1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp %s -o - | FileCheck %s
3
4; Check some LSR loop postinc
5
6; fma loop with a destination that is the same as one of the sources
7define void @fma(float* noalias nocapture readonly %A, float* noalias nocapture readonly %B, float* noalias nocapture %C, i32 %n) {
8; CHECK-LABEL: fma:
9; CHECK:       @ %bb.0: @ %entry
10; CHECK-NEXT:    .save {r4, r5, r6, lr}
11; CHECK-NEXT:    push {r4, r5, r6, lr}
12; CHECK-NEXT:    cmp r3, #1
13; CHECK-NEXT:    blt .LBB0_8
14; CHECK-NEXT:  @ %bb.1: @ %for.body.preheader
15; CHECK-NEXT:    cmp r3, #3
16; CHECK-NEXT:    bhi .LBB0_3
17; CHECK-NEXT:  @ %bb.2:
18; CHECK-NEXT:    mov.w r12, #0
19; CHECK-NEXT:    b .LBB0_6
20; CHECK-NEXT:  .LBB0_3: @ %vector.ph
21; CHECK-NEXT:    bic r12, r3, #3
22; CHECK-NEXT:    movs r5, #1
23; CHECK-NEXT:    sub.w r6, r12, #4
24; CHECK-NEXT:    mov r4, r0
25; CHECK-NEXT:    add.w lr, r5, r6, lsr #2
26; CHECK-NEXT:    mov r5, r1
27; CHECK-NEXT:    mov r6, r2
28; CHECK-NEXT:  .LBB0_4: @ %vector.body
29; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
30; CHECK-NEXT:    vldrw.u32 q0, [r4], #16
31; CHECK-NEXT:    vldrw.u32 q1, [r5], #16
32; CHECK-NEXT:    vldrw.u32 q2, [r6]
33; CHECK-NEXT:    vfma.f32 q2, q1, q0
34; CHECK-NEXT:    vstrb.8 q2, [r6], #16
35; CHECK-NEXT:    le lr, .LBB0_4
36; CHECK-NEXT:  @ %bb.5: @ %middle.block
37; CHECK-NEXT:    cmp r12, r3
38; CHECK-NEXT:    it eq
39; CHECK-NEXT:    popeq {r4, r5, r6, pc}
40; CHECK-NEXT:  .LBB0_6: @ %for.body.preheader12
41; CHECK-NEXT:    sub.w lr, r3, r12
42; CHECK-NEXT:    add.w r0, r0, r12, lsl #2
43; CHECK-NEXT:    add.w r1, r1, r12, lsl #2
44; CHECK-NEXT:    add.w r2, r2, r12, lsl #2
45; CHECK-NEXT:  .LBB0_7: @ %for.body
46; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
47; CHECK-NEXT:    vldmia r0!, {s0}
48; CHECK-NEXT:    vldmia r1!, {s2}
49; CHECK-NEXT:    vldr s4, [r2]
50; CHECK-NEXT:    vfma.f32 s4, s2, s0
51; CHECK-NEXT:    vstmia r2!, {s4}
52; CHECK-NEXT:    le lr, .LBB0_7
53; CHECK-NEXT:  .LBB0_8: @ %for.cond.cleanup
54; CHECK-NEXT:    pop {r4, r5, r6, pc}
55entry:
56  %cmp8 = icmp sgt i32 %n, 0
57  br i1 %cmp8, label %for.body.preheader, label %for.cond.cleanup
58
59for.body.preheader:                               ; preds = %entry
60  %min.iters.check = icmp ult i32 %n, 4
61  br i1 %min.iters.check, label %for.body.preheader12, label %vector.ph
62
63for.body.preheader12:                             ; preds = %middle.block, %for.body.preheader
64  %i.09.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]
65  br label %for.body
66
67vector.ph:                                        ; preds = %for.body.preheader
68  %n.vec = and i32 %n, -4
69  br label %vector.body
70
71vector.body:                                      ; preds = %vector.body, %vector.ph
72  %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
73  %0 = getelementptr inbounds float, float* %A, i32 %index
74  %1 = bitcast float* %0 to <4 x float>*
75  %wide.load = load <4 x float>, <4 x float>* %1, align 4
76  %2 = getelementptr inbounds float, float* %B, i32 %index
77  %3 = bitcast float* %2 to <4 x float>*
78  %wide.load10 = load <4 x float>, <4 x float>* %3, align 4
79  %4 = fmul fast <4 x float> %wide.load10, %wide.load
80  %5 = getelementptr inbounds float, float* %C, i32 %index
81  %6 = bitcast float* %5 to <4 x float>*
82  %wide.load11 = load <4 x float>, <4 x float>* %6, align 4
83  %7 = fadd fast <4 x float> %wide.load11, %4
84  %8 = bitcast float* %5 to <4 x float>*
85  store <4 x float> %7, <4 x float>* %8, align 4
86  %index.next = add i32 %index, 4
87  %9 = icmp eq i32 %index.next, %n.vec
88  br i1 %9, label %middle.block, label %vector.body
89
90middle.block:                                     ; preds = %vector.body
91  %cmp.n = icmp eq i32 %n.vec, %n
92  br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader12
93
94for.cond.cleanup:                                 ; preds = %for.body, %middle.block, %entry
95  ret void
96
97for.body:                                         ; preds = %for.body.preheader12, %for.body
98  %i.09 = phi i32 [ %inc, %for.body ], [ %i.09.ph, %for.body.preheader12 ]
99  %arrayidx = getelementptr inbounds float, float* %A, i32 %i.09
100  %10 = load float, float* %arrayidx, align 4
101  %arrayidx1 = getelementptr inbounds float, float* %B, i32 %i.09
102  %11 = load float, float* %arrayidx1, align 4
103  %mul = fmul fast float %11, %10
104  %arrayidx2 = getelementptr inbounds float, float* %C, i32 %i.09
105  %12 = load float, float* %arrayidx2, align 4
106  %add = fadd fast float %12, %mul
107  store float %add, float* %arrayidx2, align 4
108  %inc = add nuw nsw i32 %i.09, 1
109  %exitcond = icmp eq i32 %inc, %n
110  br i1 %exitcond, label %for.cond.cleanup, label %for.body
111}
112
113
114; Same as above but tail predicated
115; FIXME: The postinc here is put on the load, not the store. An extra mov is needed in the loop because of it.
116define void @fma_tailpred(float* noalias nocapture readonly %A, float* noalias nocapture readonly %B, float* noalias nocapture %C, i32 %n) {
117; CHECK-LABEL: fma_tailpred:
118; CHECK:       @ %bb.0: @ %entry
119; CHECK-NEXT:    .save {r4, lr}
120; CHECK-NEXT:    push {r4, lr}
121; CHECK-NEXT:    .vsave {d8, d9}
122; CHECK-NEXT:    vpush {d8, d9}
123; CHECK-NEXT:    cmp r3, #1
124; CHECK-NEXT:    blt .LBB1_3
125; CHECK-NEXT:  @ %bb.1: @ %vector.ph
126; CHECK-NEXT:    add.w r12, r3, #3
127; CHECK-NEXT:    mov.w lr, #1
128; CHECK-NEXT:    bic r12, r12, #3
129; CHECK-NEXT:    adr r4, .LCPI1_0
130; CHECK-NEXT:    sub.w r12, r12, #4
131; CHECK-NEXT:    vldrw.u32 q0, [r4]
132; CHECK-NEXT:    add.w lr, lr, r12, lsr #2
133; CHECK-NEXT:    sub.w r12, r3, #1
134; CHECK-NEXT:    movs r3, #0
135; CHECK-NEXT:    vdup.32 q1, r12
136; CHECK-NEXT:  .LBB1_2: @ %vector.body
137; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
138; CHECK-NEXT:    vdup.32 q2, r3
139; CHECK-NEXT:    adds r3, #4
140; CHECK-NEXT:    vorr q2, q2, q0
141; CHECK-NEXT:    vpttt.u32 cs, q1, q2
142; CHECK-NEXT:    vldrwt.u32 q2, [r0], #16
143; CHECK-NEXT:    vldrwt.u32 q3, [r1], #16
144; CHECK-NEXT:    vldrwt.u32 q4, [r2]
145; CHECK-NEXT:    vfma.f32 q4, q3, q2
146; CHECK-NEXT:    vpst
147; CHECK-NEXT:    vstrwt.32 q4, [r2], #16
148; CHECK-NEXT:    le lr, .LBB1_2
149; CHECK-NEXT:  .LBB1_3: @ %for.cond.cleanup
150; CHECK-NEXT:    vpop {d8, d9}
151; CHECK-NEXT:    pop {r4, pc}
152; CHECK-NEXT:    .p2align 4
153; CHECK-NEXT:  @ %bb.4:
154; CHECK-NEXT:  .LCPI1_0:
155; CHECK-NEXT:    .long 0 @ 0x0
156; CHECK-NEXT:    .long 1 @ 0x1
157; CHECK-NEXT:    .long 2 @ 0x2
158; CHECK-NEXT:    .long 3 @ 0x3
159entry:
160  %cmp8 = icmp sgt i32 %n, 0
161  br i1 %cmp8, label %vector.ph, label %for.cond.cleanup
162
163vector.ph:                                        ; preds = %entry
164  %n.rnd.up = add i32 %n, 3
165  %n.vec = and i32 %n.rnd.up, -4
166  %trip.count.minus.1 = add i32 %n, -1
167  %broadcast.splatinsert10 = insertelement <4 x i32> undef, i32 %trip.count.minus.1, i32 0
168  %broadcast.splat11 = shufflevector <4 x i32> %broadcast.splatinsert10, <4 x i32> undef, <4 x i32> zeroinitializer
169  br label %vector.body
170
171vector.body:                                      ; preds = %vector.body, %vector.ph
172  %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
173  %broadcast.splatinsert = insertelement <4 x i32> undef, i32 %index, i32 0
174  %broadcast.splat = shufflevector <4 x i32> %broadcast.splatinsert, <4 x i32> undef, <4 x i32> zeroinitializer
175  %induction = or <4 x i32> %broadcast.splat, <i32 0, i32 1, i32 2, i32 3>
176  %0 = getelementptr inbounds float, float* %A, i32 %index
177  %1 = icmp ule <4 x i32> %induction, %broadcast.splat11
178  %2 = bitcast float* %0 to <4 x float>*
179  %wide.masked.load = call <4 x float> @llvm.masked.load.v4f32.p0v4f32(<4 x float>* %2, i32 4, <4 x i1> %1, <4 x float> undef)
180  %3 = getelementptr inbounds float, float* %B, i32 %index
181  %4 = bitcast float* %3 to <4 x float>*
182  %wide.masked.load12 = call <4 x float> @llvm.masked.load.v4f32.p0v4f32(<4 x float>* %4, i32 4, <4 x i1> %1, <4 x float> undef)
183  %5 = fmul fast <4 x float> %wide.masked.load12, %wide.masked.load
184  %6 = getelementptr inbounds float, float* %C, i32 %index
185  %7 = bitcast float* %6 to <4 x float>*
186  %wide.masked.load13 = call <4 x float> @llvm.masked.load.v4f32.p0v4f32(<4 x float>* %7, i32 4, <4 x i1> %1, <4 x float> undef)
187  %8 = fadd fast <4 x float> %wide.masked.load13, %5
188  %9 = bitcast float* %6 to <4 x float>*
189  call void @llvm.masked.store.v4f32.p0v4f32(<4 x float> %8, <4 x float>* %9, i32 4, <4 x i1> %1)
190  %index.next = add i32 %index, 4
191  %10 = icmp eq i32 %index.next, %n.vec
192  br i1 %10, label %for.cond.cleanup, label %vector.body
193
