Re: help on adding new insns like min/max

Re: help on adding new insns like min/max

[Dale Johannesen]

On Monday, October 21, 2002, at 03:07  PM, Ming Ouyang wrote:

> Hi everyone,
>
> I need to add new insns support to our gcc 3.0 to support min
> and max insns, but it seems that I can't describe such function-level
> insns on RTL level, probably I need to change something in c-common.c
> or something like it?

The patterns smin<type>, umin<type>, smax<type>, umax<type> are 
supported
in RTL.

help on adding new insns like min/max

[Ming Ouyang]

Hi everyone,

I need to add new insns support to our gcc 3.0 to support min 
and max insns, but it seems that I can't describe such function-level 
insns on RTL level, probably I need to change something in c-common.c 
or something like it?

Does anyone have any idea on that? Thanks a lot.

Ming


__________________________________________________
Do you Yahoo!?
Y! Web Hosting - Let the expert host your web site
http://webhosting.yahoo.com/

Re: help on adding new insns like min/max

[Jan Hubicka]

> Hi everyone,
> 
> I need to add new insns support to our gcc 3.0 to support min 
> and max insns, but it seems that I can't describe such function-level 

The min and max instructions can be discovered by GCC automatically from
"C" alternatives in if conversion pass.  Alternatively you can add new
builtins for that, but that needs gcc 3.1 do be done cleanly I believe.

Honza

> insns on RTL level, probably I need to change something in c-common.c 
> or something like it?
> 
> Does anyone have any idea on that? Thanks a lot.
> 
> Ming
> 
> 
> __________________________________________________
> Do you Yahoo!?
> Y! Web Hosting - Let the expert host your web site
> http://webhosting.yahoo.com/

Re: help on adding new insns like min/max

[Ming Ouyang]

Dale,

Thanks for your reply, but I wonder how gcc will handle code like:
c = min(a, b)
Does min/max will be treated as builtins? Or can you demonstrate
how to make gcc to use sminsi3/smaxsi3... to use min/max insns for 
code like that?
Thanks a lot.

Ming

--- Dale Johannesen <dalej@apple.com> wrote:
> 
> On Monday, October 21, 2002, at 03:07  PM, Ming Ouyang wrote:
> 
> > Hi everyone,
> >
> > I need to add new insns support to our gcc 3.0 to support min
> > and max insns, but it seems that I can't describe such function-level
> > insns on RTL level, probably I need to change something in c-common.c
> > or something like it?
> 
> The patterns smin<type>, umin<type>, smax<type>, umax<type> are 
> supported
> in RTL.
> 


__________________________________________________
Do you Yahoo!?
Y! Web Hosting - Let the expert host your web site
http://webhosting.yahoo.com/

Re: help on adding new insns like min/max

[Ming Ouyang]

I also found that just now :) Anyway, thank you very very much.

Ming

--- Andrew Pinski <apinski@apple.com> wrote:
> Actually I messed up the definition of max:
> 
> #define max(a,b) ({__typeof__(a) a1=(a); __typeof__(b) b1 = (b); 
> (a1>b1)?a1:b1})
> 
> Sorry,
> Andrew Pinski
> 
> 
> On Monday, Oct 21, 2002, at 18:03 US/Pacific, Ming Ouyang wrote:
> 
> > Andrew,
> > Thanks for your reply. __typeof__ does make 2 or more min/max in one
> > function work, but it can't resolve complicated operands problem yet,
> > can you help on that?
> > Thanks so much.
> >
> > Ming
> >
> >
> > --- Andrew Pinski <apinski@apple.com> wrote:
> >> try this, using extensions for gcc:
> >>
> >> #define max(a,b) ({__typeof__(a) a1=(a); __typeof__(b) b1 = (b);
> >> (a>b)?a:b})
> >>
> >> Thanks,
> >> Andrew Pinski
> >>
> >> On Monday, Oct 21, 2002, at 17:34 US/Pacific, Ming Ouyang wrote:
> >>
> >>> Dale,
> >>> I just found that macro doesn't always work well, for any complicated
> >>> operands like: max(a, ++b), max(a+b, c), max(a, foo(c)), gcc won't
> >>> use max insn for them, it only use max insn for max(a,b).
> >>> In addition, if I have more than one maxs/mins, it seems gcc will
> >>> not generate max/min for them, could you give me some help on that?
> >>> Thanks a lot.
> >>>
> >>> Ming
> >>>
> >>> --- Dale Johannesen <dalej@apple.com> wrote:
> >>>>
> >>>> On Monday, October 21, 2002, at 04:05  PM, Ming Ouyang wrote:
> >>>>
> >>>>> I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
> >>>>> all debug info, gcc always treats min as a function call in my
> >>>>> test.c.00.rtl through test.c.30.dbr:
> >>>>>
> >>>>> (call (mem:SI (symbol_ref/v:SI ("smin")) ))
> >>>>>
> >>>>> I think once gcc found that is a function call, it would not
> >>>>> optimize it, am I right?
> >>>>> What I should get is something like "(smin: SI ...", is that right?
> >>>>> So probably I need to change something other than xxx.md?
> >>>>> Thanks.
> >>>>
> >>>> If you're using it as a function of course gcc will treat it as
> >>>> a function.  It is not one of the standard C functions so gcc does
> >>>> not do anything special to it (and is not allowed to, that would
> >>>> break standard conformance).  The usual way to get min/max
> >>>> functionality
> >>>> in C is to put a macro definition in a header file somewhere,
> >>>> something
> >>>> like
> >>>> #define MAX(x,y) ((x>y) ? x : y)
> >>>> (watch out for ++ operators, improvement left as an exercise)
> >>>>
> >>>
> >>>
> >>> __________________________________________________
> >>> Do you Yahoo!?
> >>> Y! Web Hosting - Let the expert host your web site
> >>> http://webhosting.yahoo.com/
> >>>
> >>
> >
> >
> > __________________________________________________
> > Do you Yahoo!?
> > Y! Web Hosting - Let the expert host your web site
> > http://webhosting.yahoo.com/
> >
> 


