• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: opt < %s -loop-vectorize -force-vector-interleave=1 -force-vector-width=2 -S | FileCheck %s
2; RUN: opt < %s -loop-vectorize -force-vector-interleave=1 -force-vector-width=2 -instcombine -S | FileCheck %s --check-prefix=IND
3; RUN: opt < %s -loop-vectorize -force-vector-interleave=2 -force-vector-width=2 -instcombine -S | FileCheck %s --check-prefix=UNROLL
4; RUN: opt < %s -loop-vectorize -force-vector-interleave=2 -force-vector-width=2 -S | FileCheck %s --check-prefix=UNROLL-NO-IC
5; RUN: opt < %s -loop-vectorize -force-vector-interleave=2 -force-vector-width=4 -enable-interleaved-mem-accesses -instcombine -S | FileCheck %s --check-prefix=INTERLEAVE
6
7target 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"
8
9; Make sure that we can handle multiple integer induction variables.
10; CHECK-LABEL: @multi_int_induction(
11; CHECK: vector.body:
12; CHECK:  %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
13; CHECK:  %[[VAR:.*]] = trunc i64 %index to i32
14; CHECK:  %offset.idx = add i32 190, %[[VAR]]
15define void @multi_int_induction(i32* %A, i32 %N) {
16for.body.lr.ph:
17  br label %for.body
18
19for.body:
20  %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
21  %count.09 = phi i32 [ 190, %for.body.lr.ph ], [ %inc, %for.body ]
22  %arrayidx2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
23  store i32 %count.09, i32* %arrayidx2, align 4
24  %inc = add nsw i32 %count.09, 1
25  %indvars.iv.next = add i64 %indvars.iv, 1
26  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
27  %exitcond = icmp ne i32 %lftr.wideiv, %N
28  br i1 %exitcond, label %for.body, label %for.end
29
30for.end:
31  ret void
32}
33
34; Make sure we remove unneeded vectorization of induction variables.
35; In order for instcombine to cleanup the vectorized induction variables that we
36; create in the loop vectorizer we need to perform some form of redundancy
37; elimination to get rid of multiple uses.
38
39; IND-LABEL: scalar_use
40
41; IND:     br label %vector.body
42; IND:     vector.body:
43;   Vectorized induction variable.
44; IND-NOT:  insertelement <2 x i64>
45; IND-NOT:  shufflevector <2 x i64>
46; IND:     br {{.*}}, label %vector.body
47
48define void @scalar_use(float* %a, float %b, i64 %offset, i64 %offset2, i64 %n) {
49entry:
50  br label %for.body
51
52for.body:
53  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
54  %ind.sum = add i64 %iv, %offset
55  %arr.idx = getelementptr inbounds float, float* %a, i64 %ind.sum
56  %l1 = load float, float* %arr.idx, align 4
57  %ind.sum2 = add i64 %iv, %offset2
58  %arr.idx2 = getelementptr inbounds float, float* %a, i64 %ind.sum2
59  %l2 = load float, float* %arr.idx2, align 4
60  %m = fmul fast float %b, %l2
61  %ad = fadd fast float %l1, %m
62  store float %ad, float* %arr.idx, align 4
63  %iv.next = add nuw nsw i64 %iv, 1
64  %exitcond = icmp eq i64 %iv.next, %n
65  br i1 %exitcond, label %loopexit, label %for.body
66
67loopexit:
68  ret void
69}
70
71; Make sure we don't create a vector induction phi node that is unused.
72; Scalarize the step vectors instead.
