llvm dev - Aug 2014 - [LLVMdev] Physical register definition removed by MachineCSE

Hi,

My target has a special configuration register that many instructions 
read implicitly, configuring for example if an add instruction should 
saturate at over/underflow or not.

Now, I have a problem where the MachineCSE removes a setting of this 
configuration register in a basic block, because 
MachineCSE::isPhysDefTriviallyDead can't find any uses of it in the 
basic block where it's defined.

Similar to other special physical registers, e.g the stack pointer, this 
configuration register is marked as "reserved".

Is there anything else I need to do for the MachineCSE to leave this 
register and not remove it just because it can't find any uses in the 
current basic block?

Best regards,
Mikael Holmén

On Fri, Aug 15, 2014 at 11:18:51AM +0200, Mikael Holmén
wrote:
> Hi,
> 
> My target has a special configuration register that many
> instructions read implicitly, configuring for example if an add
> instruction should saturate at over/underflow or not.
> 
> Now, I have a problem where the MachineCSE removes a setting of this
> configuration register in a basic block, because
> MachineCSE::isPhysDefTriviallyDead can't find any uses of it in the
> basic block where it's defined.
> 
> Similar to other special physical registers, e.g the stack pointer,
> this configuration register is marked as "reserved".
> 
> Is there anything else I need to do for the MachineCSE to leave this
> register and not remove it just because it can't find any uses in
> the current basic block?
Have you added this register as an implicit use to the instructions
like add whose behavior depends on its value?

-Tom
> 
> Best regards,
> Mikael Holmén
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Hi,

On 08/15/14 16:15, Tom Stellard wrote:
> On Fri, Aug 15, 2014 at 11:18:51AM +0200, Mikael Holmén wrote:
>> Hi,
>>
>> My target has a special configuration register that many
>> instructions read implicitly, configuring for example if an add
>> instruction should saturate at over/underflow or not.
>>
>
>> Now, I have a problem where the MachineCSE removes a setting of this
>> configuration register in a basic block, because
>> MachineCSE::isPhysDefTriviallyDead can't find any uses of it in the
>> basic block where it's defined.
>>
>> Similar to other special physical registers, e.g the stack pointer,
>> this configuration register is marked as "reserved".
>>
>> Is there anything else I need to do for the MachineCSE to leave this
>> register and not remove it just because it can't find any uses in
>> the current basic block?
>
> Have you added this register as an implicit use to the instructions
> like add whose behavior depends on its value?
Yes.

So it's like

BB1:
...
def ConfReg
... <no uses of ConfReg>
branch cond BB2
branch BB3

BB2:
...
add ... <imp-use ConfReg>
...


Since there are no uses of ConfReg after the ConfReg def in BB1, 
MachineCSE::isPhysDefTriviallyDead decides ConfReg is trivially dead, 
and removes the setting.

Best Regards,
Mikael
>
> -Tom
>
>>
>> Best regards,
>> Mikael Holmén
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>

> On Aug 15, 2014, at 2:18 AM, Mikael Holmén <mikael.holmen at
ericsson.com> wrote:
> 
> Hi,
> 
> My target has a special configuration register that many instructions read
implicitly, configuring for example if an add instruction should saturate at
over/underflow or not.
> 
> Now, I have a problem where the MachineCSE removes a setting of this
configuration register in a basic block, because
MachineCSE::isPhysDefTriviallyDead can't find any uses of it in the basic
block where it's defined.
> 
> Similar to other special physical registers, e.g the stack pointer, this
configuration register is marked as "reserved".
> 
> Is there anything else I need to do for the MachineCSE to leave this
register and not remove it just because it can't find any uses in the
current basic block?
The instructions which set this control register are distinct from normal
reg-to-reg moves and such, I assume? The way that’s typically modeled is to mark
those instructions as having side effects (hasSideEffects=1) in the .td file.

Note that LLVM doesn’t have any more general notion of things like saturating
arithmetic, so it is possible (likely even) that you’ll end up with some strange
mis-optimizations when operations get optimized across one of those
instructions. That’s a wholly separate issue, though.

-Jim
> 
> Best regards,
> Mikael Holmén
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev