The machine I am targeting has some special requirements for some
operations, say:
ADD or1, ir1, r5
would add ir1 (input reg 1) and r5 and put the result in or1 (output reg
1). The point id that input and output regs have to go paired (this
meaning an addition of ir1 with whatever always goes to or1, or an in
general irX + whatever goes to orX).
AFAIK, InstrInfo.td only allow "$src = $dst" type constraints. Is it
possible to describe more complex src/dst relations, like the one I
need?
Also, I have tried making say or1 and ir1 parts of a single superreg
(op1reg), using "$src = $dst" and a pattern to wrap the operations with
EXTRACT_SUBREG/INSERT_SUBREG compounds. But in that case I have two
problems: 1) the LLVM op operates over a vector reg, thus *overwrites*
the input also (and that is not what our machine does) 2) LLVM needs to
be able to copy those regs for two-addr instruction rewritting, but
input regs are not readable and output regs are not writable.
This latter way is probably too hackish and that's why I am having so
weird problems, but I have found no other way to achieve this. Is there
one?
The machine I am targeting has some special requirements for some
operations, say:
ADD or1, ir1, r5
would add ir1 (input reg 1) and r5 and put the result in or1 (output reg
1). The point id that input and output regs have to go paired (this
meaning an addition of ir1 with whatever always goes to or1, or an in
general irX + whatever goes to orX).
AFAIK, InstrInfo.td only allow "$src = $dst" type constraints. Is it
possible to describe more complex src/dst relations, like the one I
need?
No, that is the only type of constraint supported (besides register classes).
This latter way is probably too hackish and that's why I am having so
weird problems, but I have found no other way to achieve this. Is there
one?
It sounds like you need to present a more abstract architecture to the LLVM register allocator, but it is difficult to suggest one from the description you have given.
The PBQP allocator was designed to support a very wide range of constraints, and can handle something like this easily.
Say you have 4 of these orX/irX registers, then for any pair of virtual registers used in such an add instruction you would add the following constraint matrix to the PBQP instance:
The rows and columns of this matrix reflect the storage locations that the allocator can assign, and the elements represent the cost of a specific assignment. Say the rows represent the set { or1, or2, or3, or4 } and the columns represent { ir1, ir2, ir3, ir4 }. The infinite cost elements constrain the valid assignments to matching pairs.
Representing the constraint is dead easy, the trick would be making the allocator aware of the constraint. At present the PBQP allocator only “knows” about the basic RA constraints (aliasing, classes, interference, coalescing) described in LiveIntervals, MachineRegisterInfo and TargetRegisterInfo. I have been meaning to generalize this though.
The design I have in mind is this: We add a method to TargetRegisterInfo which returns a PBQPProblemBuilder for the target architectures. I implement a base PBQPProblemBuilder by simply lifting the current construction process out of the PBQP allocator and into its own class. Anyone (such as yourself, Carlos) who wants to represent more esoteric constraints in their architecture just extends this class, calls the base class to handle all the basic constraints, then performs their own pass over the function to add in their constraints:
struct MyTargetPBQPProblemBuilder : public PBQPProblemBuilder {
PBQP::Graph* buildProblemFor(MachineFunction mf) {
PBQP::Graph g = PBQPProblemBuilder::buildProblemFor(mf);
// Add additional constraints for my architecture here.
return g;
}
}
Any thoughts or comments? I think this should be a very straightforward extension.
The PBQP allocator was designed to support a very wide range of
constraints, and can handle something like this easily.
Glad to hear this
Say you have 4 of these orX/irX registers, then for any pair of
virtual registers used in such an add instruction you would add the
following constraint matrix to the PBQP instance:
The rows and columns of this matrix reflect the storage locations that
the allocator can assign, and the elements represent the cost of a
specific assignment. Say the rows represent the set { or1, or2, or3,
or4 } and the columns represent { ir1, ir2, ir3, ir4 }. The infinite
cost elements constrain the valid assignments to matching pairs.
I think this is exactly what I need, also because in a future stage I
would need to assign weights to some register pairs, not being
impossible but being quite expensive in terms of performance.
The design I have in mind is this: We add a method to
TargetRegisterInfo which returns a PBQPProblemBuilder for the target
architectures. I implement a base PBQPProblemBuilder by simply lifting
the current construction process out of the PBQP allocator and into
its own class. Anyone (such as yourself, Carlos) who wants to
represent more esoteric constraints in their architecture just extends
this class, calls the base class to handle all the basic constraints,
then performs their own pass over the function to add in their
constraints:
struct MyTargetPBQPProblemBuilder : public PBQPProblemBuilder {
PBQP::Graph* buildProblemFor(MachineFunction *mf) {
PBQP::Graph* g = PBQPProblemBuilder::buildProblemFor(mf);
// Add additional constraints for my architecture here.
return g;
}
}
Any thoughts or comments? I think this should be a very
straightforward extension.
I find it quite suiting my problem. If this extension is added I will
surely use it.