194for.cond.cleanup:                                 ; preds = %vector.body, %entry
195  ret void
196}
197
198
199; Multiple loads of the loop with a common base
200define i8* @test(i8* nocapture readonly %input_row, i8* nocapture readonly %input_col, i16 zeroext %output_ch, i16 zeroext %num_cols, i32 %col_offset, i16 signext %activation_min, i16 zeroext %row_len, i32* nocapture readonly %bias, i8* returned %out) {
201; CHECK-LABEL: test:
202; CHECK:       @ %bb.0: @ %entry
203; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
204; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}
205; CHECK-NEXT:    .pad #20
206; CHECK-NEXT:    sub sp, #20
207; CHECK-NEXT:    cmp r3, #4
208; CHECK-NEXT:    strd r0, r1, [sp, #12] @ 8-byte Folded Spill
209; CHECK-NEXT:    bne .LBB2_8
210; CHECK-NEXT:  @ %bb.1: @ %for.cond.preheader
211; CHECK-NEXT:    cmp r2, #0
212; CHECK-NEXT:    beq .LBB2_8
213; CHECK-NEXT:  @ %bb.2: @ %for.body.lr.ph
214; CHECK-NEXT:    ldr r3, [sp, #64]
215; CHECK-NEXT:    mov.w r9, #0
216; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
217; CHECK-NEXT:    ldr.w r11, [sp, #56]
218; CHECK-NEXT:    add.w r0, r1, r3, lsl #1
219; CHECK-NEXT:    str r0, [sp, #8] @ 4-byte Spill
220; CHECK-NEXT:    adds r0, r1, r3
221; CHECK-NEXT:    str r0, [sp, #4] @ 4-byte Spill
222; CHECK-NEXT:    add.w r0, r3, r3, lsl #1
223; CHECK-NEXT:    add r0, r1
224; CHECK-NEXT:    str r0, [sp] @ 4-byte Spill
225; CHECK-NEXT:    adds r0, r3, #7
226; CHECK-NEXT:    lsrs r0, r0, #3
227; CHECK-NEXT:    b .LBB2_5
228; CHECK-NEXT:  .LBB2_3: @ in Loop: Header=BB2_5 Depth=1
229; CHECK-NEXT:    mov r10, r12
230; CHECK-NEXT:    mov r8, r12
231; CHECK-NEXT:    mov r6, r12
232; CHECK-NEXT:  .LBB2_4: @ %for.cond.cleanup23
233; CHECK-NEXT:    @ in Loop: Header=BB2_5 Depth=1
234; CHECK-NEXT:    ldr r3, [sp, #72]
235; CHECK-NEXT:    add.w r1, r8, r10
236; CHECK-NEXT:    add r1, r6
237; CHECK-NEXT:    add r1, r12
238; CHECK-NEXT:    strb.w r1, [r3, r9]
239; CHECK-NEXT:    add.w r9, r9, #1
240; CHECK-NEXT:    cmp r9, r2
241; CHECK-NEXT:    beq .LBB2_8
242; CHECK-NEXT:  .LBB2_5: @ %for.body
243; CHECK-NEXT:    @ =>This Loop Header: Depth=1
244; CHECK-NEXT:    @ Child Loop BB2_7 Depth 2
245; CHECK-NEXT:    ldr r1, [sp, #68]
246; CHECK-NEXT:    ldr.w r12, [r1, r9, lsl #2]
247; CHECK-NEXT:    subs r1, r0, r0
248; CHECK-NEXT:    ble .LBB2_3
249; CHECK-NEXT:  @ %bb.6: @ %for.body24.preheader
250; CHECK-NEXT:    @ in Loop: Header=BB2_5 Depth=1
251; CHECK-NEXT:    ldr r7, [sp, #64]
252; CHECK-NEXT:    mov r6, r12
253; CHECK-NEXT:    ldr r3, [sp, #12] @ 4-byte Reload
254; CHECK-NEXT:    dls lr, r1
255; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
256; CHECK-NEXT:    mov r8, r12
257; CHECK-NEXT:    mla r7, r9, r7, r3
258; CHECK-NEXT:    ldr r5, [sp, #8] @ 4-byte Reload
259; CHECK-NEXT:    ldrd r4, r3, [sp] @ 8-byte Folded Reload
260; CHECK-NEXT:    mov r10, r12
261; CHECK-NEXT:  .LBB2_7: @ %for.body24
262; CHECK-NEXT:    @ Parent Loop BB2_5 Depth=1
263; CHECK-NEXT:    @ => This Inner Loop Header: Depth=2
264; CHECK-NEXT:    vldrb.s16 q0, [r4], #8
265; CHECK-NEXT:    vadd.i16 q1, q0, r11
266; CHECK-NEXT:    vldrb.s16 q0, [r7], #8
267; CHECK-NEXT:    vmlava.s16 r12, q0, q1
268; CHECK-NEXT:    vldrb.s16 q1, [r5], #8
269; CHECK-NEXT:    vadd.i16 q1, q1, r11
270; CHECK-NEXT:    vmlava.s16 r6, q0, q1
271; CHECK-NEXT:    vldrb.s16 q1, [r3], #8
272; CHECK-NEXT:    vadd.i16 q1, q1, r11
273; CHECK-NEXT:    vmlava.s16 r8, q0, q1
274; CHECK-NEXT:    vldrb.s16 q1, [r1], #8
275; CHECK-NEXT:    vadd.i16 q1, q1, r11
276; CHECK-NEXT:    vmlava.s16 r10, q0, q1
277; CHECK-NEXT:    le lr, .LBB2_7
278; CHECK-NEXT:    b .LBB2_4
279; CHECK-NEXT:  .LBB2_8: @ %if.end
280; CHECK-NEXT:    ldr r0, [sp, #72]
281; CHECK-NEXT:    add sp, #20
282; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
283entry:
284  %cmp = icmp eq i16 %num_cols, 4
285  br i1 %cmp, label %for.cond.preheader, label %if.end
286
287for.cond.preheader:                               ; preds = %entry
288  %conv2 = zext i16 %output_ch to i32
289  %cmp3114 = icmp eq i16 %output_ch, 0
290  br i1 %cmp3114, label %if.end, label %for.body.lr.ph
291
292for.body.lr.ph:                                   ; preds = %for.cond.preheader
293  %conv5 = zext i16 %row_len to i32
294  %add.ptr9 = getelementptr inbounds i8, i8* %input_col, i32 %conv5
295  %mul11 = shl nuw nsw i32 %conv5, 1
296  %add.ptr12 = getelementptr inbounds i8, i8* %input_col, i32 %mul11
297  %mul14 = mul nuw nsw i32 %conv5, 3
298  %add.ptr15 = getelementptr inbounds i8, i8* %input_col, i32 %mul14
299  %add = add nuw nsw i32 %conv5, 7
300  %div = lshr i32 %add, 3
301  %conv25 = trunc i32 %col_offset to i16
302  %.splatinsert = insertelement <8 x i16> undef, i16 %conv25, i32 0
303  %.splat = shufflevector <8 x i16> %.splatinsert, <8 x i16> undef, <8 x i32> zeroinitializer
304  br label %for.body
305
306for.body:                                         ; preds = %for.cond.cleanup23, %for.body.lr.ph
307  %i_out_ch.0116 = phi i32 [ 0, %for.body.lr.ph ], [ %inc37, %for.cond.cleanup23 ]
308  %i_row_loop.0115 = phi i32 [ undef, %for.body.lr.ph ], [ %i_row_loop.1.lcssa, %for.cond.cleanup23 ]
309  %arrayidx = getelementptr inbounds i32, i32* %bias, i32 %i_out_ch.0116
310  %0 = load i32, i32* %arrayidx, align 4
311  %cmp2199 = icmp slt i32 %i_row_loop.0115, %div
312  br i1 %cmp2199, label %for.body24.preheader, label %for.cond.cleanup23
313
314for.body24.preheader:                             ; preds = %for.body
315  %mul = mul nuw nsw i32 %i_out_ch.0116, %conv5
316  %add.ptr = getelementptr inbounds i8, i8* %input_row, i32 %mul
317  br label %for.body24
318
319for.cond.cleanup23:                               ; preds = %for.body24, %for.body
320  %acc_0.0.lcssa = phi i32 [ %0, %for.body ], [ %20, %for.body24 ]
321  %acc_1.0.lcssa = phi i32 [ %0, %for.body ], [ %21, %for.body24 ]
322  %acc_2.0.lcssa = phi i32 [ %0, %for.body ], [ %22, %for.body24 ]
323  %acc_3.0.lcssa = phi i32 [ %0, %for.body ], [ %23, %for.body24 ]
324  %i_row_loop.1.lcssa = phi i32 [ %i_row_loop.0115, %for.body ], [ %div, %for.body24 ]
325  %add31 = add nsw i32 %acc_1.0.lcssa, %acc_0.0.lcssa
326  %add32 = add nsw i32 %add31, %acc_2.0.lcssa
327  %add33 = add nsw i32 %add32, %acc_3.0.lcssa
328  %conv34 = trunc i32 %add33 to i8
329  %arrayidx35 = getelementptr inbounds i8, i8* %out, i32 %i_out_ch.0116
330  store i8 %conv34, i8* %arrayidx35, align 1
331  %inc37 = add nuw nsw i32 %i_out_ch.0116, 1
332  %exitcond120 = icmp eq i32 %inc37, %conv2
333  br i1 %exitcond120, label %if.end, label %for.body
334
335for.body24:                                       ; preds = %for.body24, %for.body24.preheader
336  %ip_r0.0109 = phi i8* [ %add.ptr26, %for.body24 ], [ %add.ptr, %for.body24.preheader ]
337  %ip_c0.0108 = phi i8* [ %add.ptr27, %for.body24 ], [ %input_col, %for.body24.preheader ]
338  %ip_c1.0107 = phi i8* [ %add.ptr28, %for.body24 ], [ %add.ptr9, %for.body24.preheader ]
339  %ip_c2.0106 = phi i8* [ %add.ptr29, %for.body24 ], [ %add.ptr12, %for.body24.preheader ]
340  %i_row_loop.1105 = phi i32 [ %inc, %for.body24 ], [ %i_row_loop.0115, %for.body24.preheader ]
341  %ip_c3.0104 = phi i8* [ %add.ptr30, %for.body24 ], [ %add.ptr15, %for.body24.preheader ]
342  %acc_3.0103 = phi i32 [ %23, %for.body24 ], [ %0, %for.body24.preheader ]
343  %acc_2.0102 = phi i32 [ %22, %for.body24 ], [ %0, %for.body24.preheader ]
344  %acc_1.0101 = phi i32 [ %21, %for.body24 ], [ %0, %for.body24.preheader ]
345  %acc_0.0100 = phi i32 [ %20, %for.body24 ], [ %0, %for.body24.preheader ]
346  %1 = bitcast i8* %ip_r0.0109 to <8 x i8>*
347  %2 = load <8 x i8>, <8 x i8>* %1, align 1
348  %3 = sext <8 x i8> %2 to <8 x i16>
349  %add.ptr26 = getelementptr inbounds i8, i8* %ip_r0.0109, i32 8
350  %4 = bitcast i8* %ip_c0.0108 to <8 x i8>*
351  %5 = load <8 x i8>, <8 x i8>* %4, align 1
352  %6 = sext <8 x i8> %5 to <8 x i16>
353  %add.ptr27 = getelementptr inbounds i8, i8* %ip_c0.0108, i32 8
354  %7 = add <8 x i16> %.splat, %6
355  %8 = bitcast i8* %ip_c1.0107 to <8 x i8>*
356  %9 = load <8 x i8>, <8 x i8>* %8, align 1
357  %10 = sext <8 x i8> %9 to <8 x i16>
358  %add.ptr28 = getelementptr inbounds i8, i8* %ip_c1.0107, i32 8
359  %11 = add <8 x i16> %.splat, %10
360  %12 = bitcast i8* %ip_c2.0106 to <8 x i8>*
361  %13 = load <8 x i8>, <8 x i8>* %12, align 1
362  %14 = sext <8 x i8> %13 to <8 x i16>
363  %add.ptr29 = getelementptr inbounds i8, i8* %ip_c2.0106, i32 8
364  %15 = add <8 x i16> %.splat, %14
365  %16 = bitcast i8* %ip_c3.0104 to <8 x i8>*
366  %17 = load <8 x i8>, <8 x i8>* %16, align 1
367  %18 = sext <8 x i8> %17 to <8 x i16>
368  %add.ptr30 = getelementptr inbounds i8, i8* %ip_c3.0104, i32 8
369  %19 = add <8 x i16> %.splat, %18
370  %20 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_0.0100, <8 x i16> %3, <8 x i16> %7)
