1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -scoped-noalias-aa -slp-vectorizer -mtriple=arm64-apple-darwin -enable-new-pm=false -S %s | FileCheck %s
3
4define void @needs_versioning_not_profitable(i32* %dst, i32* %src) {
5; CHECK-LABEL: @needs_versioning_not_profitable(
6; CHECK-NEXT:  entry:
7; CHECK-NEXT:    [[SRC_0:%.*]] = load i32, i32* [[SRC:%.*]], align 4
8; CHECK-NEXT:    [[R_0:%.*]] = ashr i32 [[SRC_0]], 16
9; CHECK-NEXT:    store i32 [[R_0]], i32* [[DST:%.*]], align 4
10; CHECK-NEXT:    [[SRC_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 1
11; CHECK-NEXT:    [[SRC_1:%.*]] = load i32, i32* [[SRC_GEP_1]], align 4
12; CHECK-NEXT:    [[R_1:%.*]] = ashr i32 [[SRC_1]], 16
13; CHECK-NEXT:    [[DST_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 1
14; CHECK-NEXT:    store i32 [[R_1]], i32* [[DST_GEP_1]], align 4
15; CHECK-NEXT:    ret void
16;
17entry:
18  %src.0 = load i32, i32* %src, align 4
19  %r.0 = ashr i32 %src.0, 16
20  store i32 %r.0, i32* %dst, align 4
21  %src.gep.1 = getelementptr inbounds i32, i32* %src, i64 1
22  %src.1 = load i32, i32* %src.gep.1, align 4
23  %r.1 = ashr i32 %src.1, 16
24  %dst.gep.1 = getelementptr inbounds i32, i32* %dst, i64 1
25  store i32 %r.1, i32* %dst.gep.1, align 4
26  ret void
27}
28
29define void @needs_versioning_profitable(i32* %dst, i32* %src) {
30; CHECK-LABEL: @needs_versioning_profitable(
31; CHECK-NEXT:  entry:
32; CHECK-NEXT:    [[SRC_0:%.*]] = load i32, i32* [[SRC:%.*]], align 4
33; CHECK-NEXT:    [[R_0:%.*]] = ashr i32 [[SRC_0]], 16
34; CHECK-NEXT:    store i32 [[R_0]], i32* [[DST:%.*]], align 4
35; CHECK-NEXT:    [[SRC_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 1
36; CHECK-NEXT:    [[SRC_1:%.*]] = load i32, i32* [[SRC_GEP_1]], align 4
37; CHECK-NEXT:    [[R_1:%.*]] = ashr i32 [[SRC_1]], 16
38; CHECK-NEXT:    [[DST_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 1
39; CHECK-NEXT:    store i32 [[R_1]], i32* [[DST_GEP_1]], align 4
40; CHECK-NEXT:    [[SRC_GEP_2:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 2
41; CHECK-NEXT:    [[SRC_2:%.*]] = load i32, i32* [[SRC_GEP_2]], align 4
42; CHECK-NEXT:    [[R_2:%.*]] = ashr i32 [[SRC_2]], 16
43; CHECK-NEXT:    [[DST_GEP_2:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 2
44; CHECK-NEXT:    store i32 [[R_2]], i32* [[DST_GEP_2]], align 4
45; CHECK-NEXT:    [[SRC_GEP_3:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 3
46; CHECK-NEXT:    [[SRC_3:%.*]] = load i32, i32* [[SRC_GEP_3]], align 4
47; CHECK-NEXT:    [[R_3:%.*]] = ashr i32 [[SRC_3]], 16
48; CHECK-NEXT:    [[DST_GEP_3:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 3
49; CHECK-NEXT:    store i32 [[R_3]], i32* [[DST_GEP_3]], align 4
50; CHECK-NEXT:    ret void
51;
52entry:
53  %src.0 = load i32, i32* %src, align 4
