Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / 7af74ee89a01bbba230bbf18bb1cb455b5f056b0 / . / llvm / test / Transforms / LoopPredication / reverse.ll
blob: 95bfe20c04e37159da61ed14aa0153eb1cea1dcc [file] [log] [blame]
Eric Christophercee313d2019-04-17 04:52:47 +0000[diff] [blame]1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -S -loop-predication -loop-predication-enable-count-down-loop=true < %s 2>&1 | FileCheck %s
3; RUN: opt -S -passes='require<scalar-evolution>,loop(loop-predication)' -loop-predication-enable-count-down-loop=true < %s 2>&1 | FileCheck %s
4
5declare void @llvm.experimental.guard(i1, ...)
6
7define i32 @signed_reverse_loop_n_to_lower_limit(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) {
8; CHECK-LABEL: @signed_reverse_loop_n_to_lower_limit(
9; CHECK-NEXT: entry:
10; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N:%.*]], 0
11; CHECK-NEXT: br i1 [[TMP5]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
12; CHECK: loop.preheader:
13; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
14; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[TMP0]], [[LENGTH:%.*]]
15; CHECK-NEXT: [[TMP2:%.*]] = icmp sge i32 [[LOWERLIMIT:%.*]], 1
16; CHECK-NEXT: [[TMP3:%.*]] = and i1 [[TMP1]], [[TMP2]]
17; CHECK-NEXT: br label [[LOOP:%.*]]
18; CHECK: loop:
19; CHECK-NEXT: [[LOOP_ACC:%.*]] = phi i32 [ [[LOOP_ACC_NEXT:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
20; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[LOOP]] ], [ [[N]], [[LOOP_PREHEADER]] ]
21; CHECK-NEXT: [[I_NEXT]] = add nsw i32 [[I]], -1
22; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP3]], i32 9) [ "deopt"() ]
23; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I_NEXT]] to i64
24; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, i32* [[ARRAY:%.*]], i64 [[I_I64]]
25; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, i32* [[ARRAY_I_PTR]], align 4
26; CHECK-NEXT: [[LOOP_ACC_NEXT]] = add i32 [[LOOP_ACC]], [[ARRAY_I]]
27; CHECK-NEXT: [[CONTINUE:%.*]] = icmp sgt i32 [[I]], [[LOWERLIMIT]]
28; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
29; CHECK: exit.loopexit:
30; CHECK-NEXT: [[LOOP_ACC_NEXT_LCSSA:%.*]] = phi i32 [ [[LOOP_ACC_NEXT]], [[LOOP]] ]
31; CHECK-NEXT: br label [[EXIT]]
32; CHECK: exit:
33; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[LOOP_ACC_NEXT_LCSSA]], [[EXIT_LOOPEXIT]] ]
34; CHECK-NEXT: ret i32 [[RESULT]]
35;
36entry:
37 %tmp5 = icmp eq i32 %n, 0
38 br i1 %tmp5, label %exit, label %loop.preheader
39
40loop.preheader:
41 br label %loop
42
43loop:
44 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ]
45 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ]
46 %i.next = add nsw i32 %i, -1
47 %within.bounds = icmp ult i32 %i.next, %length
48 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ]
49 %i.i64 = zext i32 %i.next to i64
50 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64
51 %array.i = load i32, i32* %array.i.ptr, align 4
52 %loop.acc.next = add i32 %loop.acc, %array.i
53 %continue = icmp sgt i32 %i, %lowerlimit
54 br i1 %continue, label %loop, label %exit
55
56exit:
57 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ]
58 ret i32 %result
59}
60
61define i32 @unsigned_reverse_loop_n_to_lower_limit(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) {
62; CHECK-LABEL: @unsigned_reverse_loop_n_to_lower_limit(
63; CHECK-NEXT: entry:
64; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N:%.*]], 0
65; CHECK-NEXT: br i1 [[TMP5]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
66; CHECK: loop.preheader:
67; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
68; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[TMP0]], [[LENGTH:%.*]]
69; CHECK-NEXT: [[TMP2:%.*]] = icmp uge i32 [[LOWERLIMIT:%.*]], 1
70; CHECK-NEXT: [[TMP3:%.*]] = and i1 [[TMP1]], [[TMP2]]
71; CHECK-NEXT: br label [[LOOP:%.*]]
72; CHECK: loop:
73; CHECK-NEXT: [[LOOP_ACC:%.*]] = phi i32 [ [[LOOP_ACC_NEXT:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
74; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[LOOP]] ], [ [[N]], [[LOOP_PREHEADER]] ]
75; CHECK-NEXT: [[I_NEXT]] = add nsw i32 [[I]], -1
76; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP3]], i32 9) [ "deopt"() ]
77; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I_NEXT]] to i64
78; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, i32* [[ARRAY:%.*]], i64 [[I_I64]]
79; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, i32* [[ARRAY_I_PTR]], align 4
80; CHECK-NEXT: [[LOOP_ACC_NEXT]] = add i32 [[LOOP_ACC]], [[ARRAY_I]]
81; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i32 [[I]], [[LOWERLIMIT]]
82; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
83; CHECK: exit.loopexit:
84; CHECK-NEXT: [[LOOP_ACC_NEXT_LCSSA:%.*]] = phi i32 [ [[LOOP_ACC_NEXT]], [[LOOP]] ]
85; CHECK-NEXT: br label [[EXIT]]
86; CHECK: exit:
87; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[LOOP_ACC_NEXT_LCSSA]], [[EXIT_LOOPEXIT]] ]
88; CHECK-NEXT: ret i32 [[RESULT]]
89;
90entry:
91 %tmp5 = icmp eq i32 %n, 0
92 br i1 %tmp5, label %exit, label %loop.preheader
93
94loop.preheader:
95 br label %loop
96
97loop:
98 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ]
99 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ]
100 %i.next = add nsw i32 %i, -1
101 %within.bounds = icmp ult i32 %i.next, %length
102 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ]
103 %i.i64 = zext i32 %i.next to i64
104 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64
105 %array.i = load i32, i32* %array.i.ptr, align 4
106 %loop.acc.next = add i32 %loop.acc, %array.i
107 %continue = icmp ugt i32 %i, %lowerlimit
108 br i1 %continue, label %loop, label %exit
109
110exit:
111 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ]
112 ret i32 %result
113}
114
115
116; if we predicated the loop, the guard will definitely fail and we will
117; deoptimize early on.
118define i32 @unsigned_reverse_loop_n_to_0(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) {
119; CHECK-LABEL: @unsigned_reverse_loop_n_to_0(
120; CHECK-NEXT: entry:
121; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N:%.*]], 0
122; CHECK-NEXT: br i1 [[TMP5]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
123; CHECK: loop.preheader:
124; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
125; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[TMP0]], [[LENGTH:%.*]]
126; CHECK-NEXT: [[TMP2:%.*]] = and i1 [[TMP1]], false
127; CHECK-NEXT: br label [[LOOP:%.*]]
128; CHECK: loop:
129; CHECK-NEXT: [[LOOP_ACC:%.*]] = phi i32 [ [[LOOP_ACC_NEXT:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
130; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[LOOP]] ], [ [[N]], [[LOOP_PREHEADER]] ]
131; CHECK-NEXT: [[I_NEXT]] = add nsw i32 [[I]], -1
132; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP2]], i32 9) [ "deopt"() ]
133; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I_NEXT]] to i64
134; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, i32* [[ARRAY:%.*]], i64 [[I_I64]]
135; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, i32* [[ARRAY_I_PTR]], align 4
136; CHECK-NEXT: [[LOOP_ACC_NEXT]] = add i32 [[LOOP_ACC]], [[ARRAY_I]]
137; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i32 [[I]], 0
138; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
139; CHECK: exit.loopexit:
140; CHECK-NEXT: [[LOOP_ACC_NEXT_LCSSA:%.*]] = phi i32 [ [[LOOP_ACC_NEXT]], [[LOOP]] ]
141; CHECK-NEXT: br label [[EXIT]]
142; CHECK: exit:
143; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[LOOP_ACC_NEXT_LCSSA]], [[EXIT_LOOPEXIT]] ]
144; CHECK-NEXT: ret i32 [[RESULT]]
145;
146entry:
147 %tmp5 = icmp eq i32 %n, 0
148 br i1 %tmp5, label %exit, label %loop.preheader
149
150loop.preheader:
151 br label %loop
152
153loop:
154 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ]
155 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ]
156 %i.next = add nsw i32 %i, -1
157 %within.bounds = icmp ult i32 %i.next, %length
158 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ]
159 %i.i64 = zext i32 %i.next to i64
160 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64
161 %array.i = load i32, i32* %array.i.ptr, align 4
162 %loop.acc.next = add i32 %loop.acc, %array.i
163 %continue = icmp ugt i32 %i, 0
164 br i1 %continue, label %loop, label %exit
165
166exit:
167 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ]
168 ret i32 %result
169}
170
171; do not loop predicate when the range has step -1 and latch has step 1.