371  %21 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_1.0101, <8 x i16> %3, <8 x i16> %11)
372  %22 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_2.0102, <8 x i16> %3, <8 x i16> %15)
373  %23 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_3.0103, <8 x i16> %3, <8 x i16> %19)
374  %inc = add nsw i32 %i_row_loop.1105, 1
375  %exitcond = icmp eq i32 %inc, %div
376  br i1 %exitcond, label %for.cond.cleanup23, label %for.body24
377
378if.end:                                           ; preds = %for.cond.cleanup23, %for.cond.preheader, %entry
379  ret i8* %out
380}
381
382; Same as above with optsize
383define i8* @test_optsize(i8* nocapture readonly %input_row, i8* nocapture readonly %input_col, i16 zeroext %output_ch, i16 zeroext %num_cols, i32 %col_offset, i16 signext %activation_min, i16 zeroext %row_len, i32* nocapture readonly %bias, i8* returned %out) optsize {
384; CHECK-LABEL: test_optsize:
385; CHECK:       @ %bb.0: @ %entry
386; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
387; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}
388; CHECK-NEXT:    .pad #20
389; CHECK-NEXT:    sub sp, #20
390; CHECK-NEXT:    cmp r3, #4
391; CHECK-NEXT:    strd r0, r1, [sp, #12] @ 8-byte Folded Spill
392; CHECK-NEXT:    bne .LBB3_8
393; CHECK-NEXT:  @ %bb.1: @ %for.cond.preheader
394; CHECK-NEXT:    cmp r2, #0
395; CHECK-NEXT:    beq .LBB3_8
396; CHECK-NEXT:  @ %bb.2: @ %for.body.lr.ph
397; CHECK-NEXT:    ldr r3, [sp, #64]
398; CHECK-NEXT:    mov.w r9, #0
399; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
400; CHECK-NEXT:    ldr.w r11, [sp, #56]
401; CHECK-NEXT:    add.w r0, r1, r3, lsl #1
402; CHECK-NEXT:    str r0, [sp, #8] @ 4-byte Spill
403; CHECK-NEXT:    adds r0, r1, r3
404; CHECK-NEXT:    str r0, [sp, #4] @ 4-byte Spill
405; CHECK-NEXT:    add.w r0, r3, r3, lsl #1
406; CHECK-NEXT:    add r0, r1
407; CHECK-NEXT:    str r0, [sp] @ 4-byte Spill
408; CHECK-NEXT:    adds r0, r3, #7
409; CHECK-NEXT:    lsrs r0, r0, #3
410; CHECK-NEXT:  .LBB3_3: @ %for.body
411; CHECK-NEXT:    @ =>This Loop Header: Depth=1
412; CHECK-NEXT:    @ Child Loop BB3_5 Depth 2
413; CHECK-NEXT:    ldr r1, [sp, #68]
414; CHECK-NEXT:    ldr.w r12, [r1, r9, lsl #2]
415; CHECK-NEXT:    subs r1, r0, r0
416; CHECK-NEXT:    ble .LBB3_6
417; CHECK-NEXT:  @ %bb.4: @ %for.body24.preheader
418; CHECK-NEXT:    @ in Loop: Header=BB3_3 Depth=1
419; CHECK-NEXT:    ldr r7, [sp, #64]
420; CHECK-NEXT:    mov r6, r12
421; CHECK-NEXT:    ldr r3, [sp, #12] @ 4-byte Reload
422; CHECK-NEXT:    dls lr, r1
423; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
424; CHECK-NEXT:    mov r8, r12
425; CHECK-NEXT:    mla r7, r9, r7, r3
426; CHECK-NEXT:    ldr r5, [sp, #8] @ 4-byte Reload
427; CHECK-NEXT:    ldrd r4, r3, [sp] @ 8-byte Folded Reload
428; CHECK-NEXT:    mov r10, r12
429; CHECK-NEXT:  .LBB3_5: @ %for.body24
430; CHECK-NEXT:    @ Parent Loop BB3_3 Depth=1
431; CHECK-NEXT:    @ => This Inner Loop Header: Depth=2
432; CHECK-NEXT:    vldrb.s16 q0, [r4], #8
433; CHECK-NEXT:    vadd.i16 q1, q0, r11
434; CHECK-NEXT:    vldrb.s16 q0, [r7], #8
435; CHECK-NEXT:    vmlava.s16 r12, q0, q1
436; CHECK-NEXT:    vldrb.s16 q1, [r5], #8
437; CHECK-NEXT:    vadd.i16 q1, q1, r11
438; CHECK-NEXT:    vmlava.s16 r6, q0, q1
439; CHECK-NEXT:    vldrb.s16 q1, [r3], #8
440; CHECK-NEXT:    vadd.i16 q1, q1, r11
441; CHECK-NEXT:    vmlava.s16 r8, q0, q1
442; CHECK-NEXT:    vldrb.s16 q1, [r1], #8
443; CHECK-NEXT:    vadd.i16 q1, q1, r11
444; CHECK-NEXT:    vmlava.s16 r10, q0, q1
445; CHECK-NEXT:    le lr, .LBB3_5
446; CHECK-NEXT:    b .LBB3_7
447; CHECK-NEXT:  .LBB3_6: @ in Loop: Header=BB3_3 Depth=1
448; CHECK-NEXT:    mov r10, r12
449; CHECK-NEXT:    mov r8, r12
450; CHECK-NEXT:    mov r6, r12
451; CHECK-NEXT:  .LBB3_7: @ %for.cond.cleanup23
452; CHECK-NEXT:    @ in Loop: Header=BB3_3 Depth=1
453; CHECK-NEXT:    ldr r3, [sp, #72]
454; CHECK-NEXT:    add.w r1, r8, r10
455; CHECK-NEXT:    add r1, r6
456; CHECK-NEXT:    add r1, r12
457; CHECK-NEXT:    strb.w r1, [r3, r9]
458; CHECK-NEXT:    add.w r9, r9, #1
459; CHECK-NEXT:    cmp r9, r2
460; CHECK-NEXT:    bne .LBB3_3
461; CHECK-NEXT:  .LBB3_8: @ %if.end
462; CHECK-NEXT:    ldr r0, [sp, #72]
463; CHECK-NEXT:    add sp, #20
464; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
465entry:
466  %cmp = icmp eq i16 %num_cols, 4
467  br i1 %cmp, label %for.cond.preheader, label %if.end
468
469for.cond.preheader:                               ; preds = %entry
470  %conv2 = zext i16 %output_ch to i32
471  %cmp3114 = icmp eq i16 %output_ch, 0
472  br i1 %cmp3114, label %if.end, label %for.body.lr.ph
473
474for.body.lr.ph:                                   ; preds = %for.cond.preheader
475  %conv5 = zext i16 %row_len to i32
476  %add.ptr9 = getelementptr inbounds i8, i8* %input_col, i32 %conv5
477  %mul11 = shl nuw nsw i32 %conv5, 1
478  %add.ptr12 = getelementptr inbounds i8, i8* %input_col, i32 %mul11
479  %mul14 = mul nuw nsw i32 %conv5, 3
480  %add.ptr15 = getelementptr inbounds i8, i8* %input_col, i32 %mul14
481  %add = add nuw nsw i32 %conv5, 7
482  %div = lshr i32 %add, 3
483  %conv25 = trunc i32 %col_offset to i16
484  %.splatinsert = insertelement <8 x i16> undef, i16 %conv25, i32 0
485  %.splat = shufflevector <8 x i16> %.splatinsert, <8 x i16> undef, <8 x i32> zeroinitializer
486  br label %for.body
487
488for.body:                                         ; preds = %for.cond.cleanup23, %for.body.lr.ph
489  %i_out_ch.0116 = phi i32 [ 0, %for.body.lr.ph ], [ %inc37, %for.cond.cleanup23 ]
490  %i_row_loop.0115 = phi i32 [ undef, %for.body.lr.ph ], [ %i_row_loop.1.lcssa, %for.cond.cleanup23 ]
491  %arrayidx = getelementptr inbounds i32, i32* %bias, i32 %i_out_ch.0116
492  %0 = load i32, i32* %arrayidx, align 4
493  %cmp2199 = icmp slt i32 %i_row_loop.0115, %div
494  br i1 %cmp2199, label %for.body24.preheader, label %for.cond.cleanup23
495
496for.body24.preheader:                             ; preds = %for.body
497  %mul = mul nuw nsw i32 %i_out_ch.0116, %conv5
498  %add.ptr = getelementptr inbounds i8, i8* %input_row, i32 %mul
499  br label %for.body24
500
501for.cond.cleanup23:                               ; preds = %for.body24, %for.body
502  %acc_0.0.lcssa = phi i32 [ %0, %for.body ], [ %20, %for.body24 ]
503  %acc_1.0.lcssa = phi i32 [ %0, %for.body ], [ %21, %for.body24 ]
504  %acc_2.0.lcssa = phi i32 [ %0, %for.body ], [ %22, %for.body24 ]
505  %acc_3.0.lcssa = phi i32 [ %0, %for.body ], [ %23, %for.body24 ]
506  %i_row_loop.1.lcssa = phi i32 [ %i_row_loop.0115, %for.body ], [ %div, %for.body24 ]
507  %add31 = add nsw i32 %acc_1.0.lcssa, %acc_0.0.lcssa
508  %add32 = add nsw i32 %add31, %acc_2.0.lcssa
509  %add33 = add nsw i32 %add32, %acc_3.0.lcssa
510  %conv34 = trunc i32 %add33 to i8
511  %arrayidx35 = getelementptr inbounds i8, i8* %out, i32 %i_out_ch.0116
512  store i8 %conv34, i8* %arrayidx35, align 1
513  %inc37 = add nuw nsw i32 %i_out_ch.0116, 1
514  %exitcond120 = icmp eq i32 %inc37, %conv2
515  br i1 %exitcond120, label %if.end, label %for.body
516
517for.body24:                                       ; preds = %for.body24, %for.body24.preheader
518  %ip_r0.0109 = phi i8* [ %add.ptr26, %for.body24 ], [ %add.ptr, %for.body24.preheader ]
519  %ip_c0.0108 = phi i8* [ %add.ptr27, %for.body24 ], [ %input_col, %for.body24.preheader ]
520  %ip_c1.0107 = phi i8* [ %add.ptr28, %for.body24 ], [ %add.ptr9, %for.body24.preheader ]
521  %ip_c2.0106 = phi i8* [ %add.ptr29, %for.body24 ], [ %add.ptr12, %for.body24.preheader ]
522  %i_row_loop.1105 = phi i32 [ %inc, %for.body24 ], [ %i_row_loop.0115, %for.body24.preheader ]
523  %ip_c3.0104 = phi i8* [ %add.ptr30, %for.body24 ], [ %add.ptr15, %for.body24.preheader ]
524  %acc_3.0103 = phi i32 [ %23, %for.body24 ], [ %0, %for.body24.preheader ]
525  %acc_2.0102 = phi i32 [ %22, %for.body24 ], [ %0, %for.body24.preheader ]
526  %acc_1.0101 = phi i32 [ %21, %for.body24 ], [ %0, %for.body24.preheader ]
527  %acc_0.0100 = phi i32 [ %20, %for.body24 ], [ %0, %for.body24.preheader ]
528  %1 = bitcast i8* %ip_r0.0109 to <8 x i8>*
529  %2 = load <8 x i8>, <8 x i8>* %1, align 1
530  %3 = sext <8 x i8> %2 to <8 x i16>
531  %add.ptr26 = getelementptr inbounds i8, i8* %ip_r0.0109, i32 8
532  %4 = bitcast i8* %ip_c0.0108 to <8 x i8>*
533  %5 = load <8 x i8>, <8 x i8>* %4, align 1
534  %6 = sext <8 x i8> %5 to <8 x i16>
535  %add.ptr27 = getelementptr inbounds i8, i8* %ip_c0.0108, i32 8
536  %7 = add <8 x i16> %.splat, %6
537  %8 = bitcast i8* %ip_c1.0107 to <8 x i8>*
538  %9 = load <8 x i8>, <8 x i8>* %8, align 1
539  %10 = sext <8 x i8> %9 to <8 x i16>
540  %add.ptr28 = getelementptr inbounds i8, i8* %ip_c1.0107, i32 8
541  %11 = add <8 x i16> %.splat, %10
542  %12 = bitcast i8* %ip_c2.0106 to <8 x i8>*
