1 ; RUN: opt -S -loop-fusion -debug-only=loop-fusion -disable-output < %s 2>&1 | FileCheck %s
4 @B = common global [1024 x i32] zeroinitializer, align 16
6 ; CHECK that the two candidates for fusion are placed into separate candidate
7 ; sets because they are not control flow equivalent.
9 ; CHECK: Performing Loop Fusion on function non_cfe
10 ; CHECK: Fusion Candidates:
11 ; CHECK: *** Fusion Candidate Set ***
13 ; CHECK: ****************************
14 ; CHECK: *** Fusion Candidate Set ***
15 ; CHECK: bb20.preheader
16 ; CHECK: ****************************
17 ; CHECK: Loop Fusion complete
18 define void @non_cfe(i32* noalias %arg) {
22 bb5: ; preds = %bb14, %bb
23 %indvars.iv2 = phi i64 [ %indvars.iv.next3, %bb14 ], [ 0, %bb ]
24 %.01 = phi i32 [ 0, %bb ], [ %tmp15, %bb14 ]
25 %exitcond4 = icmp ne i64 %indvars.iv2, 100
26 br i1 %exitcond4, label %bb7, label %bb16
29 %tmp = add nsw i32 %.01, -3
30 %tmp8 = add nuw nsw i64 %indvars.iv2, 3
31 %tmp9 = trunc i64 %tmp8 to i32
32 %tmp10 = mul nsw i32 %tmp, %tmp9
33 %tmp11 = trunc i64 %indvars.iv2 to i32
34 %tmp12 = srem i32 %tmp10, %tmp11
35 %tmp13 = getelementptr inbounds i32, i32* %arg, i64 %indvars.iv2
36 store i32 %tmp12, i32* %tmp13, align 4
40 %indvars.iv.next3 = add nuw nsw i64 %indvars.iv2, 1
41 %tmp15 = add nuw nsw i32 %.01, 1
45 %tmp17 = load i32, i32* %arg, align 4
46 %tmp18 = icmp slt i32 %tmp17, 0
47 br i1 %tmp18, label %bb20, label %bb33
49 bb20: ; preds = %bb30, %bb16
50 %indvars.iv = phi i64 [ %indvars.iv.next, %bb30 ], [ 0, %bb16 ]
51 %.0 = phi i32 [ 0, %bb16 ], [ %tmp31, %bb30 ]
52 %exitcond = icmp ne i64 %indvars.iv, 100
53 br i1 %exitcond, label %bb22, label %bb33
56 %tmp23 = add nsw i32 %.0, -3
57 %tmp24 = add nuw nsw i64 %indvars.iv, 3
58 %tmp25 = trunc i64 %tmp24 to i32
59 %tmp26 = mul nsw i32 %tmp23, %tmp25
60 %tmp27 = trunc i64 %indvars.iv to i32
61 %tmp28 = srem i32 %tmp26, %tmp27
62 %tmp29 = getelementptr inbounds [1024 x i32], [1024 x i32]* @B, i64 0, i64 %indvars.iv
63 store i32 %tmp28, i32* %tmp29, align 4
67 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
68 %tmp31 = add nuw nsw i32 %.0, 1
71 bb33: ; preds = %bb20, %bb16
75 ; Check that fusion detects the two canddates are not adjacent (the exit block
76 ; of the first candidate is not the preheader of the second candidate).
78 ; CHECK: Performing Loop Fusion on function non_adjacent
79 ; CHECK: Fusion Candidates:
80 ; CHECK: *** Fusion Candidate Set ***
81 ; CHECK-NEXT: [[LOOP1PREHEADER:bb[0-9]*]]
82 ; CHECK-NEXT: [[LOOP2PREHEADER:bb[0-9]*]]
83 ; CHECK-NEXT: ****************************
84 ; CHECK: Attempting fusion on Candidate Set:
85 ; CHECK-NEXT: [[LOOP1PREHEADER]]
86 ; CHECK-NEXT: [[LOOP2PREHEADER]]
87 ; CHECK: Fusion candidates are not adjacent. Not fusing.
