1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s  -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -prefer-inloop-reductions -force-reduction-intrinsics -dce -instcombine -S | FileCheck %s
3
4target 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"
5
6define i32 @reduction_sum_single(i32* noalias nocapture %A) {
7; CHECK-LABEL: @reduction_sum_single(
8; CHECK-NEXT:  entry:
9; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
10; CHECK:       vector.ph:
11; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
12; CHECK:       vector.body:
13; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
14; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
15; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
16; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
17; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
18; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
19; CHECK-NEXT:    [[TMP3]] = add i32 [[TMP2]], [[VEC_PHI]]
20; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
21; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
22; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP0:!llvm.loop !.*]]
23; CHECK:       middle.block:
24; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
25; CHECK:       scalar.ph:
26; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
27; CHECK:       .lr.ph:
28; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP2:!llvm.loop !.*]]
29; CHECK:       ._crit_edge:
30; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP3]], [[MIDDLE_BLOCK]] ]
31; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
32;
33entry:
34  br label %.lr.ph
35
36.lr.ph:                                           ; preds = %entry, %.lr.ph
37  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
38  %sum.02 = phi i32 [ %l7, %.lr.ph ], [ 0, %entry ]
39  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
40  %l3 = load i32, i32* %l2, align 4
41  %l7 = add i32 %sum.02, %l3
42  %indvars.iv.next = add i64 %indvars.iv, 1
43  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
44  %exitcond = icmp eq i32 %lftr.wideiv, 256
45  br i1 %exitcond, label %._crit_edge, label %.lr.ph
46
47._crit_edge:                                      ; preds = %.lr.ph
48  %sum.0.lcssa = phi i32 [ %l7, %.lr.ph ]
49  ret i32 %sum.0.lcssa
50}
51
52define i32 @reduction_sum(i32* noalias nocapture %A, i32* noalias nocapture %B) {
53; CHECK-LABEL: @reduction_sum(
54; CHECK-NEXT:  entry:
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 i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
60; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
61; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
62; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
63; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
64; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
65; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
66; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
67; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
68; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[VEC_IND2]])
69; CHECK-NEXT:    [[TMP5:%.*]] = add i32 [[TMP4]], [[VEC_PHI]]
70; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
71; CHECK-NEXT:    [[TMP7:%.*]] = add i32 [[TMP6]], [[TMP5]]
72; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD1]])
73; CHECK-NEXT:    [[TMP9]] = add i32 [[TMP8]], [[TMP7]]
74; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
75; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
76; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
77; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP4:!llvm.loop !.*]]
78; CHECK:       middle.block:
79; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
80; CHECK:       scalar.ph:
81; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
82; CHECK:       .lr.ph:
83; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP5:!llvm.loop !.*]]
84; CHECK:       ._crit_edge:
85; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
86; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
87;
88entry:
89  br label %.lr.ph
90
91.lr.ph:                                           ; preds = %entry, %.lr.ph
92  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
93  %sum.02 = phi i32 [ %l9, %.lr.ph ], [ 0, %entry ]
94  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
95  %l3 = load i32, i32* %l2, align 4
96  %l4 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
97  %l5 = load i32, i32* %l4, align 4
98  %l6 = trunc i64 %indvars.iv to i32
99  %l7 = add i32 %sum.02, %l6
100  %l8 = add i32 %l7, %l3
101  %l9 = add i32 %l8, %l5
102  %indvars.iv.next = add i64 %indvars.iv, 1
103  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
104  %exitcond = icmp eq i32 %lftr.wideiv, 256
105  br i1 %exitcond, label %._crit_edge, label %.lr.ph
106
107._crit_edge:                                      ; preds = %.lr.ph
108  %sum.0.lcssa = phi i32 [ %l9, %.lr.ph ]
109  ret i32 %sum.0.lcssa
110}
111
112define i32 @reduction_sum_const(i32* noalias nocapture %A) {
113; CHECK-LABEL: @reduction_sum_const(
114; CHECK-NEXT:  entry:
115; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
116; CHECK:       vector.ph:
117; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
118; CHECK:       vector.body:
119; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
120; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP4:%.*]], [[VECTOR_BODY]] ]
121; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
122; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
123; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
124; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
125; CHECK-NEXT:    [[TMP3:%.*]] = add i32 [[TMP2]], [[VEC_PHI]]
126; CHECK-NEXT:    [[TMP4]] = add i32 [[TMP3]], 12
127; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
128; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
129; CHECK-NEXT:    br i1 [[TMP5]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP6:!llvm.loop !.*]]
130; CHECK:       middle.block:
131; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
132; CHECK:       scalar.ph:
133; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
134; CHECK:       .lr.ph:
135; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP7:!llvm.loop !.*]]
136; CHECK:       ._crit_edge:
137; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP4]], [[MIDDLE_BLOCK]] ]
138; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
139;
140entry:
141  br label %.lr.ph
142
143.lr.ph:                                           ; preds = %entry, %.lr.ph
144  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
145  %sum.02 = phi i32 [ %l9, %.lr.ph ], [ 0, %entry ]
146  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
147  %l3 = load i32, i32* %l2, align 4
148  %l7 = add i32 %sum.02, %l3
149  %l9 = add i32 %l7, 3
150  %indvars.iv.next = add i64 %indvars.iv, 1
151  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
152  %exitcond = icmp eq i32 %lftr.wideiv, 256
153  br i1 %exitcond, label %._crit_edge, label %.lr.ph
154
155._crit_edge:                                      ; preds = %.lr.ph
156  %sum.0.lcssa = phi i32 [ %l9, %.lr.ph ]
157  ret i32 %sum.0.lcssa
158}
159
160define i32 @reduction_prod(i32* noalias nocapture %A, i32* noalias nocapture %B) {
161; CHECK-LABEL: @reduction_prod(
162; CHECK-NEXT:  entry:
163; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
164; CHECK:       vector.ph:
165; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
166; CHECK:       vector.body:
167; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
168; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 1, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
169; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
170; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
171; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
172; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
173; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
174; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
175; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
176; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> [[VEC_IND2]])
177; CHECK-NEXT:    [[TMP5:%.*]] = mul i32 [[TMP4]], [[VEC_PHI]]
178; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD]])
179; CHECK-NEXT:    [[TMP7:%.*]] = mul i32 [[TMP6]], [[TMP5]]
180; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD1]])
181; CHECK-NEXT:    [[TMP9]] = mul i32 [[TMP8]], [[TMP7]]
182; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
183; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
184; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
185; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP8:!llvm.loop !.*]]
186; CHECK:       middle.block:
187; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
188; CHECK:       scalar.ph:
189; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
190; CHECK:       .lr.ph:
191; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP9:!llvm.loop !.*]]
192; CHECK:       ._crit_edge:
193; CHECK-NEXT:    [[PROD_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
194; CHECK-NEXT:    ret i32 [[PROD_0_LCSSA]]
195;
196entry:
197  br label %.lr.ph
198
199.lr.ph:                                           ; preds = %entry, %.lr.ph
200  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
201  %prod.02 = phi i32 [ %l9, %.lr.ph ], [ 1, %entry ]
202  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
203  %l3 = load i32, i32* %l2, align 4
204  %l4 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
205  %l5 = load i32, i32* %l4, align 4
206  %l6 = trunc i64 %indvars.iv to i32
207  %l7 = mul i32 %prod.02, %l6
208  %l8 = mul i32 %l7, %l3
209  %l9 = mul i32 %l8, %l5
210  %indvars.iv.next = add i64 %indvars.iv, 1
211  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
212  %exitcond = icmp eq i32 %lftr.wideiv, 256
213  br i1 %exitcond, label %._crit_edge, label %.lr.ph
214
215._crit_edge:                                      ; preds = %.lr.ph
216  %prod.0.lcssa = phi i32 [ %l9, %.lr.ph ]
217  ret i32 %prod.0.lcssa
218}
219
220define i32 @reduction_mix(i32* noalias nocapture %A, i32* noalias nocapture %B) {
221; CHECK-LABEL: @reduction_mix(
222; CHECK-NEXT:  entry:
223; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
224; CHECK:       vector.ph:
225; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
226; CHECK:       vector.body:
227; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
228; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
229; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
230; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
231; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
232; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
233; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
234; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
235; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
236; CHECK-NEXT:    [[TMP4:%.*]] = mul nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
237; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[VEC_IND2]])
238; CHECK-NEXT:    [[TMP6:%.*]] = add i32 [[TMP5]], [[VEC_PHI]]
239; CHECK-NEXT:    [[TMP7:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
240; CHECK-NEXT:    [[TMP8]] = add i32 [[TMP7]], [[TMP6]]
241; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
242; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
243; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
244; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP10:!llvm.loop !.*]]
245; CHECK:       middle.block:
246; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
247; CHECK:       scalar.ph:
248; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
249; CHECK:       .lr.ph:
250; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP11:!llvm.loop !.*]]
251; CHECK:       ._crit_edge:
252; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP8]], [[MIDDLE_BLOCK]] ]
253; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
254;
255entry:
256  br label %.lr.ph
257
258.lr.ph:                                           ; preds = %entry, %.lr.ph
259  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
260  %sum.02 = phi i32 [ %l9, %.lr.ph ], [ 0, %entry ]
261  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
262  %l3 = load i32, i32* %l2, align 4
263  %l4 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
264  %l5 = load i32, i32* %l4, align 4
265  %l6 = mul nsw i32 %l5, %l3
266  %l7 = trunc i64 %indvars.iv to i32
267  %l8 = add i32 %sum.02, %l7
268  %l9 = add i32 %l8, %l6
269  %indvars.iv.next = add i64 %indvars.iv, 1
270  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
271  %exitcond = icmp eq i32 %lftr.wideiv, 256
272  br i1 %exitcond, label %._crit_edge, label %.lr.ph
273
274._crit_edge:                                      ; preds = %.lr.ph
275  %sum.0.lcssa = phi i32 [ %l9, %.lr.ph ]
276  ret i32 %sum.0.lcssa
277}
278
279define i32 @reduction_mul(i32* noalias nocapture %A, i32* noalias nocapture %B) {
280; CHECK-LABEL: @reduction_mul(
281; CHECK-NEXT:  entry:
282; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
283; CHECK:       vector.ph:
284; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
285; CHECK:       vector.body:
286; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
287; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 19, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
288; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
289; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
290; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
291; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
292; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
293; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
294; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD]])
295; CHECK-NEXT:    [[TMP5:%.*]] = mul i32 [[TMP4]], [[VEC_PHI]]
296; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> [[WIDE_LOAD1]])
297; CHECK-NEXT:    [[TMP7]] = mul i32 [[TMP6]], [[TMP5]]
298; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
299; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
300; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP12:!llvm.loop !.*]]
301; CHECK:       middle.block:
302; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
303; CHECK:       scalar.ph:
304; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
305; CHECK:       .lr.ph:
306; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP13:!llvm.loop !.*]]
307; CHECK:       ._crit_edge:
308; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
309; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
310;
311entry:
312  br label %.lr.ph
313
314.lr.ph:                                           ; preds = %entry, %.lr.ph
315  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
316  %sum.02 = phi i32 [ %l7, %.lr.ph ], [ 19, %entry ]
317  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
318  %l3 = load i32, i32* %l2, align 4
319  %l4 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
320  %l5 = load i32, i32* %l4, align 4
321  %l6 = mul i32 %sum.02, %l3
322  %l7 = mul i32 %l6, %l5
323  %indvars.iv.next = add i64 %indvars.iv, 1
324  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
325  %exitcond = icmp eq i32 %lftr.wideiv, 256
326  br i1 %exitcond, label %._crit_edge, label %.lr.ph
327
328._crit_edge:                                      ; preds = %.lr.ph
329  %sum.0.lcssa = phi i32 [ %l7, %.lr.ph ]
330  ret i32 %sum.0.lcssa
331}
332
333define i32 @start_at_non_zero(i32* nocapture %in, i32* nocapture %coeff, i32* nocapture %out) {
334; CHECK-LABEL: @start_at_non_zero(
335; CHECK-NEXT:  entry:
336; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
337; CHECK:       vector.ph:
338; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
339; CHECK:       vector.body:
340; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
341; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 120, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
342; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[IN:%.*]], i64 [[INDEX]]
343; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
344; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
345; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[COEFF:%.*]], i64 [[INDEX]]
346; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
347; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
348; CHECK-NEXT:    [[TMP4:%.*]] = mul nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
349; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
350; CHECK-NEXT:    [[TMP6]] = add i32 [[TMP5]], [[VEC_PHI]]
351; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
352; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
353; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP14:!llvm.loop !.*]]
354; CHECK:       middle.block:
355; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
356; CHECK:       scalar.ph:
357; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
358; CHECK:       for.body:
359; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP15:!llvm.loop !.*]]
360; CHECK:       for.end:
361; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
362; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
363;
364entry:
365  br label %for.body
366
367for.body:                                         ; preds = %entry, %for.body
368  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
369  %sum.09 = phi i32 [ %add, %for.body ], [ 120, %entry ]
370  %arrayidx = getelementptr inbounds i32, i32* %in, i64 %indvars.iv
371  %l0 = load i32, i32* %arrayidx, align 4
372  %arrayidx2 = getelementptr inbounds i32, i32* %coeff, i64 %indvars.iv
373  %l1 = load i32, i32* %arrayidx2, align 4
374  %mul = mul nsw i32 %l1, %l0
375  %add = add nsw i32 %mul, %sum.09
376  %indvars.iv.next = add i64 %indvars.iv, 1
377  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
378  %exitcond = icmp eq i32 %lftr.wideiv, 256
379  br i1 %exitcond, label %for.end, label %for.body
380
381for.end:                                          ; preds = %for.body, %entry
382  %sum.0.lcssa = phi i32 [ %add, %for.body ]
383  ret i32 %sum.0.lcssa
384}
385
386define i32 @reduction_and(i32* nocapture %A, i32* nocapture %B) {
387; CHECK-LABEL: @reduction_and(
388; CHECK-NEXT:  entry:
389; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
390; CHECK:       vector.ph:
391; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
392; CHECK:       vector.body:
393; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
394; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ -1, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
395; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
396; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
397; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
398; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
399; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
400; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
401; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> [[WIDE_LOAD]])
402; CHECK-NEXT:    [[TMP5:%.*]] = and i32 [[TMP4]], [[VEC_PHI]]
403; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> [[WIDE_LOAD1]])
404; CHECK-NEXT:    [[TMP7]] = and i32 [[TMP6]], [[TMP5]]
405; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
406; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
407; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP16:!llvm.loop !.*]]
408; CHECK:       middle.block:
409; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
410; CHECK:       scalar.ph:
411; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
412; CHECK:       for.body:
413; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP17:!llvm.loop !.*]]
414; CHECK:       for.end:
415; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
416; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
417;
418entry:
419  br label %for.body
420
421for.body:                                         ; preds = %entry, %for.body
422  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
423  %result.08 = phi i32 [ %and, %for.body ], [ -1, %entry ]
424  %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
425  %l0 = load i32, i32* %arrayidx, align 4
426  %arrayidx2 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
427  %l1 = load i32, i32* %arrayidx2, align 4
428  %add = and i32 %result.08, %l0
429  %and = and i32 %add, %l1
430  %indvars.iv.next = add i64 %indvars.iv, 1
431  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
432  %exitcond = icmp eq i32 %lftr.wideiv, 256
433  br i1 %exitcond, label %for.end, label %for.body
434
435for.end:                                          ; preds = %for.body, %entry
436  %result.0.lcssa = phi i32 [ %and, %for.body ]
437  ret i32 %result.0.lcssa
438}
439
440define i32 @reduction_or(i32* nocapture %A, i32* nocapture %B) {
441; CHECK-LABEL: @reduction_or(
442; CHECK-NEXT:  entry:
443; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
444; CHECK:       vector.ph:
445; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
446; CHECK:       vector.body:
447; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
448; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
449; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
450; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
451; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
452; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
453; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
454; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
455; CHECK-NEXT:    [[TMP4:%.*]] = add nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
456; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> [[TMP4]])
457; CHECK-NEXT:    [[TMP6]] = or i32 [[TMP5]], [[VEC_PHI]]
458; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
459; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
460; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP18:!llvm.loop !.*]]
461; CHECK:       middle.block:
462; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
463; CHECK:       scalar.ph:
464; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
465; CHECK:       for.body:
466; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP19:!llvm.loop !.*]]
467; CHECK:       for.end:
468; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
469; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
470;
471entry:
472  br label %for.body
473
474for.body:                                         ; preds = %entry, %for.body
475  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
476  %result.08 = phi i32 [ %or, %for.body ], [ 0, %entry ]
477  %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
478  %l0 = load i32, i32* %arrayidx, align 4
479  %arrayidx2 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
480  %l1 = load i32, i32* %arrayidx2, align 4
481  %add = add nsw i32 %l1, %l0
482  %or = or i32 %add, %result.08
483  %indvars.iv.next = add i64 %indvars.iv, 1
484  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
485  %exitcond = icmp eq i32 %lftr.wideiv, 256
486  br i1 %exitcond, label %for.end, label %for.body
487
488for.end:                                          ; preds = %for.body, %entry
489  %result.0.lcssa = phi i32 [ %or, %for.body ]
490  ret i32 %result.0.lcssa
491}
492
493define i32 @reduction_xor(i32* nocapture %A, i32* nocapture %B) {
494; CHECK-LABEL: @reduction_xor(
495; CHECK-NEXT:  entry:
496; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
