Gitiles
Code Review Sign In
gerrit-public.fairphone.software / toolchain / llvm-project / 186a7f81e813a59b68fadfe7ea90c348bb760c56 / . / llvm / test / CodeGen / Thumb2 / mve-vecreduce-loops.ll
blob: 539f760642c26b8e7189d38f2447dd9d73f9dfc7 [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
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp -tail-predication=enabled -verify-machineinstrs %s -o - | FileCheck %s
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]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
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]735; CHECK-NEXT: csel r2, r2, r1, lt
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]736; CHECK-NEXT: le lr, .LBB7_8
737; CHECK-NEXT: .LBB7_9: @ %for.cond.cleanup
738; CHECK-NEXT: mov r0, r2
739; CHECK-NEXT: pop {r7, pc}
740entry:
741 %cmp6 = icmp sgt i32 %n, 0
742 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
743
744for.body.preheader: ; preds = %entry
745 %min.iters.check = icmp ult i32 %n, 4
746 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
747
748vector.ph: ; preds = %for.body.preheader
749 %n.vec = and i32 %n, -4
750 br label %vector.body
751
752vector.body: ; preds = %vector.body, %vector.ph
753 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
754 %vec.phi = phi <4 x i32> [ <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647>, %vector.ph ], [ %3, %vector.body ]
755 %0 = getelementptr inbounds i32, i32* %x, i32 %index
756 %1 = bitcast i32* %0 to <4 x i32>*
757 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
758 %2 = icmp slt <4 x i32> %vec.phi, %wide.load
759 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
760 %index.next = add i32 %index, 4
761 %4 = icmp eq i32 %index.next, %n.vec
762 br i1 %4, label %middle.block, label %vector.body
763
764middle.block: ; preds = %vector.body
765 %5 = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> %3)
766 %cmp.n = icmp eq i32 %n.vec, %n
767 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
768
769for.body.preheader1: ; preds = %middle.block, %for.body.preheader
770 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
771 %r.07.ph = phi i32 [ 2147483647, %for.body.preheader ], [ %5, %middle.block ]
772 br label %for.body
773
774for.body: ; preds = %for.body.preheader1, %for.body
775 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
776 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
777 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
778 %6 = load i32, i32* %arrayidx, align 4
779 %c = icmp slt i32 %r.07, %6
780 %add = select i1 %c, i32 %r.07, i32 %6
781 %inc = add nuw nsw i32 %i.08, 1
782 %exitcond = icmp eq i32 %inc, %n
783 br i1 %exitcond, label %for.cond.cleanup, label %for.body
784
785for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
786 %r.0.lcssa = phi i32 [ 2147483647, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
787 ret i32 %r.0.lcssa
788}
789
790define i32 @smin_i32_inloop(i32* nocapture readonly %x, i32 %n) {
791; CHECK-LABEL: smin_i32_inloop:
792; CHECK: @ %bb.0: @ %entry
793; CHECK-NEXT: .save {r4, lr}
794; CHECK-NEXT: push {r4, lr}
795; CHECK-NEXT: cmp r1, #1
796; CHECK-NEXT: blt .LBB8_3
797; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
798; CHECK-NEXT: mov r12, r0
799; CHECK-NEXT: mvn r0, #-2147483648
800; CHECK-NEXT: cmp r1, #4
801; CHECK-NEXT: bhs .LBB8_4
802; CHECK-NEXT: @ %bb.2:
803; CHECK-NEXT: movs r3, #0
804; CHECK-NEXT: b .LBB8_7
805; CHECK-NEXT: .LBB8_3:
806; CHECK-NEXT: mvn r0, #-2147483648
807; CHECK-NEXT: b .LBB8_9
808; CHECK-NEXT: .LBB8_4: @ %vector.ph
809; CHECK-NEXT: bic r3, r1, #3
810; CHECK-NEXT: movs r2, #1
811; CHECK-NEXT: sub.w lr, r3, #4
812; CHECK-NEXT: add.w lr, r2, lr, lsr #2
813; CHECK-NEXT: mov r2, r12
814; CHECK-NEXT: dls lr, lr
815; CHECK-NEXT: .LBB8_5: @ %vector.body
816; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
817; CHECK-NEXT: vldrw.u32 q0, [r2], #16
818; CHECK-NEXT: mvn r4, #-2147483648
819; CHECK-NEXT: vminv.s32 r4, q0
820; CHECK-NEXT: cmp r0, r4
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]821; CHECK-NEXT: csel r0, r0, r4, lt
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]822; CHECK-NEXT: le lr, .LBB8_5
823; CHECK-NEXT: @ %bb.6: @ %middle.block
824; CHECK-NEXT: cmp r3, r1
825; CHECK-NEXT: it eq
826; CHECK-NEXT: popeq {r4, pc}
827; CHECK-NEXT: .LBB8_7: @ %for.body.preheader1
828; CHECK-NEXT: sub.w lr, r1, r3
829; CHECK-NEXT: add.w r1, r12, r3, lsl #2
830; CHECK-NEXT: dls lr, lr
831; CHECK-NEXT: .LBB8_8: @ %for.body
832; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
833; CHECK-NEXT: ldr r2, [r1], #4
834; CHECK-NEXT: cmp r0, r2
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]835; CHECK-NEXT: csel r0, r0, r2, lt
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]836; CHECK-NEXT: le lr, .LBB8_8
837; CHECK-NEXT: .LBB8_9: @ %for.cond.cleanup
838; CHECK-NEXT: pop {r4, pc}
839entry:
840 %cmp6 = icmp sgt i32 %n, 0
841 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
842
843for.body.preheader: ; preds = %entry
844 %min.iters.check = icmp ult i32 %n, 4
845 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
846
847vector.ph: ; preds = %for.body.preheader
848 %n.vec = and i32 %n, -4
849 br label %vector.body
850
851vector.body: ; preds = %vector.body, %vector.ph
852 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
853 %vec.phi = phi i32 [ 2147483647, %vector.ph ], [ %3, %vector.body ]
854 %0 = getelementptr inbounds i32, i32* %x, i32 %index
855 %1 = bitcast i32* %0 to <4 x i32>*
856 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
857 %l5 = call i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32> %wide.load)
858 %2 = icmp slt i32 %vec.phi, %l5
859 %3 = select i1 %2, i32 %vec.phi, i32 %l5
860 %index.next = add i32 %index, 4
861 %4 = icmp eq i32 %index.next, %n.vec
862 br i1 %4, label %middle.block, label %vector.body
863
864middle.block: ; preds = %vector.body
865 %5 = phi i32 [ %3, %vector.body ]
866 %cmp.n = icmp eq i32 %n.vec, %n
867 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
868
869for.body.preheader1: ; preds = %middle.block, %for.body.preheader
870 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
871 %r.07.ph = phi i32 [ 2147483647, %for.body.preheader ], [ %5, %middle.block ]
872 br label %for.body
873
874for.body: ; preds = %for.body.preheader1, %for.body
875 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
876 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
877 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
878 %6 = load i32, i32* %arrayidx, align 4
879 %c = icmp slt i32 %r.07, %6
880 %add = select i1 %c, i32 %r.07, i32 %6
881 %inc = add nuw nsw i32 %i.08, 1
882 %exitcond = icmp eq i32 %inc, %n
883 br i1 %exitcond, label %for.cond.cleanup, label %for.body
884
885for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
886 %r.0.lcssa = phi i32 [ 2147483647, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
887 ret i32 %r.0.lcssa
888}
889
890define i32 @smax_i32(i32* nocapture readonly %x, i32 %n) {
891; CHECK-LABEL: smax_i32:
892; CHECK: @ %bb.0: @ %entry
893; CHECK-NEXT: .save {r7, lr}
894; CHECK-NEXT: push {r7, lr}
895; CHECK-NEXT: cmp r1, #1
896; CHECK-NEXT: blt .LBB9_3
897; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
898; CHECK-NEXT: cmp r1, #4
899; CHECK-NEXT: bhs .LBB9_4
900; CHECK-NEXT: @ %bb.2:
901; CHECK-NEXT: mov.w r2, #-2147483648
902; CHECK-NEXT: movs r3, #0
903; CHECK-NEXT: b .LBB9_7
904; CHECK-NEXT: .LBB9_3:
905; CHECK-NEXT: mov.w r2, #-2147483648
906; CHECK-NEXT: b .LBB9_9
907; CHECK-NEXT: .LBB9_4: @ %vector.ph
908; CHECK-NEXT: bic r3, r1, #3
909; CHECK-NEXT: movs r2, #1
910; CHECK-NEXT: sub.w r12, r3, #4
911; CHECK-NEXT: vmov.i32 q0, #0x80000000
912; CHECK-NEXT: add.w lr, r2, r12, lsr #2
913; CHECK-NEXT: mov r2, r0
914; CHECK-NEXT: dls lr, lr
915; CHECK-NEXT: .LBB9_5: @ %vector.body
916; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
917; CHECK-NEXT: vldrw.u32 q1, [r2], #16
918; CHECK-NEXT: vmax.s32 q0, q0, q1
919; CHECK-NEXT: le lr, .LBB9_5
920; CHECK-NEXT: @ %bb.6: @ %middle.block
921; CHECK-NEXT: mov.w r2, #-2147483648
922; CHECK-NEXT: cmp r3, r1
923; CHECK-NEXT: vmaxv.s32 r2, q0
924; CHECK-NEXT: beq .LBB9_9
925; CHECK-NEXT: .LBB9_7: @ %for.body.preheader1
926; CHECK-NEXT: sub.w lr, r1, r3
927; CHECK-NEXT: add.w r0, r0, r3, lsl #2
928; CHECK-NEXT: dls lr, lr
929; CHECK-NEXT: .LBB9_8: @ %for.body
930; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
931; CHECK-NEXT: ldr r1, [r0], #4
932; CHECK-NEXT: cmp r2, r1
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]933; CHECK-NEXT: csel r2, r2, r1, gt
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]934; CHECK-NEXT: le lr, .LBB9_8
935; CHECK-NEXT: .LBB9_9: @ %for.cond.cleanup
936; CHECK-NEXT: mov r0, r2
937; CHECK-NEXT: pop {r7, pc}
938entry:
939 %cmp6 = icmp sgt i32 %n, 0
940 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
941
942for.body.preheader: ; preds = %entry
943 %min.iters.check = icmp ult i32 %n, 4
944 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
945
946vector.ph: ; preds = %for.body.preheader
947 %n.vec = and i32 %n, -4
948 br label %vector.body
949
950vector.body: ; preds = %vector.body, %vector.ph
951 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
952 %vec.phi = phi <4 x i32> [ <i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648>, %vector.ph ], [ %3, %vector.body ]
953 %0 = getelementptr inbounds i32, i32* %x, i32 %index
954 %1 = bitcast i32* %0 to <4 x i32>*
955 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
956 %2 = icmp sgt <4 x i32> %vec.phi, %wide.load
957 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
958 %index.next = add i32 %index, 4
959 %4 = icmp eq i32 %index.next, %n.vec
960 br i1 %4, label %middle.block, label %vector.body
961
962middle.block: ; preds = %vector.body
963 %5 = call i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32> %3)
964 %cmp.n = icmp eq i32 %n.vec, %n
965 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
966
967for.body.preheader1: ; preds = %middle.block, %for.body.preheader
968 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
969 %r.07.ph = phi i32 [ -2147483648, %for.body.preheader ], [ %5, %middle.block ]
970 br label %for.body
971