88 ; CHECK: Loop Fusion complete
89 define void @non_adjacent(i32* noalias %arg) {
93 bb3: ; preds = %bb11, %bb
94 %.01 = phi i64 [ 0, %bb ], [ %tmp12, %bb11 ]
95 %exitcond2 = icmp ne i64 %.01, 100
96 br i1 %exitcond2, label %bb5, label %bb4
102 %tmp = add nsw i64 %.01, -3
103 %tmp6 = add nuw nsw i64 %.01, 3
104 %tmp7 = mul nsw i64 %tmp, %tmp6
105 %tmp8 = srem i64 %tmp7, %.01
106 %tmp9 = trunc i64 %tmp8 to i32
107 %tmp10 = getelementptr inbounds i32, i32* %arg, i64 %.01
108 store i32 %tmp9, i32* %tmp10, align 4
112 %tmp12 = add nuw nsw i64 %.01, 1
118 bb14: ; preds = %bb23, %bb13
119 %.0 = phi i64 [ 0, %bb13 ], [ %tmp24, %bb23 ]
120 %exitcond = icmp ne i64 %.0, 100
121 br i1 %exitcond, label %bb16, label %bb15
123 bb15: ; preds = %bb14
126 bb16: ; preds = %bb14
127 %tmp17 = add nsw i64 %.0, -3
128 %tmp18 = add nuw nsw i64 %.0, 3
129 %tmp19 = mul nsw i64 %tmp17, %tmp18
130 %tmp20 = srem i64 %tmp19, %.0
131 %tmp21 = trunc i64 %tmp20 to i32
132 %tmp22 = getelementptr inbounds [1024 x i32], [1024 x i32]* @B, i64 0, i64 %.0
133 store i32 %tmp21, i32* %tmp22, align 4
136 bb23: ; preds = %bb16
137 %tmp24 = add nuw nsw i64 %.0, 1
140 bb25: ; preds = %bb15
144 ; Check that the different bounds are detected and prevent fusion.
146 ; CHECK: Performing Loop Fusion on function different_bounds
147 ; CHECK: Fusion Candidates:
148 ; CHECK: *** Fusion Candidate Set ***
149 ; CHECK-NEXT: [[LOOP1PREHEADER:bb[0-9]*]]
150 ; CHECK-NEXT: [[LOOP2PREHEADER:bb[0-9]*]]
151 ; CHECK-NEXT: ****************************
152 ; CHECK: Attempting fusion on Candidate Set:
153 ; CHECK-NEXT: [[LOOP1PREHEADER]]
154 ; CHECK-NEXT: [[LOOP2PREHEADER]]
155 ; CHECK: Fusion candidates do not have identical trip counts. Not fusing.
156 ; CHECK: Loop Fusion complete
157 define void @different_bounds(i32* noalias %arg) {
161 bb3: ; preds = %bb11, %bb
162 %.01 = phi i64 [ 0, %bb ], [ %tmp12, %bb11 ]
163 %exitcond2 = icmp ne i64 %.01, 100
164 br i1 %exitcond2, label %bb5, label %bb4
170 %tmp = add nsw i64 %.01, -3
171 %tmp6 = add nuw nsw i64 %.01, 3
172 %tmp7 = mul nsw i64 %tmp, %tmp6
173 %tmp8 = srem i64 %tmp7, %.01
174 %tmp9 = trunc i64 %tmp8 to i32
175 %tmp10 = getelementptr inbounds i32, i32* %arg, i64 %.01
176 store i32 %tmp9, i32* %tmp10, align 4
180 %tmp12 = add nuw nsw i64 %.01, 1
186 bb14: ; preds = %bb23, %bb13
187 %.0 = phi i64 [ 0, %bb13 ], [ %tmp24, %bb23 ]
188 %exitcond = icmp ne i64 %.0, 200
189 br i1 %exitcond, label %bb16, label %bb15
191 bb15: ; preds = %bb14
194 bb16: ; preds = %bb14
195 %tmp17 = add nsw i64 %.0, -3
196 %tmp18 = add nuw nsw i64 %.0, 3
197 %tmp19 = mul nsw i64 %tmp17, %tmp18
198 %tmp20 = srem i64 %tmp19, %.0
199 %tmp21 = trunc i64 %tmp20 to i32
200 %tmp22 = getelementptr inbounds [1024 x i32], [1024 x i32]* @B, i64 0, i64 %.0
201 store i32 %tmp21, i32* %tmp22, align 4
204 bb23: ; preds = %bb16
205 %tmp24 = add nuw nsw i64 %.0, 1
208 bb25: ; preds = %bb15
212 ; Check that the negative dependence between the two candidates is identified
213 ; and prevents fusion.
