1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2;
3; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=enabled %s -o - | \
4; RUN:   FileCheck %s --check-prefix=ENABLED
5;
6; Forcing tail-predication should not be necessary here, so we check the same
7; ENABLED label as the run above:
8; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=force-enabled %s -o - | \
9; RUN:   FileCheck %s  --check-prefix=ENABLED
10;
11; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=enabled-no-reductions %s -o - | \
12; RUN:   FileCheck %s --check-prefix=NOREDUCTIONS
13;
14; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve -tail-predication=force-enabled-no-reductions %s -o - | \
15; RUN:   FileCheck %s --check-prefix=NOREDUCTIONS
16
17define dso_local void @varying_outer_2d_reduction(i16* nocapture readonly %Input, i16* nocapture %Output, i16 signext %Size, i16 signext %N, i16 signext %Scale) local_unnamed_addr {
18; ENABLED-LABEL: varying_outer_2d_reduction:
19; ENABLED:       @ %bb.0: @ %entry
20; ENABLED-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, lr}
21; ENABLED-NEXT:    sub sp, #4
22; ENABLED-NEXT:    cmp r3, #1
23; ENABLED-NEXT:    str r0, [sp] @ 4-byte Spill
24; ENABLED-NEXT:    blt .LBB0_8
25; ENABLED-NEXT:  @ %bb.1: @ %for.body.lr.ph
26; ENABLED-NEXT:    ldr r0, [sp, #36]
27; ENABLED-NEXT:    add.w r12, r2, #3
28; ENABLED-NEXT:    ldr.w r10, [sp] @ 4-byte Reload
29; ENABLED-NEXT:    mov.w r8, #0
30; ENABLED-NEXT:    mov r9, r12
31; ENABLED-NEXT:    uxth r0, r0
32; ENABLED-NEXT:    rsbs r5, r0, #0
33; ENABLED-NEXT:    b .LBB0_4
34; ENABLED-NEXT:  .LBB0_2: @ in Loop: Header=BB0_4 Depth=1
35; ENABLED-NEXT:    movs r0, #0
36; ENABLED-NEXT:  .LBB0_3: @ %for.end
37; ENABLED-NEXT:    @ in Loop: Header=BB0_4 Depth=1
38; ENABLED-NEXT:    lsrs r0, r0, #16
39; ENABLED-NEXT:    sub.w r9, r9, #1
40; ENABLED-NEXT:    strh.w r0, [r1, r8, lsl #1]
41; ENABLED-NEXT:    add.w r8, r8, #1
42; ENABLED-NEXT:    add.w r10, r10, #2
43; ENABLED-NEXT:    cmp r8, r3
44; ENABLED-NEXT:    beq .LBB0_8
45; ENABLED-NEXT:  .LBB0_4: @ %for.body
46; ENABLED-NEXT:    @ =>This Loop Header: Depth=1
47; ENABLED-NEXT:    @ Child Loop BB0_6 Depth 2
48; ENABLED-NEXT:    cmp r2, r8
49; ENABLED-NEXT:    ble .LBB0_2
50; ENABLED-NEXT:  @ %bb.5: @ %vector.ph
51; ENABLED-NEXT:    @ in Loop: Header=BB0_4 Depth=1
52; ENABLED-NEXT:    bic r0, r9, #3
53; ENABLED-NEXT:    movs r7, #1
54; ENABLED-NEXT:    subs r0, #4
55; ENABLED-NEXT:    sub.w r4, r2, r8
56; ENABLED-NEXT:    vmov.i32 q1, #0x0
