[llvm] [SLP]Reduce number of alternate instruction, where possible (PR #123360)

[Alexey Bataev via llvm-commits]

================
@@ -9797,6 +10119,13 @@ void BoUpSLP::transformNodes() {
       reorderGatherNode(E);
   }
 
+  bool ForceLoadGather =
+      count_if(VectorizableTree, [](const std::unique_ptr<TreeEntry> &TE) {
+        return TE->isGather() && TE->hasState() &&
+               TE->getOpcode() == Instruction::Load &&
+               TE->getVectorFactor() < 16;
----------------
alexey-bataev wrote:

That's a simple estimation. If there is only 2 gather nodes with loads and they have 8 or less instructions, better to try to analyze all loads together rather than check them separately. It may allow better vectorization.

https://github.com/llvm/llvm-project/pull/123360