Re: [PATCH] loading float member of parameter stored via int registers

[Jiufu Guo <guojiufu@linux.ibm.com>]

Hi,

Segher Boessenkool <segher@kernel.crashing.org> writes:

> Hi!
>
> On Fri, Dec 23, 2022 at 08:36:36PM +0800, Jiufu Guo wrote:
>> It seems some limitations there. e.g. 1. "subreg:DF on DI register"
>> may not work well on pseudo,
>
> It is perfectly normal:
>   A hard register may be accessed in various modes throughout one
>   function, but each pseudo register is given a natural mode
>   and is accessed only in that mode.  When it is necessary to describe
>   an access to a pseudo register using a nonnatural mode, a @code{subreg}
>   expression is used.
>
> and:
>   @code{subreg} expressions are used to refer to a register in a machine
>   mode other than its natural one, or to refer to one register of
>   a multi-part @code{reg} that actually refers to several registers.
>
>   Each pseudo register has a natural mode.  If it is necessary to
>   operate on it in a different mode, the register must be
>   enclosed in a @code{subreg}.
>
> and we even have:
>   @item hard registers
>   It is seldom necessary to wrap hard registers in @code{subreg}s; such
>   registers would normally reduce to a single @code{reg} rtx.  This use of
>   @code{subreg}s is discouraged and may not be supported in the future.
>
Thanks so much for detailed explaination!

>> and 2. to convert high-part:DI to SF,
>> a "shift/rotate" is needed, and then we need to "emit shift insn"
>> in cse. I may need to update this patch.
>
> Hrm.  The machine insns to do this is just mtvsrd;xscvspdpn, but for
> converting the lowpart it is mtvsrws;xscvspdpn (this needs p9 or
> later).  We should arrive at those patterns, and we should try to not
> go via the more expensive formulations with shifts, which don't describe
> the hardware well, and which overestimate the cost of it.
Yes, understant!
>
> None of this belongs in generic code at all imo.  At expand time it
> should be expanded to something that works and can be optimised well,
> so not anything with :BLK (which has to be put in memory, something with
> unbounded size cannot be put in registers), not anything specifically
> tailored to any cpu, something nice and regular.  Using a subreg (of a
> pseudo!) is the standard way of writing a bitcast.
>
> So generic code would do a  (subreg:SF (reg:SI) 0)  to express a 32-bit
> integer bitcast to an IEEE SP number, and our machine description should
> make it work nicely.
Right!  So, I'm thinking a way: in generic code, we may generated
"shift+(subreg:SF (reg:SI) 0)"; and at somewhere (maybe in combiner),
using "mtvsr.." to replace the "shift+subreg".

BR,
Jeff (Jiufu)

>
>
> Segher