54  %r.0 = ashr i32 %src.0, 16
55  store i32 %r.0, i32* %dst, align 4
56  %src.gep.1 = getelementptr inbounds i32, i32* %src, i64 1
57  %src.1 = load i32, i32* %src.gep.1, align 4
58  %r.1 = ashr i32 %src.1, 16
59  %dst.gep.1 = getelementptr inbounds i32, i32* %dst, i64 1
60  store i32 %r.1, i32* %dst.gep.1, align 4
61  %src.gep.2 = getelementptr inbounds i32, i32* %src, i64 2
62  %src.2 = load i32, i32* %src.gep.2, align 4
63  %r.2 = ashr i32 %src.2, 16
64  %dst.gep.2 = getelementptr inbounds i32, i32* %dst, i64 2
65  store i32 %r.2, i32* %dst.gep.2, align 4
66  %src.gep.3 = getelementptr inbounds i32, i32* %src, i64 3
67  %src.3 = load i32, i32* %src.gep.3, align 4
68  %r.3 = ashr i32 %src.3, 16
69  %dst.gep.3 = getelementptr inbounds i32, i32* %dst, i64 3
70  store i32 %r.3, i32* %dst.gep.3, align 4
71
72  ret void
73}
74
75
76define void @no_version(i32* nocapture %dst, i32* nocapture readonly %src) {
77; CHECK-LABEL: @no_version(
78; CHECK-NEXT:  entry:
79; CHECK-NEXT:    [[SRC_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[SRC:%.*]], i64 1
80; CHECK-NEXT:    [[TMP0:%.*]] = bitcast i32* [[SRC]] to <2 x i32>*
81; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x i32>, <2 x i32>* [[TMP0]], align 4
82; CHECK-NEXT:    [[TMP2:%.*]] = ashr <2 x i32> [[TMP1]], <i32 16, i32 16>
83; CHECK-NEXT:    [[DST_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[DST:%.*]], i64 1
84; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[DST]] to <2 x i32>*
85; CHECK-NEXT:    store <2 x i32> [[TMP2]], <2 x i32>* [[TMP3]], align 4
86; CHECK-NEXT:    ret void
87;
88entry:
89  %src.0 = load i32, i32* %src, align 4
90  %src.gep.1 = getelementptr inbounds i32, i32* %src, i64 1
91  %src.1 = load i32, i32* %src.gep.1, align 4
92  %r.0 = ashr i32 %src.0, 16
93  %r.1 = ashr i32 %src.1, 16
94  %dst.gep.1 = getelementptr inbounds i32, i32* %dst, i64 1
95  store i32 %r.0, i32* %dst, align 4
96  store i32 %r.1, i32* %dst.gep.1, align 4
97  ret void
98}
99
100define void @version_multiple(i32* nocapture %out_block, i32* nocapture readonly %counter) {
101; CHECK-LABEL: @version_multiple(
102; CHECK-NEXT:  entry:
103; CHECK-NEXT:    [[TMP0:%.*]] = load i32, i32* [[COUNTER:%.*]], align 4
104; CHECK-NEXT:    [[TMP1:%.*]] = load i32, i32* [[OUT_BLOCK:%.*]], align 4
105; CHECK-NEXT:    [[XOR:%.*]] = xor i32 [[TMP1]], [[TMP0]]
106; CHECK-NEXT:    store i32 [[XOR]], i32* [[OUT_BLOCK]], align 4
107; CHECK-NEXT:    [[ARRAYIDX_1:%.*]] = getelementptr inbounds i32, i32* [[COUNTER]], i64 1
108; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32* [[ARRAYIDX_1]], align 4
109; CHECK-NEXT:    [[ARRAYIDX2_1:%.*]] = getelementptr inbounds i32, i32* [[OUT_BLOCK]], i64 1
110; CHECK-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX2_1]], align 4