__________________________________________________
Do you Yahoo!?
Y! Web Hosting - Let the expert host your web site
http://webhosting.yahoo.com/

Re: help on adding new insns like min/max

[Andrew Pinski]

Actually I messed up the definition of max:

#define max(a,b) ({__typeof__(a) a1=(a); __typeof__(b) b1 = (b); 
(a1>b1)?a1:b1})

Sorry,
Andrew Pinski


On Monday, Oct 21, 2002, at 18:03 US/Pacific, Ming Ouyang wrote:

> Andrew,
> Thanks for your reply. __typeof__ does make 2 or more min/max in one
> function work, but it can't resolve complicated operands problem yet,
> can you help on that?
> Thanks so much.
>
> Ming
>
>
> --- Andrew Pinski <apinski@apple.com> wrote:
>> try this, using extensions for gcc:
>>
>> #define max(a,b) ({__typeof__(a) a1=(a); __typeof__(b) b1 = (b);
>> (a>b)?a:b})
>>
>> Thanks,
>> Andrew Pinski
>>
>> On Monday, Oct 21, 2002, at 17:34 US/Pacific, Ming Ouyang wrote:
>>
>>> Dale,
>>> I just found that macro doesn't always work well, for any complicated
>>> operands like: max(a, ++b), max(a+b, c), max(a, foo(c)), gcc won't
>>> use max insn for them, it only use max insn for max(a,b).
>>> In addition, if I have more than one maxs/mins, it seems gcc will
>>> not generate max/min for them, could you give me some help on that?
>>> Thanks a lot.
>>>
>>> Ming
>>>
>>> --- Dale Johannesen <dalej@apple.com> wrote:
>>>>
>>>> On Monday, October 21, 2002, at 04:05  PM, Ming Ouyang wrote:
>>>>
>>>>> I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
>>>>> all debug info, gcc always treats min as a function call in my
>>>>> test.c.00.rtl through test.c.30.dbr:
>>>>>
>>>>> (call (mem:SI (symbol_ref/v:SI ("smin")) ))
>>>>>
>>>>> I think once gcc found that is a function call, it would not
>>>>> optimize it, am I right?
>>>>> What I should get is something like "(smin: SI ...", is that right?
>>>>> So probably I need to change something other than xxx.md?
>>>>> Thanks.
>>>>
>>>> If you're using it as a function of course gcc will treat it as
>>>> a function.  It is not one of the standard C functions so gcc does
>>>> not do anything special to it (and is not allowed to, that would
>>>> break standard conformance).  The usual way to get min/max
>>>> functionality
>>>> in C is to put a macro definition in a header file somewhere,
>>>> something
>>>> like
>>>> #define MAX(x,y) ((x>y) ? x : y)
>>>> (watch out for ++ operators, improvement left as an exercise)
>>>>
>>>
>>>
>>> __________________________________________________
>>> Do you Yahoo!?
>>> Y! Web Hosting - Let the expert host your web site
>>> http://webhosting.yahoo.com/
>>>
>>
>
>
> __________________________________________________
> Do you Yahoo!?
> Y! Web Hosting - Let the expert host your web site
> http://webhosting.yahoo.com/
>

Re: help on adding new insns like min/max

[Ming Ouyang]

Andrew,
Thanks for your reply. __typeof__ does make 2 or more min/max in one
function work, but it can't resolve complicated operands problem yet,
can you help on that? 
Thanks so much.

