[llvm-dev] Question about MachineCSE optimization

[Yuchao (Michael) via llvm-dev]

Hi

 

I am trying to fix a bug in MachineCSE and have one question about following code:

 

*********************************************

// Go through implicit defs of CSMI and MI, if a def is not dead at MI,

// we should make sure it is not dead at CSMI.

if (MO.isImplicit() && !MO.isDead() && CSMI->getOperand(i).isDead())

  ImplicitDefsToUpdate.push_back(i);

 

// Keep track of implicit defs of CSMI and MI, to clear possibly

// made-redundant kill flags.

if (MO.isImplicit() && !MO.isDead() && OldReg == NewReg)

  ImplicitDefs.push_back(OldReg);

 

if (OldReg == NewReg) {

  --NumDefs;

  continue;

}

**********************************************

 

From Above source code, it seems that CSE does not handle instruction which

have one operand that is ‘implicit+isDead+isDef’.

 

Base on my understanding, implicit dead register is mean clobbered, so it is not safe to

do CSE. Is this right?

 

Best wishes

Michael

 

[mbraun via llvm-dev]

As far as I understand it (though I didn’t write the code so I may be missing something) this is about making sure liveness flags (kill, dead) are updated correctly when merging instruction. The situations to consider are probably:

1)

OP_X …, implicit-def %X

with

OP_X …, implicit-def dead %X

The merged instruction probably should not have a dead flag set. The 2nd if appears to be about this situation:

2)

OP_X …, implicit-def dead %X

Use killed %X

…

OP_X …, implicit-def %X    ; this gets merged with the previous OP_X

Use %X

Once we merge the 2nd OP_X with the first one, we have to drop the `killed` flags on the %X users.

Though admittedly I am at a loss right now why this is only performed for implicit operands and not for all operands…

- Matthias

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[mbraun via llvm-dev]

On Oct 21, 2018, at 2:48 PM, mbraun via llvm-dev <llvm...@lists.llvm.org> wrote:

As far as I understand it (though I didn’t write the code so I may be missing something) this is about making sure liveness flags (kill, dead) are updated correctly when merging instruction. The situations to consider are probably:

1)

OP_X …, implicit-def %X

with

OP_X …, implicit-def dead %X

The merged instruction probably should not have a dead flag set. The 2nd if appears to be about this situation:

2)

OP_X …, implicit-def dead %X

Use killed %X

…

OP_X …, implicit-def %X    ; this gets merged with the previous OP_X

Use %X

This example should of course read:

OP_X …, implicit-def %X

Use killed %X

…

OP_X …, implicit-def %X    ; this gets merged with the previous OP_X

Use %X

Once we merge the 2nd OP_X with the first one, we have to drop the `killed` flags on all users of %X between the first OP_X and the position where we had the 2nd OP_X before it was merged.

[Yuchao (Michael) via llvm-dev]

Hi mbraun

 

Thanks for your information.

 

What about following situation:

3)

OP_X …, implicit-def dead %X 

with

OP_X …, implicit-def dead %X

 

Between two instructions, there may be another instruction(not OP_X) which also def %X,

Is it  correct to do CSE in this situation?

 

Best wishes

Michael

[Matthias Braun via llvm-dev]

Yes

OP_X ..., implicit-def dead %X

... potentially more defs/uses of %X

OP_X ..., implicit-def dead %X

is good to merge. Also note that:

%X = DEF

OP_X ..., implicit-def dead %X

use %X

is not valid! (The def overrides the register value but the "dead" flag promises noone will read it, even though in this case another use comes below).

- Matthias

[Yuchao (Michael) via llvm-dev]

As far as I know, dead register is not only means none would read  it , it may means register is clobbered.

 

http://lists.llvm.org/pipermail/llvm-dev/2018-February/120994.html

[Matthias Braun via llvm-dev]

On Oct 22, 2018, at 11:29 PM, Yuchao (Michael) via llvm-dev <llvm...@lists.llvm.org> wrote:

As far as I know, dead register is not only means none would read  it , it may means register is clobbered.

Yes, any definition "clobbers" a register.