497; CHECK:       vector.ph:
498; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
499; CHECK:       vector.body:
500; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
501; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
502; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
503; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
504; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
505; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
506; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
507; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
508; CHECK-NEXT:    [[TMP4:%.*]] = add nsw <4 x i32> [[WIDE_LOAD1]], [[WIDE_LOAD]]
509; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> [[TMP4]])
510; CHECK-NEXT:    [[TMP6]] = xor i32 [[TMP5]], [[VEC_PHI]]
511; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
512; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
513; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP20:!llvm.loop !.*]]
514; CHECK:       middle.block:
515; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
516; CHECK:       scalar.ph:
517; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
518; CHECK:       for.body:
519; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP21:!llvm.loop !.*]]
520; CHECK:       for.end:
521; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
522; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
523;
524entry:
525  br label %for.body
526
527for.body:                                         ; preds = %entry, %for.body
528  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
529  %result.08 = phi i32 [ %xor, %for.body ], [ 0, %entry ]
530  %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
531  %l0 = load i32, i32* %arrayidx, align 4
532  %arrayidx2 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
533  %l1 = load i32, i32* %arrayidx2, align 4
534  %add = add nsw i32 %l1, %l0
535  %xor = xor i32 %add, %result.08
536  %indvars.iv.next = add i64 %indvars.iv, 1
537  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
538  %exitcond = icmp eq i32 %lftr.wideiv, 256
539  br i1 %exitcond, label %for.end, label %for.body
540
541for.end:                                          ; preds = %for.body, %entry
542  %result.0.lcssa = phi i32 [ %xor, %for.body ]
543  ret i32 %result.0.lcssa
544}
545
546define float @reduction_fadd(float* nocapture %A, float* nocapture %B) {
547; CHECK-LABEL: @reduction_fadd(
548; CHECK-NEXT:  entry:
549; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
550; CHECK:       vector.ph:
551; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
552; CHECK:       vector.body:
553; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
554; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi float [ 0.000000e+00, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
555; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds float, float* [[A:%.*]], i64 [[INDEX]]
556; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float* [[TMP0]] to <4 x float>*
557; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP1]], align 4
558; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, float* [[B:%.*]], i64 [[INDEX]]
559; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float* [[TMP2]] to <4 x float>*
560; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x float>, <4 x float>* [[TMP3]], align 4
561; CHECK-NEXT:    [[TMP4:%.*]] = call fast float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[WIDE_LOAD]])
562; CHECK-NEXT:    [[TMP5:%.*]] = fadd float [[TMP4]], [[VEC_PHI]]
563; CHECK-NEXT:    [[TMP6:%.*]] = call fast float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[WIDE_LOAD1]])
564; CHECK-NEXT:    [[TMP7]] = fadd float [[TMP6]], [[TMP5]]
565; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
566; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
567; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP22:!llvm.loop !.*]]
568; CHECK:       middle.block:
569; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
570; CHECK:       scalar.ph:
571; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
572; CHECK:       for.body:
573; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP23:!llvm.loop !.*]]
574; CHECK:       for.end:
575; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi float [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
576; CHECK-NEXT:    ret float [[RESULT_0_LCSSA]]
577;
578entry:
579  br label %for.body
580
581for.body:                                         ; preds = %entry, %for.body
582  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
583  %result.08 = phi float [ %fadd, %for.body ], [ 0.0, %entry ]
584  %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
585  %l0 = load float, float* %arrayidx, align 4
586  %arrayidx2 = getelementptr inbounds float, float* %B, i64 %indvars.iv
587  %l1 = load float, float* %arrayidx2, align 4
588  %add = fadd fast float %result.08, %l0
589  %fadd = fadd fast float %add, %l1
590  %indvars.iv.next = add i64 %indvars.iv, 1
591  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
592  %exitcond = icmp eq i32 %lftr.wideiv, 256
593  br i1 %exitcond, label %for.end, label %for.body
594
595for.end:                                          ; preds = %for.body, %entry
596  %result.0.lcssa = phi float [ %fadd, %for.body ]
597  ret float %result.0.lcssa
598}
599
600define float @reduction_fmul(float* nocapture %A, float* nocapture %B) {
601; CHECK-LABEL: @reduction_fmul(
602; CHECK-NEXT:  entry:
603; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
604; CHECK:       vector.ph:
605; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
606; CHECK:       vector.body:
607; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
608; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi float [ 0.000000e+00, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
609; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds float, float* [[A:%.*]], i64 [[INDEX]]
610; CHECK-NEXT:    [[TMP1:%.*]] = bitcast float* [[TMP0]] to <4 x float>*
611; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP1]], align 4
612; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds float, float* [[B:%.*]], i64 [[INDEX]]
613; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float* [[TMP2]] to <4 x float>*
614; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x float>, <4 x float>* [[TMP3]], align 4
615; CHECK-NEXT:    [[TMP4:%.*]] = call fast float @llvm.vector.reduce.fmul.v4f32(float 1.000000e+00, <4 x float> [[WIDE_LOAD]])
616; CHECK-NEXT:    [[TMP5:%.*]] = fmul float [[TMP4]], [[VEC_PHI]]
617; CHECK-NEXT:    [[TMP6:%.*]] = call fast float @llvm.vector.reduce.fmul.v4f32(float 1.000000e+00, <4 x float> [[WIDE_LOAD1]])
618; CHECK-NEXT:    [[TMP7]] = fmul float [[TMP6]], [[TMP5]]
619; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
620; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
621; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP24:!llvm.loop !.*]]
622; CHECK:       middle.block:
623; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
624; CHECK:       scalar.ph:
625; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
626; CHECK:       for.body:
627; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP25:!llvm.loop !.*]]
628; CHECK:       for.end:
629; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi float [ undef, [[FOR_BODY]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
630; CHECK-NEXT:    ret float [[RESULT_0_LCSSA]]
631;
632entry:
633  br label %for.body
634
635for.body:                                         ; preds = %entry, %for.body
636  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
637  %result.08 = phi float [ %fmul, %for.body ], [ 0.0, %entry ]
638  %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
