llvm dev - Jan 2010 - [LLVMdev] 答复: [LLVMdev] How to bind a register variable with a given general purpose register?

Hi everyone,

There are 16 GPRs in my RISC, but in fact GPR13 is read-only and connected
to output of an A/D converter.
It would be very convenient if i could bind a register variable with GPR13.

Because i am a newbie i don't know how my llvm backend can support that.

I plan to implement it as below.

A.  first declare a global variable in c-code
     int  ADC  asm("GPR13");
B.  If backend finds a variable is loaded from "GPR13"  use GPR13
instead.
C.  backend can't allocate GPR13 to other variable

Is it a foolish method? Is there a better one? Please give me some guidance

Thanks
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20100104/4fe5ac84/attachment.html>

hi zhu,

i think you should map the peripheral registers to data memory space 
instead of register file, then mark that memory address as "volatile".
like said the adc register was mapped to address 0xc0000, and then the 
corresponding c source will goes like this:
//define the register
#define ADCREG (*((volatile unsigned int *)0xc0000))
//read the register
a = ADCREG

--ether

On 2010-1-4 10:59, Heyu Zhu wrote:
> Hi everyone,
> There are 16 GPRs in my RISC, but in fact GPR13 is read-only and 
> connected to output of an A/D converter.
> It would be very convenient if i could bind a register variable with 
> GPR13.
> Because i am a newbie i don't know how my llvm backend can support
that.
> I plan to implement it as below.
> A. first declare a global variable in c-code
> int ADC asm("GPR13");
> B. If backend finds a variable is loaded from "GPR13" use GPR13
instead.
> C. backend can't allocate GPR13 to other variable
> Is it a foolish method? Is there a better one? Please give me some 
> guidance
> Thanks
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>

hi zhu,

i am not sure if your c frontend support "int ADC
asm("GPR13");"
i think you could:

1. add a attribute "GPR13"(or a more meaningful name like
"adcreg") and
the corresponding handler, so you code "int ADC
asm("GPR13");" became
"int ADC __attribute__((GPR13));"
2. add a intrinsic functions like "llvm.zhu.readadcreg()".
3. map any read to the variable marked with __attribute__((GPR13)) to
llvm.zhu.readadcreg() instead of "load" instruction
4. assign gpr13 to a special register class like "gpr13class" instead
of
gprclass, so gpr13 will not be allocated as other gpr.
5. lower instrinisic functions "llvm.zhu.readadcreg()" to register
node
"gpr13" in your backend.

regards

--ether

On 2010-1-4 18:01, Demon(Xiangyang) Zhu 朱向阳 wrote:
> Hi Ether,
>
> The hardware had been fixed now. 
> If map it to memory space, it will cost another instruction cycle to
execute 'a = ADCREG',
> Algorithm accesses the A/D convert very frequently. To get higher precision
at lower frequency
> A/D output is connected to R13 directly. For the moment algorithm code is
something like
> assemble and difficult to maintain.  I want to setup a c environment for
it, but i don't know
> how to  bind a register variable with a given general purpose register.
>
> Thanks!
>
> ________________________________________
> 发件人: llvmdev-bounces at cs.uiuc.edu [llvmdev-bounces at cs.uiuc.edu] 代表
ether [etherzhhb at gmail.com]
> 发送时间: 2010年1月4日 0:36
> 收件人: llvmdev at cs.uiuc.edu
> 主题: Re: [LLVMdev] How to bind a register variable with a given general
purpose register?
>
> hi zhu,
>
> i think you should map the peripheral registers to data memory space
> instead of register file, then mark that memory address as
"volatile".
> like said the adc register was mapped to address 0xc0000, and then the
> corresponding c source will goes like this:
> //define the register
> #define ADCREG (*((volatile unsigned int *)0xc0000))
> //read the register
> a = ADCREG
>
> --ether
>
> On 2010-1-4 10:59, Heyu Zhu wrote:
>   
>> Hi everyone,
>> There are 16 GPRs in my RISC, but in fact GPR13 is read-only and
>> connected to output of an A/D converter.
>> It would be very convenient if i could bind a register variable with
>> GPR13.
>> Because i am a newbie i don't know how my llvm backend can support
that.
>> I plan to implement it as below.
>> A. first declare a global variable in c-code
>> int ADC asm("GPR13");
>> B. If backend finds a variable is loaded from "GPR13" use
GPR13 instead.
>> C. backend can't allocate GPR13 to other variable
>> Is it a foolish method? Is there a better one? Please give me some
>> guidance
>> Thanks
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>>     
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Hello
> Is it a foolish method? Is there a better one? Please give me some guidance
Inline assembler is your friend. Also, do not forget to mark this
register as unallocable, so codegen won't use / clobber it.

