Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / deb72ce29860f61fe91ddcf97e89abfc9544cf42 / . / llvm / test / CodeGen / Thumb2 / mve-vecreduce-loops.ll
blob: 29e441e3e90cf99868ef6bfb62d6f0c5ea8a3982 [file] [log] [blame]
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp -verify-machineinstrs %s -o - | FileCheck %s
3
4define i32 @add_i32(i32* nocapture readonly %x, i32 %n) {
5; CHECK-LABEL: add_i32:
6; CHECK: @ %bb.0: @ %entry
7; CHECK-NEXT: .save {r7, lr}
8; CHECK-NEXT: push {r7, lr}
9; CHECK-NEXT: cmp r1, #1
10; CHECK-NEXT: blt .LBB0_3
11; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
12; CHECK-NEXT: mov r12, r0
13; CHECK-NEXT: cmp r1, #4
14; CHECK-NEXT: bhs .LBB0_4
15; CHECK-NEXT: @ %bb.2:
16; CHECK-NEXT: movs r3, #0
17; CHECK-NEXT: movs r0, #0
18; CHECK-NEXT: b .LBB0_7
19; CHECK-NEXT: .LBB0_3:
20; CHECK-NEXT: movs r0, #0
21; CHECK-NEXT: b .LBB0_9
22; CHECK-NEXT: .LBB0_4: @ %vector.ph
23; CHECK-NEXT: bic r3, r1, #3
24; CHECK-NEXT: movs r2, #1
25; CHECK-NEXT: subs r0, r3, #4
26; CHECK-NEXT: add.w lr, r2, r0, lsr #2
27; CHECK-NEXT: movs r0, #0
28; CHECK-NEXT: mov r2, r12
29; CHECK-NEXT: dls lr, lr
30; CHECK-NEXT: .LBB0_5: @ %vector.body
31; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
32; CHECK-NEXT: vldrw.u32 q0, [r2], #16
33; CHECK-NEXT: vaddva.u32 r0, q0
34; CHECK-NEXT: le lr, .LBB0_5
35; CHECK-NEXT: @ %bb.6: @ %middle.block
36; CHECK-NEXT: cmp r3, r1
37; CHECK-NEXT: it eq
38; CHECK-NEXT: popeq {r7, pc}
39; CHECK-NEXT: .LBB0_7: @ %for.body.preheader1
40; CHECK-NEXT: sub.w lr, r1, r3
41; CHECK-NEXT: add.w r1, r12, r3, lsl #2
42; CHECK-NEXT: dls lr, lr
43; CHECK-NEXT: .LBB0_8: @ %for.body
44; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
45; CHECK-NEXT: ldr r2, [r1], #4
46; CHECK-NEXT: add r0, r2
47; CHECK-NEXT: le lr, .LBB0_8
48; CHECK-NEXT: .LBB0_9: @ %for.cond.cleanup
49; CHECK-NEXT: pop {r7, pc}
50entry:
51 %cmp6 = icmp sgt i32 %n, 0
52 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
53
54for.body.preheader: ; preds = %entry
55 %min.iters.check = icmp ult i32 %n, 4
56 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
57
58vector.ph: ; preds = %for.body.preheader
59 %n.vec = and i32 %n, -4
60 br label %vector.body
61
62vector.body: ; preds = %vector.body, %vector.ph
63 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
64 %vec.phi = phi i32 [ 0, %vector.ph ], [ %3, %vector.body ]
65 %0 = getelementptr inbounds i32, i32* %x, i32 %index
66 %1 = bitcast i32* %0 to <4 x i32>*
67 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
68 %2 = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> %wide.load)
69 %3 = add i32 %2, %vec.phi
70 %index.next = add i32 %index, 4
71 %4 = icmp eq i32 %index.next, %n.vec
72 br i1 %4, label %middle.block, label %vector.body
73
74middle.block: ; preds = %vector.body
75 %cmp.n = icmp eq i32 %n.vec, %n
76 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
77
78for.body.preheader1: ; preds = %middle.block, %for.body.preheader
79 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
80 %r.07.ph = phi i32 [ 0, %for.body.preheader ], [ %3, %middle.block ]
81 br label %for.body
82
83for.body: ; preds = %for.body.preheader1, %for.body
84 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
85 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
86 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
87 %5 = load i32, i32* %arrayidx, align 4
88 %add = add nsw i32 %5, %r.07
89 %inc = add nuw nsw i32 %i.08, 1
90 %exitcond = icmp eq i32 %inc, %n
91 br i1 %exitcond, label %for.cond.cleanup, label %for.body
92
93for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
94 %r.0.lcssa = phi i32 [ 0, %entry ], [ %3, %middle.block ], [ %add, %for.body ]
95 ret i32 %r.0.lcssa
96}
97
98define i32 @mul_i32(i32* nocapture readonly %x, i32 %n) {
99; CHECK-LABEL: mul_i32:
100; CHECK: @ %bb.0: @ %entry
101; CHECK-NEXT: .save {r7, lr}
102; CHECK-NEXT: push {r7, lr}
103; CHECK-NEXT: movs r2, #1
104; CHECK-NEXT: cmp r1, #1
105; CHECK-NEXT: blt .LBB1_8
106; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
107; CHECK-NEXT: cmp r1, #4
108; CHECK-NEXT: bhs .LBB1_3
109; CHECK-NEXT: @ %bb.2:
110; CHECK-NEXT: mov.w r12, #0
111; CHECK-NEXT: b .LBB1_6
112; CHECK-NEXT: .LBB1_3: @ %vector.ph
113; CHECK-NEXT: bic r12, r1, #3
114; CHECK-NEXT: vmov.i32 q0, #0x1
115; CHECK-NEXT: sub.w r3, r12, #4
116; CHECK-NEXT: add.w lr, r2, r3, lsr #2
117; CHECK-NEXT: mov r2, r0
118; CHECK-NEXT: dls lr, lr
119; CHECK-NEXT: .LBB1_4: @ %vector.body
120; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
121; CHECK-NEXT: vldrw.u32 q1, [r2], #16
122; CHECK-NEXT: vmul.i32 q0, q1, q0
123; CHECK-NEXT: le lr, .LBB1_4
124; CHECK-NEXT: @ %bb.5: @ %middle.block
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]125; CHECK-NEXT: vmov r2, s3
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]126; CHECK-NEXT: cmp r12, r1
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]127; CHECK-NEXT: vmov r3, s2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]128; CHECK-NEXT: mul lr, r3, r2
129; CHECK-NEXT: vmov r3, s1
130; CHECK-NEXT: vmov r2, s0
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]131; CHECK-NEXT: mul r2, r3, r2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]132; CHECK-NEXT: mul r2, r2, lr
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]133; CHECK-NEXT: beq .LBB1_8
134; CHECK-NEXT: .LBB1_6: @ %for.body.preheader1
135; CHECK-NEXT: sub.w lr, r1, r12
136; CHECK-NEXT: add.w r0, r0, r12, lsl #2
137; CHECK-NEXT: dls lr, lr
138; CHECK-NEXT: .LBB1_7: @ %for.body
139; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
140; CHECK-NEXT: ldr r1, [r0], #4
141; CHECK-NEXT: muls r2, r1, r2
142; CHECK-NEXT: le lr, .LBB1_7
143; CHECK-NEXT: .LBB1_8: @ %for.cond.cleanup
144; CHECK-NEXT: mov r0, r2
145; CHECK-NEXT: pop {r7, pc}
146entry:
147 %cmp6 = icmp sgt i32 %n, 0
148 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
149
150for.body.preheader: ; preds = %entry
151 %min.iters.check = icmp ult i32 %n, 4
152 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
153
154vector.ph: ; preds = %for.body.preheader
155 %n.vec = and i32 %n, -4
156 br label %vector.body
157
158vector.body: ; preds = %vector.body, %vector.ph
159 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
160 %vec.phi = phi <4 x i32> [ <i32 1, i32 1, i32 1, i32 1>, %vector.ph ], [ %2, %vector.body ]
161 %0 = getelementptr inbounds i32, i32* %x, i32 %index
162 %1 = bitcast i32* %0 to <4 x i32>*
163 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
164 %2 = mul <4 x i32> %wide.load, %vec.phi
165 %index.next = add i32 %index, 4
166 %3 = icmp eq i32 %index.next, %n.vec
167 br i1 %3, label %middle.block, label %vector.body
168
169middle.block: ; preds = %vector.body
170 %4 = call i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32> %2)
171 %cmp.n = icmp eq i32 %n.vec, %n
172 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
173
174for.body.preheader1: ; preds = %middle.block, %for.body.preheader
175 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
176 %r.07.ph = phi i32 [ 1, %for.body.preheader ], [ %4, %middle.block ]
177 br label %for.body
178
179for.body: ; preds = %for.body.preheader1, %for.body
180 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
181 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
182 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
183 %5 = load i32, i32* %arrayidx, align 4
184 %add = mul nsw i32 %5, %r.07
185 %inc = add nuw nsw i32 %i.08, 1
186 %exitcond = icmp eq i32 %inc, %n
187 br i1 %exitcond, label %for.cond.cleanup, label %for.body
188
189for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
190 %r.0.lcssa = phi i32 [ 1, %entry ], [ %4, %middle.block ], [ %add, %for.body ]
191 ret i32 %r.0.lcssa
192}
193
194define i32 @and_i32(i32* nocapture readonly %x, i32 %n) {
195; CHECK-LABEL: and_i32:
196; CHECK: @ %bb.0: @ %entry
197; CHECK-NEXT: .save {r7, lr}
198; CHECK-NEXT: push {r7, lr}
199; CHECK-NEXT: cmp r1, #1
200; CHECK-NEXT: blt .LBB2_3
201; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
202; CHECK-NEXT: cmp r1, #4
203; CHECK-NEXT: bhs .LBB2_4
204; CHECK-NEXT: @ %bb.2:
205; CHECK-NEXT: mov.w r2, #-1
206; CHECK-NEXT: movs r3, #0
207; CHECK-NEXT: b .LBB2_7
208; CHECK-NEXT: .LBB2_3:
209; CHECK-NEXT: mov.w r2, #-1
210; CHECK-NEXT: b .LBB2_9
211; CHECK-NEXT: .LBB2_4: @ %vector.ph
212; CHECK-NEXT: bic r3, r1, #3
213; CHECK-NEXT: movs r2, #1
214; CHECK-NEXT: sub.w r12, r3, #4
215; CHECK-NEXT: vmov.i8 q0, #0xff
216; CHECK-NEXT: add.w lr, r2, r12, lsr #2
217; CHECK-NEXT: mov r2, r0
218; CHECK-NEXT: dls lr, lr
