[llvm] [LV] Allow reductions with partial alias masking (PR #200813)

[Benjamin Maxwell via llvm-commits]

================
@@ -352,5 +352,302 @@ exit:
   ret void
 }
 
+define i8 @add_reduction(ptr %a, ptr %b, i64 %n) {
+; CHECK-LABEL: define i8 @add_reduction(
+; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i64 [[N:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[A2:%.*]] = ptrtoaddr ptr [[A]] to i64
+; CHECK-NEXT:    [[B1:%.*]] = ptrtoaddr ptr [[B]] to i64
+; CHECK-NEXT:    br label %[[VECTOR_MEMCHECK:.*]]
+; CHECK:       [[VECTOR_MEMCHECK]]:
+; CHECK-NEXT:    [[ALIAS_MASK:%.*]] = call <4 x i1> @llvm.loop.dependence.war.mask.v4i1.i64(i64 [[A2]], i64 [[B1]], i64 1)
+; CHECK-NEXT:    [[TMP9:%.*]] = zext <4 x i1> [[ALIAS_MASK]] to <4 x i64>
+; CHECK-NEXT:    [[NUM_ACTIVE_LANES:%.*]] = call i64 @llvm.vector.reduce.add.v4i64(<4 x i64> [[TMP9]])
+; CHECK-NEXT:    [[VF_IS_SCALAR:%.*]] = icmp ule i64 [[NUM_ACTIVE_LANES]], 1
+; CHECK-NEXT:    [[TMP11:%.*]] = sub i64 -1, [[N]]
+; CHECK-NEXT:    [[VF_STEP_OVERFLOW:%.*]] = icmp ult i64 [[TMP11]], [[NUM_ACTIVE_LANES]]
+; CHECK-NEXT:    [[DIFF_CHECK:%.*]] = or i1 [[VF_IS_SCALAR]], [[VF_STEP_OVERFLOW]]
+; CHECK-NEXT:    br i1 [[DIFF_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = sub i64 [[N]], 1
+; CHECK-NEXT:    [[TMP12:%.*]] = sub i64 [[NUM_ACTIVE_LANES]], 1
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i64 [[N]], [[TMP12]]
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N_RND_UP]], [[NUM_ACTIVE_LANES]]
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N_RND_UP]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i64> poison, i64 [[TRIP_COUNT_MINUS_1]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT]], <4 x i64> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <4 x i64> poison, i64 [[NUM_ACTIVE_LANES]], i64 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT6:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT5]], <4 x i64> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i8> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP13:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_IND3:%.*]] = phi <4 x i64> [ <i64 0, i64 1, i64 2, i64 3>, %[[VECTOR_PH]] ], [ [[VEC_IND_NEXT8:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[HEADER_MASK:%.*]] = icmp ule <4 x i64> [[VEC_IND3]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT:    [[CLAMPED_HEADER_MASK:%.*]] = and <4 x i1> [[HEADER_MASK]], [[ALIAS_MASK]]
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[INDEX]]
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <4 x i8> @llvm.masked.load.v4i8.p0(ptr align 1 [[TMP2]], <4 x i1> [[CLAMPED_HEADER_MASK]], <4 x i8> poison)
+; CHECK-NEXT:    [[TMP3:%.*]] = add <4 x i8> [[VEC_PHI]], [[WIDE_MASKED_LOAD]]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[INDEX]]
+; CHECK-NEXT:    call void @llvm.masked.store.v4i8.p0(<4 x i8> [[WIDE_MASKED_LOAD]], ptr align 1 [[TMP4]], <4 x i1> [[CLAMPED_HEADER_MASK]])
+; CHECK-NEXT:    [[TMP13]] = select <4 x i1> [[CLAMPED_HEADER_MASK]], <4 x i8> [[TMP3]], <4 x i8> [[VEC_PHI]]
+; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], [[NUM_ACTIVE_LANES]]
+; CHECK-NEXT:    [[VEC_IND_NEXT8]] = add nuw <4 x i64> [[VEC_IND3]], [[BROADCAST_SPLAT6]]
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP15:%.*]] = call i8 @llvm.vector.reduce.add.v4i8(<4 x i8> [[TMP13]])
+; CHECK-NEXT:    br [[EXIT:label %.*]]
+; CHECK:       [[SCALAR_PH]]:
+;
+entry:
+  br label %for.body
+
+for.body:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
+  %sum.0 = phi i8 [ 0, %entry ], [ %sum.1, %for.body ]
+  %gep.a = getelementptr inbounds i8, ptr %a, i64 %iv
+  %load.a = load i8, ptr %gep.a, align 1
+  %sum.1 = add i8 %sum.0, %load.a
+  %gep.b = getelementptr inbounds i8, ptr %b, i64 %iv
+  store i8 %load.a, ptr %gep.b, align 1
----------------
MacDue wrote:

It's done so the loop requires diff checks/alias masking 

https://github.com/llvm/llvm-project/pull/200813