[llvm] [WIP][Support] Always scale InstructionCost::Value (PR #178962)

[Ryan Buchner via llvm-commits]

bababuck wrote:

> Yes, that is the case. The big regressions on sqlite3 are correlated with code size increases around the same size: https://llvm-compile-time-tracker.com/compare.php?from=f90cbc61afa44d59efca2af3f2814e49aad6c517&to=f519057dac0cc5f542bc485f1fb96794d356bbf8&stat=size-text As long as the additional vectorization there is profitable, this is fine.

In the LLVM IR tests, there were a few tests that changed vectorization behavior because of slight costing behavior (e.g. before `(1 / 2) + 1 = 1`, but now `(1 / 2) + 1 > 1`) in cases where the vectorization was right on the edge of being/not-being profitable.

 I am trying to reproduce the increase in `squlite` to confirm it is due to the same reason. For the `compile-time-tracker`, the documentation says `stage1-O3` uses `native x86_64 target`. Would you be able to share which CPU the server is running on so I can properly set `march` to reproduce this behavior locally? I have tried a few different `march` settings without success.

https://github.com/llvm/llvm-project/pull/178962