[llvm] [ARM][SLP] Fix cost function for SLP Vectorization of ZExt/SExt (PR #122713)

[Nashe Mncube via llvm-commits]

nasherm wrote:

My most recent patch fixes the regression with a one line change to ARMTTI. From what I understand, MVE vectors greater than 128 bits aren't handled very well. It appears that the changes to SLP revealed a bug where we were allowing extensions of the type ` vadd v4i32 -> vadd v4i64`. 

@davemgreen does my fix make sense with this reasoning? I'm skeptical as I suspect I'm missing something which justifies the original code which allowed this extension.

https://github.com/llvm/llvm-project/pull/122713