73;
74; for (int i = 0; i < n; ++i)
75;   sum += a[i];
76;
77; CHECK-LABEL: @scalarize_induction_variable_01(
78; CHECK: vector.body:
79; CHECK:   %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
80; CHECK:   %[[i0:.+]] = add i64 %index, 0
81; CHECK:   %[[i1:.+]] = add i64 %index, 1
82; CHECK:   getelementptr inbounds i64, i64* %a, i64 %[[i0]]
83; CHECK:   getelementptr inbounds i64, i64* %a, i64 %[[i1]]
84;
85; UNROLL-NO-IC-LABEL: @scalarize_induction_variable_01(
86; UNROLL-NO-IC: vector.body:
87; UNROLL-NO-IC:   %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
88; UNROLL-NO-IC:   %[[i0:.+]] = add i64 %index, 0
89; UNROLL-NO-IC:   %[[i1:.+]] = add i64 %index, 1
90; UNROLL-NO-IC:   %[[i2:.+]] = add i64 %index, 2
91; UNROLL-NO-IC:   %[[i3:.+]] = add i64 %index, 3
92; UNROLL-NO-IC:   getelementptr inbounds i64, i64* %a, i64 %[[i0]]
93; UNROLL-NO-IC:   getelementptr inbounds i64, i64* %a, i64 %[[i1]]
94; UNROLL-NO-IC:   getelementptr inbounds i64, i64* %a, i64 %[[i2]]
95; UNROLL-NO-IC:   getelementptr inbounds i64, i64* %a, i64 %[[i3]]
96;
97; IND-LABEL: @scalarize_induction_variable_01(
98; IND:     vector.body:
99; IND:       %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
100; IND-NOT:   add i64 {{.*}}, 2
101; IND:       getelementptr inbounds i64, i64* %a, i64 %index
102;
103; UNROLL-LABEL: @scalarize_induction_variable_01(
104; UNROLL:     vector.body:
105; UNROLL:       %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
106; UNROLL-NOT:   add i64 {{.*}}, 4
107; UNROLL:       %[[g1:.+]] = getelementptr inbounds i64, i64* %a, i64 %index
108; UNROLL:       getelementptr i64, i64* %[[g1]], i64 2
109
110define i64 @scalarize_induction_variable_01(i64 *%a, i64 %n) {
111entry:
112  br label %for.body
113
114for.body:
115  %i = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
116  %sum = phi i64 [ %2, %for.body ], [ 0, %entry ]
117  %0 = getelementptr inbounds i64, i64* %a, i64 %i
118  %1 = load i64, i64* %0, align 8
119  %2 = add i64 %1, %sum
120  %i.next = add nuw nsw i64 %i, 1
121  %cond = icmp slt i64 %i.next, %n
122  br i1 %cond, label %for.body, label %for.end
123
124for.end:
125  %3  = phi i64 [ %2, %for.body ]
126  ret i64 %3
127}
128
129; Make sure we scalarize the step vectors used for the pointer arithmetic. We
130; can't easily simplify vectorized step vectors.
131;
132; float s = 0;
133; for (int i ; 0; i < n; i += 8)
134;   s += (a[i] + b[i] + 1.0f);
135;
136; CHECK-LABEL: @scalarize_induction_variable_02(
137; CHECK: vector.body:
138; CHECK:   %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
139; CHECK:   %offset.idx = shl i64 %index, 3
140; CHECK:   %[[i0:.+]] = add i64 %offset.idx, 0
141; CHECK:   %[[i1:.+]] = add i64 %offset.idx, 8
142; CHECK:   getelementptr inbounds float, float* %a, i64 %[[i0]]
143; CHECK:   getelementptr inbounds float, float* %a, i64 %[[i1]]
144; CHECK:   getelementptr inbounds float, float* %b, i64 %[[i0]]
145; CHECK:   getelementptr inbounds float, float* %b, i64 %[[i1]]