172define i32 @reverse_loop_range_step_increment(i32 %n, i32* %array, i32 %length) {
173; CHECK-LABEL: @reverse_loop_range_step_increment(
174; CHECK-NEXT: entry:
175; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N:%.*]], 0
176; CHECK-NEXT: br i1 [[TMP5]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
177; CHECK: loop.preheader:
178; CHECK-NEXT: br label [[LOOP:%.*]]
179; CHECK: loop:
180; CHECK-NEXT: [[LOOP_ACC:%.*]] = phi i32 [ [[LOOP_ACC_NEXT:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
181; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[LOOP]] ], [ [[N]], [[LOOP_PREHEADER]] ]
182; CHECK-NEXT: [[IRC:%.*]] = phi i32 [ [[I_INC:%.*]], [[LOOP]] ], [ 1, [[LOOP_PREHEADER]] ]
183; CHECK-NEXT: [[I_INC]] = add nuw nsw i32 [[IRC]], 1
184; CHECK-NEXT: [[WITHIN_BOUNDS:%.*]] = icmp ult i32 [[IRC]], [[LENGTH:%.*]]
185; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[WITHIN_BOUNDS]], i32 9) [ "deopt"() ]
186; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[IRC]] to i64
187; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, i32* [[ARRAY:%.*]], i64 [[I_I64]]
188; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, i32* [[ARRAY_I_PTR]], align 4
189; CHECK-NEXT: [[I_NEXT]] = add nsw i32 [[I]], -1
190; CHECK-NEXT: [[LOOP_ACC_NEXT]] = add i32 [[LOOP_ACC]], [[ARRAY_I]]
191; CHECK-NEXT: [[CONTINUE:%.*]] = icmp ugt i32 [[I]], 65534
192; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
193; CHECK: exit.loopexit:
194; CHECK-NEXT: [[LOOP_ACC_NEXT_LCSSA:%.*]] = phi i32 [ [[LOOP_ACC_NEXT]], [[LOOP]] ]
195; CHECK-NEXT: br label [[EXIT]]
196; CHECK: exit:
197; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[LOOP_ACC_NEXT_LCSSA]], [[EXIT_LOOPEXIT]] ]
198; CHECK-NEXT: ret i32 [[RESULT]]
199;
200entry:
201 %tmp5 = icmp eq i32 %n, 0
202 br i1 %tmp5, label %exit, label %loop.preheader
203
204loop.preheader:
205 br label %loop
206
207loop:
208 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ]
209 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ]
210 %irc = phi i32 [ %i.inc, %loop ], [ 1, %loop.preheader ]
211 %i.inc = add nuw nsw i32 %irc, 1
212 %within.bounds = icmp ult i32 %irc, %length
213 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ]
214 %i.i64 = zext i32 %irc to i64
215 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64
216 %array.i = load i32, i32* %array.i.ptr, align 4
217 %i.next = add nsw i32 %i, -1
218 %loop.acc.next = add i32 %loop.acc, %array.i
219 %continue = icmp ugt i32 %i, 65534
220 br i1 %continue, label %loop, label %exit
221
222exit:
223 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ]
224 ret i32 %result
225}
226
227define i32 @signed_reverse_loop_n_to_lower_limit_equal(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) {
228; CHECK-LABEL: @signed_reverse_loop_n_to_lower_limit_equal(
229; CHECK-NEXT: entry:
230; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N:%.*]], 0
231; CHECK-NEXT: br i1 [[TMP5]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
232; CHECK: loop.preheader:
233; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
234; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[TMP0]], [[LENGTH:%.*]]
235; CHECK-NEXT: [[TMP2:%.*]] = icmp sgt i32 [[LOWERLIMIT:%.*]], 1
236; CHECK-NEXT: [[TMP3:%.*]] = and i1 [[TMP1]], [[TMP2]]
237; CHECK-NEXT: br label [[LOOP:%.*]]
238; CHECK: loop:
239; CHECK-NEXT: [[LOOP_ACC:%.*]] = phi i32 [ [[LOOP_ACC_NEXT:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
240; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[LOOP]] ], [ [[N]], [[LOOP_PREHEADER]] ]
241; CHECK-NEXT: [[I_NEXT]] = add nsw i32 [[I]], -1
242; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP3]], i32 9) [ "deopt"() ]
243; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I_NEXT]] to i64
244; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, i32* [[ARRAY:%.*]], i64 [[I_I64]]
245; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, i32* [[ARRAY_I_PTR]], align 4
246; CHECK-NEXT: [[LOOP_ACC_NEXT]] = add i32 [[LOOP_ACC]], [[ARRAY_I]]