219; CHECK-NEXT: .LBB2_5: @ %vector.body
220; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
221; CHECK-NEXT: vldrw.u32 q1, [r2], #16
222; CHECK-NEXT: vand q0, q1, q0
223; CHECK-NEXT: le lr, .LBB2_5
224; CHECK-NEXT: @ %bb.6: @ %middle.block
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]225; CHECK-NEXT: vmov r12, s3
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]226; CHECK-NEXT: cmp r3, r1
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]227; CHECK-NEXT: vmov r2, s2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]228; CHECK-NEXT: vmov lr, s1
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]229; CHECK-NEXT: and.w r12, r12, r2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]230; CHECK-NEXT: vmov r2, s0
231; CHECK-NEXT: and.w r2, r2, lr
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]232; CHECK-NEXT: and.w r2, r2, r12
233; CHECK-NEXT: beq .LBB2_9
234; CHECK-NEXT: .LBB2_7: @ %for.body.preheader1
235; CHECK-NEXT: sub.w lr, r1, r3
236; CHECK-NEXT: add.w r0, r0, r3, lsl #2
237; CHECK-NEXT: dls lr, lr
238; CHECK-NEXT: .LBB2_8: @ %for.body
239; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
240; CHECK-NEXT: ldr r1, [r0], #4
241; CHECK-NEXT: ands r2, r1
242; CHECK-NEXT: le lr, .LBB2_8
243; CHECK-NEXT: .LBB2_9: @ %for.cond.cleanup
244; CHECK-NEXT: mov r0, r2
245; CHECK-NEXT: pop {r7, pc}
246entry:
247 %cmp6 = icmp sgt i32 %n, 0
248 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
249
250for.body.preheader: ; preds = %entry
251 %min.iters.check = icmp ult i32 %n, 4
252 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
253
254vector.ph: ; preds = %for.body.preheader
255 %n.vec = and i32 %n, -4
256 br label %vector.body
257
258vector.body: ; preds = %vector.body, %vector.ph
259 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
260 %vec.phi = phi <4 x i32> [ <i32 -1, i32 -1, i32 -1, i32 -1>, %vector.ph ], [ %2, %vector.body ]
261 %0 = getelementptr inbounds i32, i32* %x, i32 %index
262 %1 = bitcast i32* %0 to <4 x i32>*
263 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
264 %2 = and <4 x i32> %wide.load, %vec.phi
265 %index.next = add i32 %index, 4
266 %3 = icmp eq i32 %index.next, %n.vec
267 br i1 %3, label %middle.block, label %vector.body
268
269middle.block: ; preds = %vector.body
270 %4 = call i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32> %2)
271 %cmp.n = icmp eq i32 %n.vec, %n
272 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
273
274for.body.preheader1: ; preds = %middle.block, %for.body.preheader
275 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
276 %r.07.ph = phi i32 [ -1, %for.body.preheader ], [ %4, %middle.block ]
277 br label %for.body
278
279for.body: ; preds = %for.body.preheader1, %for.body
280 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
281 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
282 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
283 %5 = load i32, i32* %arrayidx, align 4
284 %add = and i32 %5, %r.07
285 %inc = add nuw nsw i32 %i.08, 1
286 %exitcond = icmp eq i32 %inc, %n
287 br i1 %exitcond, label %for.cond.cleanup, label %for.body
288
289for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
290 %r.0.lcssa = phi i32 [ -1, %entry ], [ %4, %middle.block ], [ %add, %for.body ]
291 ret i32 %r.0.lcssa
292}
293
294define i32 @or_i32(i32* nocapture readonly %x, i32 %n) {
295; CHECK-LABEL: or_i32:
296; CHECK: @ %bb.0: @ %entry
297; CHECK-NEXT: .save {r7, lr}
298; CHECK-NEXT: push {r7, lr}
299; CHECK-NEXT: cmp r1, #1
300; CHECK-NEXT: blt .LBB3_3
301; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
302; CHECK-NEXT: cmp r1, #4
303; CHECK-NEXT: bhs .LBB3_4
304; CHECK-NEXT: @ %bb.2:
305; CHECK-NEXT: movs r3, #0
306; CHECK-NEXT: movs r2, #0
307; CHECK-NEXT: b .LBB3_7
308; CHECK-NEXT: .LBB3_3:
309; CHECK-NEXT: movs r2, #0
310; CHECK-NEXT: b .LBB3_9
311; CHECK-NEXT: .LBB3_4: @ %vector.ph
312; CHECK-NEXT: bic r3, r1, #3
313; CHECK-NEXT: movs r2, #1
314; CHECK-NEXT: sub.w r12, r3, #4
315; CHECK-NEXT: vmov.i32 q0, #0x0
316; CHECK-NEXT: add.w lr, r2, r12, lsr #2
317; CHECK-NEXT: mov r2, r0
318; CHECK-NEXT: dls lr, lr
319; CHECK-NEXT: .LBB3_5: @ %vector.body
320; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
321; CHECK-NEXT: vldrw.u32 q1, [r2], #16
322; CHECK-NEXT: vorr q0, q1, q0
323; CHECK-NEXT: le lr, .LBB3_5
324; CHECK-NEXT: @ %bb.6: @ %middle.block
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]325; CHECK-NEXT: vmov r12, s3
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]326; CHECK-NEXT: cmp r3, r1
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]327; CHECK-NEXT: vmov r2, s2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]328; CHECK-NEXT: vmov lr, s1
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]329; CHECK-NEXT: orr.w r12, r12, r2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]330; CHECK-NEXT: vmov r2, s0
331; CHECK-NEXT: orr.w r2, r2, lr
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]332; CHECK-NEXT: orr.w r2, r2, r12
333; CHECK-NEXT: beq .LBB3_9
334; CHECK-NEXT: .LBB3_7: @ %for.body.preheader1
335; CHECK-NEXT: sub.w lr, r1, r3
336; CHECK-NEXT: add.w r0, r0, r3, lsl #2
337; CHECK-NEXT: dls lr, lr
338; CHECK-NEXT: .LBB3_8: @ %for.body
339; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
340; CHECK-NEXT: ldr r1, [r0], #4
341; CHECK-NEXT: orrs r2, r1
342; CHECK-NEXT: le lr, .LBB3_8
343; CHECK-NEXT: .LBB3_9: @ %for.cond.cleanup
344; CHECK-NEXT: mov r0, r2
345; CHECK-NEXT: pop {r7, pc}
346entry:
347 %cmp6 = icmp sgt i32 %n, 0
348 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
349
350for.body.preheader: ; preds = %entry
351 %min.iters.check = icmp ult i32 %n, 4
352 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
353
354vector.ph: ; preds = %for.body.preheader
355 %n.vec = and i32 %n, -4
356 br label %vector.body
357
358vector.body: ; preds = %vector.body, %vector.ph
359 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
360 %vec.phi = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %2, %vector.body ]
361 %0 = getelementptr inbounds i32, i32* %x, i32 %index
362 %1 = bitcast i32* %0 to <4 x i32>*
363 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
364 %2 = or <4 x i32> %wide.load, %vec.phi
365 %index.next = add i32 %index, 4
366 %3 = icmp eq i32 %index.next, %n.vec
367 br i1 %3, label %middle.block, label %vector.body
368
369middle.block: ; preds = %vector.body
370 %4 = call i32 @llvm.experimental.vector.reduce.or.v4i32(<4 x i32> %2)
371 %cmp.n = icmp eq i32 %n.vec, %n
372 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
373
374for.body.preheader1: ; preds = %middle.block, %for.body.preheader
375 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
376 %r.07.ph = phi i32 [ 0, %for.body.preheader ], [ %4, %middle.block ]
377 br label %for.body
378
379for.body: ; preds = %for.body.preheader1, %for.body
380 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
381 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
382 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
383 %5 = load i32, i32* %arrayidx, align 4
384 %add = or i32 %5, %r.07
385 %inc = add nuw nsw i32 %i.08, 1
386 %exitcond = icmp eq i32 %inc, %n
387 br i1 %exitcond, label %for.cond.cleanup, label %for.body
388
389for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
390 %r.0.lcssa = phi i32 [ 0, %entry ], [ %4, %middle.block ], [ %add, %for.body ]
391 ret i32 %r.0.lcssa
392}
393
394define i32 @xor_i32(i32* nocapture readonly %x, i32 %n) {
395; CHECK-LABEL: xor_i32:
396; CHECK: @ %bb.0: @ %entry
397; CHECK-NEXT: .save {r7, lr}
398; CHECK-NEXT: push {r7, lr}
399; CHECK-NEXT: cmp r1, #1
400; CHECK-NEXT: blt .LBB4_3
401; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
402; CHECK-NEXT: cmp r1, #4
403; CHECK-NEXT: bhs .LBB4_4
404; CHECK-NEXT: @ %bb.2:
405; CHECK-NEXT: movs r3, #0
406; CHECK-NEXT: movs r2, #0
407; CHECK-NEXT: b .LBB4_7
408; CHECK-NEXT: .LBB4_3:
409; CHECK-NEXT: movs r2, #0
410; CHECK-NEXT: b .LBB4_9
411; CHECK-NEXT: .LBB4_4: @ %vector.ph
412; CHECK-NEXT: bic r3, r1, #3
413; CHECK-NEXT: movs r2, #1
414; CHECK-NEXT: sub.w r12, r3, #4
415; CHECK-NEXT: vmov.i32 q0, #0x0
416; CHECK-NEXT: add.w lr, r2, r12, lsr #2
417; CHECK-NEXT: mov r2, r0
418; CHECK-NEXT: dls lr, lr
419; CHECK-NEXT: .LBB4_5: @ %vector.body
420; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
421; CHECK-NEXT: vldrw.u32 q1, [r2], #16
422; CHECK-NEXT: veor q0, q1, q0
423; CHECK-NEXT: le lr, .LBB4_5
424; CHECK-NEXT: @ %bb.6: @ %middle.block
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]425; CHECK-NEXT: vmov r12, s3