146;
147; UNROLL-NO-IC-LABEL: @scalarize_induction_variable_02(
148; UNROLL-NO-IC: vector.body:
149; UNROLL-NO-IC:   %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
150; UNROLL-NO-IC:   %offset.idx = shl i64 %index, 3
151; UNROLL-NO-IC:   %[[i0:.+]] = add i64 %offset.idx, 0
152; UNROLL-NO-IC:   %[[i1:.+]] = add i64 %offset.idx, 8
153; UNROLL-NO-IC:   %[[i2:.+]] = add i64 %offset.idx, 16
154; UNROLL-NO-IC:   %[[i3:.+]] = add i64 %offset.idx, 24
155; UNROLL-NO-IC:   getelementptr inbounds float, float* %a, i64 %[[i0]]
156; UNROLL-NO-IC:   getelementptr inbounds float, float* %a, i64 %[[i1]]
157; UNROLL-NO-IC:   getelementptr inbounds float, float* %a, i64 %[[i2]]
158; UNROLL-NO-IC:   getelementptr inbounds float, float* %a, i64 %[[i3]]
159; UNROLL-NO-IC:   getelementptr inbounds float, float* %b, i64 %[[i0]]
160; UNROLL-NO-IC:   getelementptr inbounds float, float* %b, i64 %[[i1]]
161; UNROLL-NO-IC:   getelementptr inbounds float, float* %b, i64 %[[i2]]
162; UNROLL-NO-IC:   getelementptr inbounds float, float* %b, i64 %[[i3]]
163;
164; IND-LABEL: @scalarize_induction_variable_02(
165; IND: vector.body:
166; IND:   %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
167; IND:   %[[i0:.+]] = shl i64 %index, 3
168; IND:   %[[i1:.+]] = or i64 %[[i0]], 8
169; IND:   getelementptr inbounds float, float* %a, i64 %[[i0]]
170; IND:   getelementptr inbounds float, float* %a, i64 %[[i1]]
171;
172; UNROLL-LABEL: @scalarize_induction_variable_02(
173; UNROLL: vector.body:
174; UNROLL:   %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
175; UNROLL:   %[[i0:.+]] = shl i64 %index, 3
176; UNROLL:   %[[i1:.+]] = or i64 %[[i0]], 8
177; UNROLL:   %[[i2:.+]] = or i64 %[[i0]], 16
178; UNROLL:   %[[i3:.+]] = or i64 %[[i0]], 24
179; UNROLL:   getelementptr inbounds float, float* %a, i64 %[[i0]]
180; UNROLL:   getelementptr inbounds float, float* %a, i64 %[[i1]]
181; UNROLL:   getelementptr inbounds float, float* %a, i64 %[[i2]]
182; UNROLL:   getelementptr inbounds float, float* %a, i64 %[[i3]]
183
184define float @scalarize_induction_variable_02(float* %a, float* %b, i64 %n) {
185entry:
186  br label %for.body
187
188for.body:
189  %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
190  %s = phi float [ 0.0, %entry ], [ %6, %for.body ]
191  %0 = getelementptr inbounds float, float* %a, i64 %i
192  %1 = load float, float* %0, align 4
193  %2 = getelementptr inbounds float, float* %b, i64 %i
194  %3 = load float, float* %2, align 4
195  %4 = fadd fast float %s, 1.0
196  %5 = fadd fast float %4, %1
197  %6 = fadd fast float %5, %3
198  %i.next = add nuw nsw i64 %i, 8
199  %cond = icmp slt i64 %i.next, %n
200  br i1 %cond, label %for.body, label %for.end
201
202for.end:
203  %s.lcssa = phi float [ %6, %for.body ]
204  ret float %s.lcssa
205}
206
207; Make sure we scalarize the step vectors used for the pointer arithmetic. We
208; can't easily simplify vectorized step vectors. (Interleaved accesses.)