972for.body: ; preds = %for.body.preheader1, %for.body
973 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
974 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
975 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
976 %6 = load i32, i32* %arrayidx, align 4
977 %c = icmp sgt i32 %r.07, %6
978 %add = select i1 %c, i32 %r.07, i32 %6
979 %inc = add nuw nsw i32 %i.08, 1
980 %exitcond = icmp eq i32 %inc, %n
981 br i1 %exitcond, label %for.cond.cleanup, label %for.body
982
983for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
984 %r.0.lcssa = phi i32 [ -2147483648, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
985 ret i32 %r.0.lcssa
986}
987
988define i32 @smax_i32_inloop(i32* nocapture readonly %x, i32 %n) {
989; CHECK-LABEL: smax_i32_inloop:
990; CHECK: @ %bb.0: @ %entry
991; CHECK-NEXT: .save {r4, lr}
992; CHECK-NEXT: push {r4, lr}
993; CHECK-NEXT: cmp r1, #1
994; CHECK-NEXT: blt .LBB10_3
995; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
996; CHECK-NEXT: mov r12, r0
997; CHECK-NEXT: mov.w r0, #-2147483648
998; CHECK-NEXT: cmp r1, #4
999; CHECK-NEXT: bhs .LBB10_4
1000; CHECK-NEXT: @ %bb.2:
1001; CHECK-NEXT: movs r3, #0
1002; CHECK-NEXT: b .LBB10_7
1003; CHECK-NEXT: .LBB10_3:
1004; CHECK-NEXT: mov.w r0, #-2147483648
1005; CHECK-NEXT: b .LBB10_9
1006; CHECK-NEXT: .LBB10_4: @ %vector.ph
1007; CHECK-NEXT: bic r3, r1, #3
1008; CHECK-NEXT: movs r2, #1
1009; CHECK-NEXT: sub.w lr, r3, #4
1010; CHECK-NEXT: add.w lr, r2, lr, lsr #2
1011; CHECK-NEXT: mov r2, r12
1012; CHECK-NEXT: dls lr, lr
1013; CHECK-NEXT: .LBB10_5: @ %vector.body
1014; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1015; CHECK-NEXT: vldrw.u32 q0, [r2], #16
1016; CHECK-NEXT: mov.w r4, #-2147483648
1017; CHECK-NEXT: vmaxv.s32 r4, q0
1018; CHECK-NEXT: cmp r0, r4
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1019; CHECK-NEXT: csel r0, r0, r4, gt
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1020; CHECK-NEXT: le lr, .LBB10_5
1021; CHECK-NEXT: @ %bb.6: @ %middle.block
1022; CHECK-NEXT: cmp r3, r1
1023; CHECK-NEXT: it eq
1024; CHECK-NEXT: popeq {r4, pc}
1025; CHECK-NEXT: .LBB10_7: @ %for.body.preheader1
1026; CHECK-NEXT: sub.w lr, r1, r3
1027; CHECK-NEXT: add.w r1, r12, r3, lsl #2
1028; CHECK-NEXT: dls lr, lr
1029; CHECK-NEXT: .LBB10_8: @ %for.body
1030; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1031; CHECK-NEXT: ldr r2, [r1], #4
1032; CHECK-NEXT: cmp r0, r2
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1033; CHECK-NEXT: csel r0, r0, r2, gt
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1034; CHECK-NEXT: le lr, .LBB10_8
1035; CHECK-NEXT: .LBB10_9: @ %for.cond.cleanup
1036; CHECK-NEXT: pop {r4, pc}
1037entry:
1038 %cmp6 = icmp sgt i32 %n, 0
1039 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1040
1041for.body.preheader: ; preds = %entry
1042 %min.iters.check = icmp ult i32 %n, 4
1043 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1044
1045vector.ph: ; preds = %for.body.preheader
1046 %n.vec = and i32 %n, -4
1047 br label %vector.body
1048
1049vector.body: ; preds = %vector.body, %vector.ph
1050 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1051 %vec.phi = phi i32 [ -2147483648, %vector.ph ], [ %3, %vector.body ]
1052 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1053 %1 = bitcast i32* %0 to <4 x i32>*
1054 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1055 %l5 = call i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32> %wide.load)
1056 %2 = icmp sgt i32 %vec.phi, %l5
1057 %3 = select i1 %2, i32 %vec.phi, i32 %l5
1058 %index.next = add i32 %index, 4
1059 %4 = icmp eq i32 %index.next, %n.vec
1060 br i1 %4, label %middle.block, label %vector.body
1061
1062middle.block: ; preds = %vector.body
1063 %5 = phi i32 [ %3, %vector.body ]
1064 %cmp.n = icmp eq i32 %n.vec, %n
1065 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1066
1067for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1068 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1069 %r.07.ph = phi i32 [ -2147483648, %for.body.preheader ], [ %5, %middle.block ]
1070 br label %for.body
1071
1072for.body: ; preds = %for.body.preheader1, %for.body
1073 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1074 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1075 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1076 %6 = load i32, i32* %arrayidx, align 4
1077 %c = icmp sgt i32 %r.07, %6
1078 %add = select i1 %c, i32 %r.07, i32 %6
1079 %inc = add nuw nsw i32 %i.08, 1
1080 %exitcond = icmp eq i32 %inc, %n
1081 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1082
1083for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1084 %r.0.lcssa = phi i32 [ -2147483648, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1085 ret i32 %r.0.lcssa
1086}
1087
1088define i32 @umin_i32(i32* nocapture readonly %x, i32 %n) {
1089; CHECK-LABEL: umin_i32:
1090; CHECK: @ %bb.0: @ %entry
1091; CHECK-NEXT: .save {r7, lr}
1092; CHECK-NEXT: push {r7, lr}
1093; CHECK-NEXT: cmp r1, #1
1094; CHECK-NEXT: blt .LBB11_3
1095; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1096; CHECK-NEXT: cmp r1, #4
1097; CHECK-NEXT: bhs .LBB11_4
1098; CHECK-NEXT: @ %bb.2:
1099; CHECK-NEXT: mov.w r2, #-1
1100; CHECK-NEXT: movs r3, #0
1101; CHECK-NEXT: b .LBB11_7
1102; CHECK-NEXT: .LBB11_3:
1103; CHECK-NEXT: mov.w r2, #-1
1104; CHECK-NEXT: b .LBB11_9
1105; CHECK-NEXT: .LBB11_4: @ %vector.ph
1106; CHECK-NEXT: bic r3, r1, #3
1107; CHECK-NEXT: movs r2, #1
1108; CHECK-NEXT: sub.w r12, r3, #4
1109; CHECK-NEXT: vmov.i8 q0, #0xff
1110; CHECK-NEXT: add.w lr, r2, r12, lsr #2
1111; CHECK-NEXT: mov r2, r0
1112; CHECK-NEXT: dls lr, lr
1113; CHECK-NEXT: .LBB11_5: @ %vector.body
1114; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1115; CHECK-NEXT: vldrw.u32 q1, [r2], #16
1116; CHECK-NEXT: vmin.u32 q0, q0, q1
1117; CHECK-NEXT: le lr, .LBB11_5
1118; CHECK-NEXT: @ %bb.6: @ %middle.block
1119; CHECK-NEXT: mov.w r2, #-1
1120; CHECK-NEXT: cmp r3, r1
1121; CHECK-NEXT: vminv.u32 r2, q0
1122; CHECK-NEXT: beq .LBB11_9
1123; CHECK-NEXT: .LBB11_7: @ %for.body.preheader1
1124; CHECK-NEXT: sub.w lr, r1, r3
1125; CHECK-NEXT: add.w r0, r0, r3, lsl #2
1126; CHECK-NEXT: dls lr, lr
1127; CHECK-NEXT: .LBB11_8: @ %for.body
1128; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1129; CHECK-NEXT: ldr r1, [r0], #4
1130; CHECK-NEXT: cmp r2, r1
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1131; CHECK-NEXT: csel r2, r2, r1, lo
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1132; CHECK-NEXT: le lr, .LBB11_8
1133; CHECK-NEXT: .LBB11_9: @ %for.cond.cleanup
1134; CHECK-NEXT: mov r0, r2
1135; CHECK-NEXT: pop {r7, pc}
1136entry:
1137 %cmp6 = icmp sgt i32 %n, 0
1138 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1139
1140for.body.preheader: ; preds = %entry
1141 %min.iters.check = icmp ult i32 %n, 4
1142 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1143
1144vector.ph: ; preds = %for.body.preheader
1145 %n.vec = and i32 %n, -4
1146 br label %vector.body
1147
1148vector.body: ; preds = %vector.body, %vector.ph
1149 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1150 %vec.phi = phi <4 x i32> [ <i32 -1, i32 -1, i32 -1, i32 -1>, %vector.ph ], [ %3, %vector.body ]
1151 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1152 %1 = bitcast i32* %0 to <4 x i32>*
1153 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1154 %2 = icmp ult <4 x i32> %vec.phi, %wide.load
1155 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
1156 %index.next = add i32 %index, 4
1157 %4 = icmp eq i32 %index.next, %n.vec
1158 br i1 %4, label %middle.block, label %vector.body
1159
1160middle.block: ; preds = %vector.body
1161 %5 = call i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32> %3)
1162 %cmp.n = icmp eq i32 %n.vec, %n
1163 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1164
1165for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1166 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1167 %r.07.ph = phi i32 [ -1, %for.body.preheader ], [ %5, %middle.block ]
1168 br label %for.body
1169
1170for.body: ; preds = %for.body.preheader1, %for.body
1171 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1172 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1173 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1174 %6 = load i32, i32* %arrayidx, align 4
1175 %c = icmp ult i32 %r.07, %6
1176 %add = select i1 %c, i32 %r.07, i32 %6
1177 %inc = add nuw nsw i32 %i.08, 1
1178 %exitcond = icmp eq i32 %inc, %n
1179 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1180
1181for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1182 %r.0.lcssa = phi i32 [ -1, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1183 ret i32 %r.0.lcssa
1184}
1185
1186define i32 @umin_i32_inloop(i32* nocapture readonly %x, i32 %n) {
1187; CHECK-LABEL: umin_i32_inloop:
1188; CHECK: @ %bb.0: @ %entry
1189; CHECK-NEXT: .save {r4, lr}
1190; CHECK-NEXT: push {r4, lr}
1191; CHECK-NEXT: cmp r1, #1
1192; CHECK-NEXT: blt .LBB12_3
1193; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1194; CHECK-NEXT: mov r12, r0
1195; CHECK-NEXT: mov.w r0, #-1
1196; CHECK-NEXT: cmp r1, #4
1197; CHECK-NEXT: bhs .LBB12_4
1198; CHECK-NEXT: @ %bb.2:
1199; CHECK-NEXT: movs r3, #0
1200; CHECK-NEXT: b .LBB12_7
1201; CHECK-NEXT: .LBB12_3:
1202; CHECK-NEXT: mov.w r0, #-1
1203; CHECK-NEXT: b .LBB12_9
1204; CHECK-NEXT: .LBB12_4: @ %vector.ph
1205; CHECK-NEXT: bic r3, r1, #3
1206; CHECK-NEXT: movs r2, #1
1207; CHECK-NEXT: sub.w lr, r3, #4
1208; CHECK-NEXT: add.w lr, r2, lr, lsr #2
1209; CHECK-NEXT: mov r2, r12
1210; CHECK-NEXT: dls lr, lr
1211; CHECK-NEXT: .LBB12_5: @ %vector.body
1212; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1213; CHECK-NEXT: vldrw.u32 q0, [r2], #16
1214; CHECK-NEXT: mov.w r4, #-1
1215; CHECK-NEXT: vminv.u32 r4, q0
1216; CHECK-NEXT: cmp r0, r4
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1217; CHECK-NEXT: csel r0, r0, r4, lo
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1218; CHECK-NEXT: le lr, .LBB12_5
1219; CHECK-NEXT: @ %bb.6: @ %middle.block
1220; CHECK-NEXT: cmp r3, r1
1221; CHECK-NEXT: it eq
1222; CHECK-NEXT: popeq {r4, pc}
1223; CHECK-NEXT: .LBB12_7: @ %for.body.preheader1
1224; CHECK-NEXT: sub.w lr, r1, r3