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]426; CHECK-NEXT: cmp r3, r1
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]427; CHECK-NEXT: vmov r2, s2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]428; CHECK-NEXT: vmov lr, s1
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]429; CHECK-NEXT: eor.w r12, r12, r2
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]430; CHECK-NEXT: vmov r2, s0
431; CHECK-NEXT: eor.w r2, r2, lr
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]432; CHECK-NEXT: eor.w r2, r2, r12
433; CHECK-NEXT: beq .LBB4_9
434; CHECK-NEXT: .LBB4_7: @ %for.body.preheader1
435; CHECK-NEXT: sub.w lr, r1, r3
436; CHECK-NEXT: add.w r0, r0, r3, lsl #2
437; CHECK-NEXT: dls lr, lr
438; CHECK-NEXT: .LBB4_8: @ %for.body
439; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
440; CHECK-NEXT: ldr r1, [r0], #4
441; CHECK-NEXT: eors r2, r1
442; CHECK-NEXT: le lr, .LBB4_8
443; CHECK-NEXT: .LBB4_9: @ %for.cond.cleanup
444; CHECK-NEXT: mov r0, r2
445; CHECK-NEXT: pop {r7, pc}
446entry:
447 %cmp6 = icmp sgt i32 %n, 0
448 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
449
450for.body.preheader: ; preds = %entry
451 %min.iters.check = icmp ult i32 %n, 4
452 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
453
454vector.ph: ; preds = %for.body.preheader
455 %n.vec = and i32 %n, -4
456 br label %vector.body
457
458vector.body: ; preds = %vector.body, %vector.ph
459 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
460 %vec.phi = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %2, %vector.body ]
461 %0 = getelementptr inbounds i32, i32* %x, i32 %index
462 %1 = bitcast i32* %0 to <4 x i32>*
463 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
464 %2 = xor <4 x i32> %wide.load, %vec.phi
465 %index.next = add i32 %index, 4
466 %3 = icmp eq i32 %index.next, %n.vec
467 br i1 %3, label %middle.block, label %vector.body
468
469middle.block: ; preds = %vector.body
470 %4 = call i32 @llvm.experimental.vector.reduce.xor.v4i32(<4 x i32> %2)
471 %cmp.n = icmp eq i32 %n.vec, %n
472 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
473
474for.body.preheader1: ; preds = %middle.block, %for.body.preheader
475 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
476 %r.07.ph = phi i32 [ 0, %for.body.preheader ], [ %4, %middle.block ]
477 br label %for.body
478
479for.body: ; preds = %for.body.preheader1, %for.body
480 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
481 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
482 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
483 %5 = load i32, i32* %arrayidx, align 4
484 %add = xor i32 %5, %r.07
485 %inc = add nuw nsw i32 %i.08, 1
486 %exitcond = icmp eq i32 %inc, %n
487 br i1 %exitcond, label %for.cond.cleanup, label %for.body
488
489for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
490 %r.0.lcssa = phi i32 [ 0, %entry ], [ %4, %middle.block ], [ %add, %for.body ]
491 ret i32 %r.0.lcssa
492}
493
494define float @fadd_f32(float* nocapture readonly %x, i32 %n) {
495; CHECK-LABEL: fadd_f32:
496; CHECK: @ %bb.0: @ %entry
497; CHECK-NEXT: .save {r7, lr}
498; CHECK-NEXT: push {r7, lr}
499; CHECK-NEXT: cmp r1, #1
500; CHECK-NEXT: blt .LBB5_3
501; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
502; CHECK-NEXT: cmp r1, #4
503; CHECK-NEXT: bhs .LBB5_4
504; CHECK-NEXT: @ %bb.2:
505; CHECK-NEXT: vldr s0, .LCPI5_0
506; CHECK-NEXT: movs r2, #0
507; CHECK-NEXT: b .LBB5_7
508; CHECK-NEXT: .LBB5_3:
509; CHECK-NEXT: vldr s0, .LCPI5_0
510; CHECK-NEXT: b .LBB5_9
511; CHECK-NEXT: .LBB5_4: @ %vector.ph
512; CHECK-NEXT: bic r2, r1, #3
513; CHECK-NEXT: movs r3, #1
514; CHECK-NEXT: sub.w r12, r2, #4
515; CHECK-NEXT: vmov.i32 q0, #0x0
516; CHECK-NEXT: add.w lr, r3, r12, lsr #2
517; CHECK-NEXT: mov r3, r0
518; CHECK-NEXT: dls lr, lr
519; CHECK-NEXT: .LBB5_5: @ %vector.body
520; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
521; CHECK-NEXT: vldrw.u32 q1, [r3], #16
522; CHECK-NEXT: vadd.f32 q0, q1, q0
523; CHECK-NEXT: le lr, .LBB5_5
524; CHECK-NEXT: @ %bb.6: @ %middle.block
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]525; CHECK-NEXT: vadd.f32 s4, s2, s3
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]526; CHECK-NEXT: cmp r2, r1
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]527; CHECK-NEXT: vadd.f32 s0, s0, s1
528; CHECK-NEXT: vadd.f32 s0, s0, s4
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]529; CHECK-NEXT: beq .LBB5_9
530; CHECK-NEXT: .LBB5_7: @ %for.body.preheader1
531; CHECK-NEXT: sub.w lr, r1, r2
532; CHECK-NEXT: add.w r0, r0, r2, lsl #2
533; CHECK-NEXT: dls lr, lr
534; CHECK-NEXT: .LBB5_8: @ %for.body
535; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
536; CHECK-NEXT: vldr s2, [r0]
537; CHECK-NEXT: adds r0, #4
538; CHECK-NEXT: vadd.f32 s0, s2, s0
539; CHECK-NEXT: le lr, .LBB5_8
540; CHECK-NEXT: .LBB5_9: @ %for.cond.cleanup
541; CHECK-NEXT: vmov r0, s0
542; CHECK-NEXT: pop {r7, pc}
543; CHECK-NEXT: .p2align 2
544; CHECK-NEXT: @ %bb.10:
545; CHECK-NEXT: .LCPI5_0:
546; CHECK-NEXT: .long 0x00000000 @ float 0
547entry:
548 %cmp6 = icmp sgt i32 %n, 0
549 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
550
551for.body.preheader: ; preds = %entry
552 %min.iters.check = icmp ult i32 %n, 4
553 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
554
555vector.ph: ; preds = %for.body.preheader
556 %n.vec = and i32 %n, -4
557 br label %vector.body
558
559vector.body: ; preds = %vector.body, %vector.ph
560 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
561 %vec.phi = phi <4 x float> [ zeroinitializer, %vector.ph ], [ %2, %vector.body ]
562 %0 = getelementptr inbounds float, float* %x, i32 %index
563 %1 = bitcast float* %0 to <4 x float>*
564 %wide.load = load <4 x float>, <4 x float>* %1, align 4
565 %2 = fadd fast <4 x float> %wide.load, %vec.phi
566 %index.next = add i32 %index, 4
567 %3 = icmp eq i32 %index.next, %n.vec
568 br i1 %3, label %middle.block, label %vector.body
569
570middle.block: ; preds = %vector.body
571 %4 = call fast float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float 0.000000e+00, <4 x float> %2)
572 %cmp.n = icmp eq i32 %n.vec, %n
573 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
574
575for.body.preheader1: ; preds = %middle.block, %for.body.preheader
576 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
577 %r.07.ph = phi float [ 0.000000e+00, %for.body.preheader ], [ %4, %middle.block ]
578 br label %for.body
579
580for.body: ; preds = %for.body.preheader1, %for.body
581 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
582 %r.07 = phi float [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
583 %arrayidx = getelementptr inbounds float, float* %x, i32 %i.08
584 %5 = load float, float* %arrayidx, align 4
585 %add = fadd fast float %5, %r.07
586 %inc = add nuw nsw i32 %i.08, 1
587 %exitcond = icmp eq i32 %inc, %n
588 br i1 %exitcond, label %for.cond.cleanup, label %for.body
589
590for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
591 %r.0.lcssa = phi float [ 0.000000e+00, %entry ], [ %4, %middle.block ], [ %add, %for.body ]
592 ret float %r.0.lcssa
593}
594
595define float @fmul_f32(float* nocapture readonly %x, i32 %n) {
596; CHECK-LABEL: fmul_f32:
597; CHECK: @ %bb.0: @ %entry
598; CHECK-NEXT: .save {r7, lr}
599; CHECK-NEXT: push {r7, lr}
600; CHECK-NEXT: cmp r1, #1
601; CHECK-NEXT: blt .LBB6_3
602; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
603; CHECK-NEXT: cmp r1, #4
604; CHECK-NEXT: bhs .LBB6_4
605; CHECK-NEXT: @ %bb.2:
606; CHECK-NEXT: vmov.f32 s0, #1.000000e+00
607; CHECK-NEXT: movs r2, #0
608; CHECK-NEXT: b .LBB6_7
609; CHECK-NEXT: .LBB6_3:
610; CHECK-NEXT: vmov.f32 s0, #1.000000e+00
611; CHECK-NEXT: b .LBB6_9
612; CHECK-NEXT: .LBB6_4: @ %vector.ph
613; CHECK-NEXT: bic r2, r1, #3
614; CHECK-NEXT: movs r3, #1
615; CHECK-NEXT: sub.w r12, r2, #4
616; CHECK-NEXT: vmov.f32 q0, #1.000000e+00
617; CHECK-NEXT: add.w lr, r3, r12, lsr #2
618; CHECK-NEXT: mov r3, r0
619; CHECK-NEXT: dls lr, lr
620; CHECK-NEXT: .LBB6_5: @ %vector.body
621; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
622; CHECK-NEXT: vldrw.u32 q1, [r3], #16
623; CHECK-NEXT: vmul.f32 q0, q1, q0
624; CHECK-NEXT: le lr, .LBB6_5
625; CHECK-NEXT: @ %bb.6: @ %middle.block
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]626; CHECK-NEXT: vmul.f32 s4, s2, s3
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]627; CHECK-NEXT: cmp r2, r1
David Greendeb72ce2020-06-29 13:53:19 +0100[diff] [blame^]628; CHECK-NEXT: vmul.f32 s0, s0, s1
629; CHECK-NEXT: vmul.f32 s0, s0, s4
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]630; CHECK-NEXT: beq .LBB6_9
