1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -S -loop-vectorize < %s | FileCheck %s
3
4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
5target triple = "x86_64-unknown-linux-gnu"
6
7define float @reduction_sum_float_ieee(i32 %n, float* %array) {
8; CHECK-LABEL: @reduction_sum_float_ieee(
9; CHECK-NEXT:  entry:
10; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
11; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
12; CHECK:       loop.preheader:
13; CHECK-NEXT:    br label [[LOOP:%.*]]
14; CHECK:       loop:
15; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ 0, [[LOOP_PREHEADER]] ]
16; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ 0.000000e+00, [[LOOP_PREHEADER]] ]
17; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, float* [[ARRAY:%.*]], i32 [[IDX]]
18; CHECK-NEXT:    [[VALUE:%.*]] = load float, float* [[ADDRESS]], align 4
19; CHECK-NEXT:    [[SUM_INC]] = fadd float [[SUM]], [[VALUE]]
20; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
21; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
22; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT:%.*]]
23; CHECK:       loop.exit.loopexit:
24; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ]
25; CHECK-NEXT:    br label [[LOOP_EXIT]]
26; CHECK:       loop.exit:
27; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
28; CHECK-NEXT:    ret float [[SUM_LCSSA]]
29;
30entry:
31  %entry.cond = icmp ne i32 0, 4096
32  br i1 %entry.cond, label %loop, label %loop.exit
33
34loop:
35  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
36  %sum = phi float [ 0.000000e+00, %entry ], [ %sum.inc, %loop ]
37  %address = getelementptr float, float* %array, i32 %idx
38  %value = load float, float* %address
39  %sum.inc = fadd float %sum, %value
40  %idx.inc = add i32 %idx, 1
41  %be.cond = icmp ne i32 %idx.inc, 4096
42  br i1 %be.cond, label %loop, label %loop.exit
43
44loop.exit:
45  %sum.lcssa = phi float [ %sum.inc, %loop ], [ 0.000000e+00, %entry ]
46  ret float %sum.lcssa
47}
48
49define float @reduction_sum_float_fastmath(i32 %n, float* %array) {
50; CHECK-LABEL: @reduction_sum_float_fastmath(
51; CHECK-NEXT:  entry:
52; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
53; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
54; CHECK:       loop.preheader:
55; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
56; CHECK:       vector.ph:
57; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
58; CHECK:       vector.body:
59; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
60; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
61; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
62; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
63; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[INDEX]], 4
64; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, float* [[ARRAY:%.*]], i32 [[TMP0]]
65; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr float, float* [[ARRAY]], i32 [[TMP1]]
66; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr float, float* [[TMP2]], i32 0
67; CHECK-NEXT:    [[TMP5:%.*]] = bitcast float* [[TMP4]] to <4 x float>*
68; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP5]], align 4
69; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr float, float* [[TMP2]], i32 4
70; CHECK-NEXT:    [[TMP7:%.*]] = bitcast float* [[TMP6]] to <4 x float>*
71; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, <4 x float>* [[TMP7]], align 4
72; CHECK-NEXT:    [[TMP8]] = fadd fast <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
73; CHECK-NEXT:    [[TMP9]] = fadd fast <4 x float> [[VEC_PHI1]], [[WIDE_LOAD2]]
74; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 8
75; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i32 [[INDEX_NEXT]], 4096
76; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
77; CHECK:       middle.block:
78; CHECK-NEXT:    [[BIN_RDX:%.*]] = fadd fast <4 x float> [[TMP9]], [[TMP8]]
