[llvm] [LV] Transform to handle exits in the scalar loop (PR #148626)

[Gaëtan Bossu via llvm-commits]

================
@@ -0,0 +1,226 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 5
+; RUN: opt -S < %s -p loop-vectorize -handle-early-exits-in-scalar-tail -force-vector-width=4 | FileCheck %s
+
+define i32 @simple_contains(ptr align 4 dereferenceable(100) readonly %array, i32 %elt) {
+; CHECK-LABEL: define i32 @simple_contains(
+; CHECK-SAME: ptr readonly align 4 dereferenceable(100) [[ARRAY:%.*]], i32 [[ELT:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    br label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[ELT]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[ARRAY]], align 4
+; CHECK-NEXT:    [[TMP0:%.*]] = icmp eq <4 x i32> [[WIDE_LOAD]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[TMP1:%.*]] = freeze <4 x i1> [[TMP0]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP1]])
+; CHECK-NEXT:    br i1 [[TMP2]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH_SPLIT:.*]]
+; CHECK:       [[VECTOR_PH_SPLIT]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH_SPLIT]] ], [ [[IV:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[IV]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[LD_ADDR:%.*]] = getelementptr inbounds i32, ptr [[ARRAY]], i64 [[IV]]
+; CHECK-NEXT:    [[UNCOUNTABLE_EXIT_MASK:%.*]] = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i64(i64 [[IV]], i64 24)
+; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = call <4 x i32> @llvm.masked.load.v4i32.p0(ptr align 4 [[LD_ADDR]], <4 x i1> [[UNCOUNTABLE_EXIT_MASK]], <4 x i32> poison)
----------------
gbossu wrote:

Why is the masked load actually necessary? Haven't we checked that this iteration can process VF elements safely thanks to the new checks?

Edit: it seems that the preheader checks actually do not verify the "initial" active lane mask. Do we require predication in the loop block because only because of the first iteration?

https://github.com/llvm/llvm-project/pull/148626