[llvm] [X86] Implement preferVectorizedNonPowerOfTwoTypeBreakdown() for X86 (PR #227169)

[Ryan Buchner via llvm-commits]

bababuck wrote:

> Is it just x86 that is seeing the compile time issue or is it the most prominent?

Most prominent I believe. The issue stems from the fact that `getRegUsageForType` assumes scalarization if `preferVectorizedNonPowerOfTwoTypeBreakdown()` is set to `false`. I don't fully understand why this isn't set to true for all backends (or at least most). I've only seen positive codegen changes by changing it.

https://github.com/llvm/llvm-project/pull/227169