1; RUN: opt < %s -loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -dce -instcombine -S | FileCheck %s
2; RUN: opt < %s -loop-vectorize -force-vector-width=4 -force-vector-interleave=2 -dce -instcombine -S | FileCheck %s --check-prefix=UNROLL
3; RUN: opt < %s -loop-vectorize -force-vector-width=4 -force-vector-interleave=2 -S | FileCheck %s --check-prefix=UNROLL-NO-IC
4; RUN: opt < %s -loop-vectorize -force-vector-width=1 -force-vector-interleave=2 -S | FileCheck %s --check-prefix=UNROLL-NO-VF
5; RUN: opt < %s -loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S | FileCheck %s --check-prefix=SINK-AFTER
6; RUN: opt < %s -loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S | FileCheck %s --check-prefix=NO-SINK-AFTER
7
8target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
9
10; void recurrence_1(int *a, int *b, int n) {
11;   for(int i = 0; i < n; i++)
12;     b[i] =  a[i] + a[i - 1]
13; }
14;
15; CHECK-LABEL: @recurrence_1(
16; CHECK:       vector.ph:
17; CHECK:         %vector.recur.init = insertelement <4 x i32> undef, i32 %pre_load, i32 3
18; CHECK:       vector.body:
19; CHECK:         %vector.recur = phi <4 x i32> [ %vector.recur.init, %vector.ph ], [ [[L1:%[a-zA-Z0-9.]+]], %vector.body ]
20; CHECK:         [[L1]] = load <4 x i32>
21; CHECK:         {{.*}} = shufflevector <4 x i32> %vector.recur, <4 x i32> [[L1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
22; CHECK:       middle.block:
23; CHECK:         %vector.recur.extract = extractelement <4 x i32> [[L1]], i32 3
24; CHECK:       scalar.ph:
25; CHECK:         %scalar.recur.init = phi i32 [ %vector.recur.extract, %middle.block ], [ %pre_load, %vector.memcheck ], [ %pre_load, %min.iters.checked ], [ %pre_load, %for.preheader ]
26; CHECK:       scalar.body:
27; CHECK:         %scalar.recur = phi i32 [ %scalar.recur.init, %scalar.ph ], [ {{.*}}, %scalar.body ]
28;
29; UNROLL-LABEL: @recurrence_1(
30; UNROLL:       vector.body:
31; UNROLL:         %vector.recur = phi <4 x i32> [ %vector.recur.init, %vector.ph ], [ [[L2:%[a-zA-Z0-9.]+]], %vector.body ]
32; UNROLL:         [[L1:%[a-zA-Z0-9.]+]] = load <4 x i32>
33; UNROLL:         [[L2]] = load <4 x i32>
34; UNROLL:         {{.*}} = shufflevector <4 x i32> %vector.recur, <4 x i32> [[L1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
35; UNROLL:         {{.*}} = shufflevector <4 x i32> [[L1]], <4 x i32> [[L2]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
36; UNROLL:       middle.block:
37; UNROLL:         %vector.recur.extract = extractelement <4 x i32> [[L2]], i32 3
38;
39define void @recurrence_1(i32* nocapture readonly %a, i32* nocapture %b, i32 %n) {
40entry:
41  br label %for.preheader
42
43for.preheader:
44  %arrayidx.phi.trans.insert = getelementptr inbounds i32, i32* %a, i64 0
45  %pre_load = load i32, i32* %arrayidx.phi.trans.insert
46  br label %scalar.body
47
48scalar.body:
49  %0 = phi i32 [ %pre_load, %for.preheader ], [ %1, %scalar.body ]