543  %13 = load <8 x i8>, <8 x i8>* %12, align 1
544  %14 = sext <8 x i8> %13 to <8 x i16>
545  %add.ptr29 = getelementptr inbounds i8, i8* %ip_c2.0106, i32 8
546  %15 = add <8 x i16> %.splat, %14
547  %16 = bitcast i8* %ip_c3.0104 to <8 x i8>*
548  %17 = load <8 x i8>, <8 x i8>* %16, align 1
549  %18 = sext <8 x i8> %17 to <8 x i16>
550  %add.ptr30 = getelementptr inbounds i8, i8* %ip_c3.0104, i32 8
551  %19 = add <8 x i16> %.splat, %18
552  %20 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_0.0100, <8 x i16> %3, <8 x i16> %7)
553  %21 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_1.0101, <8 x i16> %3, <8 x i16> %11)
554  %22 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_2.0102, <8 x i16> %3, <8 x i16> %15)
555  %23 = tail call i32 @llvm.arm.mve.vmldava.v8i16(i32 0, i32 0, i32 0, i32 %acc_3.0103, <8 x i16> %3, <8 x i16> %19)
556  %inc = add nsw i32 %i_row_loop.1105, 1
557  %exitcond = icmp eq i32 %inc, %div
558  br i1 %exitcond, label %for.cond.cleanup23, label %for.body24
559
560if.end:                                           ; preds = %for.cond.cleanup23, %for.cond.preheader, %entry
561  ret i8* %out
562}
563
564
565; Similar but predicated
566define i32 @arm_nn_mat_mul_core_4x_s8(i32 %row_elements, i32 %offset, i8* %row_base, i8* %col_base, i32* nocapture readnone %sum_col, i32* nocapture %output) {
567; CHECK-LABEL: arm_nn_mat_mul_core_4x_s8:
568; CHECK:       @ %bb.0: @ %entry
569; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, r10, lr}
570; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, r10, lr}
571; CHECK-NEXT:    ldr.w r12, [sp, #32]
572; CHECK-NEXT:    cmp r0, #1
573; CHECK-NEXT:    blt .LBB4_3
574; CHECK-NEXT:  @ %bb.1: @ %for.body.preheader
575; CHECK-NEXT:    add.w r5, r2, r1, lsl #1
576; CHECK-NEXT:    mov.w r8, #0
577; CHECK-NEXT:    movs r4, #0
578; CHECK-NEXT:    mov.w r10, #0
579; CHECK-NEXT:    movs r6, #0
580; CHECK-NEXT:    adds r7, r2, r1
581; CHECK-NEXT:    add.w r1, r1, r1, lsl #1
582; CHECK-NEXT:    add r1, r2
583; CHECK-NEXT:    dlstp.8 lr, r0
584; CHECK-NEXT:  .LBB4_2: @ %for.body
585; CHECK-NEXT:    @ =>This Inner Loop Header: Depth=1
586; CHECK-NEXT:    vldrb.u8 q0, [r3], #16
587; CHECK-NEXT:    vldrb.u8 q1, [r1], #16
588; CHECK-NEXT:    vmlava.s8 r10, q1, q0
589; CHECK-NEXT:    vldrb.u8 q1, [r5], #16
590; CHECK-NEXT:    vmlava.s8 r4, q1, q0
591; CHECK-NEXT:    vldrb.u8 q1, [r7], #16
592; CHECK-NEXT:    vmlava.s8 r6, q1, q0
593; CHECK-NEXT:    vldrb.u8 q1, [r2], #16
594; CHECK-NEXT:    vmlava.s8 r8, q1, q0
595; CHECK-NEXT:    letp lr, .LBB4_2
596; CHECK-NEXT:    b .LBB4_4
597; CHECK-NEXT:  .LBB4_3:
598; CHECK-NEXT:    mov.w r10, #0
599; CHECK-NEXT:    movs r4, #0
600; CHECK-NEXT:    movs r6, #0
601; CHECK-NEXT:    mov.w r8, #0
602; CHECK-NEXT:  .LBB4_4: @ %for.cond.cleanup
603; CHECK-NEXT:    movs r0, #0
604; CHECK-NEXT:    strd r8, r6, [r12]
605; CHECK-NEXT:    strd r4, r10, [r12, #8]
606; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r10, pc}
607entry:
608  %add = add nsw i32 %row_elements, 15
609  %div = sdiv i32 %add, 16
610  %cmp84 = icmp sgt i32 %row_elements, 0
611  br i1 %cmp84, label %for.body.preheader, label %for.cond.cleanup
612
613for.body.preheader:                               ; preds = %entry
614  %mul2 = mul nsw i32 %offset, 3
615  %add.ptr3 = getelementptr inbounds i8, i8* %row_base, i32 %mul2
616  %mul = shl nsw i32 %offset, 1
617  %add.ptr1 = getelementptr inbounds i8, i8* %row_base, i32 %mul
618  %add.ptr = getelementptr inbounds i8, i8* %row_base, i32 %offset
619  %0 = icmp sgt i32 %div, 1
620  %smax = select i1 %0, i32 %div, i32 1
621  br label %for.body
622
623for.cond.cleanup:                                 ; preds = %for.body, %entry
624  %acc_n.sroa.12.0.lcssa = phi i32 [ 0, %entry ], [ %15, %for.body ]
625  %acc_n.sroa.9.0.lcssa = phi i32 [ 0, %entry ], [ %12, %for.body ]
626  %acc_n.sroa.6.0.lcssa = phi i32 [ 0, %entry ], [ %9, %for.body ]
627  %acc_n.sroa.0.0.lcssa = phi i32 [ 0, %entry ], [ %6, %for.body ]
628  store i32 %acc_n.sroa.0.0.lcssa, i32* %output, align 4
629  %arrayidx19 = getelementptr inbounds i32, i32* %output, i32 1
630  store i32 %acc_n.sroa.6.0.lcssa, i32* %arrayidx19, align 4
631  %arrayidx21 = getelementptr inbounds i32, i32* %output, i32 2
632  store i32 %acc_n.sroa.9.0.lcssa, i32* %arrayidx21, align 4
633  %arrayidx23 = getelementptr inbounds i32, i32* %output, i32 3
634  store i32 %acc_n.sroa.12.0.lcssa, i32* %arrayidx23, align 4
635  ret i32 0
636
637for.body:                                         ; preds = %for.body, %for.body.preheader
638  %col_base.addr.095 = phi i8* [ %add.ptr4, %for.body ], [ %col_base, %for.body.preheader ]
639  %acc_n.sroa.0.094 = phi i32 [ %6, %for.body ], [ 0, %for.body.preheader ]
640  %acc_n.sroa.6.093 = phi i32 [ %9, %for.body ], [ 0, %for.body.preheader ]
641  %acc_n.sroa.9.092 = phi i32 [ %12, %for.body ], [ 0, %for.body.preheader ]
642  %i.091 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
643  %row_elem.090 = phi i32 [ %sub, %for.body ], [ %row_elements, %for.body.preheader ]
644  %acc_n.sroa.12.089 = phi i32 [ %15, %for.body ], [ 0, %for.body.preheader ]
645  %ip_row_3.088 = phi i8* [ %add.ptr15, %for.body ], [ %add.ptr3, %for.body.preheader ]
646  %ip_row_2.087 = phi i8* [ %add.ptr14, %for.body ], [ %add.ptr1, %for.body.preheader ]
647  %ip_row_1.086 = phi i8* [ %add.ptr13, %for.body ], [ %add.ptr, %for.body.preheader ]
648  %ip_row_0.085 = phi i8* [ %add.ptr12, %for.body ], [ %row_base, %for.body.preheader ]
649  %1 = tail call <16 x i1> @llvm.arm.mve.vctp8(i32 %row_elem.090)
650  %sub = add nsw i32 %row_elem.090, -16
651  %2 = bitcast i8* %col_base.addr.095 to <16 x i8>*
652  %3 = tail call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %2, i32 1, <16 x i1> %1, <16 x i8> zeroinitializer)
653  %add.ptr4 = getelementptr inbounds i8, i8* %col_base.addr.095, i32 16
654  %4 = bitcast i8* %ip_row_0.085 to <16 x i8>*
655  %5 = tail call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %4, i32 1, <16 x i1> %1, <16 x i8> zeroinitializer)
656  %6 = tail call i32 @llvm.arm.mve.vmldava.predicated.v16i8.v16i1(i32 0, i32 0, i32 0, i32 %acc_n.sroa.0.094, <16 x i8> %5, <16 x i8> %3, <16 x i1> %1)
657  %7 = bitcast i8* %ip_row_1.086 to <16 x i8>*
658  %8 = tail call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %7, i32 1, <16 x i1> %1, <16 x i8> zeroinitializer)
659  %9 = tail call i32 @llvm.arm.mve.vmldava.predicated.v16i8.v16i1(i32 0, i32 0, i32 0, i32 %acc_n.sroa.6.093, <16 x i8> %8, <16 x i8> %3, <16 x i1> %1)
660  %10 = bitcast i8* %ip_row_2.087 to <16 x i8>*
661  %11 = tail call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %10, i32 1, <16 x i1> %1, <16 x i8> zeroinitializer)
662  %12 = tail call i32 @llvm.arm.mve.vmldava.predicated.v16i8.v16i1(i32 0, i32 0, i32 0, i32 %acc_n.sroa.9.092, <16 x i8> %11, <16 x i8> %3, <16 x i1> %1)
663  %13 = bitcast i8* %ip_row_3.088 to <16 x i8>*
664  %14 = tail call <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>* %13, i32 1, <16 x i1> %1, <16 x i8> zeroinitializer)
665  %15 = tail call i32 @llvm.arm.mve.vmldava.predicated.v16i8.v16i1(i32 0, i32 0, i32 0, i32 %acc_n.sroa.12.089, <16 x i8> %14, <16 x i8> %3, <16 x i1> %1)
666  %add.ptr12 = getelementptr inbounds i8, i8* %ip_row_0.085, i32 16
667  %add.ptr13 = getelementptr inbounds i8, i8* %ip_row_1.086, i32 16
668  %add.ptr14 = getelementptr inbounds i8, i8* %ip_row_2.087, i32 16
669  %add.ptr15 = getelementptr inbounds i8, i8* %ip_row_3.088, i32 16
670  %inc = add nuw nsw i32 %i.091, 1
671  %exitcond = icmp eq i32 %inc, %smax
672  br i1 %exitcond, label %for.cond.cleanup, label %for.body
673}
674
675define i8* @signext(i8* %input_row, i8* %input_col, i16 zeroext %output_ch, i16 zeroext %num_cols, i32* nocapture readnone %output_shift, i32* nocapture readnone %output_mult, i32 %out_offset, i32 %col_offset, i32 %row_offset, i16 signext %activation_min, i16 signext %activation_max, i16 zeroext %row_len, i32* nocapture readonly %bias, i8* returned %out) {
676; CHECK-LABEL: signext:
677; CHECK:       @ %bb.0: @ %entry
678; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
679; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}
680; CHECK-NEXT:    .pad #28
681; CHECK-NEXT:    sub sp, #28
682; CHECK-NEXT:    add.w r12, sp, #12
683; CHECK-NEXT:    cmp r3, #4
684; CHECK-NEXT:    stm.w r12, {r0, r1, r2} @ 12-byte Folded Spill
685; CHECK-NEXT:    bne .LBB5_8
686; CHECK-NEXT:  @ %bb.1: @ %for.cond.preheader
687; CHECK-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
688; CHECK-NEXT:    cmp r0, #0
689; CHECK-NEXT:    beq .LBB5_8
690; CHECK-NEXT:  @ %bb.2: @ %for.body.lr.ph
691; CHECK-NEXT:    ldr r2, [sp, #92]
692; CHECK-NEXT:    mov.w r11, #0
693; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
694; CHECK-NEXT:    ldr r4, [sp, #76]
