1 ; RUN: opt < %s -loop-vectorize -force-vector-width=2 -force-vector-interleave=1 -S | FileCheck %s
2 ; RUN: opt < %s -loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S | FileCheck %s -check-prefix=WIDTH
3 ; RUN: opt -S -loop-vectorize -force-vector-width=4 < %s | FileCheck %s -check-prefix=RIGHTVF
4 ; RUN: opt -S -loop-vectorize -force-vector-width=8 < %s | FileCheck %s -check-prefix=WRONGVF
6 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
8 ; Vectorization with dependence checks.
10 ; No plausible dependence - can be vectorized.
11 ; for (i = 0; i < 1024; ++i)
12 ; A[i] = A[i + 1] + 1;
14 ; CHECK-LABEL: @f1_vec(
17 define void @f1_vec(i32* %A) {
22 %indvars.iv = phi i32 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
23 %indvars.iv.next = add i32 %indvars.iv, 1
24 %arrayidx = getelementptr inbounds i32, i32* %A, i32 %indvars.iv.next
25 %0 = load i32, i32* %arrayidx, align 4
26 %add1 = add nsw i32 %0, 1
27 %arrayidx3 = getelementptr inbounds i32, i32* %A, i32 %indvars.iv
28 store i32 %add1, i32* %arrayidx3, align 4
29 %exitcond = icmp ne i32 %indvars.iv.next, 1024
30 br i1 %exitcond, label %for.body, label %for.end
36 ; Plausible dependence of distance 1 - can't be vectorized.
37 ; for (i = 0; i < 1024; ++i)
40 ; CHECK-LABEL: @f2_novec(
41 ; CHECK-NOT: <2 x i32>
43 define void @f2_novec(i32* %A) {
48 %indvars.iv = phi i32 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
49 %arrayidx = getelementptr inbounds i32, i32* %A, i32 %indvars.iv
50 %0 = load i32, i32* %arrayidx, align 4
51 %add = add nsw i32 %0, 1
52 %indvars.iv.next = add i32 %indvars.iv, 1
53 %arrayidx3 = getelementptr inbounds i32, i32* %A, i32 %indvars.iv.next
54 store i32 %add, i32* %arrayidx3, align 4
55 %exitcond = icmp ne i32 %indvars.iv.next, 1024
56 br i1 %exitcond, label %for.body, label %for.end
62 ; Plausible dependence of distance 2 - can be vectorized with a width of 2.
63 ; for (i = 0; i < 1024; ++i)
66 ; CHECK-LABEL: @f3_vec_len(
70 ; WIDTH-NOT: <4 x i32>
72 define void @f3_vec_len(i32* %A) {
77 %i.01 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
78 %idxprom = sext i32 %i.01 to i64
79 %arrayidx = getelementptr inbounds i32, i32* %A, i64 %idxprom
80 %0 = load i32, i32* %arrayidx, align 4
81 %add = add nsw i32 %0, 1
82 %add1 = add nsw i32 %i.01, 2
83 %idxprom2 = sext i32 %add1 to i64
84 %arrayidx3 = getelementptr inbounds i32, i32* %A, i64 %idxprom2
85 store i32 %add, i32* %arrayidx3, align 4
86 %inc = add nsw i32 %i.01, 1
87 %cmp = icmp slt i32 %inc, 1024
88 br i1 %cmp, label %for.body, label %for.end
94 ; Plausible dependence of distance 1 - cannot be vectorized (without reordering
96 ; for (i = 0; i < 1024; ++i) {
102 ; CHECK-NOT: <2 x i32>
104 define void @f5(i32* %A, i32* %B) {
109 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
110 %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
111 %0 = load i32, i32* %arrayidx, align 4
112 %arrayidx2 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
113 store i32 %0, i32* %arrayidx2, align 4
114 %indvars.iv.next = add nsw i64 %indvars.iv, 1
115 %arrayidx4 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv.next
116 %1 = load i32, i32* %arrayidx4, align 4
117 store i32 %1, i32* %arrayidx, align 4
118 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
119 %exitcond = icmp ne i32 %lftr.wideiv, 1024
120 br i1 %exitcond, label %for.body, label %for.end
126 ; Dependence through a phi node - must not vectorize.