50  %indvars.iv = phi i64 [ 0, %for.preheader ], [ %indvars.iv.next, %scalar.body ]
51  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
52  %arrayidx32 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv.next
53  %1 = load i32, i32* %arrayidx32
54  %arrayidx34 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
55  %add35 = add i32 %1, %0
56  store i32 %add35, i32* %arrayidx34
57  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
58  %exitcond = icmp eq i32 %lftr.wideiv, %n
59  br i1 %exitcond, label %for.exit, label %scalar.body
60
61for.exit:
62  ret void
63}
64
65; int recurrence_2(int *a, int n) {
66;   int minmax;
67;   for (int i = 0; i < n; ++i)
68;     minmax = min(minmax, max(a[i] - a[i-1], 0));
69;   return minmax;
70; }
71;
72; CHECK-LABEL: @recurrence_2(
73; CHECK:       vector.ph:
74; CHECK:         %vector.recur.init = insertelement <4 x i32> undef, i32 %.pre, i32 3
75; CHECK:       vector.body:
76; CHECK:         %vector.recur = phi <4 x i32> [ %vector.recur.init, %vector.ph ], [ [[L1:%[a-zA-Z0-9.]+]], %vector.body ]
77; CHECK:         [[L1]] = load <4 x i32>
78; CHECK:         {{.*}} = shufflevector <4 x i32> %vector.recur, <4 x i32> [[L1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
79; CHECK:       middle.block:
80; CHECK:         %vector.recur.extract = extractelement <4 x i32> [[L1]], i32 3
81; CHECK:       scalar.ph:
82; CHECK:         %scalar.recur.init = phi i32 [ %vector.recur.extract, %middle.block ], [ %.pre, %min.iters.checked ], [ %.pre, %for.preheader ]
83; CHECK:       scalar.body:
84; CHECK:         %scalar.recur = phi i32 [ %scalar.recur.init, %scalar.ph ], [ {{.*}}, %scalar.body ]
85;
86; UNROLL-LABEL: @recurrence_2(
87; UNROLL:       vector.body:
88; UNROLL:         %vector.recur = phi <4 x i32> [ %vector.recur.init, %vector.ph ], [ [[L2:%[a-zA-Z0-9.]+]], %vector.body ]
89; UNROLL:         [[L1:%[a-zA-Z0-9.]+]] = load <4 x i32>
90; UNROLL:         [[L2]] = load <4 x i32>
91; UNROLL:         {{.*}} = shufflevector <4 x i32> %vector.recur, <4 x i32> [[L1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
92; UNROLL:         {{.*}} = shufflevector <4 x i32> [[L1]], <4 x i32> [[L2]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
93; UNROLL:       middle.block:
94; UNROLL:         %vector.recur.extract = extractelement <4 x i32> [[L2]], i32 3
95;
96define i32 @recurrence_2(i32* nocapture readonly %a, i32 %n) {
97entry:
98  %cmp27 = icmp sgt i32 %n, 0
99  br i1 %cmp27, label %for.preheader, label %for.cond.cleanup
100
101for.preheader:
102  %arrayidx2.phi.trans.insert = getelementptr inbounds i32, i32* %a, i64 -1
103  %.pre = load i32, i32* %arrayidx2.phi.trans.insert, align 4
104  br label %scalar.body
105
106for.cond.cleanup.loopexit:
107  %minmax.0.cond.lcssa = phi i32 [ %minmax.0.cond, %scalar.body ]
108  br label %for.cond.cleanup
109
110for.cond.cleanup:
111  %minmax.0.lcssa = phi i32 [ undef, %entry ], [ %minmax.0.cond.lcssa, %for.cond.cleanup.loopexit ]
112  ret i32 %minmax.0.lcssa
113
114scalar.body:
115  %0 = phi i32 [ %.pre, %for.preheader ], [ %1, %scalar.body ]
116  %indvars.iv = phi i64 [ 0, %for.preheader ], [ %indvars.iv.next, %scalar.body ]
117  %minmax.028 = phi i32 [ undef, %for.preheader ], [ %minmax.0.cond, %scalar.body ]
118  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
119  %1 = load i32, i32* %arrayidx, align 4
120  %sub3 = sub nsw i32 %1, %0
121  %cmp4 = icmp sgt i32 %sub3, 0
122  %cond = select i1 %cmp4, i32 %sub3, i32 0
123  %cmp5 = icmp slt i32 %minmax.028, %cond
124  %minmax.0.cond = select i1 %cmp5, i32 %minmax.028, i32 %cond
125  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
126  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
127  %exitcond = icmp eq i32 %lftr.wideiv, %n
128  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %scalar.body
129}
130
131; void recurrence_3(short *a, double *b, int n, float f, short p) {
132;   b[0] = (double)a[0] - f * (double)p;
133;   for (int i = 1; i < n; i++)
134;     b[i] = (double)a[i] - f * (double)a[i - 1];
135; }
136;
137; CHECK-LABEL: @recurrence_3(
138; CHECK:       vector.ph:
139; CHECK:         %vector.recur.init = insertelement <4 x i16> undef, i16 %0, i32 3
140; CHECK:       vector.body:
141; CHECK:         %vector.recur = phi <4 x i16> [ %vector.recur.init, %vector.ph ], [ [[L1:%[a-zA-Z0-9.]+]], %vector.body ]
142; CHECK:         [[L1]] = load <4 x i16>
143; CHECK:         {{.*}} = shufflevector <4 x i16> %vector.recur, <4 x i16> [[L1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
144; CHECK:       middle.block:
145; CHECK:         %vector.recur.extract = extractelement <4 x i16> [[L1]], i32 3
146; CHECK:       scalar.ph:
147; CHECK:         %scalar.recur.init = phi i16 [ %vector.recur.extract, %middle.block ], [ %0, %vector.memcheck ], [ %0, %min.iters.checked ], [ %0, %for.preheader ]
148; CHECK:       scalar.body:
149; CHECK:         %scalar.recur = phi i16 [ %scalar.recur.init, %scalar.ph ], [ {{.*}}, %scalar.body ]
150;
151; UNROLL-LABEL: @recurrence_3(
152; UNROLL:       vector.body:
153; UNROLL:         %vector.recur = phi <4 x i16> [ %vector.recur.init, %vector.ph ], [ [[L2:%[a-zA-Z0-9.]+]], %vector.body ]
154; UNROLL:         [[L1:%[a-zA-Z0-9.]+]] = load <4 x i16>
155; UNROLL:         [[L2]] = load <4 x i16>
156; UNROLL:         {{.*}} = shufflevector <4 x i16> %vector.recur, <4 x i16> [[L1]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
157; UNROLL:         {{.*}} = shufflevector <4 x i16> [[L1]], <4 x i16> [[L2]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
158; UNROLL:       middle.block:
159; UNROLL:         %vector.recur.extract = extractelement <4 x i16> [[L2]], i32 3
160;
161define void @recurrence_3(i16* nocapture readonly %a, double* nocapture %b, i32 %n, float %f, i16 %p) {
162entry:
163  %0 = load i16, i16* %a, align 2