695; CHECK-NEXT:    add.w r0, r1, r2, lsl #1
696; CHECK-NEXT:    str r0, [sp, #8] @ 4-byte Spill
697; CHECK-NEXT:    adds r0, r1, r2
698; CHECK-NEXT:    str r0, [sp, #4] @ 4-byte Spill
699; CHECK-NEXT:    add.w r0, r2, r2, lsl #1
700; CHECK-NEXT:    add r0, r1
701; CHECK-NEXT:    str r0, [sp] @ 4-byte Spill
702; CHECK-NEXT:    adds r0, r2, #7
703; CHECK-NEXT:    lsrs r1, r0, #3
704; CHECK-NEXT:    b .LBB5_5
705; CHECK-NEXT:  .LBB5_3: @ in Loop: Header=BB5_5 Depth=1
706; CHECK-NEXT:    mov r10, r12
707; CHECK-NEXT:    mov r8, r12
708; CHECK-NEXT:    mov r6, r12
709; CHECK-NEXT:  .LBB5_4: @ %for.cond.cleanup23
710; CHECK-NEXT:    @ in Loop: Header=BB5_5 Depth=1
711; CHECK-NEXT:    add.w r0, r8, r10
712; CHECK-NEXT:    ldr r1, [sp, #100]
713; CHECK-NEXT:    add r0, r6
714; CHECK-NEXT:    add r0, r12
715; CHECK-NEXT:    strb.w r0, [r1, r11]
716; CHECK-NEXT:    add.w r11, r11, #1
717; CHECK-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
718; CHECK-NEXT:    ldr r1, [sp, #24] @ 4-byte Reload
719; CHECK-NEXT:    cmp r11, r0
720; CHECK-NEXT:    beq .LBB5_8
721; CHECK-NEXT:  .LBB5_5: @ %for.body
722; CHECK-NEXT:    @ =>This Loop Header: Depth=1
723; CHECK-NEXT:    @ Child Loop BB5_7 Depth 2
724; CHECK-NEXT:    ldr r0, [sp, #96]
725; CHECK-NEXT:    cmp r1, r1
726; CHECK-NEXT:    str r1, [sp, #24] @ 4-byte Spill
727; CHECK-NEXT:    ldr.w r12, [r0, r11, lsl #2]
728; CHECK-NEXT:    bge .LBB5_3
729; CHECK-NEXT:  @ %bb.6: @ %for.body24.preheader
730; CHECK-NEXT:    @ in Loop: Header=BB5_5 Depth=1
731; CHECK-NEXT:    ldr.w lr, [sp, #92]
732; CHECK-NEXT:    ldr r0, [sp, #12] @ 4-byte Reload
733; CHECK-NEXT:    mov r6, r12
734; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
735; CHECK-NEXT:    mov r8, r12
736; CHECK-NEXT:    mla r3, r11, lr, r0
737; CHECK-NEXT:    mov r10, r12
738; CHECK-NEXT:    ldm.w sp, {r0, r5, r7} @ 12-byte Folded Reload
739; CHECK-NEXT:    dlstp.16 lr, lr
740; CHECK-NEXT:  .LBB5_7: @ %for.body24
741; CHECK-NEXT:    @ Parent Loop BB5_5 Depth=1
742; CHECK-NEXT:    @ => This Inner Loop Header: Depth=2
743; CHECK-NEXT:    vldrb.s16 q0, [r0], #8
744; CHECK-NEXT:    vadd.i16 q1, q0, r4
745; CHECK-NEXT:    vldrb.s16 q0, [r3], #8
746; CHECK-NEXT:    vmlava.s16 r12, q0, q1
747; CHECK-NEXT:    vldrb.s16 q1, [r7], #8
748; CHECK-NEXT:    vadd.i16 q1, q1, r4
749; CHECK-NEXT:    vmlava.s16 r6, q0, q1
750; CHECK-NEXT:    vldrb.s16 q1, [r5], #8
751; CHECK-NEXT:    vadd.i16 q1, q1, r4
752; CHECK-NEXT:    vmlava.s16 r8, q0, q1
753; CHECK-NEXT:    vldrb.s16 q1, [r1], #8
754; CHECK-NEXT:    vadd.i16 q1, q1, r4
755; CHECK-NEXT:    vmlava.s16 r10, q0, q1
756; CHECK-NEXT:    letp lr, .LBB5_7
757; CHECK-NEXT:    b .LBB5_4
758; CHECK-NEXT:  .LBB5_8: @ %if.end
759; CHECK-NEXT:    ldr r0, [sp, #100]
760; CHECK-NEXT:    add sp, #28
761; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
762entry:
763  %cmp = icmp eq i16 %num_cols, 4
764  br i1 %cmp, label %for.cond.preheader, label %if.end
765
766for.cond.preheader:                               ; preds = %entry
767  %conv2 = zext i16 %output_ch to i32
768  %cmp3127 = icmp eq i16 %output_ch, 0
769  br i1 %cmp3127, label %if.end, label %for.body.lr.ph
770
771for.body.lr.ph:                                   ; preds = %for.cond.preheader
772  %conv5 = zext i16 %row_len to i32
773  %add.ptr9 = getelementptr inbounds i8, i8* %input_col, i32 %conv5
774  %mul11 = shl nuw nsw i32 %conv5, 1
775  %add.ptr12 = getelementptr inbounds i8, i8* %input_col, i32 %mul11
776  %mul14 = mul nuw nsw i32 %conv5, 3
777  %add.ptr15 = getelementptr inbounds i8, i8* %input_col, i32 %mul14
778  %add = add nuw nsw i32 %conv5, 7
779  %div = lshr i32 %add, 3
780  %conv25 = trunc i32 %col_offset to i16
781  %.splatinsert.i = insertelement <8 x i16> undef, i16 %conv25, i32 0
782  %.splat.i = shufflevector <8 x i16> %.splatinsert.i, <8 x i16> undef, <8 x i32> zeroinitializer
783  br label %for.body
784
785for.body:                                         ; preds = %for.cond.cleanup23, %for.body.lr.ph
786  %i_out_ch.0129 = phi i32 [ 0, %for.body.lr.ph ], [ %inc37, %for.cond.cleanup23 ]
787  %i_row_loop.0128 = phi i32 [ undef, %for.body.lr.ph ], [ %i_row_loop.1.lcssa, %for.cond.cleanup23 ]
788  %arrayidx = getelementptr inbounds i32, i32* %bias, i32 %i_out_ch.0129
789  %0 = load i32, i32* %arrayidx, align 4
790  %cmp21111 = icmp slt i32 %i_row_loop.0128, %div
791  br i1 %cmp21111, label %for.body24.preheader, label %for.cond.cleanup23
792
793for.body24.preheader:                             ; preds = %for.body
794  %mul = mul nuw nsw i32 %i_out_ch.0129, %conv5
795  %add.ptr = getelementptr inbounds i8, i8* %input_row, i32 %mul
796  br label %for.body24
797
798for.cond.cleanup23:                               ; preds = %for.body24, %for.body
799  %acc_0.0.lcssa = phi i32 [ %0, %for.body ], [ %21, %for.body24 ]
800  %acc_1.0.lcssa = phi i32 [ %0, %for.body ], [ %22, %for.body24 ]
801  %acc_2.0.lcssa = phi i32 [ %0, %for.body ], [ %23, %for.body24 ]
802  %acc_3.0.lcssa = phi i32 [ %0, %for.body ], [ %24, %for.body24 ]
803  %i_row_loop.1.lcssa = phi i32 [ %i_row_loop.0128, %for.body ], [ %div, %for.body24 ]
804  %add31 = add nsw i32 %acc_1.0.lcssa, %acc_0.0.lcssa
805  %add32 = add nsw i32 %add31, %acc_2.0.lcssa
806  %add33 = add nsw i32 %add32, %acc_3.0.lcssa
807  %conv34 = trunc i32 %add33 to i8
808  %arrayidx35 = getelementptr inbounds i8, i8* %out, i32 %i_out_ch.0129
809  store i8 %conv34, i8* %arrayidx35, align 1
810  %inc37 = add nuw nsw i32 %i_out_ch.0129, 1
811  %exitcond133 = icmp eq i32 %inc37, %conv2
812  br i1 %exitcond133, label %if.end, label %for.body
813
814for.body24:                                       ; preds = %for.body24, %for.body24.preheader
815  %row_len_tmp.0122 = phi i32 [ %sub, %for.body24 ], [ %conv5, %for.body24.preheader ]
816  %ip_r0.0121 = phi i8* [ %add.ptr26, %for.body24 ], [ %add.ptr, %for.body24.preheader ]
817  %ip_c0.0120 = phi i8* [ %add.ptr27, %for.body24 ], [ %input_col, %for.body24.preheader ]
818  %ip_c1.0119 = phi i8* [ %add.ptr28, %for.body24 ], [ %add.ptr9, %for.body24.preheader ]
819  %ip_c2.0118 = phi i8* [ %add.ptr29, %for.body24 ], [ %add.ptr12, %for.body24.preheader ]
820  %i_row_loop.1117 = phi i32 [ %inc, %for.body24 ], [ %i_row_loop.0128, %for.body24.preheader ]
821  %ip_c3.0116 = phi i8* [ %add.ptr30, %for.body24 ], [ %add.ptr15, %for.body24.preheader ]
822  %acc_3.0115 = phi i32 [ %24, %for.body24 ], [ %0, %for.body24.preheader ]
823  %acc_2.0114 = phi i32 [ %23, %for.body24 ], [ %0, %for.body24.preheader ]
824  %acc_1.0113 = phi i32 [ %22, %for.body24 ], [ %0, %for.body24.preheader ]
825  %acc_0.0112 = phi i32 [ %21, %for.body24 ], [ %0, %for.body24.preheader ]
826  %1 = tail call <8 x i1> @llvm.arm.mve.vctp16(i32 %row_len_tmp.0122)
827  %sub = add nsw i32 %row_len_tmp.0122, -8
828  %2 = bitcast i8* %ip_r0.0121 to <8 x i8>*
829  %3 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %2, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
830  %4 = sext <8 x i8> %3 to <8 x i16>
831  %add.ptr26 = getelementptr inbounds i8, i8* %ip_r0.0121, i32 8
832  %5 = bitcast i8* %ip_c0.0120 to <8 x i8>*
833  %6 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %5, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
834  %7 = sext <8 x i8> %6 to <8 x i16>
835  %add.ptr27 = getelementptr inbounds i8, i8* %ip_c0.0120, i32 8
836  %8 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %7, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
837  %9 = bitcast i8* %ip_c1.0119 to <8 x i8>*
838  %10 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %9, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
839  %11 = sext <8 x i8> %10 to <8 x i16>
840  %add.ptr28 = getelementptr inbounds i8, i8* %ip_c1.0119, i32 8
841  %12 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %11, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
842  %13 = bitcast i8* %ip_c2.0118 to <8 x i8>*
843  %14 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %13, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
844  %15 = sext <8 x i8> %14 to <8 x i16>
845  %add.ptr29 = getelementptr inbounds i8, i8* %ip_c2.0118, i32 8
846  %16 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %15, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
847  %17 = bitcast i8* %ip_c3.0116 to <8 x i8>*
848  %18 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %17, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
849  %19 = sext <8 x i8> %18 to <8 x i16>
850  %add.ptr30 = getelementptr inbounds i8, i8* %ip_c3.0116, i32 8
851  %20 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %19, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
852  %21 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_0.0112, <8 x i16> %4, <8 x i16> %8, <8 x i1> %1)
