[llvm] [LV] Use speculative load intrinsics for early-exit vectorization (PR #180039)

[David Sherwood via llvm-commits]

https://github.com/david-arm commented:

I think it might be a good idea in this initial PR to have the behaviour speculative behaviour turned off by default, but can be enabled with a flag to allow writing tests. I'm very much in favour of vectorising more early exit loops, but I worry there may be some regressions since now we're effectively favouring loop vectorisation over unrolling of early exit loops. For example, for a simple std::find loop loading 64-bit elements we may unroll by a factor of 4 and see big wins, whereas the loop vectoriser might only choose a <2 x i64> or a <vscale x 2 x i64> type without interleaving.

https://github.com/llvm/llvm-project/pull/180039