1; RUN: opt < %s -debug-only=loop-vectorize -loop-vectorize -vectorizer-maximize-bandwidth -O2 -mtriple=x86_64-unknown-linux -S 2>&1 | FileCheck %s
2; RUN: opt < %s -debug-only=loop-vectorize -loop-vectorize -vectorizer-maximize-bandwidth -O2 -mtriple=x86_64-unknown-linux -mattr=+avx512f -S 2>&1 | FileCheck %s --check-prefix=AVX512F
3; REQUIRES: asserts
4
5@a = global [1024 x i8] zeroinitializer, align 16
6@b = global [1024 x i8] zeroinitializer, align 16
7
8define i32 @foo() {
9; This function has a loop of SAD pattern. Here we check when VF = 16 the
10; register usage doesn't exceed 16.
11;
12; CHECK-LABEL: foo
13; CHECK:      LV(REG): VF = 4
14; CHECK-NEXT: LV(REG): Found max usage: 4
15; CHECK:      LV(REG): VF = 8
16; CHECK-NEXT: LV(REG): Found max usage: 7
17; CHECK:      LV(REG): VF = 16
18; CHECK-NEXT: LV(REG): Found max usage: 13
19
20entry:
21  br label %for.body
22
23for.cond.cleanup:
24  %add.lcssa = phi i32 [ %add, %for.body ]
25  ret i32 %add.lcssa
26
27for.body:
28  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
29  %s.015 = phi i32 [ 0, %entry ], [ %add, %for.body ]
30  %arrayidx = getelementptr inbounds [1024 x i8], [1024 x i8]* @a, i64 0, i64 %indvars.iv
31  %0 = load i8, i8* %arrayidx, align 1
32  %conv = zext i8 %0 to i32
33  %arrayidx2 = getelementptr inbounds [1024 x i8], [1024 x i8]* @b, i64 0, i64 %indvars.iv
34  %1 = load i8, i8* %arrayidx2, align 1
35  %conv3 = zext i8 %1 to i32
36  %sub = sub nsw i32 %conv, %conv3
37  %ispos = icmp sgt i32 %sub, -1
38  %neg = sub nsw i32 0, %sub
39  %2 = select i1 %ispos, i32 %sub, i32 %neg
40  %add = add nsw i32 %2, %s.015
41  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
42  %exitcond = icmp eq i64 %indvars.iv.next, 1024
43  br i1 %exitcond, label %for.cond.cleanup, label %for.body
44}
45
46define i32 @goo() {
47; For indvars.iv used in a computating chain only feeding into getelementptr or cmp,
48; it will not have vector version and the vector register usage will not exceed the
49; available vector register number.
50; CHECK-LABEL: goo
51; CHECK:      LV(REG): VF = 4
52; CHECK-NEXT: LV(REG): Found max usage: 4
53; CHECK:      LV(REG): VF = 8
54; CHECK-NEXT: LV(REG): Found max usage: 7
55; CHECK:      LV(REG): VF = 16
56; CHECK-NEXT: LV(REG): Found max usage: 13
57entry:
58  br label %for.body
59
60for.cond.cleanup:                                 ; preds = %for.body
61  %add.lcssa = phi i32 [ %add, %for.body ]
62  ret i32 %add.lcssa
63
64for.body:                                         ; preds = %for.body, %entry
65  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
66  %s.015 = phi i32 [ 0, %entry ], [ %add, %for.body ]
67  %tmp1 = add nsw i64 %indvars.iv, 3
68  %arrayidx = getelementptr inbounds [1024 x i8], [1024 x i8]* @a, i64 0, i64 %tmp1
69  %tmp = load i8, i8* %arrayidx, align 1
70  %conv = zext i8 %tmp to i32
71  %tmp2 = add nsw i64 %indvars.iv, 2
72  %arrayidx2 = getelementptr inbounds [1024 x i8], [1024 x i8]* @b, i64 0, i64 %tmp2
73  %tmp3 = load i8, i8* %arrayidx2, align 1
74  %conv3 = zext i8 %tmp3 to i32
75  %sub = sub nsw i32 %conv, %conv3
76  %ispos = icmp sgt i32 %sub, -1
77  %neg = sub nsw i32 0, %sub
78  %tmp4 = select i1 %ispos, i32 %sub, i32 %neg
79  %add = add nsw i32 %tmp4, %s.015
80  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
81  %exitcond = icmp eq i64 %indvars.iv.next, 1024
82  br i1 %exitcond, label %for.cond.cleanup, label %for.body
83}
84
85define i64 @bar(i64* nocapture %a) {
86; CHECK-LABEL: bar
87; CHECK:       LV(REG): VF = 2
88; CHECK:       LV(REG): Found max usage: 4
89;
90entry:
91  br label %for.body
92
93for.cond.cleanup:
94  %add2.lcssa = phi i64 [ %add2, %for.body ]
95  ret i64 %add2.lcssa
96
97for.body:
98  %i.012 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
99  %s.011 = phi i64 [ 0, %entry ], [ %add2, %for.body ]
100  %arrayidx = getelementptr inbounds i64, i64* %a, i64 %i.012
101  %0 = load i64, i64* %arrayidx, align 8
102  %add = add nsw i64 %0, %i.012
103  store i64 %add, i64* %arrayidx, align 8
104  %add2 = add nsw i64 %add, %s.011
105  %inc = add nuw nsw i64 %i.012, 1
106  %exitcond = icmp eq i64 %inc, 1024
107  br i1 %exitcond, label %for.cond.cleanup, label %for.body
108}
109
110@d = external global [0 x i64], align 8
111@e = external global [0 x i32], align 4
112@c = external global [0 x i32], align 4
113
114define void @hoo(i32 %n) {
115; For c[i] = e[d[i]] in the loop, e[d[i]] is not consecutive but its index %tmp can
116; be gathered into a vector. For VF == 16, the vector version of %tmp will be <16 x i64>
117; so the max usage of AVX512 vector register will be 2.
118; AVX512F-LABEL: bar
119; AVX512F:       LV(REG): VF = 16
120; AVX512F:       LV(REG): Found max usage: 2
121;
122entry:
123  br label %for.body
124
125for.body:                                         ; preds = %for.body, %entry
126  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
127  %arrayidx = getelementptr inbounds [0 x i64], [0 x i64]* @d, i64 0, i64 %indvars.iv
128  %tmp = load i64, i64* %arrayidx, align 8
129  %arrayidx1 = getelementptr inbounds [0 x i32], [0 x i32]* @e, i64 0, i64 %tmp
130  %tmp1 = load i32, i32* %arrayidx1, align 4
131  %arrayidx3 = getelementptr inbounds [0 x i32], [0 x i32]* @c, i64 0, i64 %indvars.iv
132  store i32 %tmp1, i32* %arrayidx3, align 4
133  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
134  %exitcond = icmp eq i64 %indvars.iv.next, 10000
135  br i1 %exitcond, label %for.end, label %for.body
136
137for.end:                                          ; preds = %for.body
138  ret void
139}
140