111; CHECK-NEXT:    [[XOR_1:%.*]] = xor i32 [[TMP3]], [[TMP2]]
112; CHECK-NEXT:    store i32 [[XOR_1]], i32* [[ARRAYIDX2_1]], align 4
113; CHECK-NEXT:    [[ARRAYIDX_2:%.*]] = getelementptr inbounds i32, i32* [[COUNTER]], i64 2
114; CHECK-NEXT:    [[TMP4:%.*]] = load i32, i32* [[ARRAYIDX_2]], align 4
115; CHECK-NEXT:    [[ARRAYIDX2_2:%.*]] = getelementptr inbounds i32, i32* [[OUT_BLOCK]], i64 2
116; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[ARRAYIDX2_2]], align 4
117; CHECK-NEXT:    [[XOR_2:%.*]] = xor i32 [[TMP5]], [[TMP4]]
118; CHECK-NEXT:    store i32 [[XOR_2]], i32* [[ARRAYIDX2_2]], align 4
119; CHECK-NEXT:    [[ARRAYIDX_3:%.*]] = getelementptr inbounds i32, i32* [[COUNTER]], i64 3
120; CHECK-NEXT:    [[TMP6:%.*]] = load i32, i32* [[ARRAYIDX_3]], align 4
121; CHECK-NEXT:    [[ARRAYIDX2_3:%.*]] = getelementptr inbounds i32, i32* [[OUT_BLOCK]], i64 3
122; CHECK-NEXT:    [[TMP7:%.*]] = load i32, i32* [[ARRAYIDX2_3]], align 4
123; CHECK-NEXT:    [[XOR_3:%.*]] = xor i32 [[TMP7]], [[TMP6]]
124; CHECK-NEXT:    store i32 [[XOR_3]], i32* [[ARRAYIDX2_3]], align 4
125; CHECK-NEXT:    ret void
126;
127entry:
128  %0 = load i32, i32* %counter, align 4
129  %1 = load i32, i32* %out_block, align 4
130  %xor = xor i32 %1, %0
131  store i32 %xor, i32* %out_block, align 4
132  %arrayidx.1 = getelementptr inbounds i32, i32* %counter, i64 1
133  %2 = load i32, i32* %arrayidx.1, align 4
134  %arrayidx2.1 = getelementptr inbounds i32, i32* %out_block, i64 1
135  %3 = load i32, i32* %arrayidx2.1, align 4
136  %xor.1 = xor i32 %3, %2
137  store i32 %xor.1, i32* %arrayidx2.1, align 4
138  %arrayidx.2 = getelementptr inbounds i32, i32* %counter, i64 2
139  %4 = load i32, i32* %arrayidx.2, align 4
140  %arrayidx2.2 = getelementptr inbounds i32, i32* %out_block, i64 2
141  %5 = load i32, i32* %arrayidx2.2, align 4
142  %xor.2 = xor i32 %5, %4
143  store i32 %xor.2, i32* %arrayidx2.2, align 4
144  %arrayidx.3 = getelementptr inbounds i32, i32* %counter, i64 3
145  %6 = load i32, i32* %arrayidx.3, align 4
146  %arrayidx2.3 = getelementptr inbounds i32, i32* %out_block, i64 3
147  %7 = load i32, i32* %arrayidx2.3, align 4
148  %xor.3 = xor i32 %7, %6
149  store i32 %xor.3, i32* %arrayidx2.3, align 4
150  ret void
151}
152
153define i32 @use_outside_version_bb(i32* %dst, i32* %src, i1 %c.1) {
154; CHECK-LABEL: @use_outside_version_bb(
155; CHECK-NEXT:  entry:
156; CHECK-NEXT:    [[SRC_0:%.*]] = load i32, i32* [[SRC:%.*]], align 4
157; CHECK-NEXT:    [[R_0:%.*]] = ashr i32 [[SRC_0]], 16
158; CHECK-NEXT:    store i32 [[R_0]], i32* [[DST:%.*]], align 4
159; CHECK-NEXT:    [[SRC_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 1
160; CHECK-NEXT:    [[SRC_1:%.*]] = load i32, i32* [[SRC_GEP_1]], align 4
161; CHECK-NEXT:    [[R_1:%.*]] = ashr i32 [[SRC_1]], 16
