[M3devel] [M3commit] CVS Update: cm3

[Jay]

Actually what I showed is frequently wrong. We often use bitfield references, which seems wierd or wrong, but seems to work generally ok and produce better code. The RIGHT thing to use would be "component refs" in gcc parlance, but currently we don't and it isn't a small change. There is kind of a mismatch in the compiler architecture currently...


 - Jay (briefly/pocket-sized-computer-aka-phone) 

On Jun 6, 2012, at 3:17 AM, Dragiša Durić <dragisha at m3w.org> wrote:

> I know that much about generated code :).
> 
> "Good" thing is - not many things changed in *m3 backend since I ported pm3 to LINUX_ALPHA :)
> 
> On Jun 6, 2012, at 11:42 AM, Jay K wrote:
> 
>> 
>>> Functions that call setjmp 
>> 
>> 
>> I meant -- functions wtih TRY/EXCEPT or TRY/FINALLY. :)
>> 
>> - Jay
>> 
>> ----------------------------------------
>>> From: jay.krell at cornell.edu
>>> To: dragisha at m3w.org
>>> Date: Wed, 6 Jun 2012 09:38:18 +0000
>>> CC: jkrell at elego.de; m3devel at elegosoft.com
>>> Subject: Re: [M3devel] [M3commit] CVS Update: cm3
>>> 
>>> 
>>> 5.8.6 does allow many optimizations to occur.
>>> We turn off a very small number directly.
>>> Functions that call setjmp have optimizations inhibited by declaring all locals volatile.
>>> We don't give the compiler good type information, and we take the address of stuff more than necessary, by
>>> generating very low level code.
>>> Where you have e.g.
>>> MODULE Foo;
>>> TYPE Point =  RECORD x,y:INTEGER END;
>>> PROCEDURE GetY(VAR pt:Point):INTEGER = BEGIN RETURN pt.y; END GetY;
>>> 
>>> 
>>> We generate the equivalent of:
>>> 
>>> 
>>> typedef ptrdiff_t INTEGER;
>>> typedef char* ADDRESS;
>>> INTEGER Foo_GetY(ADDRESS pt) { return *(INTEGER*)(pt + sizeof(INTEGER)); }
>>> 
>>> 
>>> Maybe I'll wrap up 4.6, not enable it, and move on to 4.7..
>>> 
>>> 
>>> 
>>> - Jay
>>> 
>>> 
>>> ________________________________
>>>> Subject: Re: [M3devel] [M3commit] CVS Update: cm3
>>>> From: dragisha at m3w.org
>>>> Date: Wed, 6 Jun 2012 10:51:33 +0200
>>>> CC: jkrell at elego.de; m3devel at elegosoft.com
>>>> To: jay.krell at cornell.edu
>>>> 
>>>> I am using it, and I need it.
>>>> 
>>>> Does it run better/faster? I didn't test, but is it something to even
>>>> ask, these days, architectures, … ?
>>>> 
>>>> Only if you turned everything off in 5.8.6 and later, as you'r doing it
>>>> now, then probably my "-O2" default it is of no benefit at all :).
>>>> 
>>>> Generally, our "pitch" to "sell"
>>>> super-modern-ultra-blast-mega-fast-superlative-OO and everything else
>>>> you only dreamed about… And add "no CPU optimizations"… Imagine that.
>>>> 
>>>> On Jun 6, 2012, at 10:10 AM, Jay K wrote:
>>>> 
>>>> 7) Do folks out there really use the Modula-3/gcc optimizer, and notice
>>>> it produces code that runs much faster?
>>>> 
>>> 
>>                         
>