1225; CHECK-NEXT: add.w r1, r12, r3, lsl #2
1226; CHECK-NEXT: dls lr, lr
1227; CHECK-NEXT: .LBB12_8: @ %for.body
1228; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1229; CHECK-NEXT: ldr r2, [r1], #4
1230; CHECK-NEXT: cmp r0, r2
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1231; CHECK-NEXT: csel r0, r0, r2, hi
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1232; CHECK-NEXT: le lr, .LBB12_8
1233; CHECK-NEXT: .LBB12_9: @ %for.cond.cleanup
1234; CHECK-NEXT: pop {r4, pc}
1235entry:
1236 %cmp6 = icmp sgt i32 %n, 0
1237 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1238
1239for.body.preheader: ; preds = %entry
1240 %min.iters.check = icmp ult i32 %n, 4
1241 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1242
1243vector.ph: ; preds = %for.body.preheader
1244 %n.vec = and i32 %n, -4
1245 br label %vector.body
1246
1247vector.body: ; preds = %vector.body, %vector.ph
1248 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1249 %vec.phi = phi i32 [ -1, %vector.ph ], [ %3, %vector.body ]
1250 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1251 %1 = bitcast i32* %0 to <4 x i32>*
1252 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1253 %l5 = call i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32> %wide.load)
1254 %2 = icmp ult i32 %vec.phi, %l5
1255 %3 = select i1 %2, i32 %vec.phi, i32 %l5
1256 %index.next = add i32 %index, 4
1257 %4 = icmp eq i32 %index.next, %n.vec
1258 br i1 %4, label %middle.block, label %vector.body
1259
1260middle.block: ; preds = %vector.body
1261 %5 = phi i32 [ %3, %vector.body ]
1262 %cmp.n = icmp eq i32 %n.vec, %n
1263 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1264
1265for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1266 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1267 %r.07.ph = phi i32 [ -1, %for.body.preheader ], [ %5, %middle.block ]
1268 br label %for.body
1269
1270for.body: ; preds = %for.body.preheader1, %for.body
1271 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1272 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1273 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1274 %6 = load i32, i32* %arrayidx, align 4
1275 %c = icmp ugt i32 %r.07, %6
1276 %add = select i1 %c, i32 %r.07, i32 %6
1277 %inc = add nuw nsw i32 %i.08, 1
1278 %exitcond = icmp eq i32 %inc, %n
1279 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1280
1281for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1282 %r.0.lcssa = phi i32 [ -1, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1283 ret i32 %r.0.lcssa
1284}
1285
1286define i32 @umax_i32(i32* nocapture readonly %x, i32 %n) {
1287; CHECK-LABEL: umax_i32:
1288; CHECK: @ %bb.0: @ %entry
1289; CHECK-NEXT: .save {r7, lr}
1290; CHECK-NEXT: push {r7, lr}
1291; CHECK-NEXT: cmp r1, #1
1292; CHECK-NEXT: blt .LBB13_3
1293; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1294; CHECK-NEXT: cmp r1, #4
1295; CHECK-NEXT: bhs .LBB13_4
1296; CHECK-NEXT: @ %bb.2:
1297; CHECK-NEXT: movs r3, #0
1298; CHECK-NEXT: movs r2, #0
1299; CHECK-NEXT: b .LBB13_7
1300; CHECK-NEXT: .LBB13_3:
1301; CHECK-NEXT: movs r2, #0
1302; CHECK-NEXT: b .LBB13_9
1303; CHECK-NEXT: .LBB13_4: @ %vector.ph
1304; CHECK-NEXT: bic r3, r1, #3
1305; CHECK-NEXT: movs r2, #1
1306; CHECK-NEXT: sub.w r12, r3, #4
1307; CHECK-NEXT: vmov.i32 q0, #0x0
1308; CHECK-NEXT: add.w lr, r2, r12, lsr #2
1309; CHECK-NEXT: mov r2, r0
1310; CHECK-NEXT: dls lr, lr
1311; CHECK-NEXT: .LBB13_5: @ %vector.body
1312; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1313; CHECK-NEXT: vldrw.u32 q1, [r2], #16
1314; CHECK-NEXT: vmax.u32 q0, q0, q1
1315; CHECK-NEXT: le lr, .LBB13_5
1316; CHECK-NEXT: @ %bb.6: @ %middle.block
1317; CHECK-NEXT: movs r2, #0
1318; CHECK-NEXT: cmp r3, r1
1319; CHECK-NEXT: vmaxv.u32 r2, q0
1320; CHECK-NEXT: beq .LBB13_9
1321; CHECK-NEXT: .LBB13_7: @ %for.body.preheader1
1322; CHECK-NEXT: sub.w lr, r1, r3
1323; CHECK-NEXT: add.w r0, r0, r3, lsl #2
1324; CHECK-NEXT: dls lr, lr
1325; CHECK-NEXT: .LBB13_8: @ %for.body
1326; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1327; CHECK-NEXT: ldr r1, [r0], #4
1328; CHECK-NEXT: cmp r2, r1
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1329; CHECK-NEXT: csel r2, r2, r1, hi
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1330; CHECK-NEXT: le lr, .LBB13_8
1331; CHECK-NEXT: .LBB13_9: @ %for.cond.cleanup
1332; CHECK-NEXT: mov r0, r2
1333; CHECK-NEXT: pop {r7, pc}
1334entry:
1335 %cmp6 = icmp sgt i32 %n, 0
1336 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1337
1338for.body.preheader: ; preds = %entry
1339 %min.iters.check = icmp ult i32 %n, 4
1340 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1341
1342vector.ph: ; preds = %for.body.preheader
1343 %n.vec = and i32 %n, -4
1344 br label %vector.body
1345
1346vector.body: ; preds = %vector.body, %vector.ph
1347 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1348 %vec.phi = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %3, %vector.body ]
1349 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1350 %1 = bitcast i32* %0 to <4 x i32>*
1351 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1352 %2 = icmp ugt <4 x i32> %vec.phi, %wide.load
1353 %3 = select <4 x i1> %2, <4 x i32> %vec.phi, <4 x i32> %wide.load
1354 %index.next = add i32 %index, 4
1355 %4 = icmp eq i32 %index.next, %n.vec
1356 br i1 %4, label %middle.block, label %vector.body
1357
1358middle.block: ; preds = %vector.body
1359 %5 = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> %3)
1360 %cmp.n = icmp eq i32 %n.vec, %n
1361 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1362
1363for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1364 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1365 %r.07.ph = phi i32 [ 0, %for.body.preheader ], [ %5, %middle.block ]
1366 br label %for.body
1367
1368for.body: ; preds = %for.body.preheader1, %for.body
1369 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1370 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1371 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1372 %6 = load i32, i32* %arrayidx, align 4
1373 %c = icmp ugt i32 %r.07, %6
1374 %add = select i1 %c, i32 %r.07, i32 %6
1375 %inc = add nuw nsw i32 %i.08, 1
1376 %exitcond = icmp eq i32 %inc, %n
1377 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1378
1379for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1380 %r.0.lcssa = phi i32 [ 0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1381 ret i32 %r.0.lcssa
1382}
1383
1384define i32 @umax_i32_inloop(i32* nocapture readonly %x, i32 %n) {
1385; CHECK-LABEL: umax_i32_inloop:
1386; CHECK: @ %bb.0: @ %entry
1387; CHECK-NEXT: .save {r4, lr}
1388; CHECK-NEXT: push {r4, lr}
1389; CHECK-NEXT: cmp r1, #1
1390; CHECK-NEXT: blt .LBB14_8
1391; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1392; CHECK-NEXT: mov r12, r0
1393; CHECK-NEXT: movs r3, #0
1394; CHECK-NEXT: cmp r1, #4
1395; CHECK-NEXT: mov.w r0, #0
1396; CHECK-NEXT: blo .LBB14_5
1397; CHECK-NEXT: @ %bb.2: @ %vector.ph
1398; CHECK-NEXT: bic r3, r1, #3
1399; CHECK-NEXT: movs r2, #1
1400; CHECK-NEXT: subs r0, r3, #4
1401; CHECK-NEXT: add.w lr, r2, r0, lsr #2
1402; CHECK-NEXT: movs r0, #0
1403; CHECK-NEXT: mov r2, r12
1404; CHECK-NEXT: dls lr, lr
1405; CHECK-NEXT: .LBB14_3: @ %vector.body
1406; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1407; CHECK-NEXT: vldrw.u32 q0, [r2], #16
1408; CHECK-NEXT: movs r4, #0
1409; CHECK-NEXT: vmaxv.u32 r4, q0
1410; CHECK-NEXT: cmp r0, r4
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1411; CHECK-NEXT: csel r0, r0, r4, hi
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1412; CHECK-NEXT: le lr, .LBB14_3
1413; CHECK-NEXT: @ %bb.4: @ %middle.block
1414; CHECK-NEXT: cmp r3, r1
1415; CHECK-NEXT: it eq
1416; CHECK-NEXT: popeq {r4, pc}
1417; CHECK-NEXT: .LBB14_5: @ %for.body.preheader1
1418; CHECK-NEXT: sub.w lr, r1, r3
1419; CHECK-NEXT: add.w r1, r12, r3, lsl #2
1420; CHECK-NEXT: dls lr, lr
1421; CHECK-NEXT: .LBB14_6: @ %for.body
1422; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1423; CHECK-NEXT: ldr r2, [r1], #4
1424; CHECK-NEXT: cmp r0, r2
David Green146d35b2020-07-14 10:04:55 +0100[diff] [blame]1425; CHECK-NEXT: csel r0, r0, r2, hi
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]1426; CHECK-NEXT: le lr, .LBB14_6
1427; CHECK-NEXT: @ %bb.7: @ %for.cond.cleanup
1428; CHECK-NEXT: pop {r4, pc}
1429; CHECK-NEXT: .LBB14_8:
1430; CHECK-NEXT: movs r0, #0
1431; CHECK-NEXT: pop {r4, pc}
1432entry:
1433 %cmp6 = icmp sgt i32 %n, 0
1434 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1435
1436for.body.preheader: ; preds = %entry
1437 %min.iters.check = icmp ult i32 %n, 4
1438 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1439
1440vector.ph: ; preds = %for.body.preheader
1441 %n.vec = and i32 %n, -4
1442 br label %vector.body
1443
1444vector.body: ; preds = %vector.body, %vector.ph
1445 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1446 %vec.phi = phi i32 [ 0, %vector.ph ], [ %3, %vector.body ]
1447 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1448 %1 = bitcast i32* %0 to <4 x i32>*
1449 %wide.load = load <4 x i32>, <4 x i32>* %1, align 4
1450 %l5 = call i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32> %wide.load)
1451 %2 = icmp ugt i32 %vec.phi, %l5
1452 %3 = select i1 %2, i32 %vec.phi, i32 %l5
1453 %index.next = add i32 %index, 4
1454 %4 = icmp eq i32 %index.next, %n.vec
1455 br i1 %4, label %middle.block, label %vector.body
1456
1457middle.block: ; preds = %vector.body
1458 %5 = phi i32 [ %3, %vector.body ]
1459 %cmp.n = icmp eq i32 %n.vec, %n
1460 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1461
1462for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1463 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1464 %r.07.ph = phi i32 [ 0, %for.body.preheader ], [ %5, %middle.block ]
1465 br label %for.body
1466
1467for.body: ; preds = %for.body.preheader1, %for.body
1468 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1469 %r.07 = phi i32 [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1470 %arrayidx = getelementptr inbounds i32, i32* %x, i32 %i.08
1471 %6 = load i32, i32* %arrayidx, align 4
1472 %c = icmp ugt i32 %r.07, %6
1473 %add = select i1 %c, i32 %r.07, i32 %6
