[RegAllocFast] Refactor dominates algorithm for large basic block #72250
Conversation
The original brute force dominates algorithm is O(n) complexity so it is very slow for very large machine basic block which is very common with O0. This patch added InstrPosIndexes to assign index for each instruction and use it to determine dominance. The complexity is now O(1).
HaohaiWen
requested review from
phoebewang,
KanRobert,
williamweixiao and
arsenm
|
Hi, is there any more comment? I'd like to land this patch. |
| for (auto I = Start; I != End; ++I) { | ||
| LastIndex += Step; | ||
| Instr2PosIndex[&*I] = LastIndex; | ||
| } |
There was a problem hiding this comment.
Do we know when new MI being inserted? I think we can simply assgin it a value if we do, e.g.,
MachineInstr *New = BuildMI(MBB, II, ...
uint64_t NewPos = Instr2PosIndex[&*II] - 1;
if (LLVM_UNLIKELY(NewPos & (InstrDist - 1) == 0)) ...
Instr2PosIndex[New] = NewPos;
There was a problem hiding this comment.
This requires to carefully maintain this PosIndex for each insertion point. I think it's unfriendly for other users.
If we insert two instructions before same place, we need to be careful about the step to avoid duplication.
| for (const MachineInstr &MI : MBB) { | ||
| LastIndex += InstrDist; | ||
| Instr2PosIndex[&MI] = LastIndex; | ||
| } |
There was a problem hiding this comment.
Can we do this lazily, on first query?
It looks like dominates() is actually only used in very rare situations (where a block branches back to itself), so populating this map is unnecessary for most blocks.
There was a problem hiding this comment.
Thanks for suggestion. See #76275
The original brute force dominates algorithm is O(n) complexity so it is
very slow for very large machine basic block which is very common with
O0. This patch added InstrPosIndexes to assign index for each
instruction and use it to determine dominance. The complexity is now
O(1).