639  %l0 = load float, float* %arrayidx, align 4
640  %arrayidx2 = getelementptr inbounds float, float* %B, i64 %indvars.iv
641  %l1 = load float, float* %arrayidx2, align 4
642  %add = fmul fast float %result.08, %l0
643  %fmul = fmul fast float %add, %l1
644  %indvars.iv.next = add i64 %indvars.iv, 1
645  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
646  %exitcond = icmp eq i32 %lftr.wideiv, 256
647  br i1 %exitcond, label %for.end, label %for.body
648
649for.end:                                          ; preds = %for.body, %entry
650  %result.0.lcssa = phi float [ %fmul, %for.body ]
651  ret float %result.0.lcssa
652}
653
654define i32 @reduction_min(i32* nocapture %A, i32* nocapture %B) {
655; CHECK-LABEL: @reduction_min(
656; CHECK-NEXT:  entry:
657; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
658; CHECK:       vector.ph:
659; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
660; CHECK:       vector.body:
661; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
662; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 1000, [[VECTOR_PH]] ], [ [[RDX_MINMAX_SELECT:%.*]], [[VECTOR_BODY]] ]
663; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
664; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
665; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
666; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.vector.reduce.smin.v4i32(<4 x i32> [[WIDE_LOAD]])
667; CHECK-NEXT:    [[RDX_MINMAX_CMP:%.*]] = icmp slt i32 [[TMP2]], [[VEC_PHI]]
668; CHECK-NEXT:    [[RDX_MINMAX_SELECT]] = select i1 [[RDX_MINMAX_CMP]], i32 [[TMP2]], i32 [[VEC_PHI]]
669; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
670; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
671; CHECK-NEXT:    br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP26:!llvm.loop !.*]]
672; CHECK:       middle.block:
673; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
674; CHECK:       scalar.ph:
675; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
676; CHECK:       for.body:
677; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP27:!llvm.loop !.*]]
678; CHECK:       for.end:
679; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[RDX_MINMAX_SELECT]], [[MIDDLE_BLOCK]] ]
680; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
681;
682entry:
683  br label %for.body
684
685for.body:                                         ; preds = %entry, %for.body
686  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
687  %result.08 = phi i32 [ %v0, %for.body ], [ 1000, %entry ]
688  %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
689  %l0 = load i32, i32* %arrayidx, align 4
690  %c0 = icmp slt i32 %result.08, %l0
691  %v0 = select i1 %c0, i32 %result.08, i32 %l0
692  %indvars.iv.next = add i64 %indvars.iv, 1
693  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
694  %exitcond = icmp eq i32 %lftr.wideiv, 256
695  br i1 %exitcond, label %for.end, label %for.body
696
697for.end:                                          ; preds = %for.body, %entry
698  %result.0.lcssa = phi i32 [ %v0, %for.body ]
699  ret i32 %result.0.lcssa
700}
701
702define i32 @reduction_max(i32* nocapture %A, i32* nocapture %B) {
703; CHECK-LABEL: @reduction_max(
704; CHECK-NEXT:  entry:
705; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
706; CHECK:       vector.ph:
707; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
708; CHECK:       vector.body:
709; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
710; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 1000, [[VECTOR_PH]] ], [ [[RDX_MINMAX_SELECT:%.*]], [[VECTOR_BODY]] ]
711; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
712; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
713; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
714; CHECK-NEXT:    [[TMP2:%.*]] = call i32 @llvm.vector.reduce.umax.v4i32(<4 x i32> [[WIDE_LOAD]])
715; CHECK-NEXT:    [[RDX_MINMAX_CMP:%.*]] = icmp ugt i32 [[TMP2]], [[VEC_PHI]]
716; CHECK-NEXT:    [[RDX_MINMAX_SELECT]] = select i1 [[RDX_MINMAX_CMP]], i32 [[TMP2]], i32 [[VEC_PHI]]
717; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
718; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
719; CHECK-NEXT:    br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP28:!llvm.loop !.*]]
720; CHECK:       middle.block:
721; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
722; CHECK:       scalar.ph:
723; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
724; CHECK:       for.body:
725; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP29:!llvm.loop !.*]]
726; CHECK:       for.end:
727; CHECK-NEXT:    [[RESULT_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[RDX_MINMAX_SELECT]], [[MIDDLE_BLOCK]] ]
728; CHECK-NEXT:    ret i32 [[RESULT_0_LCSSA]]
729;
730entry:
731  br label %for.body
732
733for.body:                                         ; preds = %entry, %for.body
734  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
735  %result.08 = phi i32 [ %v0, %for.body ], [ 1000, %entry ]
736  %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
737  %l0 = load i32, i32* %arrayidx, align 4
738  %c0 = icmp ugt i32 %result.08, %l0
739  %v0 = select i1 %c0, i32 %result.08, i32 %l0
740  %indvars.iv.next = add i64 %indvars.iv, 1
741  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
742  %exitcond = icmp eq i32 %lftr.wideiv, 256
743  br i1 %exitcond, label %for.end, label %for.body
744
745for.end:                                          ; preds = %for.body, %entry
746  %result.0.lcssa = phi i32 [ %v0, %for.body ]
747  ret i32 %result.0.lcssa
748}
749
750; Sub we can create a reduction, but not inloop
751define i32 @reduction_sub_lhs(i32* noalias nocapture %A) {
752; CHECK-LABEL: @reduction_sub_lhs(
753; CHECK-NEXT:  entry:
754; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
755; CHECK:       vector.ph:
756; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
757; CHECK:       vector.body:
758; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
759; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ]
760; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
761; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
762; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
763; CHECK-NEXT:    [[TMP2]] = sub <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
764; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
765; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
766; CHECK-NEXT:    br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP30:!llvm.loop !.*]]
767; CHECK:       middle.block:
768; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP2]])
769; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
770; CHECK:       scalar.ph:
771; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
772; CHECK:       for.body:
773; CHECK-NEXT:    br i1 undef, label [[FOR_END]], label [[FOR_BODY]], [[LOOP31:!llvm.loop !.*]]
774; CHECK:       for.end:
775; CHECK-NEXT:    [[X_0_LCSSA:%.*]] = phi i32 [ undef, [[FOR_BODY]] ], [ [[TMP4]], [[MIDDLE_BLOCK]] ]
776; CHECK-NEXT:    ret i32 [[X_0_LCSSA]]
777;
778entry:
779  br label %for.body
780
781for.body:                                         ; preds = %entry, %for.body
782  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
783  %x.05 = phi i32 [ %sub, %for.body ], [ 0, %entry ]
784  %arrayidx = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
785  %l0 = load i32, i32* %arrayidx, align 4
786  %sub = sub nsw i32 %x.05, %l0
787  %indvars.iv.next = add i64 %indvars.iv, 1
788  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
789  %exitcond = icmp eq i32 %lftr.wideiv, 256
790  br i1 %exitcond, label %for.end, label %for.body
791
792for.end:                                          ; preds = %for.body, %entry
793  %x.0.lcssa = phi i32 [ %sub, %for.body ]
794  ret i32 %x.0.lcssa
795}
796
797; Conditional reductions with multi-input phis.