853  %22 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_1.0113, <8 x i16> %4, <8 x i16> %12, <8 x i1> %1)
854  %23 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_2.0114, <8 x i16> %4, <8 x i16> %16, <8 x i1> %1)
855  %24 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_3.0115, <8 x i16> %4, <8 x i16> %20, <8 x i1> %1)
856  %inc = add nsw i32 %i_row_loop.1117, 1
857  %exitcond = icmp eq i32 %inc, %div
858  br i1 %exitcond, label %for.cond.cleanup23, label %for.body24
859
860if.end:                                           ; preds = %for.cond.cleanup23, %for.cond.preheader, %entry
861  ret i8* %out
862}
863
864define i8* @signext_optsize(i8* %input_row, i8* %input_col, i16 zeroext %output_ch, i16 zeroext %num_cols, i32* nocapture readnone %output_shift, i32* nocapture readnone %output_mult, i32 %out_offset, i32 %col_offset, i32 %row_offset, i16 signext %activation_min, i16 signext %activation_max, i16 zeroext %row_len, i32* nocapture readonly %bias, i8* returned %out) optsize {
865; CHECK-LABEL: signext_optsize:
866; CHECK:       @ %bb.0: @ %entry
867; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
868; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}
869; CHECK-NEXT:    .pad #28
870; CHECK-NEXT:    sub sp, #28
871; CHECK-NEXT:    add.w r12, sp, #12
872; CHECK-NEXT:    cmp r3, #4
873; CHECK-NEXT:    stm.w r12, {r0, r1, r2} @ 12-byte Folded Spill
874; CHECK-NEXT:    bne .LBB6_8
875; CHECK-NEXT:  @ %bb.1: @ %for.cond.preheader
876; CHECK-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
877; CHECK-NEXT:    cmp r0, #0
878; CHECK-NEXT:    beq .LBB6_8
879; CHECK-NEXT:  @ %bb.2: @ %for.body.lr.ph
880; CHECK-NEXT:    ldr r2, [sp, #92]
881; CHECK-NEXT:    mov.w r11, #0
882; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
883; CHECK-NEXT:    ldr r4, [sp, #76]
884; CHECK-NEXT:    add.w r0, r1, r2, lsl #1
885; CHECK-NEXT:    str r0, [sp, #8] @ 4-byte Spill
886; CHECK-NEXT:    adds r0, r1, r2
887; CHECK-NEXT:    str r0, [sp, #4] @ 4-byte Spill
888; CHECK-NEXT:    add.w r0, r2, r2, lsl #1
889; CHECK-NEXT:    add r0, r1
890; CHECK-NEXT:    str r0, [sp] @ 4-byte Spill
891; CHECK-NEXT:    adds r0, r2, #7
892; CHECK-NEXT:    lsrs r1, r0, #3
893; CHECK-NEXT:  .LBB6_3: @ %for.body
894; CHECK-NEXT:    @ =>This Loop Header: Depth=1
895; CHECK-NEXT:    @ Child Loop BB6_5 Depth 2
896; CHECK-NEXT:    ldr r0, [sp, #96]
897; CHECK-NEXT:    cmp r1, r1
898; CHECK-NEXT:    str r1, [sp, #24] @ 4-byte Spill
899; CHECK-NEXT:    ldr.w r12, [r0, r11, lsl #2]
900; CHECK-NEXT:    bge .LBB6_6
901; CHECK-NEXT:  @ %bb.4: @ %for.body24.preheader
902; CHECK-NEXT:    @ in Loop: Header=BB6_3 Depth=1
903; CHECK-NEXT:    ldr.w lr, [sp, #92]
904; CHECK-NEXT:    ldr r0, [sp, #12] @ 4-byte Reload
905; CHECK-NEXT:    mov r6, r12
906; CHECK-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
907; CHECK-NEXT:    mov r8, r12
908; CHECK-NEXT:    mla r3, r11, lr, r0
909; CHECK-NEXT:    mov r10, r12
910; CHECK-NEXT:    ldm.w sp, {r0, r5, r7} @ 12-byte Folded Reload
911; CHECK-NEXT:    dlstp.16 lr, lr
912; CHECK-NEXT:  .LBB6_5: @ %for.body24
913; CHECK-NEXT:    @ Parent Loop BB6_3 Depth=1
914; CHECK-NEXT:    @ => This Inner Loop Header: Depth=2
915; CHECK-NEXT:    vldrb.s16 q0, [r0], #8
916; CHECK-NEXT:    vadd.i16 q1, q0, r4
917; CHECK-NEXT:    vldrb.s16 q0, [r3], #8
918; CHECK-NEXT:    vmlava.s16 r12, q0, q1
919; CHECK-NEXT:    vldrb.s16 q1, [r7], #8
920; CHECK-NEXT:    vadd.i16 q1, q1, r4
921; CHECK-NEXT:    vmlava.s16 r6, q0, q1
922; CHECK-NEXT:    vldrb.s16 q1, [r5], #8
923; CHECK-NEXT:    vadd.i16 q1, q1, r4
924; CHECK-NEXT:    vmlava.s16 r8, q0, q1
925; CHECK-NEXT:    vldrb.s16 q1, [r1], #8
926; CHECK-NEXT:    vadd.i16 q1, q1, r4
927; CHECK-NEXT:    vmlava.s16 r10, q0, q1
928; CHECK-NEXT:    letp lr, .LBB6_5
929; CHECK-NEXT:    b .LBB6_7
930; CHECK-NEXT:  .LBB6_6: @ in Loop: Header=BB6_3 Depth=1
931; CHECK-NEXT:    mov r10, r12
932; CHECK-NEXT:    mov r8, r12
933; CHECK-NEXT:    mov r6, r12
934; CHECK-NEXT:  .LBB6_7: @ %for.cond.cleanup23
935; CHECK-NEXT:    @ in Loop: Header=BB6_3 Depth=1
936; CHECK-NEXT:    add.w r0, r8, r10
937; CHECK-NEXT:    ldr r1, [sp, #100]
938; CHECK-NEXT:    add r0, r6
939; CHECK-NEXT:    add r0, r12
940; CHECK-NEXT:    strb.w r0, [r1, r11]
941; CHECK-NEXT:    add.w r11, r11, #1
942; CHECK-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
943; CHECK-NEXT:    ldr r1, [sp, #24] @ 4-byte Reload
944; CHECK-NEXT:    cmp r11, r0
945; CHECK-NEXT:    bne .LBB6_3
946; CHECK-NEXT:  .LBB6_8: @ %if.end
947; CHECK-NEXT:    ldr r0, [sp, #100]
948; CHECK-NEXT:    add sp, #28
949; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
950entry:
951  %cmp = icmp eq i16 %num_cols, 4
952  br i1 %cmp, label %for.cond.preheader, label %if.end
953
954for.cond.preheader:                               ; preds = %entry
955  %conv2 = zext i16 %output_ch to i32
956  %cmp3127 = icmp eq i16 %output_ch, 0
957  br i1 %cmp3127, label %if.end, label %for.body.lr.ph
958
959for.body.lr.ph:                                   ; preds = %for.cond.preheader
960  %conv5 = zext i16 %row_len to i32
961  %add.ptr9 = getelementptr inbounds i8, i8* %input_col, i32 %conv5
962  %mul11 = shl nuw nsw i32 %conv5, 1
963  %add.ptr12 = getelementptr inbounds i8, i8* %input_col, i32 %mul11
964  %mul14 = mul nuw nsw i32 %conv5, 3
965  %add.ptr15 = getelementptr inbounds i8, i8* %input_col, i32 %mul14
966  %add = add nuw nsw i32 %conv5, 7
967  %div = lshr i32 %add, 3
968  %conv25 = trunc i32 %col_offset to i16
969  %.splatinsert.i = insertelement <8 x i16> undef, i16 %conv25, i32 0
970  %.splat.i = shufflevector <8 x i16> %.splatinsert.i, <8 x i16> undef, <8 x i32> zeroinitializer
971  br label %for.body
972
973for.body:                                         ; preds = %for.cond.cleanup23, %for.body.lr.ph
974  %i_out_ch.0129 = phi i32 [ 0, %for.body.lr.ph ], [ %inc37, %for.cond.cleanup23 ]
975  %i_row_loop.0128 = phi i32 [ undef, %for.body.lr.ph ], [ %i_row_loop.1.lcssa, %for.cond.cleanup23 ]
976  %arrayidx = getelementptr inbounds i32, i32* %bias, i32 %i_out_ch.0129
977  %0 = load i32, i32* %arrayidx, align 4
978  %cmp21111 = icmp slt i32 %i_row_loop.0128, %div
979  br i1 %cmp21111, label %for.body24.preheader, label %for.cond.cleanup23
980
981for.body24.preheader:                             ; preds = %for.body
982  %mul = mul nuw nsw i32 %i_out_ch.0129, %conv5
983  %add.ptr = getelementptr inbounds i8, i8* %input_row, i32 %mul
984  br label %for.body24
985
986for.cond.cleanup23:                               ; preds = %for.body24, %for.body
987  %acc_0.0.lcssa = phi i32 [ %0, %for.body ], [ %21, %for.body24 ]
988  %acc_1.0.lcssa = phi i32 [ %0, %for.body ], [ %22, %for.body24 ]
989  %acc_2.0.lcssa = phi i32 [ %0, %for.body ], [ %23, %for.body24 ]
990  %acc_3.0.lcssa = phi i32 [ %0, %for.body ], [ %24, %for.body24 ]
991  %i_row_loop.1.lcssa = phi i32 [ %i_row_loop.0128, %for.body ], [ %div, %for.body24 ]
992  %add31 = add nsw i32 %acc_1.0.lcssa, %acc_0.0.lcssa
993  %add32 = add nsw i32 %add31, %acc_2.0.lcssa
994  %add33 = add nsw i32 %add32, %acc_3.0.lcssa
995  %conv34 = trunc i32 %add33 to i8
996  %arrayidx35 = getelementptr inbounds i8, i8* %out, i32 %i_out_ch.0129
997  store i8 %conv34, i8* %arrayidx35, align 1
998  %inc37 = add nuw nsw i32 %i_out_ch.0129, 1
999  %exitcond133 = icmp eq i32 %inc37, %conv2
1000  br i1 %exitcond133, label %if.end, label %for.body
1001
1002for.body24:                                       ; preds = %for.body24, %for.body24.preheader
1003  %row_len_tmp.0122 = phi i32 [ %sub, %for.body24 ], [ %conv5, %for.body24.preheader ]
1004  %ip_r0.0121 = phi i8* [ %add.ptr26, %for.body24 ], [ %add.ptr, %for.body24.preheader ]
1005  %ip_c0.0120 = phi i8* [ %add.ptr27, %for.body24 ], [ %input_col, %for.body24.preheader ]
1006  %ip_c1.0119 = phi i8* [ %add.ptr28, %for.body24 ], [ %add.ptr9, %for.body24.preheader ]
1007  %ip_c2.0118 = phi i8* [ %add.ptr29, %for.body24 ], [ %add.ptr12, %for.body24.preheader ]
1008  %i_row_loop.1117 = phi i32 [ %inc, %for.body24 ], [ %i_row_loop.0128, %for.body24.preheader ]
1009  %ip_c3.0116 = phi i8* [ %add.ptr30, %for.body24 ], [ %add.ptr15, %for.body24.preheader ]
1010  %acc_3.0115 = phi i32 [ %24, %for.body24 ], [ %0, %for.body24.preheader ]
1011  %acc_2.0114 = phi i32 [ %23, %for.body24 ], [ %0, %for.body24.preheader ]
1012  %acc_1.0113 = phi i32 [ %22, %for.body24 ], [ %0, %for.body24.preheader ]
1013  %acc_0.0112 = phi i32 [ %21, %for.body24 ], [ %0, %for.body24.preheader ]
1014  %1 = tail call <8 x i1> @llvm.arm.mve.vctp16(i32 %row_len_tmp.0122)
1015  %sub = add nsw i32 %row_len_tmp.0122, -8
1016  %2 = bitcast i8* %ip_r0.0121 to <8 x i8>*
1017  %3 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %2, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
