GlobalISel round table follow up: register bank select

Hi all,

this is the second email for the round table follow-up, this time regarding the issues around the greedy RegBankSelect and alternative mappings.

The issue I brought up was that because RegBankSelect goes top-down, it never looks at all available mappings for the operands when considering which of the mappings to apply to the current instruction. In our architecture we have one register bank dedicated to pointers and another one for anything else. We often see code where we have a G_PTR_ADD with a constant. Since the constant is not a pointer, we put it on the other register bank. We could put it on the address regbank and do provide alternative mappings for that, but since the greedy algorithm doesn't actually check the usage of the constant, it is always put on the other bank.

When RegBankSelect then sees the G_PTR_ADD and sees that one of its inputs is on the other register bank already, it then inserts a costly cross-bank copy instead of checking if that operand has any alternative mappings which would make the overall mapping for the current instruction cheaper.

Matt suggested that RegBankSelect should probably go bottom-up instead and I agree with him. I don't think there is a particular reason why RegBankSelect necessarily has to go top-down.

I'm not too familiar with the implementation of RegBankSelect. Would it be a big effort to make it work bottom-up instead? I'm guessing one of the biggest areas would be the check whether a cross-bank copy is needed as well as calculating the overall cost for alternative mappings as now all usages of the current instruction would have to be checked instead of the much more limited operands. How big of an impact would this have?

Cheers,

Dominik

Hi Dominik,

Thanks for sending this!

Hi all,

this is the second email for the round table follow-up, this time regarding the issues around the greedy RegBankSelect and alternative mappings.

The issue I brought up was that because RegBankSelect goes top-down, it never looks at all available mappings for the operands when considering which of the mappings to apply to the current instruction. In our architecture we have one register bank dedicated to pointers and another one for anything else. We often see code where we have a G_PTR_ADD with a constant. Since the constant is not a pointer, we put it on the other register bank. We could put it on the address regbank and do provide alternative mappings for that, but since the greedy algorithm doesn’t actually check the usage of the constant, it is always put on the other bank.

The intent behind the greedy algorithm was for the mapping to look at X instructions ahead, depending on the optimization level, when assigning one instruction.
Right now the window is simply 1 instruction and the code is not structured to allow to have more than one.

As we gain more insights on what we want RegBankSelect to do, it makes sense to redesign it.

When RegBankSelect then sees the G_PTR_ADD and sees that one of its inputs is on the other register bank already, it then inserts a costly cross-bank copy instead of checking if that operand has any alternative mappings which would make the overall mapping for the current instruction cheaper.

The idea is when a mapping is done, that means it was the best one at the time of the decision (greedy), thus we don’t challenge that.
“Reverting” an already set mapping is not that simple because yes for this particular use we are inserting costly cross copies, but there are no guarantees that replacing the mapping of the definition will not insert even more costly copies for the other uses.

E.g., consider:

A = def <— RBS starts here
= useFP A
= useInt A
= useInt A

Let’s assume that greedy works the way it is intended. I.e., it assigns A to the int register bank because there are 2 such uses vs. only 1 fp bank use:

A<int> = def
= useFP A<int> <— Next, RBS looks at this one
= useInt A<int>
= useInt A<int>

Now the regbank for useFP is not right and has to be repaired. Right now, we will insert the costly cross copy for that use:

A<int> = def
AFP<fp> = cross_copy A<int>
= useFP AFP<fp>
= useInt A<int>
= useInt A<int>

Now, if we were to change the definition of A to avoid this copy we would create two costly copy for the useInt. Actually, another question is what would we do when we look at the first useInt? Is this use allowed to change the definition of the instruction again? Do we duplicate the definition? Etc…

A<fp> = def <— reassign
= useFP AFP<fp>
AInt1<int> = cross_copy A<fp>
= useInt AInt1<int>
AInt2<int> = cross_copy A<fp>
= useInt AInt2<int>

Bottom line, if we allow to modify the assignments of the definition, the decision making is not local anymore and in particular may require to add repairing code all over the place. As a result the cost model becomes much more complicated.

Matt suggested that RegBankSelect should probably go bottom-up instead and I agree with him. I don’t think there is a particular reason why RegBankSelect necessarily has to go top-down.

The rationale for going top-down is that when you reach an instruction, all the operands are assigned so you know what would be the cost of repairing.
You could said that the problem is the same for definitions when going bottom-up. However, this is likely to be more problematic because usually you have fewer definitions than arguments on each individual instruction, therefore there is more guess work going on (e.g., top-down, you assume a cost for 1 definition, going bottom-up you have to assume a cost for 2 arguments or more precisely you would have to track a window of X instructions for 2 arguments instead of 1 definition.)

