[llvm] [LoopVectorize] Use CodeSize as the cost kind for minsize (PR #124119)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

> > Does the fix #118100?
> 
> Isn't #118100 using -Os? If I understand correctly -Oz adds the `minsize` opt, right?

Yes, it doesn't fix #118100 due to that using -Os, which adds optsize not minsize.

> Do you have any statistics/perf number on the impact of this?

Compiling the llvm-test-suite SingleSource and MultiSource benchmarks with `--target=aarch64-none-elf -Oz -fvectorize` (we need the `-fvectorize` because clang disables vectorization by default with `-Oz`) and looking at the code size change in each object, there's only two objects where this makes a difference:
```
Object                                                Before After
SingleSource/UnitTests/Vectorizer/recurrences.cpp.obj 3024   2940
MultiSource/Benchmarks/MallocBench/gs/gdevmem.c.obj   5492   5396
``` 

https://github.com/llvm/llvm-project/pull/124119