[Sdcc-user] z80 binary sizes increasing by 25%
Sometime after #9676 and up to and including #9682, I am seeing an increase in z80 binary sizes of 25% in some compiles. It seems to be connected to increase spill to the stack and higher use of index registers in accessing stack variables. I will try to look closer to nail down what it is but I am wondering if anyone else is seeing this. I can give a link to before and after list files showing asm generated by #9676 vs #9692 for a particular case of a binary increasing from 39k to 51k. https://drive.google.com/file/d/0B6XhJJ33xpOWLTJRTkgxRzNXQ2s/view?usp=sharing -- ___ Sdcc-user mailing list Sdcc-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/sdcc-user
Re: [Sdcc-user] z80 binary sizes increasing by 25%
On Sat, 30 Jul 2016, alvin albrecht wrote: Sometime after #9676 and up to and including #9682, I am seeing an increase in z80 binary sizes of 25% in some compiles. It seems to be connected to increase spill to the stack and higher use of index registers in accessing stack variables. I will try to look closer to nail down what it is but I am wondering if anyone else is seeing this. I can give a link to before and after list files showing asm generated by #9676 vs #9692 for a particular case of a binary increasing from 39k to 51k. https://drive.google.com/file/d/0B6XhJJ33xpOWLTJRTkgxRzNXQ2s/view?usp=shari ng Yes, I can reproduce this. It appears to be revision #9682 (the patch associated with bug #2467) where the spill location use increases. Erik -- ___ Sdcc-user mailing list Sdcc-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/sdcc-user
Re: [Sdcc-user] z80 binary sizes increasing by 25%
On 30.07.2016 10:02, Erik Petrich wrote: > > > On Sat, 30 Jul 2016, alvin albrecht wrote: > >> >> >> Sometime after #9676 and up to and including #9682, I am seeing an >> increase >> in z80 binary sizes of 25% in some compiles. It seems to be connected to >> increase spill to the stack and higher use of index registers in >> accessing >> stack variables. I will try to look closer to nail down what it is >> but I am >> wondering if anyone else is seeing this. >> >> >> I can give a link to before and after list files showing asm generated by >> #9676 vs #9692 for a particular case of a binary increasing from 39k >> to 51k. >> >> https://drive.google.com/file/d/0B6XhJJ33xpOWLTJRTkgxRzNXQ2s/view?usp=shari >> >> ng > > Yes, I can reproduce this. It appears to be revision #9682 (the patch > associated with bug #2467) where the spill location use increases. > > Erik Do we have a resonably small, self-contained C code example to reproduce this? Philipp signature.asc Description: OpenPGP digital signature -- ___ Sdcc-user mailing list Sdcc-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/sdcc-user
Re: [Sdcc-user] z80 binary sizes increasing by 25%
On Sat, 30 Jul 2016, Philipp Klaus Krause wrote:
> On 30.07.2016 10:02, Erik Petrich wrote:
>>
>>
>> On Sat, 30 Jul 2016, alvin albrecht wrote:
>>
>>>
>>>
>>> Sometime after #9676 and up to and including #9682, I am seeing an
>>> increase
>>> in z80 binary sizes of 25% in some compiles. It seems to be connected to
>>> increase spill to the stack and higher use of index registers in
>>> accessing
>>> stack variables. I will try to look closer to nail down what it is
>>> but I am
>>> wondering if anyone else is seeing this.
>>>
>>>
>>> I can give a link to before and after list files showing asm generated by
>>> #9676 vs #9692 for a particular case of a binary increasing from 39k
>>> to 51k.
>>>
>>> https://drive.google.com/file/d/0B6XhJJ33xpOWLTJRTkgxRzNXQ2s/view?usp=shari
>>>
>>> ng
>>
>> Yes, I can reproduce this. It appears to be revision #9682 (the patch
>> associated with bug #2467) where the spill location use increases.
>>
>> Erik
>
> Do we have a resonably small, self-contained C code example to reproduce
> this?
>
> Philipp
This is what I was looking at (adapted from near the beginning of Alvin's
example):
float RND0(void)
{
return 0;
}
float RND(float n)
{
return (n == 0.0) ? RND0() : RND0()*n+1.0;
}
The code for RND0() is just trivial filler and nothing unexpected happens
there, but RND() suddenly uses a sloc instead of registers in revision
#9682.
Erik
--
___
Sdcc-user mailing list
Sdcc-user@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sdcc-user
