[llvm] [X86][AVX] Prefer VPSRAV to VPSRA style shifts for known splats #39424 (PR #87913)

[Simon Pilgrim via llvm-commits]

================
@@ -599,6 +599,10 @@ def TuningPreferShiftShuffle : SubtargetFeature<"faster-shift-than-shuffle",
                                    "PreferLowerShuffleAsShift", "true",
                                    "Shifts are faster (or as fast) as shuffle">;
 
+def TuningPreferPerEltVectorShift : SubtargetFeature<"tuning-fast-per-element-vector-shift",
+                                        "PreferPerEltVectorShift", "true",
+                                        "Vector per element shifts are faster (1/cycle latency)">;
----------------
RKSimon wrote:

Drop "(1/cycle latency)" - we don't guarantee this

https://github.com/llvm/llvm-project/pull/87913