Improve cg_llvm codegen for simd_select_bitmask

[sayantn]

While testing out some intrinsic-based implementation in stdarch, I noticed that LLVM can't optimize it well if we use simple bitcast and truncs to get the actual i1xNs from the bitmasks. LLVM actually seems to like it better if we use shufflevectors, and for autoupgraded intrinsics that used integer returns, but were migrated to use i1xN returns, LLVM actually use this exact form (https://godbolt.org/z/YT99Phzj5)

This doesn't affect any codegen that uses vectors that are multiples-of-8-length. This is mostly targeted towards 2- and 4-width vectors

@rustbot A-LLVM A-codegen T-compiler

r? codegen

[saethlin]

Can you add a codegen test that demonstrates a case where this actually helps? I feel like the fact that a shufflevector leads to better codegen than bitcast + trunc is at the very least unintuitive. Is this recognized as an upstream bug?

r? author

[rustbot]

Failed to set assignee to author: cannot assign: response: {"message":"Not Found","documentation_url":"https://docs.github.com/rest/issues/assignees#add-assignees-to-an-issue","status":"404"}

Note: Only org members with at least the repository "read" role, users with write permissions, or people who have commented on the PR may be assigned.

[sayantn]

@rustbot author

I did observe some improved codegen with this, but unfortunately I forgot which example it was. I will try to find it out and add a test/example

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.