1474 %inc = add nuw nsw i32 %i.08, 1
1475 %exitcond = icmp eq i32 %inc, %n
1476 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1477
1478for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1479 %r.0.lcssa = phi i32 [ 0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1480 ret i32 %r.0.lcssa
1481}
1482
1483define float @fmin_f32(float* nocapture readonly %x, i32 %n) {
1484; CHECK-LABEL: fmin_f32:
1485; CHECK: @ %bb.0: @ %entry
1486; CHECK-NEXT: .save {r7, lr}
1487; CHECK-NEXT: push {r7, lr}
1488; CHECK-NEXT: cmp r1, #1
1489; CHECK-NEXT: blt .LBB15_3
1490; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1491; CHECK-NEXT: cmp r1, #4
1492; CHECK-NEXT: bhs .LBB15_4
1493; CHECK-NEXT: @ %bb.2:
1494; CHECK-NEXT: vldr s0, .LCPI15_0
1495; CHECK-NEXT: movs r2, #0
1496; CHECK-NEXT: b .LBB15_7
1497; CHECK-NEXT: .LBB15_3:
1498; CHECK-NEXT: vldr s0, .LCPI15_0
1499; CHECK-NEXT: b .LBB15_9
1500; CHECK-NEXT: .LBB15_4: @ %vector.ph
1501; CHECK-NEXT: bic r2, r1, #3
1502; CHECK-NEXT: movs r3, #1
1503; CHECK-NEXT: sub.w r12, r2, #4
1504; CHECK-NEXT: vmov.i32 q0, #0x0
1505; CHECK-NEXT: add.w lr, r3, r12, lsr #2
1506; CHECK-NEXT: mov r3, r0
1507; CHECK-NEXT: dls lr, lr
1508; CHECK-NEXT: .LBB15_5: @ %vector.body
1509; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1510; CHECK-NEXT: vldrw.u32 q1, [r3], #16
1511; CHECK-NEXT: vcmp.f32 lt, q0, q1
1512; CHECK-NEXT: vpsel q0, q0, q1
1513; CHECK-NEXT: le lr, .LBB15_5
1514; CHECK-NEXT: @ %bb.6: @ %middle.block
1515; CHECK-NEXT: vmov.f32 s4, s2
1516; CHECK-NEXT: cmp r2, r1
1517; CHECK-NEXT: vmov.f32 s5, s3
1518; CHECK-NEXT: vminnm.f32 q0, q0, q1
1519; CHECK-NEXT: vmov r3, s1
1520; CHECK-NEXT: vdup.32 q1, r3
1521; CHECK-NEXT: vminnm.f32 q0, q0, q1
1522; CHECK-NEXT: beq .LBB15_9
1523; CHECK-NEXT: .LBB15_7: @ %for.body.preheader1
1524; CHECK-NEXT: sub.w lr, r1, r2
1525; CHECK-NEXT: add.w r0, r0, r2, lsl #2
1526; CHECK-NEXT: dls lr, lr
1527; CHECK-NEXT: .LBB15_8: @ %for.body
1528; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1529; CHECK-NEXT: vldmia r0!, {s4}
1530; CHECK-NEXT: vcmp.f32 s0, s4
1531; CHECK-NEXT: vmrs APSR_nzcv, fpscr
1532; CHECK-NEXT: vselge.f32 s0, s4, s0
1533; CHECK-NEXT: le lr, .LBB15_8
1534; CHECK-NEXT: .LBB15_9: @ %for.cond.cleanup
1535; CHECK-NEXT: vmov r0, s0
1536; CHECK-NEXT: pop {r7, pc}
1537; CHECK-NEXT: .p2align 2
1538; CHECK-NEXT: @ %bb.10:
1539; CHECK-NEXT: .LCPI15_0:
1540; CHECK-NEXT: .long 0x00000000 @ float 0
1541entry:
1542 %cmp6 = icmp sgt i32 %n, 0
1543 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1544
1545for.body.preheader: ; preds = %entry
1546 %min.iters.check = icmp ult i32 %n, 4
1547 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1548
1549vector.ph: ; preds = %for.body.preheader
1550 %n.vec = and i32 %n, -4
1551 br label %vector.body
1552
1553vector.body: ; preds = %vector.body, %vector.ph
1554 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1555 %vec.phi = phi <4 x float> [ zeroinitializer, %vector.ph ], [ %3, %vector.body ]
1556 %0 = getelementptr inbounds float, float* %x, i32 %index
1557 %1 = bitcast float* %0 to <4 x float>*
1558 %wide.load = load <4 x float>, <4 x float>* %1, align 4
1559 %2 = fcmp ult <4 x float> %vec.phi, %wide.load
1560 %3 = select <4 x i1> %2, <4 x float> %vec.phi, <4 x float> %wide.load
1561 %index.next = add i32 %index, 4
1562 %4 = icmp eq i32 %index.next, %n.vec
1563 br i1 %4, label %middle.block, label %vector.body
1564
1565middle.block: ; preds = %vector.body
1566 %5 = call float @llvm.experimental.vector.reduce.fmin.v4f32(<4 x float> %3)
1567 %cmp.n = icmp eq i32 %n.vec, %n
1568 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1569
1570for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1571 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1572 %r.07.ph = phi float [ 0.0, %for.body.preheader ], [ %5, %middle.block ]
1573 br label %for.body
1574
1575for.body: ; preds = %for.body.preheader1, %for.body
1576 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1577 %r.07 = phi float [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1578 %arrayidx = getelementptr inbounds float, float* %x, i32 %i.08
1579 %6 = load float, float* %arrayidx, align 4
1580 %c = fcmp ult float %r.07, %6
1581 %add = select i1 %c, float %r.07, float %6
1582 %inc = add nuw nsw i32 %i.08, 1
1583 %exitcond = icmp eq i32 %inc, %n
1584 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1585
1586for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1587 %r.0.lcssa = phi float [ 0.0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1588 ret float %r.0.lcssa
1589}
1590
1591define float @fmax_f32(float* nocapture readonly %x, i32 %n) {
1592; CHECK-LABEL: fmax_f32:
1593; CHECK: @ %bb.0: @ %entry
1594; CHECK-NEXT: .save {r7, lr}
1595; CHECK-NEXT: push {r7, lr}
1596; CHECK-NEXT: cmp r1, #1
1597; CHECK-NEXT: blt .LBB16_3
1598; CHECK-NEXT: @ %bb.1: @ %for.body.preheader
1599; CHECK-NEXT: cmp r1, #4
1600; CHECK-NEXT: bhs .LBB16_4
1601; CHECK-NEXT: @ %bb.2:
1602; CHECK-NEXT: vldr s0, .LCPI16_0
1603; CHECK-NEXT: movs r2, #0
1604; CHECK-NEXT: b .LBB16_7
1605; CHECK-NEXT: .LBB16_3:
1606; CHECK-NEXT: vldr s0, .LCPI16_0
1607; CHECK-NEXT: b .LBB16_9
1608; CHECK-NEXT: .LBB16_4: @ %vector.ph
1609; CHECK-NEXT: bic r2, r1, #3
1610; CHECK-NEXT: movs r3, #1
1611; CHECK-NEXT: sub.w r12, r2, #4
1612; CHECK-NEXT: vmov.i32 q0, #0x0
1613; CHECK-NEXT: add.w lr, r3, r12, lsr #2
1614; CHECK-NEXT: mov r3, r0
1615; CHECK-NEXT: dls lr, lr
1616; CHECK-NEXT: .LBB16_5: @ %vector.body
1617; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1618; CHECK-NEXT: vldrw.u32 q1, [r3], #16
1619; CHECK-NEXT: vcmp.f32 lt, q1, q0
1620; CHECK-NEXT: vpsel q0, q0, q1
1621; CHECK-NEXT: le lr, .LBB16_5
1622; CHECK-NEXT: @ %bb.6: @ %middle.block
1623; CHECK-NEXT: vmov.f32 s4, s2
1624; CHECK-NEXT: cmp r2, r1
1625; CHECK-NEXT: vmov.f32 s5, s3
1626; CHECK-NEXT: vmaxnm.f32 q0, q0, q1
1627; CHECK-NEXT: vmov r3, s1
1628; CHECK-NEXT: vdup.32 q1, r3
1629; CHECK-NEXT: vmaxnm.f32 q0, q0, q1
1630; CHECK-NEXT: beq .LBB16_9
1631; CHECK-NEXT: .LBB16_7: @ %for.body.preheader1
1632; CHECK-NEXT: sub.w lr, r1, r2
1633; CHECK-NEXT: add.w r0, r0, r2, lsl #2
1634; CHECK-NEXT: dls lr, lr
1635; CHECK-NEXT: .LBB16_8: @ %for.body
1636; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
1637; CHECK-NEXT: vldmia r0!, {s4}
1638; CHECK-NEXT: vcmp.f32 s4, s0
1639; CHECK-NEXT: vmrs APSR_nzcv, fpscr
1640; CHECK-NEXT: vselge.f32 s0, s4, s0
1641; CHECK-NEXT: le lr, .LBB16_8
1642; CHECK-NEXT: .LBB16_9: @ %for.cond.cleanup
1643; CHECK-NEXT: vmov r0, s0
1644; CHECK-NEXT: pop {r7, pc}
1645; CHECK-NEXT: .p2align 2
1646; CHECK-NEXT: @ %bb.10:
1647; CHECK-NEXT: .LCPI16_0:
1648; CHECK-NEXT: .long 0x00000000 @ float 0
1649entry:
1650 %cmp6 = icmp sgt i32 %n, 0
1651 br i1 %cmp6, label %for.body.preheader, label %for.cond.cleanup
1652
1653for.body.preheader: ; preds = %entry
1654 %min.iters.check = icmp ult i32 %n, 4
1655 br i1 %min.iters.check, label %for.body.preheader1, label %vector.ph
1656
1657vector.ph: ; preds = %for.body.preheader
1658 %n.vec = and i32 %n, -4
1659 br label %vector.body
1660
1661vector.body: ; preds = %vector.body, %vector.ph
1662 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1663 %vec.phi = phi <4 x float> [ zeroinitializer, %vector.ph ], [ %3, %vector.body ]
1664 %0 = getelementptr inbounds float, float* %x, i32 %index
1665 %1 = bitcast float* %0 to <4 x float>*
1666 %wide.load = load <4 x float>, <4 x float>* %1, align 4
1667 %2 = fcmp ugt <4 x float> %vec.phi, %wide.load
1668 %3 = select <4 x i1> %2, <4 x float> %vec.phi, <4 x float> %wide.load
1669 %index.next = add i32 %index, 4
1670 %4 = icmp eq i32 %index.next, %n.vec
1671 br i1 %4, label %middle.block, label %vector.body
1672
1673middle.block: ; preds = %vector.body
1674 %5 = call float @llvm.experimental.vector.reduce.fmax.v4f32(<4 x float> %3)
1675 %cmp.n = icmp eq i32 %n.vec, %n
1676 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader1
1677
1678for.body.preheader1: ; preds = %middle.block, %for.body.preheader
1679 %i.08.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
1680 %r.07.ph = phi float [ 0.0, %for.body.preheader ], [ %5, %middle.block ]
1681 br label %for.body
1682
1683for.body: ; preds = %for.body.preheader1, %for.body
1684 %i.08 = phi i32 [ %inc, %for.body ], [ %i.08.ph, %for.body.preheader1 ]
1685 %r.07 = phi float [ %add, %for.body ], [ %r.07.ph, %for.body.preheader1 ]
1686 %arrayidx = getelementptr inbounds float, float* %x, i32 %i.08
1687 %6 = load float, float* %arrayidx, align 4
1688 %c = fcmp ugt float %r.07, %6
1689 %add = select i1 %c, float %r.07, float %6
1690 %inc = add nuw nsw i32 %i.08, 1
1691 %exitcond = icmp eq i32 %inc, %n
1692 br i1 %exitcond, label %for.cond.cleanup, label %for.body
1693
1694for.cond.cleanup: ; preds = %for.body, %middle.block, %entry
1695 %r.0.lcssa = phi float [ 0.0, %entry ], [ %5, %middle.block ], [ %add, %for.body ]
1696 ret float %r.0.lcssa
1697}
1698
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1699define i32 @add4i32(i32* noalias nocapture readonly %x, i32 %n) {
1700; CHECK-LABEL: add4i32:
1701; CHECK: @ %bb.0: @ %entry
1702; CHECK-NEXT: .save {r7, lr}
1703; CHECK-NEXT: push {r7, lr}
1704; CHECK-NEXT: cbz r1, .LBB17_4
1705; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1706; CHECK-NEXT: movs r2, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1707; CHECK-NEXT: dlstp.32 lr, r1
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1708; CHECK-NEXT: .LBB17_2: @ %vector.body
1709; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1710; CHECK-NEXT: vldrw.u32 q0, [r0], #16
1711; CHECK-NEXT: vaddva.u32 r2, q0
1712; CHECK-NEXT: letp lr, .LBB17_2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1713; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
1714; CHECK-NEXT: mov r0, r2
1715; CHECK-NEXT: pop {r7, pc}
1716; CHECK-NEXT: .LBB17_4:
1717; CHECK-NEXT: movs r2, #0
1718; CHECK-NEXT: mov r0, r2
1719; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1720entry:
1721 %cmp6.not = icmp eq i32 %n, 0
1722 br i1 %cmp6.not, label %for.cond.cleanup, label %vector.ph
1723
1724vector.ph: ; preds = %entry
1725 %n.rnd.up = add i32 %n, 3
1726 %n.vec = and i32 %n.rnd.up, -4
1727 %trip.count.minus.1 = add i32 %n, -1