164  %conv = sitofp i16 %0 to double
165  %conv1 = fpext float %f to double
166  %conv2 = sitofp i16 %p to double
167  %mul = fmul fast double %conv2, %conv1
168  %sub = fsub fast double %conv, %mul
169  store double %sub, double* %b, align 8
170  %cmp25 = icmp sgt i32 %n, 1
171  br i1 %cmp25, label %for.preheader, label %for.end
172
173for.preheader:
174  br label %scalar.body
175
176scalar.body:
177  %1 = phi i16 [ %0, %for.preheader ], [ %2, %scalar.body ]
178  %advars.iv = phi i64 [ %advars.iv.next, %scalar.body ], [ 1, %for.preheader ]
179  %arrayidx5 = getelementptr inbounds i16, i16* %a, i64 %advars.iv
180  %2 = load i16, i16* %arrayidx5, align 2
181  %conv6 = sitofp i16 %2 to double
182  %conv11 = sitofp i16 %1 to double
183  %mul12 = fmul fast double %conv11, %conv1
184  %sub13 = fsub fast double %conv6, %mul12
185  %arrayidx15 = getelementptr inbounds double, double* %b, i64 %advars.iv
186  store double %sub13, double* %arrayidx15, align 8
187  %advars.iv.next = add nuw nsw i64 %advars.iv, 1
188  %lftr.wideiv = trunc i64 %advars.iv.next to i32
189  %exitcond = icmp eq i32 %lftr.wideiv, %n
190  br i1 %exitcond, label %for.end.loopexit, label %scalar.body
191
192for.end.loopexit:
193  br label %for.end
194
195for.end:
196  ret void
197}
198
199; void PR26734(short *a, int *b, int *c, int d, short *e) {
200;   for (; d != 21; d++) {
201;     *b &= *c;
202;     *e = *a - 6;
203;     *c = *e;
204;   }
205; }
206;
207; CHECK-LABEL: @PR26734(
208; CHECK-NOT:   vector.ph:
209; CHECK:       }
210;
211define void @PR26734(i16* %a, i32* %b, i32* %c, i32 %d, i16* %e) {
212entry:
213  %cmp4 = icmp eq i32 %d, 21
214  br i1 %cmp4, label %entry.for.end_crit_edge, label %for.body.lr.ph
215
216entry.for.end_crit_edge:
217  %.pre = load i32, i32* %b, align 4
218  br label %for.end
219
220for.body.lr.ph:
221  %0 = load i16, i16* %a, align 2
222  %sub = add i16 %0, -6
223  %conv2 = sext i16 %sub to i32
224  %c.promoted = load i32, i32* %c, align 4
225  %b.promoted = load i32, i32* %b, align 4
226  br label %for.body
227
228for.body:
229  %inc7 = phi i32 [ %d, %for.body.lr.ph ], [ %inc, %for.body ]
230  %and6 = phi i32 [ %b.promoted, %for.body.lr.ph ], [ %and, %for.body ]
231  %conv25 = phi i32 [ %c.promoted, %for.body.lr.ph ], [ %conv2, %for.body ]
232  %and = and i32 %and6, %conv25
233  %inc = add nsw i32 %inc7, 1
234  %cmp = icmp eq i32 %inc, 21
235  br i1 %cmp, label %for.cond.for.end_crit_edge, label %for.body
236
237for.cond.for.end_crit_edge:
238  %and.lcssa = phi i32 [ %and, %for.body ]
239  store i32 %conv2, i32* %c, align 4
240  store i32 %and.lcssa, i32* %b, align 4
241  store i16 %sub, i16* %e, align 2
242  br label %for.end
243
244for.end:
245  ret void
246}
247
248; int PR27246() {
249;   unsigned int e, n;
250;   for (int i = 1; i < 49; ++i) {
251;     for (int k = i; k > 1; --k)
252;       e = k;
253;     n = e;
254;   }
255;   return n;
256; }
257;