162; CHECK-NEXT:    [[DST_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 1
163; CHECK-NEXT:    store i32 [[R_1]], i32* [[DST_GEP_1]], align 4
164; CHECK-NEXT:    br label [[EXIT:%.*]]
165; CHECK:       exit:
166; CHECK-NEXT:    ret i32 [[R_0]]
167;
168entry:
169  %src.0 = load i32, i32* %src, align 4
170  %r.0 = ashr i32 %src.0, 16
171  store i32 %r.0, i32* %dst, align 4
172  %src.gep.1 = getelementptr inbounds i32, i32* %src, i64 1
173  %src.1 = load i32, i32* %src.gep.1, align 4
174  %r.1 = ashr i32 %src.1, 16
175  %dst.gep.1 = getelementptr inbounds i32, i32* %dst, i64 1
176  store i32 %r.1, i32* %dst.gep.1, align 4
177  br label %exit
178
179exit:
180  ret i32 %r.0
181}
182
183define i32 @value_used_in_return(i32* %dst, i32* %src, i32 %x) {
184; CHECK-LABEL: @value_used_in_return(
185; CHECK-NEXT:  entry:
186; CHECK-NEXT:    [[SRC_0:%.*]] = load i32, i32* [[SRC:%.*]], align 4
187; CHECK-NEXT:    [[R_0:%.*]] = ashr i32 [[SRC_0]], 16
188; CHECK-NEXT:    store i32 [[R_0]], i32* [[DST:%.*]], align 4
189; CHECK-NEXT:    [[SRC_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 1
190; CHECK-NEXT:    [[SRC_1:%.*]] = load i32, i32* [[SRC_GEP_1]], align 4
191; CHECK-NEXT:    [[R_1:%.*]] = ashr i32 [[SRC_1]], 16
192; CHECK-NEXT:    [[DST_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 1
193; CHECK-NEXT:    store i32 [[R_1]], i32* [[DST_GEP_1]], align 4
194; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[X:%.*]], 20
195; CHECK-NEXT:    ret i32 [[ADD]]
196;
197entry:
198  %src.0 = load i32, i32* %src, align 4
199  %r.0 = ashr i32 %src.0, 16
200  store i32 %r.0, i32* %dst, align 4
201  %src.gep.1 = getelementptr inbounds i32, i32* %src, i64 1
202  %src.1 = load i32, i32* %src.gep.1, align 4
203  %r.1 = ashr i32 %src.1, 16
204  %dst.gep.1 = getelementptr inbounds i32, i32* %dst, i64 1
205  store i32 %r.1, i32* %dst.gep.1, align 4
206  %add = add i32 %x, 20
207  ret i32 %add
208}
209define i32 @needs_versioning2_cond_br(i32* %dst, i32* %src, i1 %c.1) {
210; CHECK-LABEL: @needs_versioning2_cond_br(
211; CHECK-NEXT:  entry:
212; CHECK-NEXT:    br i1 [[C_1:%.*]], label [[THEN:%.*]], label [[ELSE:%.*]]
213; CHECK:       then:
214; CHECK-NEXT:    [[SRC_0:%.*]] = load i32, i32* [[SRC:%.*]], align 4
215; CHECK-NEXT:    [[R_0:%.*]] = ashr i32 [[SRC_0]], 16
216; CHECK-NEXT:    store i32 [[R_0]], i32* [[DST:%.*]], align 4
217; CHECK-NEXT:    [[SRC_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 1
218; CHECK-NEXT:    [[SRC_1:%.*]] = load i32, i32* [[SRC_GEP_1]], align 4
219; CHECK-NEXT:    [[R_1:%.*]] = ashr i32 [[SRC_1]], 16
220; CHECK-NEXT:    [[DST_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 1
221; CHECK-NEXT:    store i32 [[R_1]], i32* [[DST_GEP_1]], align 4
222; CHECK-NEXT:    ret i32 10
223; CHECK:       else:
224; CHECK-NEXT:    ret i32 0