79; CHECK-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <4 x float> [[BIN_RDX]], <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
80; CHECK-NEXT:    [[BIN_RDX3:%.*]] = fadd fast <4 x float> [[BIN_RDX]], [[RDX_SHUF]]
81; CHECK-NEXT:    [[RDX_SHUF4:%.*]] = shufflevector <4 x float> [[BIN_RDX3]], <4 x float> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
82; CHECK-NEXT:    [[BIN_RDX5:%.*]] = fadd fast <4 x float> [[BIN_RDX3]], [[RDX_SHUF4]]
83; CHECK-NEXT:    [[TMP11:%.*]] = extractelement <4 x float> [[BIN_RDX5]], i32 0
84; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i32 4096, 4096
85; CHECK-NEXT:    br i1 [[CMP_N]], label [[LOOP_EXIT_LOOPEXIT:%.*]], label [[SCALAR_PH]]
86; CHECK:       scalar.ph:
87; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_PREHEADER]] ]
88; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ 0.000000e+00, [[LOOP_PREHEADER]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
89; CHECK-NEXT:    br label [[LOOP:%.*]]
90; CHECK:       loop:
91; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
92; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
93; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, float* [[ARRAY]], i32 [[IDX]]
94; CHECK-NEXT:    [[VALUE:%.*]] = load float, float* [[ADDRESS]], align 4
95; CHECK-NEXT:    [[SUM_INC]] = fadd fast float [[SUM]], [[VALUE]]
96; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
97; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
98; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT]], !llvm.loop !2
99; CHECK:       loop.exit.loopexit:
100; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
101; CHECK-NEXT:    br label [[LOOP_EXIT]]
102; CHECK:       loop.exit:
103; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
104; CHECK-NEXT:    ret float [[SUM_LCSSA]]
105;
106entry:
107  %entry.cond = icmp ne i32 0, 4096
108  br i1 %entry.cond, label %loop, label %loop.exit
109
110loop:
111  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
112  %sum = phi float [ 0.000000e+00, %entry ], [ %sum.inc, %loop ]
113  %address = getelementptr float, float* %array, i32 %idx
114  %value = load float, float* %address
115  %sum.inc = fadd fast float %sum, %value
116  %idx.inc = add i32 %idx, 1
117  %be.cond = icmp ne i32 %idx.inc, 4096
118  br i1 %be.cond, label %loop, label %loop.exit
119
120loop.exit:
121  %sum.lcssa = phi float [ %sum.inc, %loop ], [ 0.000000e+00, %entry ]
122  ret float %sum.lcssa
123}
124
125define float @reduction_sum_float_only_reassoc(i32 %n, float* %array) {
126; CHECK-LABEL: @reduction_sum_float_only_reassoc(
127; CHECK-NEXT:  entry:
128; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
129; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
130; CHECK:       loop.preheader:
131; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
132; CHECK:       vector.ph:
133; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
134; CHECK:       vector.body:
135; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
136; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
137; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
138; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
139; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[INDEX]], 4
140; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, float* [[ARRAY:%.*]], i32 [[TMP0]]
141; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr float, float* [[ARRAY]], i32 [[TMP1]]
142; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr float, float* [[TMP2]], i32 0
143; CHECK-NEXT:    [[TMP5:%.*]] = bitcast float* [[TMP4]] to <4 x float>*
144; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP5]], align 4
145; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr float, float* [[TMP2]], i32 4
146; CHECK-NEXT:    [[TMP7:%.*]] = bitcast float* [[TMP6]] to <4 x float>*
147; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, <4 x float>* [[TMP7]], align 4