57; ENABLED-NEXT:    add.w r6, r7, r0, lsr #2
58; ENABLED-NEXT:    sub.w r0, r12, r8
59; ENABLED-NEXT:    bic r0, r0, #3
60; ENABLED-NEXT:    subs r0, #4
61; ENABLED-NEXT:    add.w r0, r7, r0, lsr #2
62; ENABLED-NEXT:    mov r7, r10
63; ENABLED-NEXT:    dls lr, r0
64; ENABLED-NEXT:    ldr r0, [sp] @ 4-byte Reload
65; ENABLED-NEXT:  .LBB0_6: @ %vector.body
66; ENABLED-NEXT:    @ Parent Loop BB0_4 Depth=1
67; ENABLED-NEXT:    @ => This Inner Loop Header: Depth=2
68; ENABLED-NEXT:    vctp.32 r4
69; ENABLED-NEXT:    vmov q0, q1
70; ENABLED-NEXT:    vpstt
71; ENABLED-NEXT:    vldrht.s32 q1, [r0], #8
72; ENABLED-NEXT:    vldrht.s32 q2, [r7], #8
73; ENABLED-NEXT:    mov lr, r6
74; ENABLED-NEXT:    vmul.i32 q1, q2, q1
75; ENABLED-NEXT:    subs r6, #1
76; ENABLED-NEXT:    vshl.s32 q1, r5
77; ENABLED-NEXT:    subs r4, #4
78; ENABLED-NEXT:    vadd.i32 q1, q1, q0
79; ENABLED-NEXT:    le lr, .LBB0_6
80; ENABLED-NEXT:  @ %bb.7: @ %middle.block
81; ENABLED-NEXT:    @ in Loop: Header=BB0_4 Depth=1
82; ENABLED-NEXT:    vpsel q0, q1, q0
83; ENABLED-NEXT:    vaddv.u32 r0, q0
84; ENABLED-NEXT:    b .LBB0_3
85; ENABLED-NEXT:  .LBB0_8: @ %for.end17
86; ENABLED-NEXT:    add sp, #4
87; ENABLED-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, pc}
88;
89; NOREDUCTIONS-LABEL: varying_outer_2d_reduction:
90; NOREDUCTIONS:       @ %bb.0: @ %entry
91; NOREDUCTIONS-NEXT:    push.w {r4, r5, r6, r7, r8, r9, r10, lr}
92; NOREDUCTIONS-NEXT:    sub sp, #4
93; NOREDUCTIONS-NEXT:    cmp r3, #1
94; NOREDUCTIONS-NEXT:    str r0, [sp] @ 4-byte Spill
95; NOREDUCTIONS-NEXT:    blt .LBB0_8
96; NOREDUCTIONS-NEXT:  @ %bb.1: @ %for.body.lr.ph
97; NOREDUCTIONS-NEXT:    ldr r0, [sp, #36]
98; NOREDUCTIONS-NEXT:    add.w r12, r2, #3
99; NOREDUCTIONS-NEXT:    ldr.w r10, [sp] @ 4-byte Reload
100; NOREDUCTIONS-NEXT:    mov.w r8, #0
101; NOREDUCTIONS-NEXT:    mov r9, r12
102; NOREDUCTIONS-NEXT:    uxth r0, r0
103; NOREDUCTIONS-NEXT:    rsbs r5, r0, #0
104; NOREDUCTIONS-NEXT:    b .LBB0_4
105; NOREDUCTIONS-NEXT:  .LBB0_2: @ in Loop: Header=BB0_4 Depth=1
106; NOREDUCTIONS-NEXT:    movs r0, #0
107; NOREDUCTIONS-NEXT:  .LBB0_3: @ %for.end
108; NOREDUCTIONS-NEXT:    @ in Loop: Header=BB0_4 Depth=1
109; NOREDUCTIONS-NEXT:    lsrs r0, r0, #16
110; NOREDUCTIONS-NEXT:    sub.w r9, r9, #1
111; NOREDUCTIONS-NEXT:    strh.w r0, [r1, r8, lsl #1]
112; NOREDUCTIONS-NEXT:    add.w r8, r8, #1
113; NOREDUCTIONS-NEXT:    add.w r10, r10, #2
114; NOREDUCTIONS-NEXT:    cmp r8, r3
115; NOREDUCTIONS-NEXT:    beq .LBB0_8
