llvm dev - Jul 2009 - [LLVMdev] Kill-flag in two-address instruction tied operands

Hi,

Two-address instructions are represented as normal three-address  
instructions with a Desc bit that indicates the first two operands are  
tied together. The question is, should the second operand have a  
<kill> flag?

a: %R0<def> = MUL %R0, %R1<kill>
b: %R0<def> = MUL %R0<kill>, %R1<kill>

I think the current policy is a: There should be no kill-flag. The  
machine code verifier has code to handle the missing <kill> flag, and  
the regscavenger asserts if the <kill> flag is set.

Recently somebody (possibly myself) has been setting <kill> flags on  
my tied use operands, causing regscavenger to assert.

I just want to verify the policy regarding two-address <kill> flags:

Case a: <kill> is not allowed. I should add a check to the verifier  
for this. Find out who is setting the flags.
Case b: <kill> is mandatory. Remove special-case code from verifier,  
fix regscavenger.
Case c: <kill> is don't care. Fix regscavenger.

My vote goes to b: Then you don't need special-case code for two- 
address instructions. (Except for places that need special treatment  
of two-address instrs for other reasons).

/jakob

On Jul 14, 2009, at 12:10 PM, Jakob Stoklund Olesen wrote:
> Hi,
>
> Two-address instructions are represented as normal three-address
> instructions with a Desc bit that indicates the first two operands are
> tied together. The question is, should the second operand have a
> <kill> flag?
>
> a: %R0<def> = MUL %R0, %R1<kill>
> b: %R0<def> = MUL %R0<kill>, %R1<kill>
>
> I think the current policy is a: There should be no kill-flag. The
Right.
> machine code verifier has code to handle the missing <kill> flag, and
> the regscavenger asserts if the <kill> flag is set.
>
> Recently somebody (possibly myself) has been setting <kill> flags on
> my tied use operands, causing regscavenger to assert.
>
> I just want to verify the policy regarding two-address <kill> flags:
>
> Case a: <kill> is not allowed. I should add a check to the verifier
> for this. Find out who is setting the flags.
> Case b: <kill> is mandatory. Remove special-case code from verifier,
> fix regscavenger.
> Case c: <kill> is don't care. Fix regscavenger.
It's 'a'.
>
> My vote goes to b: Then you don't need special-case code for two-
> address instructions. (Except for places that need special treatment
> of two-address instrs for other reasons).
Two-address operands are being treated differently because they are  
indeed different, we are not about to change that.

Evan
>
> /jakob
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

On 15/07/2009, at 02.43, Evan Cheng wrote:
>> I just want to verify the policy regarding two-address <kill>
flags:
>>
>> Case a: <kill> is not allowed. I should add a check to the
verifier
>> for this. Find out who is setting the flags.
>> Case b: <kill> is mandatory. Remove special-case code from
verifier,
>> fix regscavenger.
>> Case c: <kill> is don't care. Fix regscavenger.
>
> It's 'a'.
Thanks, I'll add a check to the verifier so it barfs on kill flags.  
Currently it implements c: Don't care.

/jakob