[llvm] [LAA] Avoid unprofitable unit-stride versioning (PR #225331)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -360,17 +360,43 @@ define void @stride_as_btc(ptr noalias %p.out, ptr %p, i64 %stride) vscale_range
 ; COMPARE-LAA-MV-LABEL: define void @stride_as_btc(
 ; COMPARE-LAA-MV-SAME: ptr noalias [[P_OUT:%.*]], ptr [[P:%.*]], i64 [[STRIDE:%.*]]) #[[ATTR0]] {
 ; COMPARE-LAA-MV-NEXT:  [[SCALAR_PH:.*]]:
+; COMPARE-LAA-MV-NEXT:    [[TMP0:%.*]] = call i64 @llvm.smax.i64(i64 [[STRIDE]], i64 1)
+; COMPARE-LAA-MV-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 4
+; COMPARE-LAA-MV-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH1:.*]], label %[[VECTOR_PH:.*]]
+; COMPARE-LAA-MV:       [[VECTOR_PH]]:
+; COMPARE-LAA-MV-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 3
+; COMPARE-LAA-MV-NEXT:    [[N_VEC:%.*]] = sub i64 [[TMP0]], [[TMP1]]
+; COMPARE-LAA-MV-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i64> poison, i64 [[STRIDE]], i64 0
+; COMPARE-LAA-MV-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT]], <4 x i64> poison, <4 x i32> zeroinitializer
 ; COMPARE-LAA-MV-NEXT:    br label %[[HEADER1:.*]]
 ; COMPARE-LAA-MV:       [[HEADER1]]:
----------------
artagnon wrote:

```suggestion
; COMPARE-LAA-MV:       [[VECTOR_LOOP]]:
```

https://github.com/llvm/llvm-project/pull/225331