1728 br label %vector.body
1729
1730vector.body: ; preds = %vector.body, %vector.ph
1731 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1732 %vec.phi = phi i32 [ 0, %vector.ph ], [ %4, %vector.body ]
1733 %active.lane.mask = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %trip.count.minus.1)
1734 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1735 %1 = bitcast i32* %0 to <4 x i32>*
1736 %wide.masked.load = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %1, i32 4, <4 x i1> %active.lane.mask, <4 x i32> undef)
1737 %2 = select <4 x i1> %active.lane.mask, <4 x i32> %wide.masked.load, <4 x i32> zeroinitializer
1738 %3 = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> %2)
1739 %4 = add i32 %3, %vec.phi
1740 %index.next = add i32 %index, 4
1741 %5 = icmp eq i32 %index.next, %n.vec
1742 br i1 %5, label %for.cond.cleanup, label %vector.body
1743
1744for.cond.cleanup: ; preds = %vector.body, %entry
1745 %s.0.lcssa = phi i32 [ 0, %entry ], [ %4, %vector.body ]
1746 ret i32 %s.0.lcssa
1747}
1748
1749define i32 @mla4i32(i32* noalias nocapture readonly %x, i32* noalias nocapture readonly %y, i32 %n) {
1750; CHECK-LABEL: mla4i32:
1751; CHECK: @ %bb.0: @ %entry
1752; CHECK-NEXT: .save {r7, lr}
1753; CHECK-NEXT: push {r7, lr}
1754; CHECK-NEXT: cbz r2, .LBB18_4
1755; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1756; CHECK-NEXT: mov.w r12, #0
1757; CHECK-NEXT: dlstp.32 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1758; CHECK-NEXT: .LBB18_2: @ %vector.body
1759; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1760; CHECK-NEXT: vldrw.u32 q0, [r0], #16
1761; CHECK-NEXT: vldrw.u32 q1, [r1], #16
1762; CHECK-NEXT: vmlava.u32 r12, q1, q0
1763; CHECK-NEXT: letp lr, .LBB18_2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1764; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1765; CHECK-NEXT: mov r0, r12
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1766; CHECK-NEXT: pop {r7, pc}
1767; CHECK-NEXT: .LBB18_4:
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1768; CHECK-NEXT: mov.w r12, #0
1769; CHECK-NEXT: mov r0, r12
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1770; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1771entry:
1772 %cmp8.not = icmp eq i32 %n, 0
1773 br i1 %cmp8.not, label %for.cond.cleanup, label %vector.ph
1774
1775vector.ph: ; preds = %entry
1776 %n.rnd.up = add i32 %n, 3
1777 %n.vec = and i32 %n.rnd.up, -4
1778 %trip.count.minus.1 = add i32 %n, -1
1779 br label %vector.body
1780
1781vector.body: ; preds = %vector.body, %vector.ph
1782 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1783 %vec.phi = phi i32 [ 0, %vector.ph ], [ %7, %vector.body ]
1784 %active.lane.mask = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %trip.count.minus.1)
1785 %0 = getelementptr inbounds i32, i32* %x, i32 %index
1786 %1 = bitcast i32* %0 to <4 x i32>*
1787 %wide.masked.load = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %1, i32 4, <4 x i1> %active.lane.mask, <4 x i32> undef)
1788 %2 = getelementptr inbounds i32, i32* %y, i32 %index
1789 %3 = bitcast i32* %2 to <4 x i32>*
1790 %wide.masked.load13 = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %3, i32 4, <4 x i1> %active.lane.mask, <4 x i32> undef)
1791 %4 = mul nsw <4 x i32> %wide.masked.load13, %wide.masked.load
1792 %5 = select <4 x i1> %active.lane.mask, <4 x i32> %4, <4 x i32> zeroinitializer
1793 %6 = call i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32> %5)
1794 %7 = add i32 %6, %vec.phi
1795 %index.next = add i32 %index, 4
1796 %8 = icmp eq i32 %index.next, %n.vec
1797 br i1 %8, label %for.cond.cleanup, label %vector.body
1798
1799for.cond.cleanup: ; preds = %vector.body, %entry
1800 %s.0.lcssa = phi i32 [ 0, %entry ], [ %7, %vector.body ]
1801 ret i32 %s.0.lcssa
1802}
1803
1804define i32 @add8i32(i16* noalias nocapture readonly %x, i32 %n) {
1805; CHECK-LABEL: add8i32:
1806; CHECK: @ %bb.0: @ %entry
1807; CHECK-NEXT: .save {r7, lr}
1808; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1809; CHECK-NEXT: cbz r1, .LBB19_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1810; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1811; CHECK-NEXT: movs r2, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1812; CHECK-NEXT: dlstp.16 lr, r1
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1813; CHECK-NEXT: .LBB19_2: @ %vector.body
1814; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1815; CHECK-NEXT: vldrh.u16 q0, [r0], #16
1816; CHECK-NEXT: vaddva.s16 r2, q0
1817; CHECK-NEXT: letp lr, .LBB19_2
1818; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1819; CHECK-NEXT: mov r0, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1820; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1821; CHECK-NEXT: .LBB19_4:
1822; CHECK-NEXT: movs r2, #0
1823; CHECK-NEXT: mov r0, r2
1824; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1825entry:
1826 %cmp6.not = icmp eq i32 %n, 0
1827 br i1 %cmp6.not, label %for.cond.cleanup, label %vector.ph
1828
1829vector.ph: ; preds = %entry
1830 %n.rnd.up = add i32 %n, 7
1831 %n.vec = and i32 %n.rnd.up, -8
1832 %trip.count.minus.1 = add i32 %n, -1
1833 br label %vector.body
1834
1835vector.body: ; preds = %vector.body, %vector.ph
1836 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1837 %vec.phi = phi i32 [ 0, %vector.ph ], [ %5, %vector.body ]
1838 %active.lane.mask = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 %index, i32 %trip.count.minus.1)
1839 %0 = getelementptr inbounds i16, i16* %x, i32 %index
1840 %1 = bitcast i16* %0 to <8 x i16>*
1841 %wide.masked.load = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %1, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
1842 %2 = sext <8 x i16> %wide.masked.load to <8 x i32>
1843 %3 = select <8 x i1> %active.lane.mask, <8 x i32> %2, <8 x i32> zeroinitializer
1844 %4 = call i32 @llvm.experimental.vector.reduce.add.v8i32(<8 x i32> %3)
1845 %5 = add i32 %4, %vec.phi
1846 %index.next = add i32 %index, 8
1847 %6 = icmp eq i32 %index.next, %n.vec
1848 br i1 %6, label %for.cond.cleanup, label %vector.body
1849
1850for.cond.cleanup: ; preds = %vector.body, %entry
1851 %s.0.lcssa = phi i32 [ 0, %entry ], [ %5, %vector.body ]
1852 ret i32 %s.0.lcssa
1853}
1854
1855define i32 @mla8i32(i16* noalias nocapture readonly %x, i16* noalias nocapture readonly %y, i32 %n) {
1856; CHECK-LABEL: mla8i32:
1857; CHECK: @ %bb.0: @ %entry
1858; CHECK-NEXT: .save {r7, lr}
1859; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1860; CHECK-NEXT: cbz r2, .LBB20_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1861; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1862; CHECK-NEXT: mov.w r12, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1863; CHECK-NEXT: dlstp.16 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1864; CHECK-NEXT: .LBB20_2: @ %vector.body
1865; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1866; CHECK-NEXT: vldrh.u16 q0, [r0], #16
1867; CHECK-NEXT: vldrh.u16 q1, [r1], #16
1868; CHECK-NEXT: vmlava.s16 r12, q1, q0
1869; CHECK-NEXT: letp lr, .LBB20_2
1870; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1871; CHECK-NEXT: mov r0, r12
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1872; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1873; CHECK-NEXT: .LBB20_4:
1874; CHECK-NEXT: mov.w r12, #0
1875; CHECK-NEXT: mov r0, r12
1876; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1877entry:
1878 %cmp9.not = icmp eq i32 %n, 0
1879 br i1 %cmp9.not, label %for.cond.cleanup, label %vector.ph
1880
1881vector.ph: ; preds = %entry
1882 %n.rnd.up = add i32 %n, 7
1883 %n.vec = and i32 %n.rnd.up, -8
1884 %trip.count.minus.1 = add i32 %n, -1
1885 br label %vector.body
1886
1887vector.body: ; preds = %vector.body, %vector.ph
1888 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1889 %vec.phi = phi i32 [ 0, %vector.ph ], [ %9, %vector.body ]
1890 %active.lane.mask = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 %index, i32 %trip.count.minus.1)
1891 %0 = getelementptr inbounds i16, i16* %x, i32 %index
1892 %1 = bitcast i16* %0 to <8 x i16>*
1893 %wide.masked.load = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %1, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
1894 %2 = sext <8 x i16> %wide.masked.load to <8 x i32>
1895 %3 = getelementptr inbounds i16, i16* %y, i32 %index
1896 %4 = bitcast i16* %3 to <8 x i16>*
1897 %wide.masked.load14 = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %4, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
1898 %5 = sext <8 x i16> %wide.masked.load14 to <8 x i32>
1899 %6 = mul nsw <8 x i32> %5, %2
1900 %7 = select <8 x i1> %active.lane.mask, <8 x i32> %6, <8 x i32> zeroinitializer
1901 %8 = call i32 @llvm.experimental.vector.reduce.add.v8i32(<8 x i32> %7)
1902 %9 = add i32 %8, %vec.phi
1903 %index.next = add i32 %index, 8
1904 %10 = icmp eq i32 %index.next, %n.vec
1905 br i1 %10, label %for.cond.cleanup, label %vector.body
1906
1907for.cond.cleanup: ; preds = %vector.body, %entry
1908 %s.0.lcssa = phi i32 [ 0, %entry ], [ %9, %vector.body ]
1909 ret i32 %s.0.lcssa
1910}
1911
1912define i32 @add16i32(i8* noalias nocapture readonly %x, i32 %n) {
1913; CHECK-LABEL: add16i32:
1914; CHECK: @ %bb.0: @ %entry
1915; CHECK-NEXT: .save {r7, lr}
1916; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1917; CHECK-NEXT: cbz r1, .LBB21_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1918; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1919; CHECK-NEXT: movs r2, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1920; CHECK-NEXT: dlstp.8 lr, r1
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1921; CHECK-NEXT: .LBB21_2: @ %vector.body
1922; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1923; CHECK-NEXT: vldrb.u8 q0, [r0], #16
1924; CHECK-NEXT: vaddva.u8 r2, q0
1925; CHECK-NEXT: letp lr, .LBB21_2
1926; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1927; CHECK-NEXT: mov r0, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1928; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1929; CHECK-NEXT: .LBB21_4:
1930; CHECK-NEXT: movs r2, #0
1931; CHECK-NEXT: mov r0, r2
1932; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1933entry:
1934 %cmp6.not = icmp eq i32 %n, 0