798define float @reduction_conditional(float* %A, float* %B, float* %C, float %S) {
799; CHECK-LABEL: @reduction_conditional(
800; CHECK-NEXT:  entry:
801; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
802; CHECK:       vector.ph:
803; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <4 x float> <float undef, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00>, float [[S:%.*]], i32 0
804; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
805; CHECK:       vector.body:
806; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
807; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ [[TMP0]], [[VECTOR_PH]] ], [ [[PREDPHI3:%.*]], [[VECTOR_BODY]] ]
808; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds float, float* [[A:%.*]], i64 [[INDEX]]
809; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float* [[TMP1]] to <4 x float>*
810; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, <4 x float>* [[TMP2]], align 4
811; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds float, float* [[B:%.*]], i64 [[INDEX]]
812; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float* [[TMP3]] to <4 x float>*
813; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x float>, <4 x float>* [[TMP4]], align 4
814; CHECK-NEXT:    [[TMP5:%.*]] = fcmp ogt <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD1]]
815; CHECK-NEXT:    [[TMP6:%.*]] = fcmp ule <4 x float> [[WIDE_LOAD1]], <float 1.000000e+00, float 1.000000e+00, float 1.000000e+00, float 1.000000e+00>
816; CHECK-NEXT:    [[TMP7:%.*]] = fcmp ogt <4 x float> [[WIDE_LOAD]], <float 2.000000e+00, float 2.000000e+00, float 2.000000e+00, float 2.000000e+00>
817; CHECK-NEXT:    [[TMP8:%.*]] = and <4 x i1> [[TMP6]], [[TMP5]]
818; CHECK-NEXT:    [[TMP9:%.*]] = and <4 x i1> [[TMP7]], [[TMP8]]
819; CHECK-NEXT:    [[TMP10:%.*]] = xor <4 x i1> [[TMP7]], <i1 true, i1 true, i1 true, i1 true>
820; CHECK-NEXT:    [[TMP11:%.*]] = and <4 x i1> [[TMP8]], [[TMP10]]
821; CHECK-NEXT:    [[TMP12:%.*]] = xor <4 x i1> [[TMP5]], <i1 true, i1 true, i1 true, i1 true>
822; CHECK-NEXT:    [[PREDPHI_V:%.*]] = select <4 x i1> [[TMP9]], <4 x float> [[WIDE_LOAD1]], <4 x float> [[WIDE_LOAD]]
823; CHECK-NEXT:    [[PREDPHI:%.*]] = fadd fast <4 x float> [[VEC_PHI]], [[PREDPHI_V]]
824; CHECK-NEXT:    [[TMP13:%.*]] = or <4 x i1> [[TMP11]], [[TMP12]]
825; CHECK-NEXT:    [[PREDPHI3]] = select <4 x i1> [[TMP13]], <4 x float> [[VEC_PHI]], <4 x float> [[PREDPHI]]
826; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
827; CHECK-NEXT:    [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], 128
828; CHECK-NEXT:    br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP32:!llvm.loop !.*]]
829; CHECK:       middle.block:
830; CHECK-NEXT:    [[TMP15:%.*]] = call fast float @llvm.vector.reduce.fadd.v4f32(float -0.000000e+00, <4 x float> [[PREDPHI3]])
831; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
832; CHECK:       scalar.ph:
833; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
834; CHECK:       for.body:
835; CHECK-NEXT:    br i1 undef, label [[IF_THEN:%.*]], label [[FOR_INC:%.*]]
836; CHECK:       if.then:
837; CHECK-NEXT:    br i1 undef, label [[IF_THEN8:%.*]], label [[IF_ELSE:%.*]]
838; CHECK:       if.then8:
839; CHECK-NEXT:    br label [[FOR_INC]]
840; CHECK:       if.else:
841; CHECK-NEXT:    br i1 undef, label [[IF_THEN16:%.*]], label [[FOR_INC]]
842; CHECK:       if.then16:
843; CHECK-NEXT:    br label [[FOR_INC]]
844; CHECK:       for.inc:
845; CHECK-NEXT:    br i1 undef, label [[FOR_BODY]], label [[FOR_END]], [[LOOP33:!llvm.loop !.*]]
846; CHECK:       for.end:
847; CHECK-NEXT:    [[SUM_1_LCSSA:%.*]] = phi float [ undef, [[FOR_INC]] ], [ [[TMP15]], [[MIDDLE_BLOCK]] ]
848; CHECK-NEXT:    ret float [[SUM_1_LCSSA]]
849;
850entry:
851  br label %for.body
852
853for.body:
854  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.inc ]
855  %sum.033 = phi float [ %S, %entry ], [ %sum.1, %for.inc ]
856  %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
857  %l0 = load float, float* %arrayidx, align 4
858  %arrayidx2 = getelementptr inbounds float, float* %B, i64 %indvars.iv
859  %l1 = load float, float* %arrayidx2, align 4
860  %cmp3 = fcmp ogt float %l0, %l1
861  br i1 %cmp3, label %if.then, label %for.inc
862
863if.then:
864  %cmp6 = fcmp ogt float %l1, 1.000000e+00
865  br i1 %cmp6, label %if.then8, label %if.else
866
867if.then8:
868  %add = fadd fast float %sum.033, %l0
869  br label %for.inc
870
871if.else:
872  %cmp14 = fcmp ogt float %l0, 2.000000e+00
873  br i1 %cmp14, label %if.then16, label %for.inc
874
875if.then16:
876  %add19 = fadd fast float %sum.033, %l1
877  br label %for.inc
878
879for.inc:
880  %sum.1 = phi float [ %add, %if.then8 ], [ %add19, %if.then16 ], [ %sum.033, %if.else ], [ %sum.033, %for.body ]
881  %indvars.iv.next = add i64 %indvars.iv, 1
882  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
883  %exitcond = icmp ne i32 %lftr.wideiv, 128
884  br i1 %exitcond, label %for.body, label %for.end
885
886for.end:
887  %sum.1.lcssa = phi float [ %sum.1, %for.inc ]
888  ret float %sum.1.lcssa
889}
890
891define i32 @reduction_sum_multiuse(i32* noalias nocapture %A, i32* noalias nocapture %B) {
892; CHECK-LABEL: @reduction_sum_multiuse(