209;
210; for (int i = 0; i < n; ++i)
211;   a[i].f ^= y;
212;
213; INTERLEAVE-LABEL: @scalarize_induction_variable_03(
214; INTERLEAVE: vector.body:
215; INTERLEAVE:   %[[i0:.+]] = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
216; INTERLEAVE:   %[[i1:.+]] = or i64 %[[i0]], 1
217; INTERLEAVE:   %[[i2:.+]] = or i64 %[[i0]], 2
218; INTERLEAVE:   %[[i3:.+]] = or i64 %[[i0]], 3
219; INTERLEAVE:   %[[i4:.+]] = or i64 %[[i0]], 4
220; INTERLEAVE:   %[[i5:.+]] = or i64 %[[i0]], 5
221; INTERLEAVE:   %[[i6:.+]] = or i64 %[[i0]], 6
222; INTERLEAVE:   %[[i7:.+]] = or i64 %[[i0]], 7
223; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i0]], i32 1
224; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i1]], i32 1
225; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i2]], i32 1
226; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i3]], i32 1
227; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i4]], i32 1
228; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i5]], i32 1
229; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i6]], i32 1
230; INTERLEAVE:   getelementptr inbounds %pair, %pair* %p, i64 %[[i7]], i32 1
231
232%pair = type { i32, i32 }
233define void @scalarize_induction_variable_03(%pair *%p, i32 %y, i64 %n) {
234entry:
235  br label %for.body
236
237for.body:
238  %i  = phi i64 [ %i.next, %for.body ], [ 0, %entry ]
239  %f = getelementptr inbounds %pair, %pair* %p, i64 %i, i32 1
240  %0 = load i32, i32* %f, align 8
241  %1 = xor i32 %0, %y
242  store i32 %1, i32* %f, align 8
243  %i.next = add nuw nsw i64 %i, 1
244  %cond = icmp slt i64 %i.next, %n
245  br i1 %cond, label %for.body, label %for.end
246
247for.end:
248  ret void
249}
250
251; Make sure that the loop exit count computation does not overflow for i8 and
252; i16. The exit count of these loops is i8/i16 max + 1. If we don't cast the
253; induction variable to a bigger type the exit count computation will overflow
254; to 0.
255; PR17532
256
257; CHECK-LABEL: i8_loop
258; CHECK: icmp eq i32 {{.*}}, 256
259define i32 @i8_loop() nounwind readnone ssp uwtable {
260  br label %1
261
262; <label>:1                                       ; preds = %1, %0
263  %a.0 = phi i32 [ 1, %0 ], [ %2, %1 ]
264  %b.0 = phi i8 [ 0, %0 ], [ %3, %1 ]
265  %2 = and i32 %a.0, 4
266  %3 = add i8 %b.0, -1
267  %4 = icmp eq i8 %3, 0
268  br i1 %4, label %5, label %1
269
270; <label>:5                                       ; preds = %1
271  ret i32 %2
272}
273
274; CHECK-LABEL: i16_loop
275; CHECK: icmp eq i32 {{.*}}, 65536
276
277define i32 @i16_loop() nounwind readnone ssp uwtable {
278  br label %1
279
280; <label>:1                                       ; preds = %1, %0
281  %a.0 = phi i32 [ 1, %0 ], [ %2, %1 ]
282  %b.0 = phi i16 [ 0, %0 ], [ %3, %1 ]
283  %2 = and i32 %a.0, 4
284  %3 = add i16 %b.0, -1
285  %4 = icmp eq i16 %3, 0
286  br i1 %4, label %5, label %1
287
288; <label>:5                                       ; preds = %1
289  ret i32 %2
290}
291
292; This loop has a backedge taken count of i32_max. We need to check for this
293; condition and branch directly to the scalar loop.