631; CHECK-NEXT: .LBB6_7: @ %for.body.preheader1
632; CHECK-NEXT: sub.w lr, r1, r2
633; CHECK-NEXT: add.w r0, r0, r2, lsl #2
634; CHECK-NEXT: dls lr, lr
635; CHECK-NEXT: .LBB6_8: @ %for.body
636; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
637; CHECK-NEXT: vldr s2, [r0]
638; CHECK-NEXT: adds r0, #4
639; CHECK-NEXT: vmul.f32 s0, s2, s0
640; CHECK-NEXT: le lr, .LBB6_8
641; CHECK-NEXT: .LBB6_9: @ %for.cond.cleanup
642; CHECK-NEXT: vmov r0, s0
643; CHECK-NEXT: pop {r7, pc}
644entry:
645 %cmp6 = icmp sgt i32 %n, 0
646 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
647
648for.body.preheader: ; preds = %entry
649 %min.iters.check = icmp ult i32 %n, 4
650 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
651
652vector.ph: ; preds = %for.body.preheader
653 %n.vec = and i32 %n, -4
654 br label %vector.body
655
656vector.body: ; preds = %vector.body, %vector.ph
657 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
658 %vec.phi = phi <4 x float> [ <float 1.000000e+00, float 1.000000e+00, float 1.000000e+00, float 1.000000e+00>, %vector.ph ], [ %2, %vector.body ]
659 %0 = getelementptr inbounds float, float* %x, i32 %index
660 %1 = bitcast float* %0 to <4 x float>*
661 %wide.load = load <4 x float>, <4 x float>* %1, align 4
662 %2 = fmul fast <4 x float> %wide.load, %vec.phi
663 %index.next = add i32 %index, 4
664 %3 = icmp eq i32 %index.next, %n.vec
665 br i1 %3, label %middle.block, label %vector.body
666
667middle.block: ; preds = %vector.body
668 %4 = call fast float @llvm.experimental.vector.reduce.v2.fmul.f32.v4f32(float 1.000000e+00, <4 x float> %2)
669 %cmp.n = icmp eq i32 %n.vec, %n
670 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
671
672for.body.preheader1: ; preds = %middle.block, %for.body.preheader
673 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
674 %r.07.ph = phi float [ 1.000000e+00, %for.body.preheader ], [ %4, %middle.block ]
675 br label %for.body
676
677for.body: ; preds = %for.body.preheader1, %for.body
678 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
679 %r.07 = phi float [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
680 %arrayidx = getelementptr inbounds float, float* %x, i32 %i.08
681 %5 = load float, float* %arrayidx, align 4
682 %add = fmul fast float %5, %r.07
683 %inc = add nuw nsw i32 %i.08, 1
684 %exitcond = icmp eq i32 %inc, %n
685 br i1 %exitcond, label %for.cond.cleanup, label %for.body
686
687for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
688 %r.0.lcssa = phi float [ 1.000000e+00, %entry ], [ %4, %middle.block ], [ %add, %for.body ]
689 ret float %r.0.lcssa
690}
691
692define i32 @smin_i32(i32* nocapture readonly %x, i32 %n) {
693; CHECK-LABEL: smin_i32:
694; CHECK: @ %bb.0: @ %entry
695; CHECK-NEXT: .save {r7, lr}
696; CHECK-NEXT: push {r7, lr}
697; CHECK-NEXT: cmp r1, #1
698; CHECK-NEXT: blt .LBB7_3
699; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
700; CHECK-NEXT: cmp r1, #4
701; CHECK-NEXT: bhs .LBB7_4
702; CHECK-NEXT: @ %bb.2:
703; CHECK-NEXT: mvn r2, #-2147483648
704; CHECK-NEXT: movs r3, #0
705; CHECK-NEXT: b .LBB7_7
706; CHECK-NEXT: .LBB7_3:
707; CHECK-NEXT: mvn r2, #-2147483648
708; CHECK-NEXT: b .LBB7_9
709; CHECK-NEXT: .LBB7_4: @ %vector.ph
710; CHECK-NEXT: bic r3, r1, #3
711; CHECK-NEXT: movs r2, #1
712; CHECK-NEXT: sub.w r12, r3, #4
713; CHECK-NEXT: vmvn.i32 q0, #0x80000000
714; CHECK-NEXT: add.w lr, r2, r12, lsr #2
715; CHECK-NEXT: mov r2, r0
716; CHECK-NEXT: dls lr, lr
717; CHECK-NEXT: .LBB7_5: @ %vector.body
718; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
719; CHECK-NEXT: vldrw.u32 q1, [r2], #16
720; CHECK-NEXT: vmin.s32 q0, q0, q1
721; CHECK-NEXT: le lr, .LBB7_5
722; CHECK-NEXT: @ %bb.6: @ %middle.block
723; CHECK-NEXT: mvn r2, #-2147483648
724; CHECK-NEXT: cmp r3, r1
725; CHECK-NEXT: vminv.s32 r2, q0
726; CHECK-NEXT: beq .LBB7_9
727; CHECK-NEXT: .LBB7_7: @ %for.body.preheader1
728; CHECK-NEXT: sub.w lr, r1, r3
729; CHECK-NEXT: add.w r0, r0, r3, lsl #2
730; CHECK-NEXT: dls lr, lr
731; CHECK-NEXT: .LBB7_8: @ %for.body
732; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
733; CHECK-NEXT: ldr r1, [r0], #4
734; CHECK-NEXT: cmp r2, r1
735; CHECK-NEXT: it ge
736; CHECK-NEXT: movge r2, r1
737; CHECK-NEXT: le lr, .LBB7_8
738; CHECK-NEXT: .LBB7_9: @ %for.cond.cleanup
739; CHECK-NEXT: mov r0, r2
740; CHECK-NEXT: pop {r7, pc}
741entry:
742 %cmp6 = icmp sgt i32 %n, 0
743 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
744
745for.body.preheader: ; preds = %entry
746 %min.iters.check = icmp ult i32 %n, 4
747 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
748
749vector.ph: ; preds = %for.body.preheader
750 %n.vec = and i32 %n, -4
751 br label %vector.body
752
753vector.body: ; preds = %vector.body, %vector.ph
754 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
755 %vec.phi = phi <4 x i32> [ <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647>, %vector.ph ], [ %3, %vector.body ]
756 %0 = getelementptr inbounds i32, i32* %x, i32 %index
757 %1 = bitcast i32* %0 to <4 x i32>*
758 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
759 %2 = icmp slt <4 x i32> %vec.phi, %wide.load
760 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
761 %index.next = add i32 %index, 4
762 %4 = icmp eq i32 %index.next, %n.vec
763 br i1 %4, label %middle.block, label %vector.body
764
765middle.block: ; preds = %vector.body
766 %5 = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> %3)
767 %cmp.n = icmp eq i32 %n.vec, %n
768 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
769
770for.body.preheader1: ; preds = %middle.block, %for.body.preheader
771 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
772 %r.07.ph = phi i32 [ 2147483647, %for.body.preheader ], [ %5, %middle.block ]
773 br label %for.body
774
775for.body: ; preds = %for.body.preheader1, %for.body
776 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
777 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
778 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
779 %6 = load i32, i32* %arrayidx, align 4
780 %c = icmp slt i32 %r.07, %6
781 %add = select i1 %c, i32 %r.07, i32 %6
782 %inc = add nuw nsw i32 %i.08, 1
783 %exitcond = icmp eq i32 %inc, %n
784 br i1 %exitcond, label %for.cond.cleanup, label %for.body
785
786for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
787 %r.0.lcssa = phi i32 [ 2147483647, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
788 ret i32 %r.0.lcssa
789}
790
791define i32 @smin_i32_inloop(i32* nocapture readonly %x, i32 %n) {
792; CHECK-LABEL: smin_i32_inloop:
793; CHECK: @ %bb.0: @ %entry
794; CHECK-NEXT: .save {r4, lr}
795; CHECK-NEXT: push {r4, lr}
796; CHECK-NEXT: cmp r1, #1
797; CHECK-NEXT: blt .LBB8_3
798; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
799; CHECK-NEXT: mov r12, r0
800; CHECK-NEXT: mvn r0, #-2147483648
801; CHECK-NEXT: cmp r1, #4
802; CHECK-NEXT: bhs .LBB8_4
803; CHECK-NEXT: @ %bb.2:
804; CHECK-NEXT: movs r3, #0
805; CHECK-NEXT: b .LBB8_7
806; CHECK-NEXT: .LBB8_3:
807; CHECK-NEXT: mvn r0, #-2147483648
808; CHECK-NEXT: b .LBB8_9
809; CHECK-NEXT: .LBB8_4: @ %vector.ph
810; CHECK-NEXT: bic r3, r1, #3
811; CHECK-NEXT: movs r2, #1
812; CHECK-NEXT: sub.w lr, r3, #4
813; CHECK-NEXT: add.w lr, r2, lr, lsr #2
814; CHECK-NEXT: mov r2, r12
815; CHECK-NEXT: dls lr, lr
816; CHECK-NEXT: .LBB8_5: @ %vector.body
817; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
818; CHECK-NEXT: vldrw.u32 q0, [r2], #16
819; CHECK-NEXT: mvn r4, #-2147483648
820; CHECK-NEXT: vminv.s32 r4, q0
821; CHECK-NEXT: cmp r0, r4
822; CHECK-NEXT: it ge
823; CHECK-NEXT: movge r0, r4
824; CHECK-NEXT: le lr, .LBB8_5
825; CHECK-NEXT: @ %bb.6: @ %middle.block
826; CHECK-NEXT: cmp r3, r1
827; CHECK-NEXT: it eq
828; CHECK-NEXT: popeq {r4, pc}
829; CHECK-NEXT: .LBB8_7: @ %for.body.preheader1
830; CHECK-NEXT: sub.w lr, r1, r3
831; CHECK-NEXT: add.w r1, r12, r3, lsl #2
832; CHECK-NEXT: dls lr, lr
833; CHECK-NEXT: .LBB8_8: @ %for.body
834; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
835; CHECK-NEXT: ldr r2, [r1], #4
836; CHECK-NEXT: cmp r0, r2
837; CHECK-NEXT: it ge
838; CHECK-NEXT: movge r0, r2
839; CHECK-NEXT: le lr, .LBB8_8
840; CHECK-NEXT: .LBB8_9: @ %for.cond.cleanup
841; CHECK-NEXT: pop {r4, pc}
842entry:
843 %cmp6 = icmp sgt i32 %n, 0
844 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
845
846for.body.preheader: ; preds = %entry
847 %min.iters.check = icmp ult i32 %n, 4
848 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
849
850vector.ph: ; preds = %for.body.preheader
851 %n.vec = and i32 %n, -4
852 br label %vector.body
853
854vector.body: ; preds = %vector.body, %vector.ph