1935 br i1 %cmp6.not, label %for.cond.cleanup, label %vector.ph
1936
1937vector.ph: ; preds = %entry
1938 %n.rnd.up = add i32 %n, 15
1939 %n.vec = and i32 %n.rnd.up, -16
1940 %trip.count.minus.1 = add i32 %n, -1
1941 br label %vector.body
1942
1943vector.body: ; preds = %vector.body, %vector.ph
1944 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1945 %vec.phi = phi i32 [ 0, %vector.ph ], [ %5, %vector.body ]
1946 %active.lane.mask = call <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32 %index, i32 %trip.count.minus.1)
1947 %0 = getelementptr inbounds i8, i8* %x, i32 %index
1948 %1 = bitcast i8* %0 to <16 x i8>*
1949 %wide.masked.load = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %1, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
1950 %2 = zext <16 x i8> %wide.masked.load to <16 x i32>
1951 %3 = select <16 x i1> %active.lane.mask, <16 x i32> %2, <16 x i32> zeroinitializer
1952 %4 = call i32 @llvm.experimental.vector.reduce.add.v16i32(<16 x i32> %3)
1953 %5 = add i32 %4, %vec.phi
1954 %index.next = add i32 %index, 16
1955 %6 = icmp eq i32 %index.next, %n.vec
1956 br i1 %6, label %for.cond.cleanup, label %vector.body
1957
1958for.cond.cleanup: ; preds = %vector.body, %entry
1959 %s.0.lcssa = phi i32 [ 0, %entry ], [ %5, %vector.body ]
1960 ret i32 %s.0.lcssa
1961}
1962
1963define i32 @mla16i32(i8* noalias nocapture readonly %x, i8* noalias nocapture readonly %y, i32 %n) {
1964; CHECK-LABEL: mla16i32:
1965; CHECK: @ %bb.0: @ %entry
1966; CHECK-NEXT: .save {r7, lr}
1967; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1968; CHECK-NEXT: cbz r2, .LBB22_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1969; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1970; CHECK-NEXT: mov.w r12, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1971; CHECK-NEXT: dlstp.8 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1972; CHECK-NEXT: .LBB22_2: @ %vector.body
1973; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1974; CHECK-NEXT: vldrb.u8 q0, [r0], #16
1975; CHECK-NEXT: vldrb.u8 q1, [r1], #16
1976; CHECK-NEXT: vmlava.u8 r12, q1, q0
1977; CHECK-NEXT: letp lr, .LBB22_2
1978; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1979; CHECK-NEXT: mov r0, r12
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1980; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]1981; CHECK-NEXT: .LBB22_4:
1982; CHECK-NEXT: mov.w r12, #0
1983; CHECK-NEXT: mov r0, r12
1984; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]1985entry:
1986 %cmp9.not = icmp eq i32 %n, 0
1987 br i1 %cmp9.not, label %for.cond.cleanup, label %vector.ph
1988
1989vector.ph: ; preds = %entry
1990 %n.rnd.up = add i32 %n, 15
1991 %n.vec = and i32 %n.rnd.up, -16
1992 %trip.count.minus.1 = add i32 %n, -1
1993 br label %vector.body
1994
1995vector.body: ; preds = %vector.body, %vector.ph
1996 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
1997 %vec.phi = phi i32 [ 0, %vector.ph ], [ %9, %vector.body ]
1998 %active.lane.mask = call <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32 %index, i32 %trip.count.minus.1)
1999 %0 = getelementptr inbounds i8, i8* %x, i32 %index
2000 %1 = bitcast i8* %0 to <16 x i8>*
2001 %wide.masked.load = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %1, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2002 %2 = zext <16 x i8> %wide.masked.load to <16 x i32>
2003 %3 = getelementptr inbounds i8, i8* %y, i32 %index
2004 %4 = bitcast i8* %3 to <16 x i8>*
2005 %wide.masked.load14 = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %4, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2006 %5 = zext <16 x i8> %wide.masked.load14 to <16 x i32>
2007 %6 = mul nuw nsw <16 x i32> %5, %2
2008 %7 = select <16 x i1> %active.lane.mask, <16 x i32> %6, <16 x i32> zeroinitializer
2009 %8 = call i32 @llvm.experimental.vector.reduce.add.v16i32(<16 x i32> %7)
2010 %9 = add i32 %8, %vec.phi
2011 %index.next = add i32 %index, 16
2012 %10 = icmp eq i32 %index.next, %n.vec
2013 br i1 %10, label %for.cond.cleanup, label %vector.body
2014
2015for.cond.cleanup: ; preds = %vector.body, %entry
2016 %s.0.lcssa = phi i32 [ 0, %entry ], [ %9, %vector.body ]
2017 ret i32 %s.0.lcssa
2018}
2019
2020define signext i16 @add8i16(i16* noalias nocapture readonly %x, i32 %n) {
2021; CHECK-LABEL: add8i16:
2022; CHECK: @ %bb.0: @ %entry
2023; CHECK-NEXT: .save {r7, lr}
2024; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2025; CHECK-NEXT: cbz r1, .LBB23_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2026; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2027; CHECK-NEXT: movs r2, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2028; CHECK-NEXT: dlstp.16 lr, r1
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2029; CHECK-NEXT: .LBB23_2: @ %vector.body
2030; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2031; CHECK-NEXT: vldrh.u16 q0, [r0], #16
2032; CHECK-NEXT: vaddva.u16 r2, q0
2033; CHECK-NEXT: letp lr, .LBB23_2
2034; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2035; CHECK-NEXT: sxth r0, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2036; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2037; CHECK-NEXT: .LBB23_4:
2038; CHECK-NEXT: movs r2, #0
2039; CHECK-NEXT: sxth r0, r2
2040; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2041entry:
2042 %cmp8.not = icmp eq i32 %n, 0
2043 br i1 %cmp8.not, label %for.cond.cleanup, label %vector.ph
2044
2045vector.ph: ; preds = %entry
2046 %n.rnd.up = add i32 %n, 7
2047 %n.vec = and i32 %n.rnd.up, -8
2048 %trip.count.minus.1 = add i32 %n, -1
2049 br label %vector.body
2050
2051vector.body: ; preds = %vector.body, %vector.ph
2052 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2053 %vec.phi = phi i16 [ 0, %vector.ph ], [ %4, %vector.body ]
2054 %active.lane.mask = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 %index, i32 %trip.count.minus.1)
2055 %0 = getelementptr inbounds i16, i16* %x, i32 %index
2056 %1 = bitcast i16* %0 to <8 x i16>*
2057 %wide.masked.load = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %1, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
2058 %2 = select <8 x i1> %active.lane.mask, <8 x i16> %wide.masked.load, <8 x i16> zeroinitializer
2059 %3 = call i16 @llvm.experimental.vector.reduce.add.v8i16(<8 x i16> %2)
2060 %4 = add i16 %3, %vec.phi
2061 %index.next = add i32 %index, 8
2062 %5 = icmp eq i32 %index.next, %n.vec
2063 br i1 %5, label %for.cond.cleanup, label %vector.body
2064
2065for.cond.cleanup: ; preds = %vector.body, %entry
2066 %s.0.lcssa = phi i16 [ 0, %entry ], [ %4, %vector.body ]
2067 ret i16 %s.0.lcssa
2068}
2069
2070define signext i16 @mla8i16(i16* noalias nocapture readonly %x, i16* noalias nocapture readonly %y, i32 %n) {
2071; CHECK-LABEL: mla8i16:
2072; CHECK: @ %bb.0: @ %entry
2073; CHECK-NEXT: .save {r7, lr}
2074; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2075; CHECK-NEXT: cbz r2, .LBB24_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2076; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2077; CHECK-NEXT: mov.w r12, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2078; CHECK-NEXT: dlstp.16 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2079; CHECK-NEXT: .LBB24_2: @ %vector.body
2080; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2081; CHECK-NEXT: vldrh.u16 q0, [r0], #16
2082; CHECK-NEXT: vldrh.u16 q1, [r1], #16
2083; CHECK-NEXT: vmlava.u16 r12, q1, q0
2084; CHECK-NEXT: letp lr, .LBB24_2
2085; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2086; CHECK-NEXT: sxth.w r0, r12
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2087; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2088; CHECK-NEXT: .LBB24_4:
2089; CHECK-NEXT: mov.w r12, #0
2090; CHECK-NEXT: sxth.w r0, r12
2091; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2092entry:
2093 %cmp11.not = icmp eq i32 %n, 0
2094 br i1 %cmp11.not, label %for.cond.cleanup, label %vector.ph
2095
2096vector.ph: ; preds = %entry
2097 %n.rnd.up = add i32 %n, 7
2098 %n.vec = and i32 %n.rnd.up, -8
2099 %trip.count.minus.1 = add i32 %n, -1
2100 br label %vector.body
2101
2102vector.body: ; preds = %vector.body, %vector.ph
2103 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2104 %vec.phi = phi i16 [ 0, %vector.ph ], [ %7, %vector.body ]
2105 %active.lane.mask = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 %index, i32 %trip.count.minus.1)
2106 %0 = getelementptr inbounds i16, i16* %x, i32 %index
2107 %1 = bitcast i16* %0 to <8 x i16>*
2108 %wide.masked.load = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %1, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
2109 %2 = getelementptr inbounds i16, i16* %y, i32 %index
2110 %3 = bitcast i16* %2 to <8 x i16>*
2111 %wide.masked.load16 = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %3, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
2112 %4 = mul <8 x i16> %wide.masked.load16, %wide.masked.load
2113 %5 = select <8 x i1> %active.lane.mask, <8 x i16> %4, <8 x i16> zeroinitializer
2114 %6 = call i16 @llvm.experimental.vector.reduce.add.v8i16(<8 x i16> %5)
2115 %7 = add i16 %6, %vec.phi
2116 %index.next = add i32 %index, 8
2117 %8 = icmp eq i32 %index.next, %n.vec
2118 br i1 %8, label %for.cond.cleanup, label %vector.body
2119
2120for.cond.cleanup: ; preds = %vector.body, %entry
2121 %s.0.lcssa = phi i16 [ 0, %entry ], [ %7, %vector.body ]
2122 ret i16 %s.0.lcssa
2123}
2124
2125define signext i16 @add16i16(i8* noalias nocapture readonly %x, i32 %n) {
2126; CHECK-LABEL: add16i16:
2127; CHECK: @ %bb.0: @ %entry
2128; CHECK-NEXT: .save {r7, lr}
2129; CHECK-NEXT: push {r7, lr}
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2130; CHECK-NEXT: cbz r1, .LBB25_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2131; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2132; CHECK-NEXT: movs r2, #0
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2133; CHECK-NEXT: dlstp.8 lr, r1
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2134; CHECK-NEXT: .LBB25_2: @ %vector.body
2135; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2136; CHECK-NEXT: vldrb.u8 q0, [r0], #16
2137; CHECK-NEXT: vaddva.u8 r2, q0
2138; CHECK-NEXT: letp lr, .LBB25_2
2139; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2140; CHECK-NEXT: sxth r0, r2
