llvm-project

ecf8b039 - [AMDGPU] Recompute EXEC liveness in SIWholeQuadMode::toExact (#200866)

[AMDGPU] Recompute EXEC liveness in SIWholeQuadMode::toExact (#200866) This is required as stale liveness information can lead to an incorrect optimization in `SIOptimizeExecMaskingPreRA`. For example, when `hi16` is removed from EXEC, `optimizeElseBranch` produces an incorrect result by removing `S_AND`. This is caused by the following code: ```c++ SlotIndex StartIdx = LIS->getInstructionIndex(SaveExecMI); SlotIndex EndIdx = LIS->getInstructionIndex(*AndExecMI); for (MCRegUnit Unit : TRI->regunits(ExecReg)) { LiveRange &RegUnit = LIS->getRegUnit(Unit); if (RegUnit.find(StartIdx) != std::prev(RegUnit.find(EndIdx))) return false; } ``` When `hi16` is available there are two `RegUnit`s, one for `hi16` and one for `lo16`. In the case of `wqm.ll` test it produces two live ranges: ``` 0: [320r,320d:3)[368r,368d:2)[736r,736d:4)[832r,832d:1)[944r,944d:0) 0@944r 1@832r 2@368r 3@320r 4@736r 1: [12r,12d:1)[320r,320d:5)[368r,368d:4)[736r,736d:6)[744r,744d:0)[832r,832d:3)[944r,944d:2) 0@744r 1@12r 2@944r 3@832r 4@368r 5@320r 6@736r ``` When `hi16` is removed there is only one range: ``` 0: [320r,320d:3)[368r,368d:2)[736r,736d:4)[832r,832d:1)[944r,944d:0) 0@944r 1@832r 2@368r 3@320r 4@736r ``` If only the first range is considered the loop will finish without returning false and continue to remove `S_AND`. It is because EXEC register for `S_AND` added by `SIWholeQuadMode::toExact` is not in the range. This is incorrect. When `hi16` is present the incorrect behavior is masked by the second live range coming from `hi16`, causing the loop to return. This can be most likely explained by the fact that `hi16` is only computed after the new `S_AND` was added so contains more up to date information. Removing EXEC from `LiveIntervals` forces the live range to be recomputed.