127 ; for (i = 0; i < 1024; ++i) {
133 ; CHECK-NOT: <2 x i32>
135 define i32 @f6(i32* %a, i32 %tmp) {
140 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
141 %tmp.addr.08 = phi i32 [ %tmp, %entry ], [ %0, %for.body ]
142 %indvars.iv.next = add nsw i64 %indvars.iv, 1
143 %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv.next
144 store i32 %tmp.addr.08, i32* %arrayidx, align 4
145 %arrayidx3 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
146 %0 = load i32, i32* %arrayidx3, align 4
147 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
148 %exitcond = icmp ne i32 %lftr.wideiv, 1024
149 br i1 %exitcond, label %for.body, label %for.end
155 ; Don't vectorize true loop carried dependencies that are not a multiple of the
158 ; for (int i = ...; ++i) {
159 ; a[i] = a[i-3] + ...;
160 ; It is a bad idea to vectorize this loop because store-load forwarding will not
164 ; CHECK-LABEL: @nostoreloadforward(
165 ; CHECK-NOT: <2 x i32>
167 define void @nostoreloadforward(i32* %A) {
172 %indvars.iv = phi i64 [ 16, %entry ], [ %indvars.iv.next, %for.body ]
173 %0 = add nsw i64 %indvars.iv, -3
174 %arrayidx = getelementptr inbounds i32, i32* %A, i64 %0
175 %1 = load i32, i32* %arrayidx, align 4
176 %2 = add nsw i64 %indvars.iv, 4
177 %arrayidx2 = getelementptr inbounds i32, i32* %A, i64 %2
178 %3 = load i32, i32* %arrayidx2, align 4
179 %add3 = add nsw i32 %3, %1
180 %arrayidx5 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
181 store i32 %add3, i32* %arrayidx5, align 4
182 %indvars.iv.next = add i64 %indvars.iv, 1
183 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
184 %exitcond = icmp ne i32 %lftr.wideiv, 128
185 br i1 %exitcond, label %for.body, label %for.end
192 ; for (int i = ...; ++i) {
194 ; c[i] = a[i-3] + ...;
195 ; It is a bad idea to vectorize this loop because store-load forwarding will not
199 ; CHECK-LABEL: @nostoreloadforward2(
200 ; CHECK-NOT: <2 x i32>
202 define void @nostoreloadforward2(i32* noalias %A, i32* noalias %B, i32* noalias %C) {
207 %indvars.iv = phi i64 [ 16, %entry ], [ %indvars.iv.next, %for.body ]
208 %arrayidx = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
209 %0 = load i32, i32* %arrayidx, align 4
210 %arrayidx2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
211 store i32 %0, i32* %arrayidx2, align 4
212 %1 = add nsw i64 %indvars.iv, -3
213 %arrayidx4 = getelementptr inbounds i32, i32* %A, i64 %1
214 %2 = load i32, i32* %arrayidx4, align 4
215 %arrayidx6 = getelementptr inbounds i32, i32* %C, i64 %indvars.iv
216 store i32 %2, i32* %arrayidx6, align 4
217 %indvars.iv.next = add i64 %indvars.iv, 1
218 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
219 %exitcond = icmp ne i32 %lftr.wideiv, 128
220 br i1 %exitcond, label %for.body, label %for.end
227 ;Check the new calculation of the maximum safe distance in bits which can be vectorized.
228 ;The previous behavior did not take account that the stride was 2.
229 ;Therefore the maxVF was computed as 8 instead of 4, as the dependence distance here is 6 iterations, given by |N-(N-12)|/2.
236 ; for (j = 0; j < M - 6; ++j)
239 ; a[N - 12 - 2 * j] = 7;
244 ; RIGHTVF-LABEL: @pr34283
247 ; WRONGVF-LABLE: @pr34283
248 ; WRONGVF-NOT: <8 x i64>
250 @a = common local_unnamed_addr global [64 x i32] zeroinitializer, align 16
252 ; Function Attrs: norecurse nounwind uwtable
253 define void @pr34283() local_unnamed_addr {
258 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
259 %0 = shl i64 %indvars.iv, 1
260 %1 = sub nuw nsw i64 64, %0
261 %arrayidx = getelementptr inbounds [64 x i32], [64 x i32]* @a, i64 0, i64 %1
262 store i32 69, i32* %arrayidx, align 8
263 %2 = sub nuw nsw i64 52, %0
264 %arrayidx4 = getelementptr inbounds [64 x i32], [64 x i32]* @a, i64 0, i64 %2
265 store i32 7, i32* %arrayidx4, align 8
266 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
267 %exitcond = icmp eq i64 %indvars.iv.next, 26
268 br i1 %exitcond, label %for.end, label %for.body