1018  %4 = sext <8 x i8> %3 to <8 x i16>
1019  %add.ptr26 = getelementptr inbounds i8, i8* %ip_r0.0121, i32 8
1020  %5 = bitcast i8* %ip_c0.0120 to <8 x i8>*
1021  %6 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %5, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
1022  %7 = sext <8 x i8> %6 to <8 x i16>
1023  %add.ptr27 = getelementptr inbounds i8, i8* %ip_c0.0120, i32 8
1024  %8 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %7, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
1025  %9 = bitcast i8* %ip_c1.0119 to <8 x i8>*
1026  %10 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %9, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
1027  %11 = sext <8 x i8> %10 to <8 x i16>
1028  %add.ptr28 = getelementptr inbounds i8, i8* %ip_c1.0119, i32 8
1029  %12 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %11, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
1030  %13 = bitcast i8* %ip_c2.0118 to <8 x i8>*
1031  %14 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %13, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
1032  %15 = sext <8 x i8> %14 to <8 x i16>
1033  %add.ptr29 = getelementptr inbounds i8, i8* %ip_c2.0118, i32 8
1034  %16 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %15, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
1035  %17 = bitcast i8* %ip_c3.0116 to <8 x i8>*
1036  %18 = tail call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* %17, i32 1, <8 x i1> %1, <8 x i8> zeroinitializer)
1037  %19 = sext <8 x i8> %18 to <8 x i16>
1038  %add.ptr30 = getelementptr inbounds i8, i8* %ip_c3.0116, i32 8
1039  %20 = tail call <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16> %19, <8 x i16> %.splat.i, <8 x i1> %1, <8 x i16> undef)
1040  %21 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_0.0112, <8 x i16> %4, <8 x i16> %8, <8 x i1> %1)
1041  %22 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_1.0113, <8 x i16> %4, <8 x i16> %12, <8 x i1> %1)
1042  %23 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_2.0114, <8 x i16> %4, <8 x i16> %16, <8 x i1> %1)
1043  %24 = tail call i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32 0, i32 0, i32 0, i32 %acc_3.0115, <8 x i16> %4, <8 x i16> %20, <8 x i1> %1)
1044  %inc = add nsw i32 %i_row_loop.1117, 1
1045  %exitcond = icmp eq i32 %inc, %div
1046  br i1 %exitcond, label %for.cond.cleanup23, label %for.body24
1047
1048if.end:                                           ; preds = %for.cond.cleanup23, %for.cond.preheader, %entry
1049  ret i8* %out
1050}
1051
1052%struct.arm_cfft_instance_f32 = type { i16, float*, i16*, i16, i32*, i32*, i32*, float*, float*, float* }
1053define arm_aapcs_vfpcc void @_Z37_arm_radix4_butterfly_inverse_f32_mvePK21arm_cfft_instance_f32Pfjf(%struct.arm_cfft_instance_f32* nocapture readonly %0, float* %1, i32 %2, float %3) {
1054; CHECK-LABEL: _Z37_arm_radix4_butterfly_inverse_f32_mvePK21arm_cfft_instance_f32Pfjf:
1055; CHECK:       @ %bb.0:
1056; CHECK-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
1057; CHECK-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}
1058; CHECK-NEXT:    .pad #4
1059; CHECK-NEXT:    sub sp, #4
1060; CHECK-NEXT:    .vsave {d8, d9, d10, d11, d12, d13, d14, d15}
1061; CHECK-NEXT:    vpush {d8, d9, d10, d11, d12, d13, d14, d15}
1062; CHECK-NEXT:    .pad #40
1063; CHECK-NEXT:    sub sp, #40
1064; CHECK-NEXT:    cmp r2, #8
1065; CHECK-NEXT:    str r1, [sp, #16] @ 4-byte Spill
1066; CHECK-NEXT:    vstr s0, [sp] @ 4-byte Spill
1067; CHECK-NEXT:    mov r1, r2
1068; CHECK-NEXT:    str r2, [sp, #4] @ 4-byte Spill
1069; CHECK-NEXT:    blo .LBB7_9
1070; CHECK-NEXT:  @ %bb.1:
1071; CHECK-NEXT:    ldr r2, [sp, #4] @ 4-byte Reload
1072; CHECK-NEXT:    mov.w r11, #0
1073; CHECK-NEXT:    mov.w r12, #1
1074; CHECK-NEXT:    str r2, [sp, #12] @ 4-byte Spill
1075; CHECK-NEXT:    lsrs r1, r2, #2
1076; CHECK-NEXT:    b .LBB7_3
1077; CHECK-NEXT:  .LBB7_2: @ in Loop: Header=BB7_3 Depth=1
1078; CHECK-NEXT:    ldr r2, [sp, #8] @ 4-byte Reload
1079; CHECK-NEXT:    add.w r11, r11, #1
1080; CHECK-NEXT:    lsl.w r12, r12, #2
1081; CHECK-NEXT:    cmp r2, #7
1082; CHECK-NEXT:    asr.w r1, r2, #2
1083; CHECK-NEXT:    ble .LBB7_9
1084; CHECK-NEXT:  .LBB7_3: @ =>This Loop Header: Depth=1
1085; CHECK-NEXT:    @ Child Loop BB7_6 Depth 2
1086; CHECK-NEXT:    @ Child Loop BB7_7 Depth 3
1087; CHECK-NEXT:    str r1, [sp, #8] @ 4-byte Spill
1088; CHECK-NEXT:    cmp.w r12, #1
1089; CHECK-NEXT:    ldr r1, [sp, #12] @ 4-byte Reload
1090; CHECK-NEXT:    lsr.w r2, r1, #2
1091; CHECK-NEXT:    str r2, [sp, #12] @ 4-byte Spill
1092; CHECK-NEXT:    blt .LBB7_2
1093; CHECK-NEXT:  @ %bb.4: @ in Loop: Header=BB7_3 Depth=1
1094; CHECK-NEXT:    lsrs r2, r1, #3
1095; CHECK-NEXT:    str r2, [sp, #24] @ 4-byte Spill
1096; CHECK-NEXT:    beq .LBB7_2
1097; CHECK-NEXT:  @ %bb.5: @ %.preheader
1098; CHECK-NEXT:    @ in Loop: Header=BB7_3 Depth=1
1099; CHECK-NEXT:    ldr r2, [sp, #12] @ 4-byte Reload
1100; CHECK-NEXT:    lsls r1, r1, #1
1101; CHECK-NEXT:    movs r4, #0
1102; CHECK-NEXT:    str r1, [sp, #20] @ 4-byte Spill
1103; CHECK-NEXT:    lsl.w r10, r2, #1
1104; CHECK-NEXT:  .LBB7_6: @ Parent Loop BB7_3 Depth=1
1105; CHECK-NEXT:    @ => This Loop Header: Depth=2
1106; CHECK-NEXT:    @ Child Loop BB7_7 Depth 3
1107; CHECK-NEXT:    ldr r1, [sp, #20] @ 4-byte Reload
1108; CHECK-NEXT:    ldrd lr, r2, [r0, #16]
1109; CHECK-NEXT:    ldrd r3, r8, [r0, #24]
1110; CHECK-NEXT:    muls r1, r4, r1
1111; CHECK-NEXT:    ldr.w r2, [r2, r11, lsl #2]
1112; CHECK-NEXT:    ldrd r7, r5, [r0, #32]
1113; CHECK-NEXT:    ldr.w r3, [r3, r11, lsl #2]
1114; CHECK-NEXT:    ldr.w r6, [lr, r11, lsl #2]
1115; CHECK-NEXT:    add.w r7, r7, r2, lsl #2
1116; CHECK-NEXT:    ldr r2, [sp, #16] @ 4-byte Reload
1117; CHECK-NEXT:    add.w r5, r5, r3, lsl #2
1118; CHECK-NEXT:    add.w r1, r2, r1, lsl #2
1119; CHECK-NEXT:    add.w r3, r8, r6, lsl #2
1120; CHECK-NEXT:    ldr r6, [sp, #24] @ 4-byte Reload
1121; CHECK-NEXT:    add.w r2, r1, r10, lsl #2
1122; CHECK-NEXT:    add.w r8, r2, r10, lsl #2
1123; CHECK-NEXT:    add.w r9, r8, r10, lsl #2
1124; CHECK-NEXT:    dls lr, r6
1125; CHECK-NEXT:  .LBB7_7: @ Parent Loop BB7_3 Depth=1
1126; CHECK-NEXT:    @ Parent Loop BB7_6 Depth=2
1127; CHECK-NEXT:    @ => This Inner Loop Header: Depth=3
1128; CHECK-NEXT:    vldrw.u32 q3, [r9]
1129; CHECK-NEXT:    vldrw.u32 q4, [r2]
1130; CHECK-NEXT:    vldrw.u32 q6, [r8]
1131; CHECK-NEXT:    vldrw.u32 q7, [r1]
1132; CHECK-NEXT:    vsub.f32 q5, q4, q3
1133; CHECK-NEXT:    vsub.f32 q0, q7, q6
1134; CHECK-NEXT:    vcadd.f32 q1, q0, q5, #270
1135; CHECK-NEXT:    vcadd.f32 q2, q0, q5, #90
1136; CHECK-NEXT:    vadd.f32 q0, q4, q3
1137; CHECK-NEXT:    vadd.f32 q3, q6, q7
1138; CHECK-NEXT:    vsub.f32 q4, q3, q0
1139; CHECK-NEXT:    vadd.f32 q0, q3, q0
1140; CHECK-NEXT:    vstrb.8 q0, [r1], #16
1141; CHECK-NEXT:    vldrw.u32 q0, [r7], #16
1142; CHECK-NEXT:    vcmul.f32 q3, q0, q4, #0
1143; CHECK-NEXT:    vcmla.f32 q3, q0, q4, #90
1144; CHECK-NEXT:    vstrb.8 q3, [r2], #16
1145; CHECK-NEXT:    vldrw.u32 q0, [r3], #16
1146; CHECK-NEXT:    vcmul.f32 q3, q0, q2, #0
1147; CHECK-NEXT:    vcmla.f32 q3, q0, q2, #90
1148; CHECK-NEXT:    vstrb.8 q3, [r8], #16
1149; CHECK-NEXT:    vldrw.u32 q0, [r5], #16
1150; CHECK-NEXT:    vcmul.f32 q2, q0, q1, #0
1151; CHECK-NEXT:    vcmla.f32 q2, q0, q1, #90
1152; CHECK-NEXT:    vstrb.8 q2, [r9], #16
1153; CHECK-NEXT:    le lr, .LBB7_7
1154; CHECK-NEXT:  @ %bb.8: @ in Loop: Header=BB7_6 Depth=2
1155; CHECK-NEXT:    adds r4, #1
1156; CHECK-NEXT:    cmp r4, r12
1157; CHECK-NEXT:    bne .LBB7_6
1158; CHECK-NEXT:    b .LBB7_2
1159; CHECK-NEXT:  .LBB7_9:
1160; CHECK-NEXT:    ldr r0, [sp, #4] @ 4-byte Reload
1161; CHECK-NEXT:    lsrs r0, r0, #3
1162; CHECK-NEXT:    wls lr, r0, .LBB7_12
1163; CHECK-NEXT:  @ %bb.10:
1164; CHECK-NEXT:    adr r0, .LCPI7_0
1165; CHECK-NEXT:    vldr s0, [sp] @ 4-byte Reload
1166; CHECK-NEXT:    vldrw.u32 q1, [r0]
1167; CHECK-NEXT:    ldr r0, [sp, #16] @ 4-byte Reload
1168; CHECK-NEXT:    vadd.i32 q1, q1, r0
1169; CHECK-NEXT:    vmov r0, s0
1170; CHECK-NEXT:    vldrw.u32 q2, [q1, #64]!
1171; CHECK-NEXT:    vldrw.u32 q0, [q1, #16]
1172; CHECK-NEXT:  .LBB7_11: @ =>This Inner Loop Header: Depth=1
1173; CHECK-NEXT:    vldrw.u32 q3, [q1, #24]
1174; CHECK-NEXT:    vldrw.u32 q4, [q1, #8]
1175; CHECK-NEXT:    vsub.f32 q6, q2, q0
1176; CHECK-NEXT:    vadd.f32 q0, q2, q0
1177; CHECK-NEXT:    vsub.f32 q5, q4, q3
1178; CHECK-NEXT:    vadd.f32 q3, q4, q3
1179; CHECK-NEXT:    vcadd.f32 q7, q6, q5, #270
1180; CHECK-NEXT:    vsub.f32 q2, q0, q3
1181; CHECK-NEXT:    vmul.f32 q7, q7, r0
1182; CHECK-NEXT:    vadd.f32 q3, q0, q3
1183; CHECK-NEXT:    vstrw.32 q7, [sp, #24] @ 16-byte Spill
1184; CHECK-NEXT:    vcadd.f32 q7, q6, q5, #90
1185; CHECK-NEXT:    vmul.f32 q4, q2, r0
1186; CHECK-NEXT:    vldrw.u32 q2, [q1, #64]!
