1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp < %s -codegenprepare -S | FileCheck -check-prefix=CHECK %s
3
4; Sink the shufflevector/insertelement pair, followed by the trunc. The sunk instruction end up dead.
5define signext i8 @dead(i16* noalias nocapture readonly %s1, i16 zeroext %x, i8* noalias nocapture %d, i32 %n) {
6; CHECK-LABEL: @dead(
7; CHECK-NEXT:  entry:
8; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[N:%.*]], -8
9; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
10; CHECK:       vector.body:
11; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
12; CHECK-NEXT:    [[TMP0:%.*]] = trunc i16 [[X:%.*]] to i8
13; CHECK-NEXT:    [[L6:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i32 [[INDEX]]
14; CHECK-NEXT:    [[L7:%.*]] = bitcast i16* [[L6]] to <8 x i16>*
15; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <8 x i16>, <8 x i16>* [[L7]], align 2
16; CHECK-NEXT:    [[L8:%.*]] = trunc <8 x i16> [[WIDE_LOAD]] to <8 x i8>
17; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <8 x i8> undef, i8 [[TMP0]], i32 0
18; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> undef, <8 x i32> zeroinitializer
19; CHECK-NEXT:    [[L9:%.*]] = mul <8 x i8> [[TMP2]], [[L8]]
20; CHECK-NEXT:    [[L13:%.*]] = getelementptr inbounds i8, i8* [[D:%.*]], i32 [[INDEX]]
21; CHECK-NEXT:    [[L14:%.*]] = bitcast i8* [[L13]] to <8 x i8>*
22; CHECK-NEXT:    store <8 x i8> [[L9]], <8 x i8>* [[L14]], align 1
23; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 8
24; CHECK-NEXT:    [[L15:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
25; CHECK-NEXT:    br i1 [[L15]], label [[EXIT:%.*]], label [[VECTOR_BODY]]
26; CHECK:       exit:
27; CHECK-NEXT:    ret i8 0
28;
29entry:
30  %n.vec = and i32 %n, -8
31  %l0 = trunc i16 %x to i8
32  %l1 = insertelement <8 x i8> undef, i8 %l0, i32 0
33  %broadcast.splat26 = shufflevector <8 x i8> %l1, <8 x i8> undef, <8 x i32> zeroinitializer
34  br label %vector.body
35
36vector.body:                                      ; preds = %vector.body, %entry
37  %index = phi i32 [ 0, %entry ], [ %index.next, %vector.body ]
38  %l6 = getelementptr inbounds i16, i16* %s1, i32 %index
39  %l7 = bitcast i16* %l6 to <8 x i16>*
40  %wide.load = load <8 x i16>, <8 x i16>* %l7, align 2
41  %l8 = trunc <8 x i16> %wide.load to <8 x i8>
42  %l9 = mul <8 x i8> %broadcast.splat26, %l8
43  %l13 = getelementptr inbounds i8, i8* %d, i32 %index
44  %l14 = bitcast i8* %l13 to <8 x i8>*
45  store <8 x i8> %l9, <8 x i8>* %l14, align 1
46  %index.next = add i32 %index, 8
47  %l15 = icmp eq i32 %index.next, %n.vec
48  br i1 %l15, label %exit, label %vector.body
49
50exit:                                     ; preds = %vector.body
51  ret i8 0
52}
53
54; Same as above, but the shuffle has an extra use meaning it shouldnt be deleted
55define signext i8 @alive(i16* noalias nocapture readonly %s1, i16 zeroext %x, i8* noalias nocapture %d, i32 %n) {
56; CHECK-LABEL: @alive(
57; CHECK-NEXT:  entry:
58; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[N:%.*]], -8
59; CHECK-NEXT:    [[L0:%.*]] = trunc i16 [[X:%.*]] to i8
60; CHECK-NEXT:    [[L1:%.*]] = insertelement <8 x i8> undef, i8 [[L0]], i32 0
61; CHECK-NEXT:    [[BROADCAST_SPLAT26:%.*]] = shufflevector <8 x i8> [[L1]], <8 x i8> undef, <8 x i32> zeroinitializer
62; CHECK-NEXT:    [[L2:%.*]] = sub <8 x i8> zeroinitializer, [[BROADCAST_SPLAT26]]
63; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
64; CHECK:       vector.body:
65; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
66; CHECK-NEXT:    [[TMP0:%.*]] = trunc i16 [[X]] to i8
67; CHECK-NEXT:    [[L6:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i32 [[INDEX]]
68; CHECK-NEXT:    [[L7:%.*]] = bitcast i16* [[L6]] to <8 x i16>*
69; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <8 x i16>, <8 x i16>* [[L7]], align 2
70; CHECK-NEXT:    [[L8:%.*]] = trunc <8 x i16> [[WIDE_LOAD]] to <8 x i8>
71; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <8 x i8> undef, i8 [[TMP0]], i32 0
72; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> undef, <8 x i32> zeroinitializer
73; CHECK-NEXT:    [[L9:%.*]] = mul <8 x i8> [[TMP2]], [[L8]]
74; CHECK-NEXT:    [[L13:%.*]] = getelementptr inbounds i8, i8* [[D:%.*]], i32 [[INDEX]]
75; CHECK-NEXT:    [[L14:%.*]] = bitcast i8* [[L13]] to <8 x i8>*
76; CHECK-NEXT:    store <8 x i8> [[L9]], <8 x i8>* [[L14]], align 1
77; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 8
78; CHECK-NEXT:    [[L15:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
79; CHECK-NEXT:    br i1 [[L15]], label [[EXIT:%.*]], label [[VECTOR_BODY]]
80; CHECK:       exit:
81; CHECK-NEXT:    ret i8 0
82;
83entry:
84  %n.vec = and i32 %n, -8
85  %l0 = trunc i16 %x to i8
86  %l1 = insertelement <8 x i8> undef, i8 %l0, i32 0
87  %broadcast.splat26 = shufflevector <8 x i8> %l1, <8 x i8> undef, <8 x i32> zeroinitializer
88  %l2 = sub <8 x i8> zeroinitializer, %broadcast.splat26
89  br label %vector.body
90
91vector.body:                                      ; preds = %vector.body, %entry
92  %index = phi i32 [ 0, %entry ], [ %index.next, %vector.body ]
93  %l6 = getelementptr inbounds i16, i16* %s1, i32 %index
94  %l7 = bitcast i16* %l6 to <8 x i16>*
95  %wide.load = load <8 x i16>, <8 x i16>* %l7, align 2
96  %l8 = trunc <8 x i16> %wide.load to <8 x i8>
97  %l9 = mul <8 x i8> %broadcast.splat26, %l8
98  %l13 = getelementptr inbounds i8, i8* %d, i32 %index
99  %l14 = bitcast i8* %l13 to <8 x i8>*
100  store <8 x i8> %l9, <8 x i8>* %l14, align 1
101  %index.next = add i32 %index, 8
102  %l15 = icmp eq i32 %index.next, %n.vec
103  br i1 %l15, label %exit, label %vector.body
104
105exit:                                     ; preds = %vector.body
106  ret i8 0
107}
108