258; CHECK-LABEL: @PR27246(
259; CHECK-NOT:   vector.ph:
260; CHECK:       }
261;
262define i32 @PR27246() {
263entry:
264  br label %for.cond1.preheader
265
266for.cond1.preheader:
267  %i.016 = phi i32 [ 1, %entry ], [ %inc, %for.cond.cleanup3 ]
268  %e.015 = phi i32 [ undef, %entry ], [ %e.1.lcssa, %for.cond.cleanup3 ]
269  br label %for.cond1
270
271for.cond.cleanup:
272  %e.1.lcssa.lcssa = phi i32 [ %e.1.lcssa, %for.cond.cleanup3 ]
273  ret i32 %e.1.lcssa.lcssa
274
275for.cond1:
276  %e.1 = phi i32 [ %k.0, %for.cond1 ], [ %e.015, %for.cond1.preheader ]
277  %k.0 = phi i32 [ %dec, %for.cond1 ], [ %i.016, %for.cond1.preheader ]
278  %cmp2 = icmp sgt i32 %k.0, 1
279  %dec = add nsw i32 %k.0, -1
280  br i1 %cmp2, label %for.cond1, label %for.cond.cleanup3
281
282for.cond.cleanup3:
283  %e.1.lcssa = phi i32 [ %e.1, %for.cond1 ]
284  %inc = add nuw nsw i32 %i.016, 1
285  %exitcond = icmp eq i32 %inc, 49
286  br i1 %exitcond, label %for.cond.cleanup, label %for.cond1.preheader
287}
288
289; UNROLL-NO-IC-LABEL: @PR30183(
290; UNROLL-NO-IC:       vector.ph:
291; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> undef, i32 [[PRE_LOAD:%.*]], i32 3
292; UNROLL-NO-IC-NEXT:    br label %vector.body
293; UNROLL-NO-IC:       vector.body:
294; UNROLL-NO-IC-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
295; UNROLL-NO-IC-NEXT:    [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], %vector.ph ], [ [[TMP42:%.*]], %vector.body ]
296; UNROLL-NO-IC:         [[TMP27:%.*]] = load i32, i32* {{.*}}
297; UNROLL-NO-IC-NEXT:    [[TMP28:%.*]] = load i32, i32* {{.*}}
298; UNROLL-NO-IC-NEXT:    [[TMP29:%.*]] = load i32, i32* {{.*}}
299; UNROLL-NO-IC-NEXT:    [[TMP30:%.*]] = load i32, i32* {{.*}}
300; UNROLL-NO-IC-NEXT:    [[TMP35:%.*]] = insertelement <4 x i32> undef, i32 [[TMP27]], i32 0
301; UNROLL-NO-IC-NEXT:    [[TMP36:%.*]] = insertelement <4 x i32> [[TMP35]], i32 [[TMP28]], i32 1
302; UNROLL-NO-IC-NEXT:    [[TMP37:%.*]] = insertelement <4 x i32> [[TMP36]], i32 [[TMP29]], i32 2
303; UNROLL-NO-IC-NEXT:    [[TMP38:%.*]] = insertelement <4 x i32> [[TMP37]], i32 [[TMP30]], i32 3
304; UNROLL-NO-IC-NEXT:    [[TMP31:%.*]] = load i32, i32* {{.*}}
305; UNROLL-NO-IC-NEXT:    [[TMP32:%.*]] = load i32, i32* {{.*}}
306; UNROLL-NO-IC-NEXT:    [[TMP33:%.*]] = load i32, i32* {{.*}}
307; UNROLL-NO-IC-NEXT:    [[TMP34:%.*]] = load i32, i32* {{.*}}
308; UNROLL-NO-IC-NEXT:    [[TMP39:%.*]] = insertelement <4 x i32> undef, i32 [[TMP31]], i32 0
309; UNROLL-NO-IC-NEXT:    [[TMP40:%.*]] = insertelement <4 x i32> [[TMP39]], i32 [[TMP32]], i32 1
310; UNROLL-NO-IC-NEXT:    [[TMP41:%.*]] = insertelement <4 x i32> [[TMP40]], i32 [[TMP33]], i32 2
311; UNROLL-NO-IC-NEXT:    [[TMP42]] = insertelement <4 x i32> [[TMP41]], i32 [[TMP34]], i32 3
312; UNROLL-NO-IC-NEXT:    [[TMP43:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[TMP38]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