855 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
856 %vec.phi = phi i32 [ 2147483647, %vector.ph ], [ %3, %vector.body ]
857 %0 = getelementptr inbounds i32, i32* %x, i32 %index
858 %1 = bitcast i32* %0 to <4 x i32>*
859 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
860 %l5 = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> %wide.load)
861 %2 = icmp slt i32 %vec.phi, %l5
862 %3 = select i1 %2, i32 %vec.phi, i32 %l5
863 %index.next = add i32 %index, 4
864 %4 = icmp eq i32 %index.next, %n.vec
865 br i1 %4, label %middle.block, label %vector.body
866
867middle.block: ; preds = %vector.body
868 %5 = phi i32 [ %3, %vector.body ]
869 %cmp.n = icmp eq i32 %n.vec, %n
870 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
871
872for.body.preheader1: ; preds = %middle.block, %for.body.preheader
873 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
874 %r.07.ph = phi i32 [ 2147483647, %for.body.preheader ], [ %5, %middle.block ]
875 br label %for.body
876
877for.body: ; preds = %for.body.preheader1, %for.body
878 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
879 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
880 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
881 %6 = load i32, i32* %arrayidx, align 4
882 %c = icmp slt i32 %r.07, %6
883 %add = select i1 %c, i32 %r.07, i32 %6
884 %inc = add nuw nsw i32 %i.08, 1
885 %exitcond = icmp eq i32 %inc, %n
886 br i1 %exitcond, label %for.cond.cleanup, label %for.body
887
888for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
889 %r.0.lcssa = phi i32 [ 2147483647, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
890 ret i32 %r.0.lcssa
891}
892
893define i32 @smax_i32(i32* nocapture readonly %x, i32 %n) {
894; CHECK-LABEL: smax_i32:
895; CHECK: @ %bb.0: @ %entry
896; CHECK-NEXT: .save {r7, lr}
897; CHECK-NEXT: push {r7, lr}
898; CHECK-NEXT: cmp r1, #1
899; CHECK-NEXT: blt .LBB9_3
900; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
901; CHECK-NEXT: cmp r1, #4
902; CHECK-NEXT: bhs .LBB9_4
903; CHECK-NEXT: @ %bb.2:
904; CHECK-NEXT: mov.w r2, #-2147483648
905; CHECK-NEXT: movs r3, #0
906; CHECK-NEXT: b .LBB9_7
907; CHECK-NEXT: .LBB9_3:
908; CHECK-NEXT: mov.w r2, #-2147483648
909; CHECK-NEXT: b .LBB9_9
910; CHECK-NEXT: .LBB9_4: @ %vector.ph
911; CHECK-NEXT: bic r3, r1, #3
912; CHECK-NEXT: movs r2, #1
913; CHECK-NEXT: sub.w r12, r3, #4
914; CHECK-NEXT: vmov.i32 q0, #0x80000000
915; CHECK-NEXT: add.w lr, r2, r12, lsr #2
916; CHECK-NEXT: mov r2, r0
917; CHECK-NEXT: dls lr, lr
918; CHECK-NEXT: .LBB9_5: @ %vector.body
919; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
920; CHECK-NEXT: vldrw.u32 q1, [r2], #16
921; CHECK-NEXT: vmax.s32 q0, q0, q1
922; CHECK-NEXT: le lr, .LBB9_5
923; CHECK-NEXT: @ %bb.6: @ %middle.block
924; CHECK-NEXT: mov.w r2, #-2147483648
925; CHECK-NEXT: cmp r3, r1
926; CHECK-NEXT: vmaxv.s32 r2, q0
927; CHECK-NEXT: beq .LBB9_9
928; CHECK-NEXT: .LBB9_7: @ %for.body.preheader1
929; CHECK-NEXT: sub.w lr, r1, r3
930; CHECK-NEXT: add.w r0, r0, r3, lsl #2
931; CHECK-NEXT: dls lr, lr
932; CHECK-NEXT: .LBB9_8: @ %for.body
933; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
934; CHECK-NEXT: ldr r1, [r0], #4
935; CHECK-NEXT: cmp r2, r1
936; CHECK-NEXT: it le
937; CHECK-NEXT: movle r2, r1
938; CHECK-NEXT: le lr, .LBB9_8
939; CHECK-NEXT: .LBB9_9: @ %for.cond.cleanup
940; CHECK-NEXT: mov r0, r2
941; CHECK-NEXT: pop {r7, pc}
942entry:
943 %cmp6 = icmp sgt i32 %n, 0
944 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
945
946for.body.preheader: ; preds = %entry
947 %min.iters.check = icmp ult i32 %n, 4
948 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
949
950vector.ph: ; preds = %for.body.preheader
951 %n.vec = and i32 %n, -4
952 br label %vector.body
953
954vector.body: ; preds = %vector.body, %vector.ph
955 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
956 %vec.phi = phi <4 x i32> [ <i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648>, %vector.ph ], [ %3, %vector.body ]
957 %0 = getelementptr inbounds i32, i32* %x, i32 %index
958 %1 = bitcast i32* %0 to <4 x i32>*
959 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
960 %2 = icmp sgt <4 x i32> %vec.phi, %wide.load
961 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
962 %index.next = add i32 %index, 4
963 %4 = icmp eq i32 %index.next, %n.vec
964 br i1 %4, label %middle.block, label %vector.body
965
966middle.block: ; preds = %vector.body
967 %5 = call i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32> %3)
968 %cmp.n = icmp eq i32 %n.vec, %n
969 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
970
971for.body.preheader1: ; preds = %middle.block, %for.body.preheader
972 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
973 %r.07.ph = phi i32 [ -2147483648, %for.body.preheader ], [ %5, %middle.block ]
974 br label %for.body
975
976for.body: ; preds = %for.body.preheader1, %for.body
977 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
978 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
979 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
980 %6 = load i32, i32* %arrayidx, align 4
981 %c = icmp sgt i32 %r.07, %6
982 %add = select i1 %c, i32 %r.07, i32 %6
983 %inc = add nuw nsw i32 %i.08, 1
984 %exitcond = icmp eq i32 %inc, %n
985 br i1 %exitcond, label %for.cond.cleanup, label %for.body
986
987for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
988 %r.0.lcssa = phi i32 [ -2147483648, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
989 ret i32 %r.0.lcssa
990}
991
992define i32 @smax_i32_inloop(i32* nocapture readonly %x, i32 %n) {
993; CHECK-LABEL: smax_i32_inloop:
994; CHECK: @ %bb.0: @ %entry
995; CHECK-NEXT: .save {r4, lr}
996; CHECK-NEXT: push {r4, lr}
997; CHECK-NEXT: cmp r1, #1
998; CHECK-NEXT: blt .LBB10_3
999; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1000; CHECK-NEXT: mov r12, r0
1001; CHECK-NEXT: mov.w r0, #-2147483648
1002; CHECK-NEXT: cmp r1, #4
1003; CHECK-NEXT: bhs .LBB10_4
1004; CHECK-NEXT: @ %bb.2:
1005; CHECK-NEXT: movs r3, #0
1006; CHECK-NEXT: b .LBB10_7
1007; CHECK-NEXT: .LBB10_3:
1008; CHECK-NEXT: mov.w r0, #-2147483648
1009; CHECK-NEXT: b .LBB10_9
1010; CHECK-NEXT: .LBB10_4: @ %vector.ph
1011; CHECK-NEXT: bic r3, r1, #3
1012; CHECK-NEXT: movs r2, #1
1013; CHECK-NEXT: sub.w lr, r3, #4
1014; CHECK-NEXT: add.w lr, r2, lr, lsr #2
1015; CHECK-NEXT: mov r2, r12
1016; CHECK-NEXT: dls lr, lr
1017; CHECK-NEXT: .LBB10_5: @ %vector.body
1018; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1019; CHECK-NEXT: vldrw.u32 q0, [r2], #16
1020; CHECK-NEXT: mov.w r4, #-2147483648
1021; CHECK-NEXT: vmaxv.s32 r4, q0
1022; CHECK-NEXT: cmp r0, r4
1023; CHECK-NEXT: it le
1024; CHECK-NEXT: movle r0, r4
1025; CHECK-NEXT: le lr, .LBB10_5
1026; CHECK-NEXT: @ %bb.6: @ %middle.block
1027; CHECK-NEXT: cmp r3, r1
1028; CHECK-NEXT: it eq
1029; CHECK-NEXT: popeq {r4, pc}
1030; CHECK-NEXT: .LBB10_7: @ %for.body.preheader1
1031; CHECK-NEXT: sub.w lr, r1, r3
1032; CHECK-NEXT: add.w r1, r12, r3, lsl #2
1033; CHECK-NEXT: dls lr, lr
1034; CHECK-NEXT: .LBB10_8: @ %for.body
1035; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1036; CHECK-NEXT: ldr r2, [r1], #4
1037; CHECK-NEXT: cmp r0, r2
1038; CHECK-NEXT: it le
1039; CHECK-NEXT: movle r0, r2
1040; CHECK-NEXT: le lr, .LBB10_8
1041; CHECK-NEXT: .LBB10_9: @ %for.cond.cleanup
1042; CHECK-NEXT: pop {r4, pc}
1043entry:
1044 %cmp6 = icmp sgt i32 %n, 0
1045 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1046
1047for.body.preheader: ; preds = %entry
1048 %min.iters.check = icmp ult i32 %n, 4
1049 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1050
1051vector.ph: ; preds = %for.body.preheader
1052 %n.vec = and i32 %n, -4
1053 br label %vector.body
1054
1055vector.body: ; preds = %vector.body, %vector.ph
1056 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1057 %vec.phi = phi i32 [ -2147483648, %vector.ph ], [ %3, %vector.body ]
1058 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1059 %1 = bitcast i32* %0 to <4 x i32>*
1060 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1061 %l5 = call i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32> %wide.load)
1062 %2 = icmp sgt i32 %vec.phi, %l5
1063 %3 = select i1 %2, i32 %vec.phi, i32 %l5
1064 %index.next = add i32 %index, 4
1065 %4 = icmp eq i32 %index.next, %n.vec
1066 br i1 %4, label %middle.block, label %vector.body
1067
1068middle.block: ; preds = %vector.body
1069 %5 = phi i32 [ %3, %vector.body ]
1070 %cmp.n = icmp eq i32 %n.vec, %n
1071 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1072