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2141; CHECK-NEXT: pop {r7, pc}
2142; CHECK-NEXT: .LBB25_4:
2143; CHECK-NEXT: movs r2, #0
2144; CHECK-NEXT: sxth r0, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2145; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2146entry:
2147 %cmp8.not = icmp eq i32 %n, 0
2148 br i1 %cmp8.not, label %for.cond.cleanup, label %vector.ph
2149
2150vector.ph: ; preds = %entry
2151 %n.rnd.up = add i32 %n, 15
2152 %n.vec = and i32 %n.rnd.up, -16
2153 %trip.count.minus.1 = add i32 %n, -1
2154 br label %vector.body
2155
2156vector.body: ; preds = %vector.body, %vector.ph
2157 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2158 %vec.phi = phi i16 [ 0, %vector.ph ], [ %5, %vector.body ]
2159 %active.lane.mask = call <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32 %index, i32 %trip.count.minus.1)
2160 %0 = getelementptr inbounds i8, i8* %x, i32 %index
2161 %1 = bitcast i8* %0 to <16 x i8>*
2162 %wide.masked.load = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %1, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2163 %2 = zext <16 x i8> %wide.masked.load to <16 x i16>
2164 %3 = select <16 x i1> %active.lane.mask, <16 x i16> %2, <16 x i16> zeroinitializer
2165 %4 = call i16 @llvm.experimental.vector.reduce.add.v16i16(<16 x i16> %3)
2166 %5 = add i16 %4, %vec.phi
2167 %index.next = add i32 %index, 16
2168 %6 = icmp eq i32 %index.next, %n.vec
2169 br i1 %6, label %for.cond.cleanup, label %vector.body
2170
2171for.cond.cleanup: ; preds = %vector.body, %entry
2172 %s.0.lcssa = phi i16 [ 0, %entry ], [ %5, %vector.body ]
2173 ret i16 %s.0.lcssa
2174}
2175
2176define signext i16 @mla16i16(i8* noalias nocapture readonly %x, i8* noalias nocapture readonly %y, i32 %n) {
2177; CHECK-LABEL: mla16i16:
2178; CHECK: @ %bb.0: @ %entry
2179; CHECK-NEXT: .save {r7, lr}
2180; CHECK-NEXT: push {r7, lr}
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2181; CHECK-NEXT: cbz r2, .LBB26_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2182; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2183; CHECK-NEXT: mov.w r12, #0
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2184; CHECK-NEXT: dlstp.8 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2185; CHECK-NEXT: .LBB26_2: @ %vector.body
2186; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2187; CHECK-NEXT: vldrb.u8 q0, [r0], #16
2188; CHECK-NEXT: vldrb.u8 q1, [r1], #16
2189; CHECK-NEXT: vmlava.u8 r12, q1, q0
2190; CHECK-NEXT: letp lr, .LBB26_2
2191; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2192; CHECK-NEXT: sxth.w r0, r12
David Green186a7f82020-08-09 11:09:49 +0100[diff] [blame^]2193; CHECK-NEXT: pop {r7, pc}
2194; CHECK-NEXT: .LBB26_4:
2195; CHECK-NEXT: mov.w r12, #0
2196; CHECK-NEXT: sxth.w r0, r12
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2197; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2198entry:
2199 %cmp13.not = icmp eq i32 %n, 0
2200 br i1 %cmp13.not, label %for.cond.cleanup, label %vector.ph
2201
2202vector.ph: ; preds = %entry
2203 %n.rnd.up = add i32 %n, 15
2204 %n.vec = and i32 %n.rnd.up, -16
2205 %trip.count.minus.1 = add i32 %n, -1
2206 br label %vector.body
2207
2208vector.body: ; preds = %vector.body, %vector.ph
2209 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2210 %vec.phi = phi i16 [ 0, %vector.ph ], [ %9, %vector.body ]
2211 %active.lane.mask = call <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32 %index, i32 %trip.count.minus.1)
2212 %0 = getelementptr inbounds i8, i8* %x, i32 %index
2213 %1 = bitcast i8* %0 to <16 x i8>*
2214 %wide.masked.load = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %1, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2215 %2 = zext <16 x i8> %wide.masked.load to <16 x i16>
2216 %3 = getelementptr inbounds i8, i8* %y, i32 %index
2217 %4 = bitcast i8* %3 to <16 x i8>*
2218 %wide.masked.load18 = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %4, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2219 %5 = zext <16 x i8> %wide.masked.load18 to <16 x i16>
2220 %6 = mul nuw <16 x i16> %5, %2
2221 %7 = select <16 x i1> %active.lane.mask, <16 x i16> %6, <16 x i16> zeroinitializer
2222 %8 = call i16 @llvm.experimental.vector.reduce.add.v16i16(<16 x i16> %7)
2223 %9 = add i16 %8, %vec.phi
2224 %index.next = add i32 %index, 16
2225 %10 = icmp eq i32 %index.next, %n.vec
2226 br i1 %10, label %for.cond.cleanup, label %vector.body
2227
2228for.cond.cleanup: ; preds = %vector.body, %entry
2229 %s.0.lcssa = phi i16 [ 0, %entry ], [ %9, %vector.body ]
2230 ret i16 %s.0.lcssa
2231}
2232
2233define zeroext i8 @add16i8(i8* noalias nocapture readonly %x, i32 %n) {
2234; CHECK-LABEL: add16i8:
2235; CHECK: @ %bb.0: @ %entry
2236; CHECK-NEXT: .save {r7, lr}
2237; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2238; CHECK-NEXT: cbz r1, .LBB27_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2239; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2240; CHECK-NEXT: movs r2, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2241; CHECK-NEXT: dlstp.8 lr, r1
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2242; CHECK-NEXT: .LBB27_2: @ %vector.body
2243; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2244; CHECK-NEXT: vldrb.u8 q0, [r0], #16
2245; CHECK-NEXT: vaddva.u8 r2, q0
2246; CHECK-NEXT: letp lr, .LBB27_2
2247; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2248; CHECK-NEXT: uxtb r0, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2249; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2250; CHECK-NEXT: .LBB27_4:
2251; CHECK-NEXT: movs r2, #0
2252; CHECK-NEXT: uxtb r0, r2
2253; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2254entry:
2255 %cmp7.not = icmp eq i32 %n, 0
2256 br i1 %cmp7.not, label %for.cond.cleanup, label %vector.ph
2257
2258vector.ph: ; preds = %entry
2259 %n.rnd.up = add i32 %n, 15
2260 %n.vec = and i32 %n.rnd.up, -16
2261 %trip.count.minus.1 = add i32 %n, -1
2262 br label %vector.body
2263
2264vector.body: ; preds = %vector.body, %vector.ph
2265 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2266 %vec.phi = phi i8 [ 0, %vector.ph ], [ %4, %vector.body ]
2267 %active.lane.mask = call <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32 %index, i32 %trip.count.minus.1)
2268 %0 = getelementptr inbounds i8, i8* %x, i32 %index
2269 %1 = bitcast i8* %0 to <16 x i8>*
2270 %wide.masked.load = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %1, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2271 %2 = select <16 x i1> %active.lane.mask, <16 x i8> %wide.masked.load, <16 x i8> zeroinitializer
2272 %3 = call i8 @llvm.experimental.vector.reduce.add.v16i8(<16 x i8> %2)
2273 %4 = add i8 %3, %vec.phi
2274 %index.next = add i32 %index, 16
2275 %5 = icmp eq i32 %index.next, %n.vec
2276 br i1 %5, label %for.cond.cleanup, label %vector.body
2277
2278for.cond.cleanup: ; preds = %vector.body, %entry
2279 %s.0.lcssa = phi i8 [ 0, %entry ], [ %4, %vector.body ]
2280 ret i8 %s.0.lcssa
2281}
2282
2283define zeroext i8 @mla16i8(i8* noalias nocapture readonly %x, i8* noalias nocapture readonly %y, i32 %n) {
2284; CHECK-LABEL: mla16i8:
2285; CHECK: @ %bb.0: @ %entry
2286; CHECK-NEXT: .save {r7, lr}
2287; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2288; CHECK-NEXT: cbz r2, .LBB28_4
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2289; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2290; CHECK-NEXT: mov.w r12, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2291; CHECK-NEXT: dlstp.8 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2292; CHECK-NEXT: .LBB28_2: @ %vector.body
2293; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2294; CHECK-NEXT: vldrb.u8 q0, [r0], #16
2295; CHECK-NEXT: vldrb.u8 q1, [r1], #16
2296; CHECK-NEXT: vmlava.u8 r12, q1, q0
2297; CHECK-NEXT: letp lr, .LBB28_2
2298; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2299; CHECK-NEXT: uxtb.w r0, r12
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2300; CHECK-NEXT: pop {r7, pc}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2301; CHECK-NEXT: .LBB28_4:
2302; CHECK-NEXT: mov.w r12, #0
2303; CHECK-NEXT: uxtb.w r0, r12
2304; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2305entry:
2306 %cmp10.not = icmp eq i32 %n, 0
2307 br i1 %cmp10.not, label %for.cond.cleanup, label %vector.ph
2308
2309vector.ph: ; preds = %entry
2310 %n.rnd.up = add i32 %n, 15
2311 %n.vec = and i32 %n.rnd.up, -16
2312 %trip.count.minus.1 = add i32 %n, -1
2313 br label %vector.body
2314
2315vector.body: ; preds = %vector.body, %vector.ph
2316 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2317 %vec.phi = phi i8 [ 0, %vector.ph ], [ %7, %vector.body ]
2318 %active.lane.mask = call <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32 %index, i32 %trip.count.minus.1)
2319 %0 = getelementptr inbounds i8, i8* %x, i32 %index
2320 %1 = bitcast i8* %0 to <16 x i8>*
2321 %wide.masked.load = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %1, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2322 %2 = getelementptr inbounds i8, i8* %y, i32 %index
2323 %3 = bitcast i8* %2 to <16 x i8>*
2324 %wide.masked.load15 = call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %3, i32 1, <16 x i1> %active.lane.mask, <16 x i8> undef)
2325 %4 = mul <16 x i8> %wide.masked.load15, %wide.masked.load
2326 %5 = select <16 x i1> %active.lane.mask, <16 x i8> %4, <16 x i8> zeroinitializer
2327 %6 = call i8 @llvm.experimental.vector.reduce.add.v16i8(<16 x i8> %5)
2328 %7 = add i8 %6, %vec.phi
2329 %index.next = add i32 %index, 16
2330 %8 = icmp eq i32 %index.next, %n.vec
2331 br i1 %8, label %for.cond.cleanup, label %vector.body
2332
2333for.cond.cleanup: ; preds = %vector.body, %entry
2334 %s.0.lcssa = phi i8 [ 0, %entry ], [ %7, %vector.body ]
2335 ret i8 %s.0.lcssa
2336}
2337
2338define i64 @add4i64(i32* noalias nocapture readonly %x, i32 %n) {
2339; CHECK-LABEL: add4i64:
2340; CHECK: @ %bb.0: @ %entry
2341; CHECK-NEXT: .save {r7, lr}
2342; CHECK-NEXT: push {r7, lr}
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2343; CHECK-NEXT: cbz r1, .LBB29_3
