[PowerPC] Optimize select_cc with VSX compare and xxsel

[ecnelises]

VSX has 'Predicate Compare Instructions' to compare float values with specific predicates and save result into VSX register. They can be used to optimize select_cc of float types. For target CPUs without support for scalar comparisons, vector version will be used.

[bzEq]

Can you explain the rationale of the preference of VSX compare? We only see two cases improved in the test. Are there scenarios vector compare is better?

[ecnelises]

This is splitted from https://reviews.llvm.org/D75895. Since VSX instructions can use twice registers than VMX, it does no harm to codegen.

[nemanjai]

Also, please add a description that will reflect what the eventual commit message would be.

[ecnelises]

In previous plan, this patch adds patterns for xxsel and xscmp(eq|ge|gt)(s|d)p. I revised to make it a combine, because:

1. The instruction does not touch CR, we can't use it to lower SETCC. It seems SETCC and SELECT_CC share table of setCondCodeAction. Implementing it in legaliztion requires large change to existing SETCC handling.
2. The ideal select_cc f32/f64/f128 ... pattern won't appear in isel, instead, they are transformed in legalization and hand coded isel.
3. Implementing in Select() cannot fallback easily if target CPU doesn't meet requirement.

[ecnelises]

Gentle ping

[ecnelises]

Gentle ping

[nemanjai]

Please note that I have just approved to unblock the review. I have not reviewed it yet, simply approved it to remove my blocking change request so others can review it and I will try to find time to review it myself next week.

[arsenm]

Is this still relevant?

[nemanjai]

@lei137 Can you please close this if it isn't relevant or rebase to make it current and ask for another review if it is relevant?

[lei137]

Closing this PR as we are not certain if this is still relevant for us at this time.