[llvm] [LoopVectorize] Improve Vectorization of Low Trip Count Loops (PR #195823)

[Jack Styles via llvm-commits]

================
@@ -3066,6 +3071,36 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) {
       }
     }
 
+    // Allow cases where the ExactTC == (VF * IC) or ExactTC == (VF * IC) + 1.
+    //
+    // This produces at most 1 vector iteration, and at most 1 scalar iteration
+    // with no remainder. Later passes will eliminate the loop and leave
+    // straight-line code as the both iteration counts are statically known.
+    //
+    // If a function is marked as minsize/optsize or OptForSize is set, do not
+    // allow this form of transformation as this will increase CodeSize.
+    //
+    // For loops with small bodies, the cost model is not currently reliable
+    // enough to accurately determine if vectorization is beneficial.
+    unsigned EffectiveIC = UserIC > 0 ? UserIC : 1;
+    unsigned MaxVFForTC = llvm::bit_floor(TC.getFixedValue());
+    if (TC.getFixedValue() - MaxVFForTC <= 1 && MaxVFForTC / EffectiveIC > 1 &&
----------------
Stylie777 wrote:

I disagree, as I mentioned in #225633 I think it would be better to avoid limit it to situations where the Loop Latch is in the Loop Exiting Block. I think limiting this to where we have a scalar iteration is too limiting. See #225635.

https://github.com/llvm/llvm-project/pull/195823