-- 
With best regards, Anton Korobeynikov
Faculty of Mathematics and Mechanics, Saint Petersburg State University

or you can just add a built-in function in c frontend "readadcreg()"
and
emit it as llvm.zhu.readadcreg(), so you can just call the built-in
function to get the adc register value.

On 2010-1-4 21:27, Demon(Xiangyang) Zhu 朱向阳 wrote:
> Hi Ether,
>
> Thank you very much. I will try it soon as your description.
>
> Regards
>
> ________________________________________
> 发件人: ether [etherzhhb at gmail.com]
> 发送时间: 2010年1月4日 3:29
> 收件人: Demon(Xiangyang) Zhu 朱向阳
> 抄送: llvmdev at cs.uiuc.edu
> 主题: Re: 答复: [LLVMdev] How to bind a register variable with a given general
purpose register?
>
> hi zhu,
>
> i am not sure if your c frontend support "int ADC
asm("GPR13");"
> i think you could:
>
> 1. add a attribute "GPR13"(or a more meaningful name like
"adcreg") and
> the corresponding handler, so you code "int ADC
asm("GPR13");" became
> "int ADC __attribute__((GPR13));"
> 2. add a intrinsic functions like "llvm.zhu.readadcreg()".
> 3. map any read to the variable marked with __attribute__((GPR13)) to
> llvm.zhu.readadcreg() instead of "load" instruction
> 4. assign gpr13 to a special register class like "gpr13class"
instead of
> gprclass, so gpr13 will not be allocated as other gpr.
> 5. lower instrinisic functions "llvm.zhu.readadcreg()" to
register node
> "gpr13" in your backend.
>
> regards
>
> --ether
>
> On 2010-1-4 18:01, Demon(Xiangyang) Zhu 朱向阳 wrote:
>   
>> Hi Ether,
>>
>> The hardware had been fixed now.
>> If map it to memory space, it will cost another instruction cycle to
execute 'a = ADCREG',
>> Algorithm accesses the A/D convert very frequently. To get higher
precision at lower frequency
>> A/D output is connected to R13 directly. For the moment algorithm code
is something like
>> assemble and difficult to maintain.  I want to setup a c environment
for it, but i don't know
>> how to  bind a register variable with a given general purpose register.
>>
>> Thanks!
>>
>> ________________________________________
>> 发件人: llvmdev-bounces at cs.uiuc.edu [llvmdev-bounces at cs.uiuc.edu] 代表
ether [etherzhhb at gmail.com]
>> 发送时间: 2010年1月4日 0:36
>> 收件人: llvmdev at cs.uiuc.edu
>> 主题: Re: [LLVMdev] How to bind a register variable with a given general
purpose register?
>>
>> hi zhu,
>>
>> i think you should map the peripheral registers to data memory space
>> instead of register file, then mark that memory address as
"volatile".
>> like said the adc register was mapped to address 0xc0000, and then the
>> corresponding c source will goes like this:
>> //define the register
>> #define ADCREG (*((volatile unsigned int *)0xc0000))
>> //read the register
>> a = ADCREG
>>
>> --ether
>>
>> On 2010-1-4 10:59, Heyu Zhu wrote:
>>
>>     
>>> Hi everyone,
>>> There are 16 GPRs in my RISC, but in fact GPR13 is read-only and
>>> connected to output of an A/D converter.
>>> It would be very convenient if i could bind a register variable
with
>>> GPR13.
>>> Because i am a newbie i don't know how my llvm backend can
support that.
>>> I plan to implement it as below.
>>> A. first declare a global variable in c-code
>>> int ADC asm("GPR13");
>>> B. If backend finds a variable is loaded from "GPR13" use
GPR13 instead.
>>> C. backend can't allocate GPR13 to other variable
>>> Is it a foolish method? Is there a better one? Please give me some
>>> guidance
>>> Thanks
>>>
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>>
>>>
>>>       
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

On Sun, Jan 3, 2010 at 8:59 PM, Heyu Zhu <zhu.heyu at gmail.com>
wrote:
> Hi everyone,
>
> There are 16 GPRs in my RISC, but in fact GPR13 is read-only and connected
> to output of an A/D converter.
> It would be very convenient if i could bind a register variable with GPR13.
>
> Because i am a newbie i don't know how my llvm backend can support
that.
>
> I plan to implement it as below.
>
> A.  first declare a global variable in c-code
>      int  ADC  asm("GPR13");
> B.  If backend finds a variable is loaded from "GPR13"  use GPR13
instead.
> C.  backend can't allocate GPR13 to other variable
>
> Is it a foolish method? Is there a better one? Please give me some guidance
>
> Thanks
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
If your end goal is to allow programs to use the output of the A/D
converter, my recommendation would be to create a platform-specific
intrinsic function that returns the output of the A/D converter, and
have your back-end refuse to allocate GPR13 as a register.  The
intrinsic function would of course simply cough up the value of GPR13,
which would then be available for use in the program.