313; UNROLL-NO-IC-NEXT:    [[TMP44:%.*]] = shufflevector <4 x i32> [[TMP38]], <4 x i32> [[TMP42]], <4 x i32> <i32 3, i32 4, i32 5, i32 6>
314; UNROLL-NO-IC-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 8
315; UNROLL-NO-IC:         br i1 {{.*}}, label %middle.block, label %vector.body
316;
317define void @PR30183(i32 %pre_load, i32* %a, i32* %b, i64 %n) {
318entry:
319  br label %scalar.body
320
321scalar.body:
322  %i = phi i64 [ 0, %entry ], [ %i.next, %scalar.body ]
323  %tmp0 = phi i32 [ %pre_load, %entry ], [ %tmp2, %scalar.body ]
324  %i.next = add nuw nsw i64 %i, 2
325  %tmp1 = getelementptr inbounds i32, i32* %a, i64 %i.next
326  %tmp2 = load i32, i32* %tmp1
327  %cond = icmp eq i64 %i.next,%n
328  br i1 %cond, label %for.end, label %scalar.body
329
330for.end:
331  ret void
332}
333
334; UNROLL-NO-IC-LABEL: @constant_folded_previous_value(
335; UNROLL-NO-IC:       vector.body:
336; UNROLL-NO-IC:         [[VECTOR_RECUR:%.*]] = phi <4 x i64> [ <i64 undef, i64 undef, i64 undef, i64 0>, %vector.ph ], [ <i64 1, i64 1, i64 1, i64 1>, %vector.body ]
337; UNROLL-NO-IC-NEXT:    [[TMP0:%.*]] = shufflevector <4 x i64> [[VECTOR_RECUR]], <4 x i64> <i64 1, i64 1, i64 1, i64 1>, <4 x i32> <i32 3, i32 4, i32 5, i32 6>
338; UNROLL-NO-IC:         br i1 {{.*}}, label %middle.block, label %vector.body
339;
340define void @constant_folded_previous_value() {
341entry:
342  br label %scalar.body
343
344scalar.body:
345  %i = phi i64 [ 0, %entry ], [ %i.next, %scalar.body ]
346  %tmp2 = phi i64 [ 0, %entry ], [ %tmp3, %scalar.body ]
347  %tmp3 = add i64 0, 1
348  %i.next = add nuw nsw i64 %i, 1
349  %cond = icmp eq i64 %i.next, undef
350  br i1 %cond, label %for.end, label %scalar.body
351
352for.end:
353  ret void
354}
355
356; We vectorize this first order recurrence, by generating two
357; extracts for the phi `val.phi` - one at the last index and
358; another at the second last index. We need these 2 extracts because
359; the first order recurrence phi is used outside the loop, so we require the phi
360; itself and not its update (addx).
361; UNROLL-NO-IC-LABEL: extract_second_last_iteration
362; UNROLL-NO-IC: vector.body
363; UNROLL-NO-IC:   %step.add = add <4 x i32> %vec.ind, <i32 4, i32 4, i32 4, i32 4>
364; UNROLL-NO-IC:   %[[L1:.+]] = add <4 x i32> %vec.ind, %broadcast.splat
365; UNROLL-NO-IC:   %[[L2:.+]] = add <4 x i32> %step.add, %broadcast.splat
366; UNROLL-NO-IC:   %index.next = add i32 %index, 8
367; UNROLL-NO-IC:   icmp eq i32 %index.next, 96
368; UNROLL-NO-IC: middle.block
369; UNROLL-NO-IC:   icmp eq i32 96, 96
370; UNROLL-NO-IC:   %vector.recur.extract = extractelement <4 x i32> %[[L2]], i32 3
371; UNROLL-NO-IC:   %vector.recur.extract.for.phi = extractelement <4 x i32> %[[L2]], i32 2
372; UNROLL-NO-IC: for.end
373; UNROLL-NO-IC:   %val.phi.lcssa = phi i32 [ %scalar.recur, %for.body ], [ %vector.recur.extract.for.phi, %middle.block ]
374; Check the case when unrolled but not vectorized.
