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
5 @a = global [1024 x i8] zeroinitializer, align 16
6 @b = global [1024 x i8] zeroinitializer, align 16
9 ; This function has a loop of SAD pattern. Here we check when VF = 16 the
10 ; register usage doesn't exceed 16.
13 ; CHECK: LV(REG): VF = 8
14 ; CHECK-NEXT: LV(REG): Found max usage: 7
15 ; CHECK: LV(REG): VF = 16
16 ; CHECK-NEXT: LV(REG): Found max usage: 13
22 %add.lcssa = phi i32 [ %add, %for.body ]
26 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
27 %s.015 = phi i32 [ 0, %entry ], [ %add, %for.body ]
28 %arrayidx = getelementptr inbounds [1024 x i8], [1024 x i8]* @a, i64 0, i64 %indvars.iv
29 %0 = load i8, i8* %arrayidx, align 1
30 %conv = zext i8 %0 to i32
31 %arrayidx2 = getelementptr inbounds [1024 x i8], [1024 x i8]* @b, i64 0, i64 %indvars.iv
32 %1 = load i8, i8* %arrayidx2, align 1
33 %conv3 = zext i8 %1 to i32
34 %sub = sub nsw i32 %conv, %conv3
35 %ispos = icmp sgt i32 %sub, -1
36 %neg = sub nsw i32 0, %sub
37 %2 = select i1 %ispos, i32 %sub, i32 %neg
38 %add = add nsw i32 %2, %s.015
39 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
40 %exitcond = icmp eq i64 %indvars.iv.next, 1024
41 br i1 %exitcond, label %for.cond.cleanup, label %for.body
45 ; For indvars.iv used in a computating chain only feeding into getelementptr or cmp,
46 ; it will not have vector version and the vector register usage will not exceed the
47 ; available vector register number.
49 ; CHECK: LV(REG): VF = 8
50 ; CHECK-NEXT: LV(REG): Found max usage: 7
51 ; CHECK: LV(REG): VF = 16
52 ; CHECK-NEXT: LV(REG): Found max usage: 13
56 for.cond.cleanup: ; preds = %for.body
57 %add.lcssa = phi i32 [ %add, %for.body ]
60 for.body: ; preds = %for.body, %entry
61 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
62 %s.015 = phi i32 [ 0, %entry ], [ %add, %for.body ]
63 %tmp1 = add nsw i64 %indvars.iv, 3
64 %arrayidx = getelementptr inbounds [1024 x i8], [1024 x i8]* @a, i64 0, i64 %tmp1
65 %tmp = load i8, i8* %arrayidx, align 1
66 %conv = zext i8 %tmp to i32
67 %tmp2 = add nsw i64 %indvars.iv, 2
68 %arrayidx2 = getelementptr inbounds [1024 x i8], [1024 x i8]* @b, i64 0, i64 %tmp2
69 %tmp3 = load i8, i8* %arrayidx2, align 1
70 %conv3 = zext i8 %tmp3 to i32
71 %sub = sub nsw i32 %conv, %conv3
72 %ispos = icmp sgt i32 %sub, -1
73 %neg = sub nsw i32 0, %sub
74 %tmp4 = select i1 %ispos, i32 %sub, i32 %neg
75 %add = add nsw i32 %tmp4, %s.015
76 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
77 %exitcond = icmp eq i64 %indvars.iv.next, 1024
78 br i1 %exitcond, label %for.cond.cleanup, label %for.body
81 define i64 @bar(i64* nocapture %a) {
83 ; CHECK: LV(REG): VF = 2
84 ; CHECK: LV(REG): Found max usage: 3
90 %add2.lcssa = phi i64 [ %add2, %for.body ]
94 %i.012 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
95 %s.011 = phi i64 [ 0, %entry ], [ %add2, %for.body ]
96 %arrayidx = getelementptr inbounds i64, i64* %a, i64 %i.012
97 %0 = load i64, i64* %arrayidx, align 8
98 %add = add nsw i64 %0, %i.012
99 store i64 %add, i64* %arrayidx, align 8
100 %add2 = add nsw i64 %add, %s.011
101 %inc = add nuw nsw i64 %i.012, 1
102 %exitcond = icmp eq i64 %inc, 1024
103 br i1 %exitcond, label %for.cond.cleanup, label %for.body
106 @d = external global [0 x i64], align 8
107 @e = external global [0 x i32], align 4
108 @c = external global [0 x i32], align 4
110 define void @hoo(i32 %n) {
111 ; For c[i] = e[d[i]] in the loop, e[d[i]] is not consecutive but its index %tmp can
112 ; be gathered into a vector. For VF == 16, the vector version of %tmp will be <16 x i64>
113 ; so the max usage of AVX512 vector register will be 2.
115 ; AVX512F: LV(REG): VF = 16
116 ; AVX512F: LV(REG): Found max usage: 2
121 for.body: ; preds = %for.body, %entry
122 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
123 %arrayidx = getelementptr inbounds [0 x i64], [0 x i64]* @d, i64 0, i64 %indvars.iv
124 %tmp = load i64, i64* %arrayidx, align 8
125 %arrayidx1 = getelementptr inbounds [0 x i32], [0 x i32]* @e, i64 0, i64 %tmp
126 %tmp1 = load i32, i32* %arrayidx1, align 4
127 %arrayidx3 = getelementptr inbounds [0 x i32], [0 x i32]* @c, i64 0, i64 %indvars.iv
128 store i32 %tmp1, i32* %arrayidx3, align 4
129 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
130 %exitcond = icmp eq i64 %indvars.iv.next, 10000
131 br i1 %exitcond, label %for.end, label %for.body
133 for.end: ; preds = %for.body