294
295; CHECK-LABEL: max_i32_backedgetaken
296; CHECK:  br i1 true, label %scalar.ph, label %min.iters.checked
297
298; CHECK: middle.block:
299; CHECK:  %[[v9:.+]] = extractelement <2 x i32> %bin.rdx, i32 0
300; CHECK: scalar.ph:
301; CHECK:  %bc.resume.val = phi i32 [ 0, %middle.block ], [ 0, %[[v0:.+]] ]
302; CHECK:  %bc.merge.rdx = phi i32 [ 1, %[[v0:.+]] ], [ 1, %min.iters.checked ], [ %[[v9]], %middle.block ]
303
304define i32 @max_i32_backedgetaken() nounwind readnone ssp uwtable {
305
306  br label %1
307
308; <label>:1                                       ; preds = %1, %0
309  %a.0 = phi i32 [ 1, %0 ], [ %2, %1 ]
310  %b.0 = phi i32 [ 0, %0 ], [ %3, %1 ]
311  %2 = and i32 %a.0, 4
312  %3 = add i32 %b.0, -1
313  %4 = icmp eq i32 %3, 0
314  br i1 %4, label %5, label %1
315
316; <label>:5                                       ; preds = %1
317  ret i32 %2
318}
319
320; When generating the overflow check we must sure that the induction start value
321; is defined before the branch to the scalar preheader.
322
323; CHECK-LABEL: testoverflowcheck
324; CHECK: entry
325; CHECK: %[[LOAD:.*]] = load i8
326; CHECK: br
327
328; CHECK: scalar.ph
329; CHECK: phi i8 [ %{{.*}}, %middle.block ], [ %[[LOAD]], %entry ]
330
331@e = global i8 1, align 1
332@d = common global i32 0, align 4
333@c = common global i32 0, align 4
334define i32 @testoverflowcheck() {
335entry:
336  %.pr.i = load i8, i8* @e, align 1
337  %0 = load i32, i32* @d, align 4
338  %c.promoted.i = load i32, i32* @c, align 4
339  br label %cond.end.i
340
341cond.end.i:
342  %inc4.i = phi i8 [ %.pr.i, %entry ], [ %inc.i, %cond.end.i ]
343  %and3.i = phi i32 [ %c.promoted.i, %entry ], [ %and.i, %cond.end.i ]
344  %and.i = and i32 %0, %and3.i
345  %inc.i = add i8 %inc4.i, 1
346  %tobool.i = icmp eq i8 %inc.i, 0
347  br i1 %tobool.i, label %loopexit, label %cond.end.i
348
349loopexit:
350  ret i32 %and.i
351}
352
353; The SCEV expression of %sphi is (zext i8 {%t,+,1}<%loop> to i32)
354; In order to recognize %sphi as an induction PHI and vectorize this loop,
355; we need to convert the SCEV expression into an AddRecExpr.
356; The expression gets converted to {zext i8 %t to i32,+,1}.
357
358; CHECK-LABEL: wrappingindvars1
359; CHECK-LABEL: vector.scevcheck
360; CHECK-LABEL: vector.ph
361; CHECK: %[[START:.*]] = add <2 x i32> %{{.*}}, <i32 0, i32 1>
362; CHECK-LABEL: vector.body
363; CHECK: %[[PHI:.*]] = phi <2 x i32> [ %[[START]], %vector.ph ], [ %[[STEP:.*]], %vector.body ]
364; CHECK: %[[STEP]] = add <2 x i32> %[[PHI]], <i32 2, i32 2>
365define void @wrappingindvars1(i8 %t, i32 %len, i32 *%A) {
366 entry:
367  %st = zext i8 %t to i16
368  %ext = zext i8 %t to i32
369  %ecmp = icmp ult i16 %st, 42
370  br i1 %ecmp, label %loop, label %exit
371
372 loop:
373
374  %idx = phi i8 [ %t, %entry ], [ %idx.inc, %loop ]
375  %idx.b = phi i32 [ 0, %entry ], [ %idx.b.inc, %loop ]
376  %sphi = phi i32 [ %ext, %entry ], [%idx.inc.ext, %loop]
377
378  %ptr = getelementptr inbounds i32, i32* %A, i8 %idx
379  store i32 %sphi, i32* %ptr
380
381  %idx.inc = add i8 %idx, 1
382  %idx.inc.ext = zext i8 %idx.inc to i32
383  %idx.b.inc = add nuw nsw i32 %idx.b, 1
384
385  %c = icmp ult i32 %idx.b, %len
386  br i1 %c, label %loop, label %exit
387
388 exit:
389  ret void
390}
391
392; The SCEV expression of %sphi is (4 * (zext i8 {%t,+,1}<%loop> to i32))
393; In order to recognize %sphi as an induction PHI and vectorize this loop,
394; we need to convert the SCEV expression into an AddRecExpr.