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2344; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2345; CHECK-NEXT: movs r2, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2346; CHECK-NEXT: mov r3, r2
2347; CHECK-NEXT: dlstp.32 lr, r1
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2348; CHECK-NEXT: .LBB29_2: @ %vector.body
2349; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2350; CHECK-NEXT: vldrw.u32 q0, [r0], #16
2351; CHECK-NEXT: vaddlva.s32 r2, r3, q0
2352; CHECK-NEXT: letp lr, .LBB29_2
2353; CHECK-NEXT: b .LBB29_4
2354; CHECK-NEXT: .LBB29_3:
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2355; CHECK-NEXT: movs r2, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2356; CHECK-NEXT: mov r3, r2
2357; CHECK-NEXT: .LBB29_4: @ %for.cond.cleanup
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2358; CHECK-NEXT: mov r0, r2
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2359; CHECK-NEXT: mov r1, r3
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2360; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2361entry:
2362 %cmp6.not = icmp eq i32 %n, 0
2363 br i1 %cmp6.not, label %for.cond.cleanup, label %vector.ph
2364
2365vector.ph: ; preds = %entry
2366 %n.rnd.up = add i32 %n, 3
2367 %n.vec = and i32 %n.rnd.up, -4
2368 %trip.count.minus.1 = add i32 %n, -1
2369 br label %vector.body
2370
2371vector.body: ; preds = %vector.body, %vector.ph
2372 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2373 %vec.phi = phi i64 [ 0, %vector.ph ], [ %5, %vector.body ]
2374 %active.lane.mask = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %trip.count.minus.1)
2375 %0 = getelementptr inbounds i32, i32* %x, i32 %index
2376 %1 = bitcast i32* %0 to <4 x i32>*
2377 %wide.masked.load = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %1, i32 4, <4 x i1> %active.lane.mask, <4 x i32> undef)
2378 %2 = sext <4 x i32> %wide.masked.load to <4 x i64>
2379 %3 = select <4 x i1> %active.lane.mask, <4 x i64> %2, <4 x i64> zeroinitializer
2380 %4 = call i64 @llvm.experimental.vector.reduce.add.v4i64(<4 x i64> %3)
2381 %5 = add i64 %4, %vec.phi
2382 %index.next = add i32 %index, 4
2383 %6 = icmp eq i32 %index.next, %n.vec
2384 br i1 %6, label %for.cond.cleanup, label %vector.body
2385
2386for.cond.cleanup: ; preds = %vector.body, %entry
2387 %s.0.lcssa = phi i64 [ 0, %entry ], [ %5, %vector.body ]
2388 ret i64 %s.0.lcssa
2389}
2390
2391define i64 @mla4i64(i32* noalias nocapture readonly %x, i32* noalias nocapture readonly %y, i32 %n) {
2392; CHECK-LABEL: mla4i64:
2393; CHECK: @ %bb.0: @ %entry
2394; CHECK-NEXT: .save {r7, lr}
2395; CHECK-NEXT: push {r7, lr}
2396; CHECK-NEXT: cbz r2, .LBB30_3
2397; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2398; CHECK-NEXT: mov.w r12, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2399; CHECK-NEXT: mov r3, r12
2400; CHECK-NEXT: dlstp.32 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2401; CHECK-NEXT: .LBB30_2: @ %vector.body
2402; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2403; CHECK-NEXT: vldrw.u32 q0, [r0], #16
2404; CHECK-NEXT: vldrw.u32 q1, [r1], #16
2405; CHECK-NEXT: vmlalva.s32 r12, r3, q1, q0
2406; CHECK-NEXT: letp lr, .LBB30_2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2407; CHECK-NEXT: b .LBB30_4
2408; CHECK-NEXT: .LBB30_3:
2409; CHECK-NEXT: mov.w r12, #0
2410; CHECK-NEXT: mov r3, r12
2411; CHECK-NEXT: .LBB30_4: @ %for.cond.cleanup
2412; CHECK-NEXT: mov r0, r12
2413; CHECK-NEXT: mov r1, r3
2414; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2415entry:
2416 %cmp9.not = icmp eq i32 %n, 0
2417 br i1 %cmp9.not, label %for.cond.cleanup, label %vector.ph
2418
2419vector.ph: ; preds = %entry
2420 %n.rnd.up = add i32 %n, 3
2421 %n.vec = and i32 %n.rnd.up, -4
2422 %trip.count.minus.1 = add i32 %n, -1
2423 br label %vector.body
2424
2425vector.body: ; preds = %vector.body, %vector.ph
2426 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2427 %vec.phi = phi i64 [ 0, %vector.ph ], [ %9, %vector.body ]
2428 %active.lane.mask = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %trip.count.minus.1)
2429 %0 = getelementptr inbounds i32, i32* %x, i32 %index
2430 %1 = bitcast i32* %0 to <4 x i32>*
2431 %wide.masked.load = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %1, i32 4, <4 x i1> %active.lane.mask, <4 x i32> undef)
2432 %2 = sext <4 x i32> %wide.masked.load to <4 x i64>
2433 %3 = getelementptr inbounds i32, i32* %y, i32 %index
2434 %4 = bitcast i32* %3 to <4 x i32>*
2435 %wide.masked.load14 = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %4, i32 4, <4 x i1> %active.lane.mask, <4 x i32> undef)
2436 %5 = sext <4 x i32> %wide.masked.load14 to <4 x i64>
2437 %6 = mul nsw <4 x i64> %5, %2
2438 %7 = select <4 x i1> %active.lane.mask, <4 x i64> %6, <4 x i64> zeroinitializer
2439 %8 = call i64 @llvm.experimental.vector.reduce.add.v4i64(<4 x i64> %7)
2440 %9 = add i64 %8, %vec.phi
2441 %index.next = add i32 %index, 4
2442 %10 = icmp eq i32 %index.next, %n.vec
2443 br i1 %10, label %for.cond.cleanup, label %vector.body
2444
2445for.cond.cleanup: ; preds = %vector.body, %entry
2446 %s.0.lcssa = phi i64 [ 0, %entry ], [ %9, %vector.body ]
2447 ret i64 %s.0.lcssa
2448}
2449
2450define i64 @mla8i64(i16* noalias nocapture readonly %x, i16* noalias nocapture readonly %y, i32 %n) {
2451; CHECK-LABEL: mla8i64:
2452; CHECK: @ %bb.0: @ %entry
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2453; CHECK-NEXT: .save {r7, lr}
2454; CHECK-NEXT: push {r7, lr}
2455; CHECK-NEXT: cbz r2, .LBB31_3
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2456; CHECK-NEXT: @ %bb.1: @ %vector.ph
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2457; CHECK-NEXT: mov.w r12, #0
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2458; CHECK-NEXT: mov r3, r12
2459; CHECK-NEXT: dlstp.16 lr, r2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2460; CHECK-NEXT: .LBB31_2: @ %vector.body
2461; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2462; CHECK-NEXT: vldrh.u16 q0, [r0], #16
2463; CHECK-NEXT: vldrh.u16 q1, [r1], #16
2464; CHECK-NEXT: vmlalva.s16 r12, r3, q1, q0
2465; CHECK-NEXT: letp lr, .LBB31_2
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2466; CHECK-NEXT: b .LBB31_4
2467; CHECK-NEXT: .LBB31_3:
2468; CHECK-NEXT: mov.w r12, #0
2469; CHECK-NEXT: mov r3, r12
2470; CHECK-NEXT: .LBB31_4: @ %for.cond.cleanup
2471; CHECK-NEXT: mov r0, r12
2472; CHECK-NEXT: mov r1, r3
David Green8590e5a2020-08-09 10:57:17 +0100[diff] [blame]2473; CHECK-NEXT: pop {r7, pc}
David Green25e38c32020-08-07 17:16:56 +0100[diff] [blame]2474entry:
2475 %cmp9.not = icmp eq i32 %n, 0
2476 br i1 %cmp9.not, label %for.cond.cleanup, label %vector.ph
2477
2478vector.ph: ; preds = %entry
2479 %n.rnd.up = add i32 %n, 7
2480 %n.vec = and i32 %n.rnd.up, -8
2481 %trip.count.minus.1 = add i32 %n, -1
2482 br label %vector.body
2483
2484vector.body: ; preds = %vector.body, %vector.ph
2485 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
2486 %vec.phi = phi i64 [ 0, %vector.ph ], [ %9, %vector.body ]
2487 %active.lane.mask = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 %index, i32 %trip.count.minus.1)
2488 %0 = getelementptr inbounds i16, i16* %x, i32 %index
2489 %1 = bitcast i16* %0 to <8 x i16>*
2490 %wide.masked.load = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %1, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
2491 %2 = sext <8 x i16> %wide.masked.load to <8 x i64>
2492 %3 = getelementptr inbounds i16, i16* %y, i32 %index
2493 %4 = bitcast i16* %3 to <8 x i16>*
2494 %wide.masked.load14 = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* %4, i32 2, <8 x i1> %active.lane.mask, <8 x i16> undef)
2495 %5 = sext <8 x i16> %wide.masked.load14 to <8 x i64>
2496 %6 = mul nsw <8 x i64> %5, %2
2497 %7 = select <8 x i1> %active.lane.mask, <8 x i64> %6, <8 x i64> zeroinitializer
2498 %8 = call i64 @llvm.experimental.vector.reduce.add.v8i64(<8 x i64> %7)
2499 %9 = add i64 %8, %vec.phi
2500 %index.next = add i32 %index, 8
2501 %10 = icmp eq i32 %index.next, %n.vec
2502 br i1 %10, label %for.cond.cleanup, label %vector.body
2503
2504for.cond.cleanup: ; preds = %vector.body, %entry
2505 %s.0.lcssa = phi i64 [ 0, %entry ], [ %9, %vector.body ]
2506 ret i64 %s.0.lcssa
2507}
2508
2509declare <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32, i32) #1
2510declare <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>*, i32 immarg, <4 x i1>, <4 x i32>) #2
2511declare <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32, i32) #1
2512declare <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>*, i32 immarg, <8 x i1>, <8 x i16>) #2
2513declare i32 @llvm.experimental.vector.reduce.add.v8i32(<8 x i32>) #3
2514declare <16 x i1> @llvm.get.active.lane.mask.v16i1.i32(i32, i32) #1
2515declare <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>*, i32 immarg, <16 x i1>, <16 x i8>) #2
2516declare i32 @llvm.experimental.vector.reduce.add.v16i32(<16 x i32>) #3
2517declare i16 @llvm.experimental.vector.reduce.add.v8i16(<8 x i16>) #3
2518declare i16 @llvm.experimental.vector.reduce.add.v16i16(<16 x i16>) #3
2519declare i8 @llvm.experimental.vector.reduce.add.v16i8(<16 x i8>) #3
2520declare i64 @llvm.experimental.vector.reduce.add.v4i64(<4 x i64>) #3
2521declare i64 @llvm.experimental.vector.reduce.add.v8i64(<8 x i64>) #3
2522
David Greenc7551572020-06-09 11:04:29 +0100[diff] [blame]2523declare i32 @llvm.experimental.vector.reduce.add.v4i32(<4 x i32>)
2524declare i32 @llvm.experimental.vector.reduce.mul.v4i32(<4 x i32>)
2525declare i32 @llvm.experimental.vector.reduce.and.v4i32(<4 x i32>)
2526declare i32 @llvm.experimental.vector.reduce.or.v4i32(<4 x i32>)
2527declare i32 @llvm.experimental.vector.reduce.xor.v4i32(<4 x i32>)
2528declare float @llvm.experimental.vector.reduce.v2.fadd.f32.v4f32(float, <4 x float>)
2529declare float @llvm.experimental.vector.reduce.v2.fmul.f32.v4f32(float, <4 x float>)
2530declare i32 @llvm.experimental.vector.reduce.smin.v4i32(<4 x i32>)
2531declare i32 @llvm.experimental.vector.reduce.smax.v4i32(<4 x i32>)
2532declare i32 @llvm.experimental.vector.reduce.umin.v4i32(<4 x i32>)
2533declare i32 @llvm.experimental.vector.reduce.umax.v4i32(<4 x i32>)
2534declare float @llvm.experimental.vector.reduce.fmin.v4f32(<4 x float>)
2535declare float @llvm.experimental.vector.reduce.fmax.v4f32(<4 x float>)