893; CHECK-NEXT:  entry:
894; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
895; CHECK:       .lr.ph:
896; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[DOTLR_PH]] ], [ 0, [[ENTRY:%.*]] ]
897; CHECK-NEXT:    [[SUM_02:%.*]] = phi i32 [ [[L10:%.*]], [[DOTLR_PH]] ], [ 0, [[ENTRY]] ]
898; CHECK-NEXT:    [[L2:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]]
899; CHECK-NEXT:    [[L3:%.*]] = load i32, i32* [[L2]], align 4
900; CHECK-NEXT:    [[L6:%.*]] = trunc i64 [[INDVARS_IV]] to i32
901; CHECK-NEXT:    [[L7:%.*]] = add i32 [[SUM_02]], [[L6]]
902; CHECK-NEXT:    [[L8:%.*]] = add i32 [[L7]], [[L3]]
903; CHECK-NEXT:    [[L10]] = add i32 [[L8]], [[SUM_02]]
904; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1
905; CHECK-NEXT:    [[LFTR_WIDEIV:%.*]] = trunc i64 [[INDVARS_IV_NEXT]] to i32
906; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[LFTR_WIDEIV]], 256
907; CHECK-NEXT:    br i1 [[EXITCOND]], label [[END:%.*]], label [[DOTLR_PH]]
908; CHECK:       end:
909; CHECK-NEXT:    ret i32 [[L10]]
910;
911entry:
912  br label %.lr.ph
913
914.lr.ph:                                           ; preds = %entry, %.lr.ph
915  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
916  %sum.02 = phi i32 [ %l10, %.lr.ph ], [ 0, %entry ]
917  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
918  %l3 = load i32, i32* %l2, align 4
919  %l4 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
920  %l5 = load i32, i32* %l4, align 4
921  %l6 = trunc i64 %indvars.iv to i32
922  %l7 = add i32 %sum.02, %l6
923  %l8 = add i32 %l7, %l3
924  %l9 = add i32 %l8, %l5
925  %l10 = add i32 %l8, %sum.02
926  %indvars.iv.next = add i64 %indvars.iv, 1
927  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
928  %exitcond = icmp eq i32 %lftr.wideiv, 256
929  br i1 %exitcond, label %end, label %.lr.ph
930
931end:
932  %f1 = phi i32 [ %l10, %.lr.ph ]
933  ret i32 %f1
934}
935
936; Predicated loop, cannot (yet) use in-loop reductions.
937define i32 @reduction_predicated(i32* noalias nocapture %A, i32* noalias nocapture %B) {
938; CHECK-LABEL: @reduction_predicated(
939; CHECK-NEXT:  entry:
940; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
941; CHECK:       vector.ph:
942; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
943; CHECK:       vector.body:
944; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
945; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
946; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT3:%.*]], [[VECTOR_BODY]] ]
947; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
948; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[TMP0]] to <4 x i32>*
949; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP1]], align 4
950; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[INDEX]]
951; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>*
952; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4
953; CHECK-NEXT:    [[TMP4:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[VEC_IND2]])
954; CHECK-NEXT:    [[TMP5:%.*]] = add i32 [[TMP4]], [[VEC_PHI]]
955; CHECK-NEXT:    [[TMP6:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD]])
956; CHECK-NEXT:    [[TMP7:%.*]] = add i32 [[TMP6]], [[TMP5]]
957; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[WIDE_LOAD1]])
958; CHECK-NEXT:    [[TMP9]] = add i32 [[TMP8]], [[TMP7]]
959; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
960; CHECK-NEXT:    [[VEC_IND_NEXT3]] = add <4 x i32> [[VEC_IND2]], <i32 4, i32 4, i32 4, i32 4>
961; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
962; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP34:!llvm.loop !.*]]
963; CHECK:       middle.block:
964; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
965; CHECK:       scalar.ph:
966; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
967; CHECK:       .lr.ph:
968; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP35:!llvm.loop !.*]]
969; CHECK:       ._crit_edge:
970; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP9]], [[MIDDLE_BLOCK]] ]
971; CHECK-NEXT:    ret i32 [[SUM_0_LCSSA]]
972;
973entry:
974  br label %.lr.ph
975
976.lr.ph:                                           ; preds = %entry, %.lr.ph
977  %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
978  %sum.02 = phi i32 [ %l9, %.lr.ph ], [ 0, %entry ]
979  %l2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
980  %l3 = load i32, i32* %l2, align 4
981  %l4 = getelementptr inbounds i32, i32* %B, i64 %indvars.iv
982  %l5 = load i32, i32* %l4, align 4
983  %l6 = trunc i64 %indvars.iv to i32
984  %l7 = add i32 %sum.02, %l6
985  %l8 = add i32 %l7, %l3
986  %l9 = add i32 %l8, %l5
987  %indvars.iv.next = add i64 %indvars.iv, 1
988  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
989  %exitcond = icmp eq i32 %lftr.wideiv, 256
990  br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !6
991
992._crit_edge:                                      ; preds = %.lr.ph
993  %sum.0.lcssa = phi i32 [ %l9, %.lr.ph ]
994  ret i32 %sum.0.lcssa
995}
996
997define i8 @reduction_add_trunc(i8* noalias nocapture %A) {
998; CHECK-LABEL: @reduction_add_trunc(
999; CHECK-NEXT:  entry:
1000; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1001; CHECK:       vector.ph:
1002; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
1003; CHECK:       vector.body:
1004; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1005; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i8> [ <i8 -1, i8 0, i8 0, i8 0>, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
1006; CHECK-NEXT:    [[TMP0:%.*]] = sext i32 [[INDEX]] to i64
1007; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, i8* [[A:%.*]], i64 [[TMP0]]
1008; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[TMP1]] to <4 x i8>*
1009; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i8>, <4 x i8>* [[TMP2]], align 4
1010; CHECK-NEXT:    [[TMP3]] = add <4 x i8> [[VEC_PHI]], [[WIDE_LOAD]]
1011; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 4
1012; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 256
1013; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP36:!llvm.loop !.*]]
1014; CHECK:       middle.block:
1015; CHECK-NEXT:    [[TMP5:%.*]] = call i8 @llvm.vector.reduce.add.v4i8(<4 x i8> [[TMP3]])
1016; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
1017; CHECK:       scalar.ph:
1018; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
1019; CHECK:       .lr.ph:
1020; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP37:!llvm.loop !.*]]
1021; CHECK:       ._crit_edge:
1022; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i8 [ undef, [[DOTLR_PH]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ]
1023; CHECK-NEXT:    ret i8 [[SUM_0_LCSSA]]
1024;
1025entry:
1026  br label %.lr.ph
1027
1028.lr.ph:                                           ; preds = %entry, %.lr.ph
1029  %indvars.iv = phi i32 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
1030  %sum.02p = phi i32 [ %l9, %.lr.ph ], [ 255, %entry ]
1031  %sum.02 = and i32 %sum.02p, 255
1032  %l2 = getelementptr inbounds i8, i8* %A, i32 %indvars.iv
1033  %l3 = load i8, i8* %l2, align 4
1034  %l3e = zext i8 %l3 to i32
1035  %l9 = add i32 %sum.02, %l3e
1036  %indvars.iv.next = add i32 %indvars.iv, 1
1037  %exitcond = icmp eq i32 %indvars.iv.next, 256
1038  br i1 %exitcond, label %._crit_edge, label %.lr.ph
1039
1040._crit_edge:                                      ; preds = %.lr.ph
1041  %sum.0.lcssa = phi i32 [ %l9, %.lr.ph ]
1042  %ret = trunc i32 %sum.0.lcssa to i8
1043  ret i8 %ret
1044}
1045
1046
1047define i8 @reduction_and_trunc(i8* noalias nocapture %A) {
1048; CHECK-LABEL: @reduction_and_trunc(
1049; CHECK-NEXT:  entry:
1050; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1051; CHECK:       vector.ph:
1052; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
1053; CHECK:       vector.body:
1054; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
1055; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 255, [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
1056; CHECK-NEXT:    [[TMP0:%.*]] = and i32 [[VEC_PHI]], 255
1057; CHECK-NEXT:    [[TMP1:%.*]] = sext i32 [[INDEX]] to i64
1058; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i8, i8* [[A:%.*]], i64 [[TMP1]]
1059; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i8* [[TMP2]] to <4 x i8>*
1060; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i8>, <4 x i8>* [[TMP3]], align 4
1061; CHECK-NEXT:    [[TMP4:%.*]] = zext <4 x i8> [[WIDE_LOAD]] to <4 x i32>
1062; CHECK-NEXT:    [[TMP5:%.*]] = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> [[TMP4]])
1063; CHECK-NEXT:    [[TMP6]] = and i32 [[TMP5]], [[TMP0]]
1064; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 4
1065; CHECK-NEXT:    [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], 256
1066; CHECK-NEXT:    br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], [[LOOP38:!llvm.loop !.*]]
1067; CHECK:       middle.block:
1068; CHECK-NEXT:    br i1 true, label [[DOT_CRIT_EDGE:%.*]], label [[SCALAR_PH]]
1069; CHECK:       scalar.ph:
1070; CHECK-NEXT:    br label [[DOTLR_PH:%.*]]
1071; CHECK:       .lr.ph:
1072; CHECK-NEXT:    br i1 undef, label [[DOT_CRIT_EDGE]], label [[DOTLR_PH]], [[LOOP39:!llvm.loop !.*]]
1073; CHECK:       ._crit_edge:
1074; CHECK-NEXT:    [[SUM_0_LCSSA:%.*]] = phi i32 [ undef, [[DOTLR_PH]] ], [ [[TMP6]], [[MIDDLE_BLOCK]] ]
1075; CHECK-NEXT:    [[RET:%.*]] = trunc i32 [[SUM_0_LCSSA]] to i8
1076; CHECK-NEXT:    ret i8 [[RET]]
1077;
1078entry:
1079  br label %.lr.ph
1080
1081.lr.ph:                                           ; preds = %entry, %.lr.ph
1082  %indvars.iv = phi i32 [ %indvars.iv.next, %.lr.ph ], [ 0, %entry ]
1083  %sum.02p = phi i32 [ %l9, %.lr.ph ], [ 255, %entry ]
1084  %sum.02 = and i32 %sum.02p, 255
1085  %l2 = getelementptr inbounds i8, i8* %A, i32 %indvars.iv
1086  %l3 = load i8, i8* %l2, align 4
1087  %l3e = zext i8 %l3 to i32
1088  %l9 = and i32 %sum.02, %l3e
1089  %indvars.iv.next = add i32 %indvars.iv, 1
1090  %exitcond = icmp eq i32 %indvars.iv.next, 256
1091  br i1 %exitcond, label %._crit_edge, label %.lr.ph
1092
1093._crit_edge:                                      ; preds = %.lr.ph
1094  %sum.0.lcssa = phi i32 [ %l9, %.lr.ph ]
1095  %ret = trunc i32 %sum.0.lcssa to i8
1096  ret i8 %ret
1097}
1098
1099!6 = distinct !{!6, !7, !8}
1100!7 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}
1101!8 = !{!"llvm.loop.vectorize.enable", i1 true}
1102