395; The expression gets converted to ({4 * (zext %t to i32),+,4}).
396; CHECK-LABEL: wrappingindvars2
397; CHECK-LABEL: vector.scevcheck
398; CHECK-LABEL: vector.ph
399; CHECK: %[[START:.*]] = add <2 x i32> %{{.*}}, <i32 0, i32 4>
400; CHECK-LABEL: vector.body
401; CHECK: %[[PHI:.*]] = phi <2 x i32> [ %[[START]], %vector.ph ], [ %[[STEP:.*]], %vector.body ]
402; CHECK: %[[STEP]] = add <2 x i32> %[[PHI]], <i32 8, i32 8>
403define void @wrappingindvars2(i8 %t, i32 %len, i32 *%A) {
404
405entry:
406  %st = zext i8 %t to i16
407  %ext = zext i8 %t to i32
408  %ext.mul = mul i32 %ext, 4
409
410  %ecmp = icmp ult i16 %st, 42
411  br i1 %ecmp, label %loop, label %exit
412
413 loop:
414
415  %idx = phi i8 [ %t, %entry ], [ %idx.inc, %loop ]
416  %sphi = phi i32 [ %ext.mul, %entry ], [%mul, %loop]
417  %idx.b = phi i32 [ 0, %entry ], [ %idx.b.inc, %loop ]
418
419  %ptr = getelementptr inbounds i32, i32* %A, i8 %idx
420  store i32 %sphi, i32* %ptr
421
422  %idx.inc = add i8 %idx, 1
423  %idx.inc.ext = zext i8 %idx.inc to i32
424  %mul = mul i32 %idx.inc.ext, 4
425  %idx.b.inc = add nuw nsw i32 %idx.b, 1
426
427  %c = icmp ult i32 %idx.b, %len
428  br i1 %c, label %loop, label %exit
429
430 exit:
431  ret void
432}
433
434; Check that we generate vectorized IVs in the pre-header
435; instead of widening the scalar IV inside the loop, when
436; we know how to do that.
437; IND-LABEL: veciv
438; IND: vector.body:
439; IND: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
440; IND: %vec.ind = phi <2 x i32> [ <i32 0, i32 1>, %vector.ph ], [ %step.add, %vector.body ]
441; IND: %step.add = add <2 x i32> %vec.ind, <i32 2, i32 2>
442; IND: %index.next = add i32 %index, 2
443; IND: %[[CMP:.*]] = icmp eq i32 %index.next
444; IND: br i1 %[[CMP]]
445; UNROLL-LABEL: veciv
446; UNROLL: vector.body:
447; UNROLL: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
448; UNROLL: %vec.ind = phi <2 x i32> [ <i32 0, i32 1>, %vector.ph ], [ %step.add1, %vector.body ]
449; UNROLL: %step.add = add <2 x i32> %vec.ind, <i32 2, i32 2>
450; UNROLL: %step.add1 = add <2 x i32> %vec.ind, <i32 4, i32 4>
451; UNROLL: %index.next = add i32 %index, 4
452; UNROLL: %[[CMP:.*]] = icmp eq i32 %index.next
453; UNROLL: br i1 %[[CMP]]
454define void @veciv(i32* nocapture %a, i32 %start, i32 %k) {
455for.body.preheader:
456  br label %for.body
457
458for.body:
459  %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
460  %arrayidx = getelementptr inbounds i32, i32* %a, i32 %indvars.iv
461  store i32 %indvars.iv, i32* %arrayidx, align 4
462  %indvars.iv.next = add nuw nsw i32 %indvars.iv, 1
463  %exitcond = icmp eq i32 %indvars.iv.next, %k
464  br i1 %exitcond, label %exit, label %for.body
465
466exit:
467  ret void
468}
469
470; IND-LABEL: trunciv
471; IND: vector.body:
472; IND: %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