225;
226entry:
227  br i1 %c.1, label %then, label %else
228
229then:
230  %src.0 = load i32, i32* %src, align 4
231  %r.0 = ashr i32 %src.0, 16
232  store i32 %r.0, i32* %dst, align 4
233  %src.gep.1 = getelementptr inbounds i32, i32* %src, i64 1
234  %src.1 = load i32, i32* %src.gep.1, align 4
235  %r.1 = ashr i32 %src.1, 16
236  %dst.gep.1 = getelementptr inbounds i32, i32* %dst, i64 1
237  store i32 %r.1, i32* %dst.gep.1, align 4
238  ret i32 10
239
240
241else:
242  ret i32 0
243}
244
245define void @pointer_defined_in_bb(i32* %dst, i32** %src.p) {
246; CHECK-LABEL: @pointer_defined_in_bb(
247; CHECK-NEXT:  entry:
248; CHECK-NEXT:    [[SRC:%.*]] = load i32*, i32** [[SRC_P:%.*]], align 8
249; CHECK-NEXT:    [[SRC_0:%.*]] = load i32, i32* [[SRC]], align 4
250; CHECK-NEXT:    [[R_0:%.*]] = ashr i32 [[SRC_0]], 16
251; CHECK-NEXT:    store i32 [[R_0]], i32* [[DST:%.*]], align 4
252; CHECK-NEXT:    [[SRC_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[SRC]], i64 1
253; CHECK-NEXT:    [[SRC_1:%.*]] = load i32, i32* [[SRC_GEP_1]], align 4
254; CHECK-NEXT:    [[R_1:%.*]] = ashr i32 [[SRC_1]], 16
255; CHECK-NEXT:    [[DST_GEP_1:%.*]] = getelementptr inbounds i32, i32* [[DST]], i64 1
256; CHECK-NEXT:    store i32 [[R_1]], i32* [[DST_GEP_1]], align 4
257; CHECK-NEXT:    ret void
258;
259entry:
260  %src = load i32*, i32** %src.p
261  %src.0 = load i32, i32* %src, align 4
262  %r.0 = ashr i32 %src.0, 16
263  store i32 %r.0, i32* %dst, align 4
264  %src.gep.1 = getelementptr inbounds i32, i32* %src, i64 1
265  %src.1 = load i32, i32* %src.gep.1, align 4
266  %r.1 = ashr i32 %src.1, 16
267  %dst.gep.1 = getelementptr inbounds i32, i32* %dst, i64 1
268  store i32 %r.1, i32* %dst.gep.1, align 4
269  ret void
270}
271
272define void @clobber_same_underlying_object(i32* %this) {
273; CHECK-LABEL: @clobber_same_underlying_object(
274; CHECK-NEXT:  entry:
275; CHECK-NEXT:    [[P_3:%.*]] = getelementptr inbounds i32, i32* [[THIS:%.*]], i32 3
276; CHECK-NEXT:    store i32 10, i32* [[P_3]], align 8
277; CHECK-NEXT:    tail call void @clobber()
278; CHECK-NEXT:    [[P_4:%.*]] = getelementptr inbounds i32, i32* [[THIS]], i32 4
279; CHECK-NEXT:    [[L2:%.*]] = load i32, i32* [[P_4]], align 8
280; CHECK-NEXT:    store i32 20, i32* [[P_4]], align 8
281; CHECK-NEXT:    ret void
282;
283entry:
284  %p.3 = getelementptr inbounds i32, i32* %this, i32 3
285  store i32 10, i32* %p.3, align 8
286  tail call void @clobber()
287  %p.4 = getelementptr inbounds i32, i32* %this, i32 4
288  %l2 = load i32, i32* %p.4, align 8
289  store i32 20, i32* %p.4, align 8
290  ret void
291}
292
293declare void @clobber()
294
295define void @slp_not_beneficial(i32* %A, i32* %B) {
296; CHECK-LABEL: @slp_not_beneficial(
297; CHECK-NEXT:  bb:
298; CHECK-NEXT:    [[TMP:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i32 4
299; CHECK-NEXT:    store i32 0, i32* [[TMP]], align 8
300; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, i32* [[A]], i32 5
301; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i32 4
302; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[TMP4]], align 8
303; CHECK-NEXT:    store i32 [[TMP5]], i32* [[TMP3]], align 8
304; CHECK-NEXT:    ret void
305;
306bb:
307  %tmp = getelementptr inbounds i32, i32* %A, i32 4
308  store i32 0, i32* %tmp, align 8
309  %tmp3 = getelementptr inbounds i32, i32* %A, i32 5
310  %tmp4 = getelementptr inbounds i32, i32* %B, i32 4
311  %tmp5 = load i32, i32* %tmp4, align 8
312  store i32 %tmp5, i32* %tmp3, align 8
313  ret void
314}
315
316define void @widget(double* %ptr, double* %ptr.2) {
317; CHECK-LABEL: @widget(
318; CHECK-NEXT:  bb1:
319; CHECK-NEXT:    [[TMP3:%.*]] = load double, double* null, align 8
320; CHECK-NEXT:    [[TMP4:%.*]] = fmul double undef, [[TMP3]]
321; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds double, double* [[PTR:%.*]], i32 0
322; CHECK-NEXT:    [[TMP6:%.*]] = load double, double* [[TMP5]], align 8
323; CHECK-NEXT:    [[TMP7:%.*]] = fadd double [[TMP6]], [[TMP4]]
324; CHECK-NEXT:    store double [[TMP7]], double* [[TMP5]], align 8
325; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds double, double* [[PTR_2:%.*]], i64 0
326; CHECK-NEXT:    [[TMP9:%.*]] = load double, double* [[TMP8]], align 8
327; CHECK-NEXT:    [[TMP10:%.*]] = fmul double undef, [[TMP9]]
328; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds double, double* [[PTR]], i32 1
329; CHECK-NEXT:    [[TMP12:%.*]] = load double, double* [[TMP11]], align 8
330; CHECK-NEXT:    [[TMP13:%.*]] = fadd double [[TMP12]], [[TMP10]]
331; CHECK-NEXT:    store double [[TMP13]], double* [[TMP11]], align 8
332; CHECK-NEXT:    br label [[BB15:%.*]]
333; CHECK:       bb15:
334; CHECK-NEXT:    br label [[BB15]]
335;
336bb1:                                              ; preds = %bb
337  %tmp3 = load double, double* null, align 8
338  %tmp4 = fmul double undef, %tmp3
339  %tmp5 = getelementptr inbounds double, double* %ptr, i32 0
340  %tmp6 = load double, double* %tmp5, align 8
341  %tmp7 = fadd double %tmp6, %tmp4
342  store double %tmp7, double* %tmp5, align 8
343  %tmp8 = getelementptr inbounds double, double* %ptr.2, i64 0
344  %tmp9 = load double, double* %tmp8, align 8
345  %tmp10 = fmul double undef, %tmp9
346  %tmp11 = getelementptr inbounds double, double* %ptr, i32 1
347  %tmp12 = load double, double* %tmp11, align 8
348  %tmp13 = fadd double %tmp12, %tmp10
349  store double %tmp13, double* %tmp11, align 8
350  br label %bb15
351
352bb15:                                             ; preds = %bb15, %bb14
353  br label %bb15
354}
355