375; UNROLL-NO-VF-LABEL: extract_second_last_iteration
376; UNROLL-NO-VF: vector.body:
377; UNROLL-NO-VF:   %induction = add i32 %index, 0
378; UNROLL-NO-VF:   %induction1 = add i32 %index, 1
379; UNROLL-NO-VF:   %[[L1:.+]] = add i32 %induction, %x
380; UNROLL-NO-VF:   %[[L2:.+]] = add i32 %induction1, %x
381; UNROLL-NO-VF:   %index.next = add i32 %index, 2
382; UNROLL-NO-VF:   icmp eq i32 %index.next, 96
383; UNROLL-NO-VF: for.end:
384; UNROLL-NO-VF:   %val.phi.lcssa = phi i32 [ %scalar.recur, %for.body ], [ %[[L1]], %middle.block ]
385define i32 @extract_second_last_iteration(i32* %cval, i32 %x)  {
386entry:
387  br label %for.body
388
389for.body:
390  %inc.phi = phi i32 [ 0, %entry ], [ %inc, %for.body ]
391  %val.phi = phi i32 [ 0, %entry ], [ %addx, %for.body ]
392  %inc = add i32 %inc.phi, 1
393  %bc = zext i32 %inc.phi to i64
394  %addx = add i32 %inc.phi, %x
395  %cmp = icmp eq i32 %inc.phi, 95
396  br i1 %cmp, label %for.end, label %for.body
397
398for.end:
399  ret i32 %val.phi
400}
401
402; We vectorize this first order recurrence, with a set of insertelements for
403; each unrolled part. Make sure these insertelements are generated in-order,
404; because the shuffle of the first order recurrence will be added after the
405; insertelement of the last part UF - 1, assuming the latter appears after the
406; insertelements of all other parts.
407;
408; int PR33613(double *b, double j, int d) {
409;   int a = 0;
410;   for(int i = 0; i < 10240; i++, b+=25) {
411;     double f = b[d]; // Scalarize to form insertelements
412;     if (j * f)
413;       a++;
414;     j = f;
415;   }
416;   return a;
417; }
418;
419; UNROLL-NO-IC-LABEL: @PR33613(
420; UNROLL-NO-IC:     vector.body:
421; UNROLL-NO-IC:       [[VECTOR_RECUR:%.*]] = phi <4 x double>
422; UNROLL-NO-IC:       shufflevector <4 x double> [[VECTOR_RECUR]], <4 x double> {{.*}}, <4 x i32> <i32 3, i32 4, i32 5, i32 6>
423; UNROLL-NO-IC-NEXT:  shufflevector <4 x double> {{.*}}, <4 x double> {{.*}}, <4 x i32> <i32 3, i32 4, i32 5, i32 6>
424; UNROLL-NO-IC-NOT:   insertelement <4 x double>
425; UNROLL-NO-IC:     middle.block:
426;
427define i32 @PR33613(double* %b, double %j, i32 %d) {
428entry:
429  %idxprom = sext i32 %d to i64
430  br label %for.body
431
432for.cond.cleanup:
433  %a.1.lcssa = phi i32 [ %a.1, %for.body ]
434  ret i32 %a.1.lcssa
435
436for.body:
437  %b.addr.012 = phi double* [ %b, %entry ], [ %add.ptr, %for.body ]
438  %i.011 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]
439  %a.010 = phi i32 [ 0, %entry ], [ %a.1, %for.body ]
440  %j.addr.09 = phi double [ %j, %entry ], [ %0, %for.body ]
441  %arrayidx = getelementptr inbounds double, double* %b.addr.012, i64 %idxprom
442  %0 = load double, double* %arrayidx, align 8
443  %mul = fmul double %j.addr.09, %0
444  %tobool = fcmp une double %mul, 0.000000e+00
445  %inc = zext i1 %tobool to i32
446  %a.1 = add nsw i32 %a.010, %inc
447  %inc1 = add nuw nsw i32 %i.011, 1
448  %add.ptr = getelementptr inbounds double, double* %b.addr.012, i64 25
449  %exitcond = icmp eq i32 %inc1, 10240
450  br i1 %exitcond, label %for.cond.cleanup, label %for.body
451}
452
453; void sink_after(short *a, int n, int *b) {
454;   for(int i = 0; i < n; i++)
455;     b[i] = (a[i] * a[i + 1]);
456; }
457;
458; SINK-AFTER-LABEL: sink_after
459; Check that the sext sank after the load in the vector loop.