1073for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1074 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1075 %r.07.ph = phi i32 [ -2147483648, %for.body.preheader ], [ %5, %middle.block ]
1076 br label %for.body
1077
1078for.body: ; preds = %for.body.preheader1, %for.body
1079 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1080 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1081 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1082 %6 = load i32, i32* %arrayidx, align 4
1083 %c = icmp sgt i32 %r.07, %6
1084 %add = select i1 %c, i32 %r.07, i32 %6
1085 %inc = add nuw nsw i32 %i.08, 1
1086 %exitcond = icmp eq i32 %inc, %n
1087 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1088
1089for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1090 %r.0.lcssa = phi i32 [ -2147483648, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1091 ret i32 %r.0.lcssa
1092}
1093
1094define i32 @umin_i32(i32* nocapture readonly %x, i32 %n) {
1095; CHECK-LABEL: umin_i32:
1096; CHECK: @ %bb.0: @ %entry
1097; CHECK-NEXT: .save {r7, lr}
1098; CHECK-NEXT: push {r7, lr}
1099; CHECK-NEXT: cmp r1, #1
1100; CHECK-NEXT: blt .LBB11_3
1101; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1102; CHECK-NEXT: cmp r1, #4
1103; CHECK-NEXT: bhs .LBB11_4
1104; CHECK-NEXT: @ %bb.2:
1105; CHECK-NEXT: mov.w r2, #-1
1106; CHECK-NEXT: movs r3, #0
1107; CHECK-NEXT: b .LBB11_7
1108; CHECK-NEXT: .LBB11_3:
1109; CHECK-NEXT: mov.w r2, #-1
1110; CHECK-NEXT: b .LBB11_9
1111; CHECK-NEXT: .LBB11_4: @ %vector.ph
1112; CHECK-NEXT: bic r3, r1, #3
1113; CHECK-NEXT: movs r2, #1
1114; CHECK-NEXT: sub.w r12, r3, #4
1115; CHECK-NEXT: vmov.i8 q0, #0xff
1116; CHECK-NEXT: add.w lr, r2, r12, lsr #2
1117; CHECK-NEXT: mov r2, r0
1118; CHECK-NEXT: dls lr, lr
1119; CHECK-NEXT: .LBB11_5: @ %vector.body
1120; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1121; CHECK-NEXT: vldrw.u32 q1, [r2], #16
1122; CHECK-NEXT: vmin.u32 q0, q0, q1
1123; CHECK-NEXT: le lr, .LBB11_5
1124; CHECK-NEXT: @ %bb.6: @ %middle.block
1125; CHECK-NEXT: mov.w r2, #-1
1126; CHECK-NEXT: cmp r3, r1
1127; CHECK-NEXT: vminv.u32 r2, q0
1128; CHECK-NEXT: beq .LBB11_9
1129; CHECK-NEXT: .LBB11_7: @ %for.body.preheader1
1130; CHECK-NEXT: sub.w lr, r1, r3
1131; CHECK-NEXT: add.w r0, r0, r3, lsl #2
1132; CHECK-NEXT: dls lr, lr
1133; CHECK-NEXT: .LBB11_8: @ %for.body
1134; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1135; CHECK-NEXT: ldr r1, [r0], #4
1136; CHECK-NEXT: cmp r2, r1
1137; CHECK-NEXT: it hs
1138; CHECK-NEXT: movhs r2, r1
1139; CHECK-NEXT: le lr, .LBB11_8
1140; CHECK-NEXT: .LBB11_9: @ %for.cond.cleanup
1141; CHECK-NEXT: mov r0, r2
1142; CHECK-NEXT: pop {r7, pc}
1143entry:
1144 %cmp6 = icmp sgt i32 %n, 0
1145 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1146
1147for.body.preheader: ; preds = %entry
1148 %min.iters.check = icmp ult i32 %n, 4
1149 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1150
1151vector.ph: ; preds = %for.body.preheader
1152 %n.vec = and i32 %n, -4
1153 br label %vector.body
1154
1155vector.body: ; preds = %vector.body, %vector.ph
1156 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1157 %vec.phi = phi <4 x i32> [ <i32 -1, i32 -1, i32 -1, i32 -1>, %vector.ph ], [ %3, %vector.body ]
1158 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1159 %1 = bitcast i32* %0 to <4 x i32>*
1160 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1161 %2 = icmp ult <4 x i32> %vec.phi, %wide.load
1162 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
1163 %index.next = add i32 %index, 4
1164 %4 = icmp eq i32 %index.next, %n.vec
1165 br i1 %4, label %middle.block, label %vector.body
1166
1167middle.block: ; preds = %vector.body
1168 %5 = call i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32> %3)
1169 %cmp.n = icmp eq i32 %n.vec, %n
1170 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1171
1172for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1173 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1174 %r.07.ph = phi i32 [ -1, %for.body.preheader ], [ %5, %middle.block ]
1175 br label %for.body
1176
1177for.body: ; preds = %for.body.preheader1, %for.body
1178 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1179 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1180 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1181 %6 = load i32, i32* %arrayidx, align 4
1182 %c = icmp ult i32 %r.07, %6
1183 %add = select i1 %c, i32 %r.07, i32 %6
1184 %inc = add nuw nsw i32 %i.08, 1
1185 %exitcond = icmp eq i32 %inc, %n
1186 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1187
1188for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1189 %r.0.lcssa = phi i32 [ -1, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1190 ret i32 %r.0.lcssa
1191}
1192
1193define i32 @umin_i32_inloop(i32* nocapture readonly %x, i32 %n) {
1194; CHECK-LABEL: umin_i32_inloop:
1195; CHECK: @ %bb.0: @ %entry
1196; CHECK-NEXT: .save {r4, lr}
1197; CHECK-NEXT: push {r4, lr}
1198; CHECK-NEXT: cmp r1, #1
1199; CHECK-NEXT: blt .LBB12_3
1200; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1201; CHECK-NEXT: mov r12, r0
1202; CHECK-NEXT: mov.w r0, #-1
1203; CHECK-NEXT: cmp r1, #4
1204; CHECK-NEXT: bhs .LBB12_4
1205; CHECK-NEXT: @ %bb.2:
1206; CHECK-NEXT: movs r3, #0
1207; CHECK-NEXT: b .LBB12_7
1208; CHECK-NEXT: .LBB12_3:
1209; CHECK-NEXT: mov.w r0, #-1
1210; CHECK-NEXT: b .LBB12_9
1211; CHECK-NEXT: .LBB12_4: @ %vector.ph
1212; CHECK-NEXT: bic r3, r1, #3
1213; CHECK-NEXT: movs r2, #1
1214; CHECK-NEXT: sub.w lr, r3, #4
1215; CHECK-NEXT: add.w lr, r2, lr, lsr #2
1216; CHECK-NEXT: mov r2, r12
1217; CHECK-NEXT: dls lr, lr
1218; CHECK-NEXT: .LBB12_5: @ %vector.body
1219; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1220; CHECK-NEXT: vldrw.u32 q0, [r2], #16
1221; CHECK-NEXT: mov.w r4, #-1
1222; CHECK-NEXT: vminv.u32 r4, q0
1223; CHECK-NEXT: cmp r0, r4
1224; CHECK-NEXT: it hs
1225; CHECK-NEXT: movhs r0, r4
1226; CHECK-NEXT: le lr, .LBB12_5
1227; CHECK-NEXT: @ %bb.6: @ %middle.block
1228; CHECK-NEXT: cmp r3, r1
1229; CHECK-NEXT: it eq
1230; CHECK-NEXT: popeq {r4, pc}
1231; CHECK-NEXT: .LBB12_7: @ %for.body.preheader1
1232; CHECK-NEXT: sub.w lr, r1, r3
1233; CHECK-NEXT: add.w r1, r12, r3, lsl #2
1234; CHECK-NEXT: dls lr, lr
1235; CHECK-NEXT: .LBB12_8: @ %for.body
1236; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1237; CHECK-NEXT: ldr r2, [r1], #4
1238; CHECK-NEXT: cmp r0, r2
1239; CHECK-NEXT: it ls
1240; CHECK-NEXT: movls r0, r2
1241; CHECK-NEXT: le lr, .LBB12_8
1242; CHECK-NEXT: .LBB12_9: @ %for.cond.cleanup
1243; CHECK-NEXT: pop {r4, pc}
1244entry:
1245 %cmp6 = icmp sgt i32 %n, 0
1246 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1247
1248for.body.preheader: ; preds = %entry
1249 %min.iters.check = icmp ult i32 %n, 4
1250 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1251
1252vector.ph: ; preds = %for.body.preheader
1253 %n.vec = and i32 %n, -4
1254 br label %vector.body
1255
1256vector.body: ; preds = %vector.body, %vector.ph
1257 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1258 %vec.phi = phi i32 [ -1, %vector.ph ], [ %3, %vector.body ]
1259 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1260 %1 = bitcast i32* %0 to <4 x i32>*
1261 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1262 %l5 = call i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32> %wide.load)
1263 %2 = icmp ult i32 %vec.phi, %l5
1264 %3 = select i1 %2, i32 %vec.phi, i32 %l5
1265 %index.next = add i32 %index, 4
1266 %4 = icmp eq i32 %index.next, %n.vec
1267 br i1 %4, label %middle.block, label %vector.body
1268
1269middle.block: ; preds = %vector.body
1270 %5 = phi i32 [ %3, %vector.body ]
1271 %cmp.n = icmp eq i32 %n.vec, %n
1272 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1273
1274for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1275 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1276 %r.07.ph = phi i32 [ -1, %for.body.preheader ], [ %5, %middle.block ]
1277 br label %for.body
1278
1279for.body: ; preds = %for.body.preheader1, %for.body
1280 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1281 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1282 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1283 %6 = load i32, i32* %arrayidx, align 4
1284 %c = icmp ugt i32 %r.07, %6
1285 %add = select i1 %c, i32 %r.07, i32 %6
1286 %inc = add nuw nsw i32 %i.08, 1
1287 %exitcond = icmp eq i32 %inc, %n
1288 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1289
1290for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1291 %r.0.lcssa = phi i32 [ -1, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1292 ret i32 %r.0.lcssa
