llvm dev - Apr 2016 - Inline asm clobber registers name

Hi all,


I am currently working on AMDGPU inline assembly and encountered problem
with naming clobber registers in asm constraints. It looks like by default
LLVM tries to match register specified in constraint to register name of
register definition in .td file but not to the AsmName for this register.


For example if we have register definition:

def MYReg0 : Register<"r0", 0>;

We want to create inline assembly and add this register to clobbers list.
Inline assembly should look something like this:

                i32 asm "nop", "~{r0}" ()


We used AsmName for register MYReg0 inside clobbers list. But this
constraint fails to work because
TargetLowering::getRegForInlineAsmConstraint() tries to match register
definition name (“MYReg0”) not its AsmName (“r0”). So to make this work we
should write this assembly:

                i32 asm "nop", "~{MYReg0}" ()


I believe that this behavior is not correct. It works because in most
back-ends register definition names and AsmNames are equal (e.g. def EAX :
X86Reg<"eax", ...>) but in AMDGPU we want to have different
def-names and
AsmNames.


This might be done by changing core LLVM code or in target-specific
getRegForInlineAsmConstraint()
method. What do you suppose to be better solution?


Thanks,

Sam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160407/6e1e7239/attachment.html>

----- Original Message -----
> From: "Семен Колтон via llvm-dev" <llvm-dev at
lists.llvm.org>
> To: "via llvm-dev" <llvm-dev at lists.llvm.org>
> Sent: Thursday, April 7, 2016 10:06:21 AM
> Subject: [llvm-dev] Inline asm clobber registers name
> Hi all,
> I am currently working on AMDGPU inline assembly and encountered
> problem with naming clobber registers in asm constraints. It looks
> like by default LLVM tries to match register specified in constraint
> to register name of register definition in .td file but not to the
> AsmName for this register.
> For example if we have register definition:
> def MYReg0 : Register<"r0", 0>;
> We want to create inline assembly and add this register to clobbers
> list. Inline assembly should look something like this:
> i32 asm "nop", "~{r0}" ()
> We used AsmName for register MYReg0 inside clobbers list. But this
> constraint fails to work because
> TargetLowering::getRegForInlineAsmConstraint() tries to match
> register definition name (“MYReg0”) not its AsmName (“r0”). So to
> make this work we should write this assembly:
> i32 asm "nop", "~{MYReg0}" ()
> I believe that this behavior is not correct. It works because in most
> back-ends register definition names and AsmNames are equal ( e.g.
> def EAX : X86Reg<"eax", ...> ) but in AMDGPU we want to
have
> different def-names and AsmNames.
> This might be done by changing core LLVM code or in target-specific
> getRegForInlineAsmConstraint() method. What do you suppose to be
> better solution?
I agree. I have the following FIXME in PPCISelLowering: 

// r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit registers 
// (which we call X[0-9]+). If a 64-bit value has been requested, and a 
// 32-bit GPR has been selected, then 'upgrade' it to the 64-bit parent 
// register. 
// FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow use 
// the AsmName field from *RegisterInfo.td, then this would not be necessary. 
if (R.first && VT == MVT::i64 && Subtarget.isPPC64() && 
PPC::GPRCRegClass.contains(R.first)) 
return std::make_pair(TRI->getMatchingSuperReg(R.first, 
PPC::sub_32, &PPC::G8RCRegClass), 
&PPC::G8RCRegClass); 

-Hal 
> Thanks,
> Sam
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-- 

Hal Finkel 
Assistant Computational Scientist 
Leadership Computing Facility 
Argonne National Laboratory 
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160407/11a4ded7/attachment.html>

Hi Hal,


In that case I will start implementing it in core LLVM. What I want to do
is add table with AsmNames to TargetGenRegisterInfo.inc similar to
RegStrings table. This table could be used by getRegForInlineAsmConstraint.

I don’t want to force all back-ends to use AsmNames instead of def-names.
There is useful review for this created by Tom Stellard:
http://reviews.llvm.org/D15614.

What do you think about all this?


Thanks,

Sam

2016-04-07 18:46 GMT+03:00 Hal Finkel <hfinkel at anl.gov>:
>
>
> ------------------------------
>
> *From: *"Семен Колтон via llvm-dev" <llvm-dev at
lists.llvm.org>
> *To: *"via llvm-dev" <llvm-dev at lists.llvm.org>
> *Sent: *Thursday, April 7, 2016 10:06:21 AM
> *Subject: *[llvm-dev] Inline asm clobber registers name
>
> Hi all,
>
>
> I am currently working on AMDGPU inline assembly and encountered problem
> with naming clobber registers in asm constraints. It looks like by default
> LLVM tries to match register specified in constraint to register name of
> register definition in .td file but not to the AsmName for this register.
>
>
> For example if we have register definition:
>
> def MYReg0 : Register<"r0", 0>;
>
> We want to create inline assembly and add this register to clobbers list.
> Inline assembly should look something like this:
>
>                 i32 asm "nop", "~{r0}" ()
>
>
> We used AsmName for register MYReg0 inside clobbers list. But this
> constraint fails to work because
> TargetLowering::getRegForInlineAsmConstraint() tries to match register
> definition name (“MYReg0”) not its AsmName (“r0”). So to make this work we
> should write this assembly:
>
>                 i32 asm "nop", "~{MYReg0}" ()
>
>
> I believe that this behavior is not correct. It works because in most
> back-ends register definition names and AsmNames are equal (e.g. def EAX
> : X86Reg<"eax", ...>) but in AMDGPU we want to have
different def-names
> and AsmNames.
>
>
> This might be done by changing core LLVM code or in target-specific
getRegForInlineAsmConstraint()
> method. What do you suppose to be better solution?
>
> I agree. I have the following FIXME in PPCISelLowering:
>
>   // r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit
> registers
>   // (which we call X[0-9]+). If a 64-bit value has been requested, and a
>   // 32-bit GPR has been selected, then 'upgrade' it to the 64-bit
parent
>   // register.
>   // FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow
> use
>   // the AsmName field from *RegisterInfo.td, then this would not be
> necessary.
>   if (R.first && VT == MVT::i64 && Subtarget.isPPC64()
&&
>       PPC::GPRCRegClass.contains(R.first))
>     return std::make_pair(TRI->getMatchingSuperReg(R.first,
>                             PPC::sub_32, &PPC::G8RCRegClass),
>                           &PPC::G8RCRegClass);
>
>  -Hal
>
>
> Thanks,
>
> Sam
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
>
>
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160408/5839a6bd/attachment-0001.html>