Any idea why some 32bits architectures return in register value larger than the register ?

Any idea why some 32bits architectures return in register value larger than the register ?

[William Tambe]

When looking at backends for moxie and ft32, which are cpus for which
the return register is 32bits, TARGET_ RETURN_IN_MEMORY is implemented
such that the function value is returned in a register even when its
size is 64bits. ie:

https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/ft32/ft32.c#L63

https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L60

Any idea why it is implemented that way ?

Re: Any idea why some 32bits architectures return in register value larger than the register ?

[Jeff Law]

On 4/2/19 2:30 PM, William Tambe wrote:
> When looking at backends for moxie and ft32, which are cpus for which
> the return register is 32bits, TARGET_ RETURN_IN_MEMORY is implemented
> such that the function value is returned in a register even when its
> size is 64bits. ie:
> 
> https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/ft32/ft32.c#L63
> 
> https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L60
> 
> Any idea why it is implemented that way ?
> 
Could be that's how the ABI is defined.  It's often the case that ABI's
define double-word objects as being passed and returned in register
pairs for efficiency.

jeff

Re: Any idea why some 32bits architectures return in register value larger than the register ?

[William Tambe]

is that still ok even when TARGET_FUNCTION_VALUE, TARGET_LIBCALL_VALUE
and TARGET_FUNCTION_VALUE_REGNO_P are defined for a single register
that can only hold a single-word object ? ie:

https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L75
https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L86
https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L94

On Tue, Apr 2, 2019 at 3:37 PM Jeff Law <law@redhat.com> wrote:
>
> On 4/2/19 2:30 PM, William Tambe wrote:
> > When looking at backends for moxie and ft32, which are cpus for which
> > the return register is 32bits, TARGET_ RETURN_IN_MEMORY is implemented
> > such that the function value is returned in a register even when its
> > size is 64bits. ie:
> >
> > https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/ft32/ft32.c#L63
> >
> > https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L60
> >
> > Any idea why it is implemented that way ?
> >
> Could be that's how the ABI is defined.  It's often the case that ABI's
> define double-word objects as being passed and returned in register
> pairs for efficiency.
>
> jeff

Re: Any idea why some 32bits architectures return in register value larger than the register ?

[Jeff Law]

On 4/2/19 2:45 PM, William Tambe wrote:
> is that still ok even when TARGET_FUNCTION_VALUE, TARGET_LIBCALL_VALUE
> and TARGET_FUNCTION_VALUE_REGNO_P are defined for a single register
> that can only hold a single-word object ? ie:
> 
> https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L75
> https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L86
> https://github.com/gcc-mirror/gcc/blob/gcc-8_3_0-release/gcc/config/moxie/moxie.c#L94
Note that in the first two examples we generate a REG rtx with the given
mode.  If the given mode is larger than a single word, that's fine.  It
means a register pair (or quad or octa depending on the size of the mode).

So on a typical 32bit target you might have

(reg:SI 0)  for returning 32 bit objects
(reg:DI 0)  for returning 64 bit objects
(reg:TI 0)  for returning 128 bit objects

Jeff