247; CHECK-NEXT: [[CONTINUE:%.*]] = icmp sge i32 [[I]], [[LOWERLIMIT]]
248; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
249; CHECK: exit.loopexit:
250; CHECK-NEXT: [[LOOP_ACC_NEXT_LCSSA:%.*]] = phi i32 [ [[LOOP_ACC_NEXT]], [[LOOP]] ]
251; CHECK-NEXT: br label [[EXIT]]
252; CHECK: exit:
253; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[LOOP_ACC_NEXT_LCSSA]], [[EXIT_LOOPEXIT]] ]
254; CHECK-NEXT: ret i32 [[RESULT]]
255;
256entry:
257 %tmp5 = icmp eq i32 %n, 0
258 br i1 %tmp5, label %exit, label %loop.preheader
259
260loop.preheader:
261 br label %loop
262
263loop:
264 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ]
265 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ]
266 %i.next = add nsw i32 %i, -1
267 %within.bounds = icmp ult i32 %i.next, %length
268 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ]
269 %i.i64 = zext i32 %i.next to i64
270 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64
271 %array.i = load i32, i32* %array.i.ptr, align 4
272 %loop.acc.next = add i32 %loop.acc, %array.i
273 %continue = icmp sge i32 %i, %lowerlimit
274 br i1 %continue, label %loop, label %exit
275
276exit:
277 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ]
278 ret i32 %result
279}
280
281define i32 @unsigned_reverse_loop_n_to_lower_limit_equal(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) {
282; CHECK-LABEL: @unsigned_reverse_loop_n_to_lower_limit_equal(
283; CHECK-NEXT: entry:
284; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N:%.*]], 0
285; CHECK-NEXT: br i1 [[TMP5]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
286; CHECK: loop.preheader:
287; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
288; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[TMP0]], [[LENGTH:%.*]]
289; CHECK-NEXT: [[TMP2:%.*]] = icmp ugt i32 [[LOWERLIMIT:%.*]], 1
290; CHECK-NEXT: [[TMP3:%.*]] = and i1 [[TMP1]], [[TMP2]]
291; CHECK-NEXT: br label [[LOOP:%.*]]
292; CHECK: loop:
293; CHECK-NEXT: [[LOOP_ACC:%.*]] = phi i32 [ [[LOOP_ACC_NEXT:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
294; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[LOOP]] ], [ [[N]], [[LOOP_PREHEADER]] ]
295; CHECK-NEXT: [[I_NEXT]] = add nsw i32 [[I]], -1
296; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP3]], i32 9) [ "deopt"() ]
297; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I_NEXT]] to i64
298; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, i32* [[ARRAY:%.*]], i64 [[I_I64]]
299; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, i32* [[ARRAY_I_PTR]], align 4
300; CHECK-NEXT: [[LOOP_ACC_NEXT]] = add i32 [[LOOP_ACC]], [[ARRAY_I]]
301; CHECK-NEXT: [[CONTINUE:%.*]] = icmp uge i32 [[I]], [[LOWERLIMIT]]
302; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
303; CHECK: exit.loopexit:
304; CHECK-NEXT: [[LOOP_ACC_NEXT_LCSSA:%.*]] = phi i32 [ [[LOOP_ACC_NEXT]], [[LOOP]] ]
305; CHECK-NEXT: br label [[EXIT]]
306; CHECK: exit:
307; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[LOOP_ACC_NEXT_LCSSA]], [[EXIT_LOOPEXIT]] ]
308; CHECK-NEXT: ret i32 [[RESULT]]
309;
310entry:
311 %tmp5 = icmp eq i32 %n, 0
312 br i1 %tmp5, label %exit, label %loop.preheader
313
314loop.preheader:
315 br label %loop
316
317loop:
318 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ]
319 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ]
320 %i.next = add nsw i32 %i, -1
321 %within.bounds = icmp ult i32 %i.next, %length
322 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ]
323 %i.i64 = zext i32 %i.next to i64
324 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64
325 %array.i = load i32, i32* %array.i.ptr, align 4
326 %loop.acc.next = add i32 %loop.acc, %array.i
327 %continue = icmp uge i32 %i, %lowerlimit
328 br i1 %continue, label %loop, label %exit
329
330exit:
331 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ]
332 ret i32 %result
333}
334
335
336; if we predicated the loop, the guard will definitely fail and we will
337; deoptimize early on.