473; IND: %[[VECIND:.*]] = phi <2 x i32> [ <i32 0, i32 1>, %vector.ph ], [ %[[STEPADD:.*]], %vector.body ]
474; IND: %[[STEPADD]] = add <2 x i32> %[[VECIND]], <i32 2, i32 2>
475; IND: %index.next = add i64 %index, 2
476; IND: %[[CMP:.*]] = icmp eq i64 %index.next
477; IND: br i1 %[[CMP]]
478define void @trunciv(i32* nocapture %a, i32 %start, i64 %k) {
479for.body.preheader:
480  br label %for.body
481
482for.body:
483  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
484  %trunc.iv = trunc i64 %indvars.iv to i32
485  %arrayidx = getelementptr inbounds i32, i32* %a, i32 %trunc.iv
486  store i32 %trunc.iv, i32* %arrayidx, align 4
487  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
488  %exitcond = icmp eq i64 %indvars.iv.next, %k
489  br i1 %exitcond, label %exit, label %for.body
490
491exit:
492  ret void
493}
494
495; IND-LABEL: nonprimary
496; IND-LABEL: vector.ph
497; IND: %[[INSERT:.*]] = insertelement <2 x i32> undef, i32 %i, i32 0
498; IND: %[[SPLAT:.*]] = shufflevector <2 x i32> %[[INSERT]], <2 x i32> undef, <2 x i32> zeroinitializer
499; IND: %[[START:.*]] = add <2 x i32> %[[SPLAT]], <i32 0, i32 42>
500; IND-LABEL: vector.body:
501; IND: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
502; IND: %vec.ind = phi <2 x i32> [ %[[START]], %vector.ph ], [ %step.add, %vector.body ]
503; IND: %step.add = add <2 x i32> %vec.ind, <i32 84, i32 84>
504; IND: %index.next = add i32 %index, 2
505; IND: %[[CMP:.*]] = icmp eq i32 %index.next
506; IND: br i1 %[[CMP]]
507; UNROLL-LABEL: nonprimary
508; UNROLL-LABEL: vector.ph
509; UNROLL: %[[INSERT:.*]] = insertelement <2 x i32> undef, i32 %i, i32 0
510; UNROLL: %[[SPLAT:.*]] = shufflevector <2 x i32> %[[INSERT]], <2 x i32> undef, <2 x i32> zeroinitializer
511; UNROLL: %[[START:.*]] = add <2 x i32> %[[SPLAT]], <i32 0, i32 42>
512; UNROLL-LABEL: vector.body:
513; UNROLL: %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
514; UNROLL: %vec.ind = phi <2 x i32> [ %[[START]], %vector.ph ], [ %step.add1, %vector.body ]
515; UNROLL: %step.add = add <2 x i32> %vec.ind, <i32 84, i32 84>
516; UNROLL: %step.add1 = add <2 x i32> %vec.ind, <i32 168, i32 168>
517; UNROLL: %index.next = add i32 %index, 4
518; UNROLL: %[[CMP:.*]] = icmp eq i32 %index.next
519; UNROLL: br i1 %[[CMP]]
520define void @nonprimary(i32* nocapture %a, i32 %start, i32 %i, i32 %k) {
521for.body.preheader:
522  br label %for.body
523
524for.body:
525  %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ %i, %for.body.preheader ]
526  %arrayidx = getelementptr inbounds i32, i32* %a, i32 %indvars.iv
527  store i32 %indvars.iv, i32* %arrayidx, align 4
528  %indvars.iv.next = add nuw nsw i32 %indvars.iv, 42
529  %exitcond = icmp eq i32 %indvars.iv.next, %k
530  br i1 %exitcond, label %exit, label %for.body
531
532exit:
533  ret void
534}
535