[llvm] [LVI] Skip self loops in `solveBlockValueNonLocal` (PR #127763)

[Yingwei Zheng via llvm-commits]

dtcxzyw wrote:

Seems like the comptime overhead is due to more optimizations performed in CVP.
```
Improvements:
  correlated-value-propagation.NumSRems 1242 -> 1957 +57.57%
  correlated-value-propagation.NumReturns 203 -> 237 +16.75%
  correlated-value-propagation.NumAShrsConverted 4035 -> 4424 +9.64%
  correlated-value-propagation.NumSICmps 48127 -> 51578 +7.17%
  correlated-value-propagation.NumSMinMax 4646 -> 4871 +4.84%
  correlated-value-propagation.NumSIToFP 1557 -> 1631 +4.75%
  instcombine.NegatorNumNegationsFoundInCache 4056 -> 4245 +4.66%
  correlated-value-propagation.NumSubNUW 27953 -> 29215 +4.51%
  correlated-value-propagation.NumSDivs 17512 -> 18290 +4.44%
  correlated-value-propagation.NumMinMax 10215 -> 10647 +4.23%
```


https://github.com/llvm/llvm-project/pull/127763