338define i32 @unsigned_reverse_loop_n_to_1(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) {
339; CHECK-LABEL: @unsigned_reverse_loop_n_to_1(
340; CHECK-NEXT: entry:
341; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[N:%.*]], 0
342; CHECK-NEXT: br i1 [[TMP5]], label [[EXIT:%.*]], label [[LOOP_PREHEADER:%.*]]
343; CHECK: loop.preheader:
344; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[N]], -1
345; CHECK-NEXT: [[TMP1:%.*]] = icmp ult i32 [[TMP0]], [[LENGTH:%.*]]
346; CHECK-NEXT: [[TMP2:%.*]] = and i1 [[TMP1]], false
347; CHECK-NEXT: br label [[LOOP:%.*]]
348; CHECK: loop:
349; CHECK-NEXT: [[LOOP_ACC:%.*]] = phi i32 [ [[LOOP_ACC_NEXT:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
350; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[LOOP]] ], [ [[N]], [[LOOP_PREHEADER]] ]
351; CHECK-NEXT: [[I_NEXT]] = add nsw i32 [[I]], -1
352; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP2]], i32 9) [ "deopt"() ]
353; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I_NEXT]] to i64
354; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, i32* [[ARRAY:%.*]], i64 [[I_I64]]
355; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, i32* [[ARRAY_I_PTR]], align 4
356; CHECK-NEXT: [[LOOP_ACC_NEXT]] = add i32 [[LOOP_ACC]], [[ARRAY_I]]
357; CHECK-NEXT: [[CONTINUE:%.*]] = icmp uge i32 [[I]], 1
358; CHECK-NEXT: br i1 [[CONTINUE]], label [[LOOP]], label [[EXIT_LOOPEXIT:%.*]]
359; CHECK: exit.loopexit:
360; CHECK-NEXT: [[LOOP_ACC_NEXT_LCSSA:%.*]] = phi i32 [ [[LOOP_ACC_NEXT]], [[LOOP]] ]
361; CHECK-NEXT: br label [[EXIT]]
362; CHECK: exit:
363; CHECK-NEXT: [[RESULT:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[LOOP_ACC_NEXT_LCSSA]], [[EXIT_LOOPEXIT]] ]
364; CHECK-NEXT: ret i32 [[RESULT]]
365;
366entry:
367 %tmp5 = icmp eq i32 %n, 0
368 br i1 %tmp5, label %exit, label %loop.preheader
369
370loop.preheader:
371 br label %loop
372
373loop:
374 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ]
375 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ]
376 %i.next = add nsw i32 %i, -1
377 %within.bounds = icmp ult i32 %i.next, %length
378 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ]
379 %i.i64 = zext i32 %i.next to i64
380 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64
381 %array.i = load i32, i32* %array.i.ptr, align 4
382 %loop.acc.next = add i32 %loop.acc, %array.i
383 %continue = icmp uge i32 %i, 1
384 br i1 %continue, label %loop, label %exit
385
386exit:
387 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ]
388 ret i32 %result
389}
390