Ming


--- Andrew Pinski <apinski@apple.com> wrote:
> try this, using extensions for gcc:
> 
> #define max(a,b) ({__typeof__(a) a1=(a); __typeof__(b) b1 = (b);  
> (a>b)?a:b})
> 
> Thanks,
> Andrew Pinski
> 
> On Monday, Oct 21, 2002, at 17:34 US/Pacific, Ming Ouyang wrote:
> 
> > Dale,
> > I just found that macro doesn't always work well, for any complicated
> > operands like: max(a, ++b), max(a+b, c), max(a, foo(c)), gcc won't
> > use max insn for them, it only use max insn for max(a,b).
> > In addition, if I have more than one maxs/mins, it seems gcc will
> > not generate max/min for them, could you give me some help on that?
> > Thanks a lot.
> >
> > Ming
> >
> > --- Dale Johannesen <dalej@apple.com> wrote:
> >>
> >> On Monday, October 21, 2002, at 04:05  PM, Ming Ouyang wrote:
> >>
> >>> I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
> >>> all debug info, gcc always treats min as a function call in my
> >>> test.c.00.rtl through test.c.30.dbr:
> >>>
> >>> (call (mem:SI (symbol_ref/v:SI ("smin")) ))
> >>>
> >>> I think once gcc found that is a function call, it would not
> >>> optimize it, am I right?
> >>> What I should get is something like "(smin: SI ...", is that right?
> >>> So probably I need to change something other than xxx.md?
> >>> Thanks.
> >>
> >> If you're using it as a function of course gcc will treat it as
> >> a function.  It is not one of the standard C functions so gcc does
> >> not do anything special to it (and is not allowed to, that would
> >> break standard conformance).  The usual way to get min/max 
> >> functionality
> >> in C is to put a macro definition in a header file somewhere, 
> >> something
> >> like
> >> #define MAX(x,y) ((x>y) ? x : y)
> >> (watch out for ++ operators, improvement left as an exercise)
> >>
> >
> >
> > __________________________________________________
> > Do you Yahoo!?
> > Y! Web Hosting - Let the expert host your web site
> > http://webhosting.yahoo.com/
> >
> 


__________________________________________________
Do you Yahoo!?
Y! Web Hosting - Let the expert host your web site
http://webhosting.yahoo.com/

Re: help on adding new insns like min/max

[Andrew Pinski]

try this, using extensions for gcc:

#define max(a,b) ({__typeof__(a) a1=(a); __typeof__(b) b1 = (b);  
(a>b)?a:b})

Thanks,
Andrew Pinski

On Monday, Oct 21, 2002, at 17:34 US/Pacific, Ming Ouyang wrote:

> Dale,
> I just found that macro doesn't always work well, for any complicated
> operands like: max(a, ++b), max(a+b, c), max(a, foo(c)), gcc won't
> use max insn for them, it only use max insn for max(a,b).
> In addition, if I have more than one maxs/mins, it seems gcc will
> not generate max/min for them, could you give me some help on that?
> Thanks a lot.
>
> Ming
>
> --- Dale Johannesen <dalej@apple.com> wrote:
>>
>> On Monday, October 21, 2002, at 04:05  PM, Ming Ouyang wrote:
>>
>>> I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
>>> all debug info, gcc always treats min as a function call in my
>>> test.c.00.rtl through test.c.30.dbr:
>>>
>>> (call (mem:SI (symbol_ref/v:SI ("smin")) ))
>>>
>>> I think once gcc found that is a function call, it would not
>>> optimize it, am I right?
>>> What I should get is something like "(smin: SI ...", is that right?
>>> So probably I need to change something other than xxx.md?
>>> Thanks.
>>
>> If you're using it as a function of course gcc will treat it as
>> a function.  It is not one of the standard C functions so gcc does
>> not do anything special to it (and is not allowed to, that would
>> break standard conformance).  The usual way to get min/max 
>> functionality
>> in C is to put a macro definition in a header file somewhere, 
>> something
>> like
>> #define MAX(x,y) ((x>y) ? x : y)
>> (watch out for ++ operators, improvement left as an exercise)
>>
>
>
> __________________________________________________
> Do you Yahoo!?
> Y! Web Hosting - Let the expert host your web site
> http://webhosting.yahoo.com/
>

Re: help on adding new insns like min/max

[Ming Ouyang]

Dale,
I just found that macro doesn't always work well, for any complicated
operands like: max(a, ++b), max(a+b, c), max(a, foo(c)), gcc won't
use max insn for them, it only use max insn for max(a,b). 
In addition, if I have more than one maxs/mins, it seems gcc will 
not generate max/min for them, could you give me some help on that? 
Thanks a lot.

