[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]]:
-; COMPARE-LAA-MV-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[HEADER1]] ]
+; COMPARE-LAA-MV-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[HEADER1]] ]
+; COMPARE-LAA-MV-NEXT:    [[VEC_IND:%.*]] = phi <4 x i64> [ <i64 0, i64 1, i64 2, i64 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], %[[HEADER1]] ]
+; COMPARE-LAA-MV-NEXT:    [[TMP2:%.*]] = mul <4 x i64> [[VEC_IND]], [[BROADCAST_SPLAT]]
+; COMPARE-LAA-MV-NEXT:    [[WIDE_GEP:%.*]] = getelementptr i64, ptr [[P]], <4 x i64> [[TMP2]]
+; COMPARE-LAA-MV-NEXT:    [[WIDE_MASKED_GATHER:%.*]] = call <4 x i64> @llvm.masked.gather.v4i64.v4p0(<4 x ptr> align 8 [[WIDE_GEP]], <4 x i1> splat (i1 true), <4 x i64> poison)
+; COMPARE-LAA-MV-NEXT:    [[TMP3:%.*]] = getelementptr i64, ptr [[P_OUT]], i64 [[INDEX]]
+; COMPARE-LAA-MV-NEXT:    store <4 x i64> [[WIDE_MASKED_GATHER]], ptr [[TMP3]], align 8
+; COMPARE-LAA-MV-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; COMPARE-LAA-MV-NEXT:    [[VEC_IND_NEXT]] = add nsw <4 x i64> [[VEC_IND]], splat (i64 4)
+; COMPARE-LAA-MV-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; COMPARE-LAA-MV-NEXT:    br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[HEADER1]], !llvm.loop [[LOOP4:![0-9]+]]
+; COMPARE-LAA-MV:       [[MIDDLE_BLOCK]]:
+; COMPARE-LAA-MV-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]
+; COMPARE-LAA-MV-NEXT:    br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH1]]
+; COMPARE-LAA-MV:       [[SCALAR_PH1]]:
+; COMPARE-LAA-MV-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[SCALAR_PH]] ]
+; COMPARE-LAA-MV-NEXT:    br label %[[HEADER:.*]]
+; COMPARE-LAA-MV:       [[HEADER]]:
----------------
artagnon wrote:

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

I think reset-variable-names with UTC will have the same effect, but introduce unrelated changes? We can just tweak it by hand for now?

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