215 ; CHECK: Performing Loop Fusion on function negative_dependence
216 ; CHECK: Fusion Candidates:
217 ; CHECK: *** Fusion Candidate Set ***
218 ; CHECK-NEXT: [[LOOP1PREHEADER:bb[0-9]*]]
219 ; CHECK-NEXT: [[LOOP2PREHEADER:bb[0-9]*]]
220 ; CHECK-NEXT: ****************************
221 ; CHECK: Attempting fusion on Candidate Set:
222 ; CHECK-NEXT: [[LOOP1PREHEADER]]
223 ; CHECK-NEXT: [[LOOP2PREHEADER]]
224 ; CHECK: Memory dependencies do not allow fusion!
225 ; CHECK: Loop Fusion complete
226 define void @negative_dependence(i32* noalias %arg) {
230 bb5: ; preds = %bb9, %bb
231 %indvars.iv2 = phi i64 [ %indvars.iv.next3, %bb9 ], [ 0, %bb ]
232 %exitcond4 = icmp ne i64 %indvars.iv2, 100
233 br i1 %exitcond4, label %bb7, label %bb11
236 %tmp = getelementptr inbounds i32, i32* %arg, i64 %indvars.iv2
237 %tmp8 = trunc i64 %indvars.iv2 to i32
238 store i32 %tmp8, i32* %tmp, align 4
242 %indvars.iv.next3 = add nuw nsw i64 %indvars.iv2, 1
245 bb11: ; preds = %bb18, %bb5
246 %indvars.iv = phi i64 [ %indvars.iv.next, %bb18 ], [ 0, %bb5 ]
247 %exitcond = icmp ne i64 %indvars.iv, 100
248 br i1 %exitcond, label %bb13, label %bb19
250 bb13: ; preds = %bb11
251 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
252 %tmp14 = getelementptr inbounds i32, i32* %arg, i64 %indvars.iv.next
253 %tmp15 = load i32, i32* %tmp14, align 4
254 %tmp16 = shl nsw i32 %tmp15, 1
255 %tmp17 = getelementptr inbounds [1024 x i32], [1024 x i32]* @B, i64 0, i64 %indvars.iv
256 store i32 %tmp16, i32* %tmp17, align 4
259 bb18: ; preds = %bb13
262 bb19: ; preds = %bb11
266 ; Check for values defined in Loop 0 and used in Loop 1.
267 ; It is not safe to fuse in this case, because the second loop has
268 ; a use of %.01.lcssa which is defined in the body of loop 0. The
269 ; first loop must execute completely in order to compute the correct
270 ; value of %.01.lcssa to be used in the second loop.
272 ; CHECK: Performing Loop Fusion on function sumTest
273 ; CHECK: Fusion Candidates:
274 ; CHECK: *** Fusion Candidate Set ***
275 ; CHECK-NEXT: [[LOOP1PREHEADER:bb[0-9]*]]
276 ; CHECK-NEXT: [[LOOP2PREHEADER:bb[0-9]*]]
277 ; CHECK-NEXT: ****************************
278 ; CHECK: Attempting fusion on Candidate Set:
279 ; CHECK-NEXT: [[LOOP1PREHEADER]]
280 ; CHECK-NEXT: [[LOOP2PREHEADER]]
281 ; CHECK: Memory dependencies do not allow fusion!