Ming

--- Dale Johannesen <dalej@apple.com> wrote:
> 
> On Monday, October 21, 2002, at 04:05  PM, Ming Ouyang wrote:
> 
> > I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
> > all debug info, gcc always treats min as a function call in my
> > test.c.00.rtl through test.c.30.dbr:
> >
> > (call (mem:SI (symbol_ref/v:SI ("smin")) ))
> >
> > I think once gcc found that is a function call, it would not
> > optimize it, am I right?
> > What I should get is something like "(smin: SI ...", is that right?
> > So probably I need to change something other than xxx.md?
> > Thanks.
> 
> If you're using it as a function of course gcc will treat it as
> a function.  It is not one of the standard C functions so gcc does
> not do anything special to it (and is not allowed to, that would
> break standard conformance).  The usual way to get min/max functionality
> in C is to put a macro definition in a header file somewhere, something 
> like
> #define MAX(x,y) ((x>y) ? x : y)
> (watch out for ++ operators, improvement left as an exercise)
> 


__________________________________________________
Do you Yahoo!?
Y! Web Hosting - Let the expert host your web site
http://webhosting.yahoo.com/

Re: help on adding new insns like min/max

[Ming Ouyang]

Yes, you are totally right and I'm misled by the abs implemention in
our gcc compiler (I don't know who added support for the abs insn),
it can support "abs(a)" and will translate it into one single abs 
insn. Now it seems that guy did something really complicated.
Thank you so much :)

Ming

--- Dale Johannesen <dalej@apple.com> wrote:
> 
> On Monday, October 21, 2002, at 04:05  PM, Ming Ouyang wrote:
> 
> > I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
> > all debug info, gcc always treats min as a function call in my
> > test.c.00.rtl through test.c.30.dbr:
> >
> > (call (mem:SI (symbol_ref/v:SI ("smin")) ))
> >
> > I think once gcc found that is a function call, it would not
> > optimize it, am I right?
> > What I should get is something like "(smin: SI ...", is that right?
> > So probably I need to change something other than xxx.md?
> > Thanks.
> 
> If you're using it as a function of course gcc will treat it as
> a function.  It is not one of the standard C functions so gcc does
> not do anything special to it (and is not allowed to, that would
> break standard conformance).  The usual way to get min/max functionality
> in C is to put a macro definition in a header file somewhere, something 
> like
> #define MAX(x,y) ((x>y) ? x : y)
> (watch out for ++ operators, improvement left as an exercise)
> 


__________________________________________________
Do you Yahoo!?
Y! Web Hosting - Let the expert host your web site
http://webhosting.yahoo.com/

Re: help on adding new insns like min/max

[Dale Johannesen]

On Monday, October 21, 2002, at 04:05  PM, Ming Ouyang wrote:

> I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
> all debug info, gcc always treats min as a function call in my
> test.c.00.rtl through test.c.30.dbr:
>
> (call (mem:SI (symbol_ref/v:SI ("smin")) ))
>
> I think once gcc found that is a function call, it would not
> optimize it, am I right?
> What I should get is something like "(smin: SI ...", is that right?
> So probably I need to change something other than xxx.md?
> Thanks.

If you're using it as a function of course gcc will treat it as
a function.  It is not one of the standard C functions so gcc does
not do anything special to it (and is not allowed to, that would
break standard conformance).  The usual way to get min/max functionality
in C is to put a macro definition in a header file somewhere, something 
like
#define MAX(x,y) ((x>y) ? x : y)
(watch out for ++ operators, improvement left as an exercise)

Re: help on adding new insns like min/max

[Ming Ouyang]

I tried my gcc with definition of sminsi3/smaxsi3 and I dumpped
all debug info, gcc always treats min as a function call in my
test.c.00.rtl through test.c.30.dbr:

(call (mem:SI (symbol_ref/v:SI ("smin")) ))

I think once gcc found that is a function call, it would not 
optimize it, am I right?
What I should get is something like "(smin: SI ...", is that right?
So probably I need to change something other than xxx.md?
Thanks.

--- Dale Johannesen <dalej@apple.com> wrote:
> 
> On Monday, October 21, 2002, at 03:18  PM, Ming Ouyang wrote:
> 
> > Dale,
> >
> > Thanks for your reply, but I wonder how gcc will handle code like:
> > c = min(a, b)
> > Does min/max will be treated as builtins? Or can you demonstrate
> > how to make gcc to use sminsi3/smaxsi3... to use min/max insns for
> > code like that?
> 
> The optimization phases should create min/max from simpler RTL.
> See noce_try_minmax in ifcvt.c.  That won't necessarily handle
> whatever case you want to handle, but it will give you an idea.
> 


__________________________________________________
Do you Yahoo!?
Y! Web Hosting - Let the expert host your web site
http://webhosting.yahoo.com/