1293}
1294
1295define i32 @umax_i32(i32* nocapture readonly %x, i32 %n) {
1296; CHECK-LABEL: umax_i32:
1297; CHECK: @ %bb.0: @ %entry
1298; CHECK-NEXT: .save {r7, lr}
1299; CHECK-NEXT: push {r7, lr}
1300; CHECK-NEXT: cmp r1, #1
1301; CHECK-NEXT: blt .LBB13_3
1302; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1303; CHECK-NEXT: cmp r1, #4
1304; CHECK-NEXT: bhs .LBB13_4
1305; CHECK-NEXT: @ %bb.2:
1306; CHECK-NEXT: movs r3, #0
1307; CHECK-NEXT: movs r2, #0
1308; CHECK-NEXT: b .LBB13_7
1309; CHECK-NEXT: .LBB13_3:
1310; CHECK-NEXT: movs r2, #0
1311; CHECK-NEXT: b .LBB13_9
1312; CHECK-NEXT: .LBB13_4: @ %vector.ph
1313; CHECK-NEXT: bic r3, r1, #3
1314; CHECK-NEXT: movs r2, #1
1315; CHECK-NEXT: sub.w r12, r3, #4
1316; CHECK-NEXT: vmov.i32 q0, #0x0
1317; CHECK-NEXT: add.w lr, r2, r12, lsr #2
1318; CHECK-NEXT: mov r2, r0
1319; CHECK-NEXT: dls lr, lr
1320; CHECK-NEXT: .LBB13_5: @ %vector.body
1321; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1322; CHECK-NEXT: vldrw.u32 q1, [r2], #16
1323; CHECK-NEXT: vmax.u32 q0, q0, q1
1324; CHECK-NEXT: le lr, .LBB13_5
1325; CHECK-NEXT: @ %bb.6: @ %middle.block
1326; CHECK-NEXT: movs r2, #0
1327; CHECK-NEXT: cmp r3, r1
1328; CHECK-NEXT: vmaxv.u32 r2, q0
1329; CHECK-NEXT: beq .LBB13_9
1330; CHECK-NEXT: .LBB13_7: @ %for.body.preheader1
1331; CHECK-NEXT: sub.w lr, r1, r3
1332; CHECK-NEXT: add.w r0, r0, r3, lsl #2
1333; CHECK-NEXT: dls lr, lr
1334; CHECK-NEXT: .LBB13_8: @ %for.body
1335; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1336; CHECK-NEXT: ldr r1, [r0], #4
1337; CHECK-NEXT: cmp r2, r1
1338; CHECK-NEXT: it ls
1339; CHECK-NEXT: movls r2, r1
1340; CHECK-NEXT: le lr, .LBB13_8
1341; CHECK-NEXT: .LBB13_9: @ %for.cond.cleanup
1342; CHECK-NEXT: mov r0, r2
1343; CHECK-NEXT: pop {r7, pc}
1344entry:
1345 %cmp6 = icmp sgt i32 %n, 0
1346 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1347
1348for.body.preheader: ; preds = %entry
1349 %min.iters.check = icmp ult i32 %n, 4
1350 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1351
1352vector.ph: ; preds = %for.body.preheader
1353 %n.vec = and i32 %n, -4
1354 br label %vector.body
1355
1356vector.body: ; preds = %vector.body, %vector.ph
1357 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1358 %vec.phi = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %3, %vector.body ]
1359 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1360 %1 = bitcast i32* %0 to <4 x i32>*
1361 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1362 %2 = icmp ugt <4 x i32> %vec.phi, %wide.load
1363 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
1364 %index.next = add i32 %index, 4
1365 %4 = icmp eq i32 %index.next, %n.vec
1366 br i1 %4, label %middle.block, label %vector.body
1367
1368middle.block: ; preds = %vector.body
1369 %5 = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> %3)
1370 %cmp.n = icmp eq i32 %n.vec, %n
1371 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1372
1373for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1374 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1375 %r.07.ph = phi i32 [ 0, %for.body.preheader ], [ %5, %middle.block ]
1376 br label %for.body
1377
1378for.body: ; preds = %for.body.preheader1, %for.body
1379 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1380 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1381 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1382 %6 = load i32, i32* %arrayidx, align 4
1383 %c = icmp ugt i32 %r.07, %6
1384 %add = select i1 %c, i32 %r.07, i32 %6
1385 %inc = add nuw nsw i32 %i.08, 1
1386 %exitcond = icmp eq i32 %inc, %n
1387 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1388
1389for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1390 %r.0.lcssa = phi i32 [ 0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1391 ret i32 %r.0.lcssa
1392}
1393
1394define i32 @umax_i32_inloop(i32* nocapture readonly %x, i32 %n) {
1395; CHECK-LABEL: umax_i32_inloop:
1396; CHECK: @ %bb.0: @ %entry
1397; CHECK-NEXT: .save {r4, lr}
1398; CHECK-NEXT: push {r4, lr}
1399; CHECK-NEXT: cmp r1, #1
1400; CHECK-NEXT: blt .LBB14_8
1401; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1402; CHECK-NEXT: mov r12, r0
1403; CHECK-NEXT: movs r3, #0
1404; CHECK-NEXT: cmp r1, #4
1405; CHECK-NEXT: mov.w r0, #0
1406; CHECK-NEXT: blo .LBB14_5
1407; CHECK-NEXT: @ %bb.2: @ %vector.ph
1408; CHECK-NEXT: bic r3, r1, #3
1409; CHECK-NEXT: movs r2, #1
1410; CHECK-NEXT: subs r0, r3, #4
1411; CHECK-NEXT: add.w lr, r2, r0, lsr #2
1412; CHECK-NEXT: movs r0, #0
1413; CHECK-NEXT: mov r2, r12
1414; CHECK-NEXT: dls lr, lr
1415; CHECK-NEXT: .LBB14_3: @ %vector.body
1416; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1417; CHECK-NEXT: vldrw.u32 q0, [r2], #16
1418; CHECK-NEXT: movs r4, #0
1419; CHECK-NEXT: vmaxv.u32 r4, q0
1420; CHECK-NEXT: cmp r0, r4
1421; CHECK-NEXT: it ls
1422; CHECK-NEXT: movls r0, r4
1423; CHECK-NEXT: le lr, .LBB14_3
1424; CHECK-NEXT: @ %bb.4: @ %middle.block
1425; CHECK-NEXT: cmp r3, r1
1426; CHECK-NEXT: it eq
1427; CHECK-NEXT: popeq {r4, pc}
1428; CHECK-NEXT: .LBB14_5: @ %for.body.preheader1
1429; CHECK-NEXT: sub.w lr, r1, r3
1430; CHECK-NEXT: add.w r1, r12, r3, lsl #2
1431; CHECK-NEXT: dls lr, lr
1432; CHECK-NEXT: .LBB14_6: @ %for.body
1433; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1434; CHECK-NEXT: ldr r2, [r1], #4
1435; CHECK-NEXT: cmp r0, r2
1436; CHECK-NEXT: it ls
1437; CHECK-NEXT: movls r0, r2
1438; CHECK-NEXT: le lr, .LBB14_6
1439; CHECK-NEXT: @ %bb.7: @ %for.cond.cleanup
1440; CHECK-NEXT: pop {r4, pc}
1441; CHECK-NEXT: .LBB14_8:
1442; CHECK-NEXT: movs r0, #0
1443; CHECK-NEXT: pop {r4, pc}
1444entry:
1445 %cmp6 = icmp sgt i32 %n, 0
1446 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1447
1448for.body.preheader: ; preds = %entry
1449 %min.iters.check = icmp ult i32 %n, 4
1450 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1451
1452vector.ph: ; preds = %for.body.preheader
1453 %n.vec = and i32 %n, -4
1454 br label %vector.body
1455
1456vector.body: ; preds = %vector.body, %vector.ph
1457 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1458 %vec.phi = phi i32 [ 0, %vector.ph ], [ %3, %vector.body ]
1459 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1460 %1 = bitcast i32* %0 to <4 x i32>*
1461 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1462 %l5 = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> %wide.load)
1463 %2 = icmp ugt i32 %vec.phi, %l5
1464 %3 = select i1 %2, i32 %vec.phi, i32 %l5
1465 %index.next = add i32 %index, 4
1466 %4 = icmp eq i32 %index.next, %n.vec
1467 br i1 %4, label %middle.block, label %vector.body
1468
1469middle.block: ; preds = %vector.body
1470 %5 = phi i32 [ %3, %vector.body ]
1471 %cmp.n = icmp eq i32 %n.vec, %n
1472 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1473
1474for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1475 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1476 %r.07.ph = phi i32 [ 0, %for.body.preheader ], [ %5, %middle.block ]
1477 br label %for.body
1478
1479for.body: ; preds = %for.body.preheader1, %for.body
1480 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1481 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1482 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1483 %6 = load i32, i32* %arrayidx, align 4
1484 %c = icmp ugt i32 %r.07, %6
1485 %add = select i1 %c, i32 %r.07, i32 %6
1486 %inc = add nuw nsw i32 %i.08, 1
1487 %exitcond = icmp eq i32 %inc, %n
1488 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1489
1490for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1491 %r.0.lcssa = phi i32 [ 0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1492 ret i32 %r.0.lcssa
1493}
1494
1495define float @fmin_f32(float* nocapture readonly %x, i32 %n) {
1496; CHECK-LABEL: fmin_f32:
1497; CHECK: @ %bb.0: @ %entry
1498; CHECK-NEXT: .save {r7, lr}
1499; CHECK-NEXT: push {r7, lr}
1500; CHECK-NEXT: cmp r1, #1
1501; CHECK-NEXT: blt .LBB15_3
1502; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1503; CHECK-NEXT: cmp r1, #4
1504; CHECK-NEXT: bhs .LBB15_4
1505; CHECK-NEXT: @ %bb.2:
1506; CHECK-NEXT: vldr s0, .LCPI15_0
1507; CHECK-NEXT: movs r2, #0
1508; CHECK-NEXT: b .LBB15_7
1509; CHECK-NEXT: .LBB15_3:
1510; CHECK-NEXT: vldr s0, .LCPI15_0
1511; CHECK-NEXT: b .LBB15_9
1512; CHECK-NEXT: .LBB15_4: @ %vector.ph
1513; CHECK-NEXT: bic r2, r1, #3
1514; CHECK-NEXT: movs r3, #1
1515; CHECK-NEXT: sub.w r12, r2, #4
1516; CHECK-NEXT: vmov.i32 q0, #0x0
1517; CHECK-NEXT: add.w lr, r3, r12, lsr #2
1518; CHECK-NEXT: mov r3, r0
1519; CHECK-NEXT: dls lr, lr
1520; CHECK-NEXT: .LBB15_5: @ %vector.body
