12:02:54PM -0500, Peter Bergner wrote:
>>>> On 9/2/22 11:31 AM, Segher Boessenkool wrote:
>>>>> (Did you also look at non-MMA VIEW_CONVERT_EXPR uses btw?)
>>>> 
>>>> I did.  It seemed they were all related to pointers to vectors and I remember
>>>> you mentioning that as one of the reasons for using VIEW_CONVERT_EXPR over
>>>> NOP_EXPR, so I left them alone to be safe.
>>> 
>>> Huh?  I have no idea what you mean here.
>>> 
>>> Casting from one pointer type to another never needs it.  Casting from a
>>> scalar integer type to a pointer type not either AFAIKi.  But I am not a
>>> Gimple expert, all this might be wrong, it isn't documented anywbere :-(
>> 
>> NOP_EXPR is for conversions between types with the same kind
>> (and pointer-to-integer and integer-to-pointer
>> conversions when pointer and integer are of the same size).
>> When used on vectors it converts the vector elements.  When you want to
>> re-interpret V4SI as V4SF you need VIEW_CONVERT_EXPR (bit_cast),
>> likewise V4SI interpreted as V16QI needs that.
>> 
>> Think of VIEW_CONVERT as bit_cast and NOP_EXPR as conversion.
>> Of course for some conversions (like unsigned int to int) you can also
>> use a VIEW_CONVERT since it's semantically the same.  In those cases
>> we canonicalize to NOP_EXPR via folding.
> 
> Thanks for the explanation!
> 
> About that last point...  You say VIEW_CONVERT_EXPR is folded to
> NOP_EXPR where possible.  Does that happen in all cases / can we depend
> on that?  So that in target code like what started this thread the only
> real difference is documentation of intent?  (Which never is unimportant
> of course!)

You should be able to depend on that, yes.

Richard 

> 
> 
> Segher