282 ; CHECK: Loop Fusion complete
283 define i32 @sumTest(i32* noalias %arg) {
287 bb6: ; preds = %bb9, %bb
288 %indvars.iv3 = phi i64 [ %indvars.iv.next4, %bb9 ], [ 0, %bb ]
289 %.01 = phi i32 [ 0, %bb ], [ %tmp11, %bb9 ]
290 %exitcond5 = icmp ne i64 %indvars.iv3, 100
291 br i1 %exitcond5, label %bb9, label %bb13
294 %tmp = getelementptr inbounds i32, i32* %arg, i64 %indvars.iv3
295 %tmp10 = load i32, i32* %tmp, align 4
296 %tmp11 = add nsw i32 %.01, %tmp10
297 %indvars.iv.next4 = add nuw nsw i64 %indvars.iv3, 1
300 bb13: ; preds = %bb20, %bb6
301 %.01.lcssa = phi i32 [ %.01, %bb6 ], [ %.01.lcssa, %bb20 ]
302 %indvars.iv = phi i64 [ %indvars.iv.next, %bb20 ], [ 0, %bb6 ]
303 %exitcond = icmp ne i64 %indvars.iv, 100
304 br i1 %exitcond, label %bb15, label %bb14
306 bb14: ; preds = %bb13
309 bb15: ; preds = %bb13
310 %tmp16 = getelementptr inbounds i32, i32* %arg, i64 %indvars.iv
311 %tmp17 = load i32, i32* %tmp16, align 4
312 %tmp18 = sdiv i32 %tmp17, %.01.lcssa
313 %tmp19 = getelementptr inbounds [1024 x i32], [1024 x i32]* @B, i64 0, i64 %indvars.iv
314 store i32 %tmp18, i32* %tmp19, align 4
317 bb20: ; preds = %bb15
318 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
321 bb21: ; preds = %bb14
325 ; Similar to sumTest above. The first loop computes %add and must
326 ; complete before it is used in the second loop. Thus, these two loops
327 ; also cannot be fused.
329 ; CHECK: Performing Loop Fusion on function test
330 ; CHECK: Fusion Candidates:
331 ; CHECK: *** Fusion Candidate Set ***
332 ; CHECK-NEXT: [[LOOP1PREHEADER:for.body[0-9]*.preheader]]
333 ; CHECK-NEXT: [[LOOP2PREHEADER:for.body[0-9]*.preheader]]
334 ; CHECK-NEXT: ****************************
335 ; CHECK: Attempting fusion on Candidate Set:
336 ; CHECK-NEXT: [[LOOP1PREHEADER]]
337 ; CHECK-NEXT: [[LOOP2PREHEADER]]
338 ; CHECK: Memory dependencies do not allow fusion!
339 ; CHECK: Loop Fusion complete
340 define float @test(float* nocapture %a, i32 %n) {
342 %conv = zext i32 %n to i64
343 %cmp32 = icmp eq i32 %n, 0
344 br i1 %cmp32, label %for.cond.cleanup7, label %for.body
346 for.body: ; preds = %for.body, %entry
347 %i.034 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
348 %sum1.033 = phi float [ %add, %for.body ], [ 0.000000e+00, %entry ]
349 %idxprom = trunc i64 %i.034 to i32
350 %arrayidx = getelementptr inbounds float, float* %a, i32 %idxprom
351 %0 = load float, float* %arrayidx, align 4
352 %add = fadd float %sum1.033, %0
353 %inc = add nuw nsw i64 %i.034, 1
354 %cmp = icmp ult i64 %inc, %conv
355 br i1 %cmp, label %for.body, label %for.body8
357 for.body8: ; preds = %for.body, %for.body8
358 %i2.031 = phi i64 [ %inc14, %for.body8 ], [ 0, %for.body ]
359 %idxprom9 = trunc i64 %i2.031 to i32
360 %arrayidx10 = getelementptr inbounds float, float* %a, i32 %idxprom9
361 %1 = load float, float* %arrayidx10, align 4
362 %div = fdiv float %1, %add
363 store float %div, float* %arrayidx10, align 4
364 %inc14 = add nuw nsw i64 %i2.031, 1
365 %cmp5 = icmp ult i64 %inc14, %conv
366 br i1 %cmp5, label %for.body8, label %for.cond.cleanup7
368 for.cond.cleanup7: ; preds = %for.body8, %entry
369 %sum1.0.lcssa36 = phi float [ 0.000000e+00, %entry ], [ %add, %for.body8 ]
370 ret float %sum1.0.lcssa36