I’m not too familiar with the implementation of RegBankSelect. Would it be a big effort to make it work bottom-up instead? I’m guessing one of the biggest areas would be the check whether a cross-bank copy is needed as well as calculating the overall cost for alternative mappings as now all usages of the current instruction would have to be checked instead of the much more limited operands. How big of an impact would this have?

It’s been a while since I looked at the implementation but I would expect this to be significant.

I think we should step back and check what we want before investing any time in some rewrite. For instance, I don’t see what bottom-up fundamentally gives us. It seems like a workaround to me.
That said, it would work either way!

Cheers,
-Quentin

Hi Quentin,

thanks for picking up the conversation!

I think we should step back and check what we want before investing any time in some rewrite.

That is a very fair point and I might have been getting ahead of myself in my last email.
What I would like to see from RegBankSelect is to produce the mapping with the overall lowest cost. Keeping track of all different combinations of mappings will certainly be non-trivial however, so I wonder if there is a smart way to do this without spending too much compilation time.

Ideally for instructions with no operands (like G_CONSTANT) it could also check whether a cross-bank copy is actually worth it or if it would be more beneficial to simply rematerialize the instruction on the required bank. For such instructions this information should already be available as part of the cost-modelling in RegBankSelect: we could simply compare the cost of a mapping on the required bank vs. the cost of a cross-bank copy.

Would you see this as a valid direction for RegBankSelect?

Best regards,

Dominik

Hi Dominik,

Hi Quentin,

thanks for picking up the conversation!

I think we should step back and check what we want before investing any time in some rewrite.

That is a very fair point and I might have been getting ahead of myself in my last email.
What I would like to see from RegBankSelect is to produce the mapping with the overall lowest cost. Keeping track of all different combinations of mappings will certainly be non-trivial however, so I wonder if there is a smart way to do this without spending too much compilation time.

Ideally for instructions with no operands (like G_CONSTANT) it could also check whether a cross-bank copy is actually worth it or if it would be more beneficial to simply rematerialize the instruction on the required bank. For such instructions this information should already be available as part of the cost-modelling in RegBankSelect: we could simply compare the cost of a mapping on the required bank vs. the cost of a cross-bank copy.

Would you see this as a valid direction for RegBankSelect?

Yes, I think this is a valid direction. I actually think we shouldn’t restrict ourselves to instructions with no operands. We could always duplicate next to the original instruction, i.e., we wouldn’t have any issue materializing the arguments.
The one thing that needs work is the cost model.

Cheers,
-Quentin

Hi Quentin,

Sure, I guess that would also work as long we don’t have to introduce copies for the operands. Theoretically this could even try to duplicate the operands if overall it’s still cheaper than a cross-bank copy but that is probably some pandoras box just waiting to be opened :slight_smile:

Hi Quentin,

Sure, I guess that would also work as long we don’t have to introduce copies for the operands. Theoretically this could even try to duplicate the operands if overall it’s still cheaper than a cross-bank copy but that is probably some pandoras box just waiting to be opened :slight_smile:

Hehe!
Ideally we would compute the costs globally and do the code transformation only once. But I am guessing this is wishful thinking.

The one thing that needs work is the cost model.

What would you say does the cost model need in its current state?

I don’t remember how everything works so take it with a grain of salt.
I’d say we would need to take into account how much it would cost to duplicate the definitions and compute the repairing cost with those duplications.
The thing is it may become compile time intensive pretty quickly if we want to consider all the possibilities.

As a starter we could focus on the best mapping for the current instruction and try to match the desired regbanks from each operand and only evaluate that cost against just plain repairing.

In any case, I think it would require a non-trivial amount of work.

By the way, does this have anything to do with what Matt was talking about during the round table? Unfortunately I don’t remember exactly what his issues with RegBankSelect were but I’d be interested to know whether what we talked about would also benefit him.

Partially. IIRC Matt’s problem was also how we apply the mapping and how, right now, we have to abuse the observer mechanism to do what we want.

Hi all,

I was pondering on how regbank-select could find the optimal assignment, and came up with the following algorithm:

Hi Gabriel,

Such approach matches what I had in mind for a “global” cost model which would have been a higher tier than greedy and fast.
Actually, the idea was that “global" would only be a variant of greedy where the window for the computed costs would have been infinitely large (right now the only available window is 1 instruction).

If we decide to go with such approach, I think we need to keep it as a possible variant, as opposed to the uniquely available approach because I suspect the compile time implications are high.

Cheers,
-Quentin