1 ; RUN: opt -S -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=1 < %s | FileCheck %s
3 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
5 ; Check the scenario where we have an unknown Stride, which happens to also be
6 ; the loop iteration count, so if we specialize the loop for the Stride==1 case,
7 ; this also implies that the loop will iterate no more than a single iteration,
8 ; as in the following example:
12 ; for(unsigned int k=0;k<N;k++) {
16 ; We check here that the following runtime scev guard for Stride==1 is NOT generated:
18 ; %ident.check = icmp ne i32 %N, 1
19 ; %0 = or i1 false, %ident.check
20 ; br i1 %0, label %scalar.ph, label %vector.ph
21 ; Instead the loop is vectorized with an unknown stride.
24 ; CHECK: for.body.lr.ph
25 ; CHECK-NOT: %ident.check = icmp ne i32 %N, 1
26 ; CHECK-NOT: %{{[0-9]+}} = or i1 false, %ident.check
27 ; CHECK-NOT: br i1 %{{[0-9]+}}, label %scalar.ph, label %vector.ph
35 define i32 @foo1(i32 %N, ptr nocapture readnone %A, ptr nocapture readonly %B, i32 %i, i32 %j) {
37 %cmp8 = icmp eq i32 %N, 0
38 br i1 %cmp8, label %for.end, label %for.body.lr.ph
44 %tmp.010 = phi i32 [ 0, %for.body.lr.ph ], [ %add1, %for.body ]
45 %k.09 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
46 %mul = mul i32 %k.09, %N
47 %add = add i32 %mul, %j
48 %arrayidx = getelementptr inbounds i16, ptr %B, i32 %add
49 %0 = load i16, ptr %arrayidx, align 2
50 %conv = sext i16 %0 to i32
51 %add1 = add nsw i32 %tmp.010, %conv
52 %inc = add nuw i32 %k.09, 1
53 %exitcond = icmp eq i32 %inc, %N
54 br i1 %exitcond, label %for.end.loopexit, label %for.body
57 %add1.lcssa = phi i32 [ %add1, %for.body ]
61 %tmp.0.lcssa = phi i32 [ 0, %entry ], [ %add1.lcssa, %for.end.loopexit ]
66 ; Check the same, but also where the Stride and the loop iteration count
67 ; are not of the same data type.
71 ; for(unsigned int k=0;k<N;k++) {
75 ; We check here that the following runtime scev guard for Stride==1 is NOT generated:
77 ; %ident.check = icmp ne i16 %N, 1
78 ; %0 = or i1 false, %ident.check
79 ; br i1 %0, label %scalar.ph, label %vector.ph
83 ; CHECK: vector.scevcheck:
84 ; CHECK-NEXT: add nsw i32 %conv, -1
85 ; CHECK-NEXT: [[NEG:%.+]] = sub i32 0, %conv
86 ; CHECK-NOT: %ident.check = icmp ne i16 %N, 1
87 ; CHECK-NOT: %{{[0-9]+}} = or i1 false, %ident.check
88 ; CHECK-NOT: br i1 %{{[0-9]+}}, label %scalar.ph, label %vector.ph
95 define i32 @foo2(i16 zeroext %N, ptr nocapture readnone %A, ptr nocapture readonly %B, i32 %i, i32 %j) {
97 %conv = zext i16 %N to i32
98 %cmp11 = icmp eq i16 %N, 0
99 br i1 %cmp11, label %for.end, label %for.body.lr.ph
105 %tmp.013 = phi i32 [ 0, %for.body.lr.ph ], [ %add4, %for.body ]
106 %k.012 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
107 %mul = mul nuw i32 %k.012, %conv
108 %add = add i32 %mul, %j
109 %arrayidx = getelementptr inbounds i16, ptr %B, i32 %add
110 %0 = load i16, ptr %arrayidx, align 2
111 %conv3 = sext i16 %0 to i32
112 %add4 = add nsw i32 %tmp.013, %conv3
113 %inc = add nuw nsw i32 %k.012, 1
114 %exitcond = icmp eq i32 %inc, %conv
115 br i1 %exitcond, label %for.end.loopexit, label %for.body
118 %add4.lcssa = phi i32 [ %add4, %for.body ]
122 %tmp.0.lcssa = phi i32 [ 0, %entry ], [ %add4.lcssa, %for.end.loopexit ]