1521; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1522; CHECK-NEXT: vldrw.u32 q1, [r3], #16
1523; CHECK-NEXT: vcmp.f32 lt, q0, q1
1524; CHECK-NEXT: vpsel q0, q0, q1
1525; CHECK-NEXT: le lr, .LBB15_5
1526; CHECK-NEXT: @ %bb.6: @ %middle.block
1527; CHECK-NEXT: vmov.f32 s4, s2
1528; CHECK-NEXT: cmp r2, r1
1529; CHECK-NEXT: vmov.f32 s5, s3
1530; CHECK-NEXT: vminnm.f32 q0, q0, q1
1531; CHECK-NEXT: vmov r3, s1
1532; CHECK-NEXT: vdup.32 q1, r3
1533; CHECK-NEXT: vminnm.f32 q0, q0, q1
1534; CHECK-NEXT: beq .LBB15_9
1535; CHECK-NEXT: .LBB15_7: @ %for.body.preheader1
1536; CHECK-NEXT: sub.w lr, r1, r2
1537; CHECK-NEXT: add.w r0, r0, r2, lsl #2
1538; CHECK-NEXT: dls lr, lr
1539; CHECK-NEXT: .LBB15_8: @ %for.body
1540; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1541; CHECK-NEXT: vldmia r0!, {s4}
1542; CHECK-NEXT: vcmp.f32 s0, s4
1543; CHECK-NEXT: vmrs APSR_nzcv, fpscr
1544; CHECK-NEXT: vselge.f32 s0, s4, s0
1545; CHECK-NEXT: le lr, .LBB15_8
1546; CHECK-NEXT: .LBB15_9: @ %for.cond.cleanup
1547; CHECK-NEXT: vmov r0, s0
1548; CHECK-NEXT: pop {r7, pc}
1549; CHECK-NEXT: .p2align 2
1550; CHECK-NEXT: @ %bb.10:
1551; CHECK-NEXT: .LCPI15_0:
1552; CHECK-NEXT: .long 0x00000000 @ float 0
1553entry:
1554 %cmp6 = icmp sgt i32 %n, 0
1555 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1556
1557for.body.preheader: ; preds = %entry
1558 %min.iters.check = icmp ult i32 %n, 4
1559 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1560
1561vector.ph: ; preds = %for.body.preheader
1562 %n.vec = and i32 %n, -4
1563 br label %vector.body
1564
1565vector.body: ; preds = %vector.body, %vector.ph
1566 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1567 %vec.phi = phi <4 x float> [ zeroinitializer, %vector.ph ], [ %3, %vector.body ]
1568 %0 = getelementptr inbounds float, float* %x, i32 %index
1569 %1 = bitcast float* %0 to <4 x float>*
1570 %wide.load = load <4 x float>, <4 x float>* %1, align 4
1571 %2 = fcmp ult <4 x float> %vec.phi, %wide.load
1572 %3 = select <4 x i1> %2, <4 x float> %vec.phi, <4 x float> %wide.load
1573 %index.next = add i32 %index, 4
1574 %4 = icmp eq i32 %index.next, %n.vec
1575 br i1 %4, label %middle.block, label %vector.body
1576
1577middle.block: ; preds = %vector.body
1578 %5 = call float @llvm.experimental.vector.reduce.fmin.v4f32(<4 x float> %3)
1579 %cmp.n = icmp eq i32 %n.vec, %n
1580 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1581
1582for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1583 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1584 %r.07.ph = phi float [ 0.0, %for.body.preheader ], [ %5, %middle.block ]
1585 br label %for.body
1586
1587for.body: ; preds = %for.body.preheader1, %for.body
1588 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1589 %r.07 = phi float [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1590 %arrayidx = getelementptr inbounds float, float* %x, i32 %i.08
1591 %6 = load float, float* %arrayidx, align 4
1592 %c = fcmp ult float %r.07, %6
1593 %add = select i1 %c, float %r.07, float %6
1594 %inc = add nuw nsw i32 %i.08, 1
1595 %exitcond = icmp eq i32 %inc, %n
1596 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1597
1598for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1599 %r.0.lcssa = phi float [ 0.0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1600 ret float %r.0.lcssa
1601}
1602
1603define float @fmax_f32(float* nocapture readonly %x, i32 %n) {
1604; CHECK-LABEL: fmax_f32:
1605; CHECK: @ %bb.0: @ %entry
1606; CHECK-NEXT: .save {r7, lr}
1607; CHECK-NEXT: push {r7, lr}
1608; CHECK-NEXT: cmp r1, #1
1609; CHECK-NEXT: blt .LBB16_3
1610; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1611; CHECK-NEXT: cmp r1, #4
1612; CHECK-NEXT: bhs .LBB16_4
1613; CHECK-NEXT: @ %bb.2:
1614; CHECK-NEXT: vldr s0, .LCPI16_0
1615; CHECK-NEXT: movs r2, #0
1616; CHECK-NEXT: b .LBB16_7
1617; CHECK-NEXT: .LBB16_3:
1618; CHECK-NEXT: vldr s0, .LCPI16_0
1619; CHECK-NEXT: b .LBB16_9
1620; CHECK-NEXT: .LBB16_4: @ %vector.ph
1621; CHECK-NEXT: bic r2, r1, #3
1622; CHECK-NEXT: movs r3, #1
1623; CHECK-NEXT: sub.w r12, r2, #4
1624; CHECK-NEXT: vmov.i32 q0, #0x0
1625; CHECK-NEXT: add.w lr, r3, r12, lsr #2
1626; CHECK-NEXT: mov r3, r0
1627; CHECK-NEXT: dls lr, lr
1628; CHECK-NEXT: .LBB16_5: @ %vector.body
1629; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1630; CHECK-NEXT: vldrw.u32 q1, [r3], #16
1631; CHECK-NEXT: vcmp.f32 lt, q1, q0
1632; CHECK-NEXT: vpsel q0, q0, q1
1633; CHECK-NEXT: le lr, .LBB16_5
1634; CHECK-NEXT: @ %bb.6: @ %middle.block
1635; CHECK-NEXT: vmov.f32 s4, s2
1636; CHECK-NEXT: cmp r2, r1
1637; CHECK-NEXT: vmov.f32 s5, s3
1638; CHECK-NEXT: vmaxnm.f32 q0, q0, q1
1639; CHECK-NEXT: vmov r3, s1
1640; CHECK-NEXT: vdup.32 q1, r3
1641; CHECK-NEXT: vmaxnm.f32 q0, q0, q1
1642; CHECK-NEXT: beq .LBB16_9
1643; CHECK-NEXT: .LBB16_7: @ %for.body.preheader1
1644; CHECK-NEXT: sub.w lr, r1, r2
1645; CHECK-NEXT: add.w r0, r0, r2, lsl #2
1646; CHECK-NEXT: dls lr, lr
1647; CHECK-NEXT: .LBB16_8: @ %for.body
1648; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1649; CHECK-NEXT: vldmia r0!, {s4}
1650; CHECK-NEXT: vcmp.f32 s4, s0
1651; CHECK-NEXT: vmrs APSR_nzcv, fpscr
1652; CHECK-NEXT: vselge.f32 s0, s4, s0
1653; CHECK-NEXT: le lr, .LBB16_8
1654; CHECK-NEXT: .LBB16_9: @ %for.cond.cleanup
1655; CHECK-NEXT: vmov r0, s0
1656; CHECK-NEXT: pop {r7, pc}
1657; CHECK-NEXT: .p2align 2
1658; CHECK-NEXT: @ %bb.10:
1659; CHECK-NEXT: .LCPI16_0:
1660; CHECK-NEXT: .long 0x00000000 @ float 0
1661entry:
1662 %cmp6 = icmp sgt i32 %n, 0
1663 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1664
1665for.body.preheader: ; preds = %entry
1666 %min.iters.check = icmp ult i32 %n, 4
1667 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1668
1669vector.ph: ; preds = %for.body.preheader
1670 %n.vec = and i32 %n, -4
1671 br label %vector.body
1672
1673vector.body: ; preds = %vector.body, %vector.ph
1674 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1675 %vec.phi = phi <4 x float> [ zeroinitializer, %vector.ph ], [ %3, %vector.body ]
1676 %0 = getelementptr inbounds float, float* %x, i32 %index
1677 %1 = bitcast float* %0 to <4 x float>*
1678 %wide.load = load <4 x float>, <4 x float>* %1, align 4
1679 %2 = fcmp ugt <4 x float> %vec.phi, %wide.load
1680 %3 = select <4 x i1> %2, <4 x float> %vec.phi, <4 x float> %wide.load
1681 %index.next = add i32 %index, 4
1682 %4 = icmp eq i32 %index.next, %n.vec
1683 br i1 %4, label %middle.block, label %vector.body
1684
1685middle.block: ; preds = %vector.body
1686 %5 = call float @llvm.experimental.vector.reduce.fmax.v4f32(<4 x float> %3)
1687 %cmp.n = icmp eq i32 %n.vec, %n
1688 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1689
1690for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1691 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1692 %r.07.ph = phi float [ 0.0, %for.body.preheader ], [ %5, %middle.block ]
1693 br label %for.body
1694
1695for.body: ; preds = %for.body.preheader1, %for.body
1696 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1697 %r.07 = phi float [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1698 %arrayidx = getelementptr inbounds float, float* %x, i32 %i.08
1699 %6 = load float, float* %arrayidx, align 4
1700 %c = fcmp ugt float %r.07, %6
1701 %add = select i1 %c, float %r.07, float %6
1702 %inc = add nuw nsw i32 %i.08, 1
1703 %exitcond = icmp eq i32 %inc, %n
1704 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1705
1706for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1707 %r.0.lcssa = phi float [ 0.0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1708 ret float %r.0.lcssa
1709}
1710
1711declare i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32>)
1712declare i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32>)
1713declare i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32>)
1714declare i32 @llvm.experimental.vector.reduce.or.v4i32(<4 x i32>)
1715declare i32 @llvm.experimental.vector.reduce.xor.v4i32(<4 x i32>)
1716declare float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float, <4 x float>)
1717declare float @llvm.experimental.vector.reduce.v2.fmul.f32.v4f32(float, <4 x float>)
1718declare i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32>)
1719declare i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32>)
1720declare i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32>)
1721declare i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32>)
1722declare float @llvm.experimental.vector.reduce.fmin.v4f32(<4 x float>)
1723declare float @llvm.experimental.vector.reduce.fmax.v4f32(<4 x float>)