Re: [Path,AVR]: Improve loading of 32-bit constants

[Georg-Johann Lay <avr@gjlay.de>]

Denis Chertykov wrote:
> 2011/7/6 Georg-Johann Lay <avr@gjlay.de>:
>> Denis Chertykov wrote:
>>> 2011/7/6 Georg-Johann Lay <avr@gjlay.de>:
>>>> For loading a 32-bit constant in a register, there is room for
>>>> improvement:
>>>>
>>>> * SF can be handled the same way as SI and therefore the patch
>>>>  adds a peep2 to produce a *reload_insf analogon to *reload_insi.
>>>>
>>>> * If the destination register overlaps NO_LD_REGS, values already
>>>>  loaded into some other byte can be reused by a simple MOV.
>>>>  This is helpful then moving values like, e.g. -2, -100 etc. because
>>>>  all high bytes are 0xff.
>>>>
>>>> * 0.0f can be directly moved to memory.
>>>>
>>>> * The mov insns contain "!d" constraint. I see no reason to make "d"
>>>>  expensive and discourage use of d-regs.  A "*d" to hide is better
>>>>  because it does it neither puts additional pressure on "d" nor
>>>>  discourages "d".
>>>>
>>> I would like to have a real code examples.
>>>
>>> Denis.
>> Hi Denis.
>>
>> Attached you find a small C file and the asm that is generated by new
>> and old versions (-Os -mmcu=atmega88 -S -dp).
>>
>> I took away some regs as potential clobbers (or -fno-peephole2) to
>> show the effect of high register pressure.  Bit even if a clobber was
>> available you can see that the new version is smarter in reusing
>> values, e.g. note the loading of -1L to r22-r25.
> 
> I have asked about example of *d instead of !d.
> Just svn GCC with *d vs svn GCC !d.
> 
> 
> Denis.

Ah, I couldn't depict that from your question.

I thought it could help in cases like these:

long z;

void inc (long y)
{
    z += y;
}

that gets compiled with -Os to

inc:
	push r16
	push r17
/* prologue: function */
/* frame size = 0 */
/* stack size = 2 */
.L__stack_usage = 2
	lds r16,z
	lds r17,z+1
	lds r18,z+2
	lds r19,z+3
	add r16,r22
	adc r17,r23
	adc r18,r24
	adc r19,r25
	sts z,r16
	sts z+1,r17
	sts z+2,r18
	sts z+3,r19
/* epilogue start */
	pop r17
	pop r16
	ret

But with the *d the code is still the same and R16 chosen instead of
better R18.  Maybe that's an IRA issue.

Looking again at the "*d" resp. "!d", I think the alternative is
superfluous because there is a "r" alternative and "d" is a subset of
"r", so allocator can always switch to "r" if it does not like or see "d".

I think we con remove that alternative, it's just confusing.

Johann