1187; CHECK-NEXT:    vmul.f32 q5, q7, r0
1188; CHECK-NEXT:    vmul.f32 q3, q3, r0
1189; CHECK-NEXT:    vldrw.u32 q0, [q1, #16]
1190; CHECK-NEXT:    vstrw.32 q3, [q1, #-64]
1191; CHECK-NEXT:    vstrw.32 q4, [q1, #-56]
1192; CHECK-NEXT:    vstrw.32 q5, [q1, #-48]
1193; CHECK-NEXT:    vldrw.u32 q3, [sp, #24] @ 16-byte Reload
1194; CHECK-NEXT:    vstrw.32 q3, [q1, #-40]
1195; CHECK-NEXT:    le lr, .LBB7_11
1196; CHECK-NEXT:  .LBB7_12:
1197; CHECK-NEXT:    add sp, #40
1198; CHECK-NEXT:    vpop {d8, d9, d10, d11, d12, d13, d14, d15}
1199; CHECK-NEXT:    add sp, #4
1200; CHECK-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}
1201; CHECK-NEXT:    .p2align 4
1202; CHECK-NEXT:  @ %bb.13:
1203; CHECK-NEXT:  .LCPI7_0:
1204; CHECK-NEXT:    .long 4294967232 @ 0xffffffc0
1205; CHECK-NEXT:    .long 4294967236 @ 0xffffffc4
1206; CHECK-NEXT:    .long 4294967264 @ 0xffffffe0
1207; CHECK-NEXT:    .long 4294967268 @ 0xffffffe4
1208  %5 = icmp ugt i32 %2, 7
1209  br i1 %5, label %6, label %26
1210
12116: ; preds = %4
1212  %7 = lshr i32 %2, 2
1213  %8 = getelementptr inbounds %struct.arm_cfft_instance_f32, %struct.arm_cfft_instance_f32* %0, i32 0, i32 7
1214  %9 = getelementptr inbounds %struct.arm_cfft_instance_f32, %struct.arm_cfft_instance_f32* %0, i32 0, i32 4
1215  %10 = getelementptr inbounds %struct.arm_cfft_instance_f32, %struct.arm_cfft_instance_f32* %0, i32 0, i32 8
1216  %11 = getelementptr inbounds %struct.arm_cfft_instance_f32, %struct.arm_cfft_instance_f32* %0, i32 0, i32 5
1217  %12 = getelementptr inbounds %struct.arm_cfft_instance_f32, %struct.arm_cfft_instance_f32* %0, i32 0, i32 9
1218  %13 = getelementptr inbounds %struct.arm_cfft_instance_f32, %struct.arm_cfft_instance_f32* %0, i32 0, i32 6
1219  br label %14
1220
122114: ; preds = %6, %40
1222  %15 = phi i32 [ %2, %6 ], [ %19, %40 ]
1223  %16 = phi i32 [ %7, %6 ], [ %43, %40 ]
1224  %17 = phi i32 [ 1, %6 ], [ %41, %40 ]
1225  %18 = phi i32 [ 0, %6 ], [ %42, %40 ]
1226  %19 = lshr i32 %15, 2
1227  %20 = icmp sgt i32 %17, 0
1228  br i1 %20, label %21, label %40
1229
123021: ; preds = %14
1231  %22 = shl i32 %15, 1
1232  %23 = shl nuw nsw i32 %19, 1
1233  %24 = lshr i32 %15, 3
1234  %25 = icmp eq i32 %24, 0
1235  br i1 %25, label %40, label %45
1236
123726: ; preds = %40, %4
1238  %27 = ptrtoint float* %1 to i32
1239  %28 = insertelement <4 x i32> undef, i32 %27, i32 0
1240  %29 = shufflevector <4 x i32> %28, <4 x i32> undef, <4 x i32> zeroinitializer
1241  %30 = add <4 x i32> %29, <i32 -64, i32 -60, i32 -32, i32 -28>
1242  %31 = tail call { <4 x float>, <4 x i32> } @llvm.arm.mve.vldr.gather.base.wb.v4f32.v4i32(<4 x i32> %30, i32 64)
1243  %32 = extractvalue { <4 x float>, <4 x i32> } %31, 1
1244  %33 = lshr i32 %2, 3
1245  %34 = icmp eq i32 %33, 0
1246  br i1 %34, label %141, label %35
1247
124835: ; preds = %26
1249  %36 = tail call <4 x float> @llvm.arm.mve.vldr.gather.base.v4f32.v4i32(<4 x i32> %32, i32 16)
1250  %37 = extractvalue { <4 x float>, <4 x i32> } %31, 0
1251  %38 = insertelement <4 x float> undef, float %3, i32 0
1252  %39 = shufflevector <4 x float> %38, <4 x float> undef, <4 x i32> zeroinitializer
1253  br label %116
1254
125540: ; preds = %113, %21, %14
1256  %41 = shl i32 %17, 2
1257  %42 = add nuw nsw i32 %18, 1
1258  %43 = ashr i32 %16, 2
1259  %44 = icmp sgt i32 %16, 7
1260  br i1 %44, label %14, label %26
1261
126245: ; preds = %21, %113
1263  %46 = phi i32 [ %114, %113 ], [ 0, %21 ]
1264  %47 = load float*, float** %8, align 4
1265  %48 = load i32*, i32** %9, align 4
1266  %49 = getelementptr inbounds i32, i32* %48, i32 %18
1267  %50 = load i32, i32* %49, align 4
1268  %51 = getelementptr inbounds float, float* %47, i32 %50
1269  %52 = load float*, float** %10, align 4
1270  %53 = load i32*, i32** %11, align 4
1271  %54 = getelementptr inbounds i32, i32* %53, i32 %18
1272  %55 = load i32, i32* %54, align 4
1273  %56 = getelementptr inbounds float, float* %52, i32 %55
1274  %57 = load float*, float** %12, align 4
1275  %58 = load i32*, i32** %13, align 4
1276  %59 = getelementptr inbounds i32, i32* %58, i32 %18
1277  %60 = load i32, i32* %59, align 4
1278  %61 = getelementptr inbounds float, float* %57, i32 %60
1279  %62 = mul i32 %22, %46
1280  %63 = getelementptr inbounds float, float* %1, i32 %62
1281  %64 = getelementptr inbounds float, float* %63, i32 %23
1282  %65 = getelementptr inbounds float, float* %64, i32 %23
1283  %66 = getelementptr inbounds float, float* %65, i32 %23
1284  br label %67
1285
128667: ; preds = %45, %67
1287  %68 = phi float* [ %63, %45 ], [ %89, %67 ]
1288  %69 = phi float* [ %65, %45 ], [ %103, %67 ]
1289  %70 = phi float* [ %66, %45 ], [ %110, %67 ]
1290  %71 = phi float* [ %64, %45 ], [ %96, %67 ]
1291  %72 = phi float* [ %61, %45 ], [ %107, %67 ]
1292  %73 = phi float* [ %56, %45 ], [ %93, %67 ]
1293  %74 = phi float* [ %51, %45 ], [ %100, %67 ]
1294  %75 = phi i32 [ %24, %45 ], [ %111, %67 ]
1295  %76 = bitcast float* %69 to <4 x float>*
1296  %77 = bitcast float* %68 to <4 x float>*
1297  %78 = load <4 x float>, <4 x float>* %76, align 4
1298  %79 = load <4 x float>, <4 x float>* %77, align 4
1299  %80 = bitcast float* %71 to <4 x float>*
1300  %81 = load <4 x float>, <4 x float>* %80, align 4
1301  %82 = bitcast float* %70 to <4 x float>*
1302  %83 = load <4 x float>, <4 x float>* %82, align 4
1303  %84 = fadd <4 x float> %78, %79
1304  %85 = fsub <4 x float> %79, %78
1305  %86 = fadd <4 x float> %81, %83
1306  %87 = fsub <4 x float> %81, %83
1307  %88 = fadd <4 x float> %84, %86
1308  store <4 x float> %88, <4 x float>* %77, align 4
1309  %89 = getelementptr inbounds float, float* %68, i32 4
1310  %90 = fsub <4 x float> %84, %86
1311  %91 = bitcast float* %73 to <4 x float>*
1312  %92 = load <4 x float>, <4 x float>* %91, align 4
1313  %93 = getelementptr inbounds float, float* %73, i32 4
1314  %94 = tail call <4 x float> @llvm.arm.mve.vcmulq.v4f32(i32 0, <4 x float> %92, <4 x float> %90)
1315  %95 = tail call <4 x float> @llvm.arm.mve.vcmlaq.v4f32(i32 1, <4 x float> %94, <4 x float> %92, <4 x float> %90)
1316  store <4 x float> %95, <4 x float>* %80, align 4
1317  %96 = getelementptr inbounds float, float* %71, i32 4
1318  %97 = tail call <4 x float> @llvm.arm.mve.vcaddq.v4f32(i32 1, i32 0, <4 x float> %85, <4 x float> %87)
1319  %98 = bitcast float* %74 to <4 x float>*
1320  %99 = load <4 x float>, <4 x float>* %98, align 4
1321  %100 = getelementptr inbounds float, float* %74, i32 4
1322  %101 = tail call <4 x float> @llvm.arm.mve.vcmulq.v4f32(i32 0, <4 x float> %99, <4 x float> %97)
1323  %102 = tail call <4 x float> @llvm.arm.mve.vcmlaq.v4f32(i32 1, <4 x float> %101, <4 x float> %99, <4 x float> %97)
1324  store <4 x float> %102, <4 x float>* %76, align 4
1325  %103 = getelementptr inbounds float, float* %69, i32 4
1326  %104 = tail call <4 x float> @llvm.arm.mve.vcaddq.v4f32(i32 1, i32 1, <4 x float> %85, <4 x float> %87)
1327  %105 = bitcast float* %72 to <4 x float>*
1328  %106 = load <4 x float>, <4 x float>* %105, align 4
1329  %107 = getelementptr inbounds float, float* %72, i32 4
1330  %108 = tail call <4 x float> @llvm.arm.mve.vcmulq.v4f32(i32 0, <4 x float> %106, <4 x float> %104)
1331  %109 = tail call <4 x float> @llvm.arm.mve.vcmlaq.v4f32(i32 1, <4 x float> %108, <4 x float> %106, <4 x float> %104)
1332  store <4 x float> %109, <4 x float>* %82, align 4
1333  %110 = getelementptr inbounds float, float* %70, i32 4
1334  %111 = add nsw i32 %75, -1
1335  %112 = icmp eq i32 %111, 0
1336  br i1 %112, label %113, label %67
1337
1338113: ; preds = %67
1339  %114 = add nuw nsw i32 %46, 1
1340  %115 = icmp eq i32 %114, %17
1341  br i1 %115, label %40, label %45
1342
1343116: ; preds = %35, %116
1344  %117 = phi <4 x i32> [ %32, %35 ], [ %128, %116 ]
1345  %118 = phi i32 [ %33, %35 ], [ %139, %116 ]
1346  %119 = phi <4 x float> [ %36, %35 ], [ %130, %116 ]
1347  %120 = phi <4 x float> [ %37, %35 ], [ %129, %116 ]
1348  %121 = fadd <4 x float> %120, %119
1349  %122 = fsub <4 x float> %120, %119
1350  %123 = tail call <4 x float> @llvm.arm.mve.vldr.gather.base.v4f32.v4i32(<4 x i32> %117, i32 8)
1351  %124 = tail call <4 x float> @llvm.arm.mve.vldr.gather.base.v4f32.v4i32(<4 x i32> %117, i32 24)
1352  %125 = fadd <4 x float> %123, %124
1353  %126 = fsub <4 x float> %123, %124
1354  %127 = tail call { <4 x float>, <4 x i32> } @llvm.arm.mve.vldr.gather.base.wb.v4f32.v4i32(<4 x i32> %117, i32 64)
1355  %128 = extractvalue { <4 x float>, <4 x i32> } %127, 1
1356  %129 = extractvalue { <4 x float>, <4 x i32> } %127, 0
1357  %130 = tail call <4 x float> @llvm.arm.mve.vldr.gather.base.v4f32.v4i32(<4 x i32> %128, i32 16)
1358  %131 = fadd <4 x float> %121, %125
1359  %132 = fmul <4 x float> %39, %131
1360  tail call void @llvm.arm.mve.vstr.scatter.base.v4i32.v4f32(<4 x i32> %128, i32 -64, <4 x float> %132)
1361  %133 = fsub <4 x float> %121, %125
1362  %134 = fmul <4 x float> %39, %133
1363  tail call void @llvm.arm.mve.vstr.scatter.base.v4i32.v4f32(<4 x i32> %128, i32 -56, <4 x float> %134)
1364  %135 = tail call <4 x float> @llvm.arm.mve.vcaddq.v4f32(i32 1, i32 0, <4 x float> %122, <4 x float> %126)
1365  %136 = fmul <4 x float> %39, %135
1366  tail call void @llvm.arm.mve.vstr.scatter.base.v4i32.v4f32(<4 x i32> %128, i32 -48, <4 x float> %136)
1367  %137 = tail call <4 x float> @llvm.arm.mve.vcaddq.v4f32(i32 1, i32 1, <4 x float> %122, <4 x float> %126)
1368  %138 = fmul <4 x float> %39, %137
1369  tail call void @llvm.arm.mve.vstr.scatter.base.v4i32.v4f32(<4 x i32> %128, i32 -40, <4 x float> %138)
1370  %139 = add nsw i32 %118, -1
1371  %140 = icmp eq i32 %139, 0
1372  br i1 %140, label %141, label %116
1373
1374141: ; preds = %116, %26
1375  ret void
1376}
1377
1378declare <16 x i1> @llvm.arm.mve.vctp8(i32)
1379declare <8 x i1> @llvm.arm.mve.vctp16(i32)
1380declare i32 @llvm.arm.mve.pred.v2i.v16i1(<16 x i1>)
1381declare <4 x float> @llvm.masked.load.v4f32.p0v4f32(<4 x float>*, i32 immarg, <4 x i1>, <4 x float>)
1382declare <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>*, i32, <8 x i1>, <8 x i8>)
1383declare <16 x i8> @llvm.masked.load.v16i8.p0v16i8(<16 x i8>*, i32 immarg, <16 x i1>, <16 x i8>)
1384declare void @llvm.masked.store.v4f32.p0v4f32(<4 x float>, <4 x float>*, i32 immarg, <4 x i1>)
1385declare i32 @llvm.vector.reduce.add.v16i8(<16 x i32> %ext4)
1386declare i32 @llvm.arm.mve.vmldava.v8i16(i32, i32, i32, i32, <8 x i16>, <8 x i16>)
1387declare i32 @llvm.arm.mve.vmldava.predicated.v16i8.v16i1(i32, i32, i32, i32, <16 x i8>, <16 x i8>, <16 x i1>)
1388declare i32 @llvm.arm.mve.vmldava.predicated.v8i16.v8i1(i32, i32, i32, i32, <8 x i16>, <8 x i16>, <8 x i1>)
1389declare <8 x i16> @llvm.arm.mve.add.predicated.v8i16.v8i1(<8 x i16>, <8 x i16>, <8 x i1>, <8 x i16>)
1390
1391declare <4 x float> @llvm.arm.mve.vcmulq.v4f32(i32, <4 x float>, <4 x float>)
1392declare <4 x float> @llvm.arm.mve.vcmlaq.v4f32(i32, <4 x float>, <4 x float>, <4 x float>)
1393declare <4 x float> @llvm.arm.mve.vcaddq.v4f32(i32, i32, <4 x float>, <4 x float>)
1394declare { <4 x float>, <4 x i32> } @llvm.arm.mve.vldr.gather.base.wb.v4f32.v4i32(<4 x i32>, i32)
1395declare <4 x float> @llvm.arm.mve.vldr.gather.base.v4f32.v4i32(<4 x i32>, i32)
1396declare void @llvm.arm.mve.vstr.scatter.base.v4i32.v4f32(<4 x i32>, i32, <4 x float>)
1397