460; SINK-AFTER: vector.body
461; SINK-AFTER:   %vector.recur = phi <4 x i16> [ %vector.recur.init, %vector.ph ], [ %wide.load, %vector.body ]
462; SINK-AFTER:   %wide.load = load <4 x i16>
463; SINK-AFTER:   %[[VSHUF:.+]] = shufflevector <4 x i16> %vector.recur, <4 x i16> %wide.load, <4 x i32> <i32 3, i32 4, i32 5, i32 6>
464; SINK-AFTER:   %[[VCONV:.+]] = sext <4 x i16> %[[VSHUF]] to <4 x i32>
465; SINK-AFTER:   %[[VCONV3:.+]] = sext <4 x i16> %wide.load to <4 x i32>
466; SINK-AFTER:   mul nsw <4 x i32> %[[VCONV3]], %[[VCONV]]
467; Check also that the sext sank after the load in the scalar loop.
468; SINK-AFTER: for.body
469; SINK-AFTER:   %scalar.recur = phi i16 [ %scalar.recur.init, %scalar.ph ], [ %[[LOAD:.+]], %for.body ]
470; SINK-AFTER:   %[[LOAD]] = load i16, i16* %arrayidx2
471; SINK-AFTER:   %[[CONV:.+]] = sext i16 %scalar.recur to i32
472; SINK-AFTER:   %[[CONV3:.+]] = sext i16 %[[LOAD]] to i32
473; SINK-AFTER:   %mul = mul nsw i32 %[[CONV3]], %[[CONV]]
474;
475define void @sink_after(i16* %a, i32* %b, i64 %n) {
476entry:
477  %.pre = load i16, i16* %a
478  br label %for.body
479
480for.body:
481  %0 = phi i16 [ %.pre, %entry ], [ %1, %for.body ]
482  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
483  %conv = sext i16 %0 to i32
484  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
485  %arrayidx2 = getelementptr inbounds i16, i16* %a, i64 %indvars.iv.next
486  %1 = load i16, i16* %arrayidx2
487  %conv3 = sext i16 %1 to i32
488  %mul = mul nsw i32 %conv3, %conv
489  %arrayidx5 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
490  store i32 %mul, i32* %arrayidx5
491  %exitcond = icmp eq i64 %indvars.iv.next, %n
492  br i1 %exitcond, label %for.end, label %for.body
493
494for.end:
495  ret void
496}
497
498; void no_sink_after(short *a, int n, int *b) {
499;   for(int i = 0; i < n; i++)
500;     b[i] = ((a[i] + 2) * a[i + 1]);
501; }
502;
503; NO-SINK-AFTER-LABEL: no_sink_after
504; NO-SINK-AFTER-NOT:   vector.ph:
505; NO-SINK-AFTER:       }
506;
507define void @no_sink_after(i16* %a, i32* %b, i64 %n) {
508entry:
509  %.pre = load i16, i16* %a
510  br label %for.body
511
512for.body:
513  %0 = phi i16 [ %.pre, %entry ], [ %1, %for.body ]
514  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
515  %conv = sext i16 %0 to i32
516  %add = add nsw i32 %conv, 2
517  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
518  %arrayidx2 = getelementptr inbounds i16, i16* %a, i64 %indvars.iv.next
519  %1 = load i16, i16* %arrayidx2
520  %conv3 = sext i16 %1 to i32
521  %mul = mul nsw i32 %add, %conv3
522  %arrayidx5 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
523  store i32 %mul, i32* %arrayidx5
524  %exitcond = icmp eq i64 %indvars.iv.next, %n
525  br i1 %exitcond, label %for.end, label %for.body
526
527for.end:
528  ret void
529}
530