148; CHECK-NEXT:    [[TMP8]] = fadd reassoc <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
149; CHECK-NEXT:    [[TMP9]] = fadd reassoc <4 x float> [[VEC_PHI1]], [[WIDE_LOAD2]]
150; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 8
151; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i32 [[INDEX_NEXT]], 4096
152; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !4
153; CHECK:       middle.block:
154; CHECK-NEXT:    [[BIN_RDX:%.*]] = fadd reassoc <4 x float> [[TMP9]], [[TMP8]]
155; CHECK-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <4 x float> [[BIN_RDX]], <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
156; CHECK-NEXT:    [[BIN_RDX3:%.*]] = fadd reassoc <4 x float> [[BIN_RDX]], [[RDX_SHUF]]
157; CHECK-NEXT:    [[RDX_SHUF4:%.*]] = shufflevector <4 x float> [[BIN_RDX3]], <4 x float> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
158; CHECK-NEXT:    [[BIN_RDX5:%.*]] = fadd reassoc <4 x float> [[BIN_RDX3]], [[RDX_SHUF4]]
159; CHECK-NEXT:    [[TMP11:%.*]] = extractelement <4 x float> [[BIN_RDX5]], i32 0
160; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i32 4096, 4096
161; CHECK-NEXT:    br i1 [[CMP_N]], label [[LOOP_EXIT_LOOPEXIT:%.*]], label [[SCALAR_PH]]
162; CHECK:       scalar.ph:
163; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_PREHEADER]] ]
164; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ 0.000000e+00, [[LOOP_PREHEADER]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
165; CHECK-NEXT:    br label [[LOOP:%.*]]
166; CHECK:       loop:
167; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
168; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
169; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, float* [[ARRAY]], i32 [[IDX]]
170; CHECK-NEXT:    [[VALUE:%.*]] = load float, float* [[ADDRESS]], align 4
171; CHECK-NEXT:    [[SUM_INC]] = fadd reassoc float [[SUM]], [[VALUE]]
172; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
173; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
174; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT]], !llvm.loop !5
175; CHECK:       loop.exit.loopexit:
176; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
177; CHECK-NEXT:    br label [[LOOP_EXIT]]
178; CHECK:       loop.exit:
179; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
180; CHECK-NEXT:    ret float [[SUM_LCSSA]]
181;
182entry:
183  %entry.cond = icmp ne i32 0, 4096
184  br i1 %entry.cond, label %loop, label %loop.exit
185
186loop:
187  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
188  %sum = phi float [ 0.000000e+00, %entry ], [ %sum.inc, %loop ]
189  %address = getelementptr float, float* %array, i32 %idx
190  %value = load float, float* %address
191  %sum.inc = fadd reassoc float %sum, %value
192  %idx.inc = add i32 %idx, 1
193  %be.cond = icmp ne i32 %idx.inc, 4096
194  br i1 %be.cond, label %loop, label %loop.exit
195
196loop.exit:
197  %sum.lcssa = phi float [ %sum.inc, %loop ], [ 0.000000e+00, %entry ]
198  ret float %sum.lcssa
199}
200
201define float @reduction_sum_float_only_reassoc_and_contract(i32 %n, float* %array) {
202; CHECK-LABEL: @reduction_sum_float_only_reassoc_and_contract(
203; CHECK-NEXT:  entry:
204; CHECK-NEXT:    [[ENTRY_COND:%.*]] = icmp ne i32 0, 4096
205; CHECK-NEXT:    br i1 [[ENTRY_COND]], label [[LOOP_PREHEADER:%.*]], label [[LOOP_EXIT:%.*]]
206; CHECK:       loop.preheader:
207; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
208; CHECK:       vector.ph:
209; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
210; CHECK:       vector.body:
211; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
212; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
213; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
214; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
215; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[INDEX]], 4
216; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr float, float* [[ARRAY:%.*]], i32 [[TMP0]]
217; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr float, float* [[ARRAY]], i32 [[TMP1]]
218; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr float, float* [[TMP2]], i32 0
219; CHECK-NEXT:    [[TMP5:%.*]] = bitcast float* [[TMP4]] to <4 x float>*
220; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP5]], align 4
221; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr float, float* [[TMP2]], i32 4
222; CHECK-NEXT:    [[TMP7:%.*]] = bitcast float* [[TMP6]] to <4 x float>*
223; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, <4 x float>* [[TMP7]], align 4
224; CHECK-NEXT:    [[TMP8]] = fadd reassoc contract <4 x float> [[VEC_PHI]], [[WIDE_LOAD]]
225; CHECK-NEXT:    [[TMP9]] = fadd reassoc contract <4 x float> [[VEC_PHI1]], [[WIDE_LOAD2]]
226; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 8
227; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i32 [[INDEX_NEXT]], 4096
228; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !6
229; CHECK:       middle.block:
230; CHECK-NEXT:    [[BIN_RDX:%.*]] = fadd reassoc contract <4 x float> [[TMP9]], [[TMP8]]
231; CHECK-NEXT:    [[RDX_SHUF:%.*]] = shufflevector <4 x float> [[BIN_RDX]], <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
232; CHECK-NEXT:    [[BIN_RDX3:%.*]] = fadd reassoc contract <4 x float> [[BIN_RDX]], [[RDX_SHUF]]
233; CHECK-NEXT:    [[RDX_SHUF4:%.*]] = shufflevector <4 x float> [[BIN_RDX3]], <4 x float> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
234; CHECK-NEXT:    [[BIN_RDX5:%.*]] = fadd reassoc contract <4 x float> [[BIN_RDX3]], [[RDX_SHUF4]]
235; CHECK-NEXT:    [[TMP11:%.*]] = extractelement <4 x float> [[BIN_RDX5]], i32 0
236; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i32 4096, 4096
237; CHECK-NEXT:    br i1 [[CMP_N]], label [[LOOP_EXIT_LOOPEXIT:%.*]], label [[SCALAR_PH]]
238; CHECK:       scalar.ph:
239; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 4096, [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_PREHEADER]] ]
240; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ 0.000000e+00, [[LOOP_PREHEADER]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
241; CHECK-NEXT:    br label [[LOOP:%.*]]
242; CHECK:       loop:
243; CHECK-NEXT:    [[IDX:%.*]] = phi i32 [ [[IDX_INC:%.*]], [[LOOP]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
244; CHECK-NEXT:    [[SUM:%.*]] = phi float [ [[SUM_INC:%.*]], [[LOOP]] ], [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ]
245; CHECK-NEXT:    [[ADDRESS:%.*]] = getelementptr float, float* [[ARRAY]], i32 [[IDX]]
246; CHECK-NEXT:    [[VALUE:%.*]] = load float, float* [[ADDRESS]], align 4
247; CHECK-NEXT:    [[SUM_INC]] = fadd reassoc contract float [[SUM]], [[VALUE]]
248; CHECK-NEXT:    [[IDX_INC]] = add i32 [[IDX]], 1
249; CHECK-NEXT:    [[BE_COND:%.*]] = icmp ne i32 [[IDX_INC]], 4096
250; CHECK-NEXT:    br i1 [[BE_COND]], label [[LOOP]], label [[LOOP_EXIT_LOOPEXIT]], !llvm.loop !7
251; CHECK:       loop.exit.loopexit:
252; CHECK-NEXT:    [[SUM_INC_LCSSA:%.*]] = phi float [ [[SUM_INC]], [[LOOP]] ], [ [[TMP11]], [[MIDDLE_BLOCK]] ]
253; CHECK-NEXT:    br label [[LOOP_EXIT]]
254; CHECK:       loop.exit:
255; CHECK-NEXT:    [[SUM_LCSSA:%.*]] = phi float [ 0.000000e+00, [[ENTRY:%.*]] ], [ [[SUM_INC_LCSSA]], [[LOOP_EXIT_LOOPEXIT]] ]
256; CHECK-NEXT:    ret float [[SUM_LCSSA]]
257;
258entry:
259  %entry.cond = icmp ne i32 0, 4096
260  br i1 %entry.cond, label %loop, label %loop.exit
261
262loop:
263  %idx = phi i32 [ 0, %entry ], [ %idx.inc, %loop ]
264  %sum = phi float [ 0.000000e+00, %entry ], [ %sum.inc, %loop ]
265  %address = getelementptr float, float* %array, i32 %idx
266  %value = load float, float* %address
267  %sum.inc = fadd reassoc contract float %sum, %value
268  %idx.inc = add i32 %idx, 1
269  %be.cond = icmp ne i32 %idx.inc, 4096
270  br i1 %be.cond, label %loop, label %loop.exit
271
272loop.exit:
273  %sum.lcssa = phi float [ %sum.inc, %loop ], [ 0.000000e+00, %entry ]
274  ret float %sum.lcssa
275}
276