116; NOREDUCTIONS-NEXT:  .LBB0_4: @ %for.body
117; NOREDUCTIONS-NEXT:    @ =>This Loop Header: Depth=1
118; NOREDUCTIONS-NEXT:    @ Child Loop BB0_6 Depth 2
119; NOREDUCTIONS-NEXT:    cmp r2, r8
120; NOREDUCTIONS-NEXT:    ble .LBB0_2
121; NOREDUCTIONS-NEXT:  @ %bb.5: @ %vector.ph
122; NOREDUCTIONS-NEXT:    @ in Loop: Header=BB0_4 Depth=1
123; NOREDUCTIONS-NEXT:    bic r0, r9, #3
124; NOREDUCTIONS-NEXT:    movs r7, #1
125; NOREDUCTIONS-NEXT:    subs r0, #4
126; NOREDUCTIONS-NEXT:    sub.w r4, r2, r8
127; NOREDUCTIONS-NEXT:    vmov.i32 q1, #0x0
128; NOREDUCTIONS-NEXT:    add.w r6, r7, r0, lsr #2
129; NOREDUCTIONS-NEXT:    sub.w r0, r12, r8
130; NOREDUCTIONS-NEXT:    bic r0, r0, #3
131; NOREDUCTIONS-NEXT:    subs r0, #4
132; NOREDUCTIONS-NEXT:    add.w r0, r7, r0, lsr #2
133; NOREDUCTIONS-NEXT:    mov r7, r10
134; NOREDUCTIONS-NEXT:    dls lr, r0
135; NOREDUCTIONS-NEXT:    ldr r0, [sp] @ 4-byte Reload
136; NOREDUCTIONS-NEXT:  .LBB0_6: @ %vector.body
137; NOREDUCTIONS-NEXT:    @ Parent Loop BB0_4 Depth=1
138; NOREDUCTIONS-NEXT:    @ => This Inner Loop Header: Depth=2
139; NOREDUCTIONS-NEXT:    vctp.32 r4
140; NOREDUCTIONS-NEXT:    vmov q0, q1
141; NOREDUCTIONS-NEXT:    vpstt
142; NOREDUCTIONS-NEXT:    vldrht.s32 q1, [r0], #8
143; NOREDUCTIONS-NEXT:    vldrht.s32 q2, [r7], #8
144; NOREDUCTIONS-NEXT:    mov lr, r6
145; NOREDUCTIONS-NEXT:    vmul.i32 q1, q2, q1
146; NOREDUCTIONS-NEXT:    subs r6, #1
147; NOREDUCTIONS-NEXT:    vshl.s32 q1, r5
148; NOREDUCTIONS-NEXT:    subs r4, #4
149; NOREDUCTIONS-NEXT:    vadd.i32 q1, q1, q0
150; NOREDUCTIONS-NEXT:    le lr, .LBB0_6
151; NOREDUCTIONS-NEXT:  @ %bb.7: @ %middle.block
152; NOREDUCTIONS-NEXT:    @ in Loop: Header=BB0_4 Depth=1
153; NOREDUCTIONS-NEXT:    vpsel q0, q1, q0
154; NOREDUCTIONS-NEXT:    vaddv.u32 r0, q0
155; NOREDUCTIONS-NEXT:    b .LBB0_3
156; NOREDUCTIONS-NEXT:  .LBB0_8: @ %for.end17
157; NOREDUCTIONS-NEXT:    add sp, #4
158; NOREDUCTIONS-NEXT:    pop.w {r4, r5, r6, r7, r8, r9, r10, pc}
159entry:
160  %conv = sext i16 %N to i32
161  %cmp36 = icmp sgt i16 %N, 0
162  br i1 %cmp36, label %for.body.lr.ph, label %for.end17
163
164for.body.lr.ph:                                   ; preds = %entry
165  %conv2 = sext i16 %Size to i32
166  %conv1032 = zext i16 %Scale to i32
167  %i = add i32 %conv2, 3
168  br label %for.body
169
170for.body:                                         ; preds = %for.end, %for.body.lr.ph
171  %lsr.iv51 = phi i32 [ %lsr.iv.next, %for.end ], [ %i, %for.body.lr.ph ]
172  %lsr.iv46 = phi i16* [ %scevgep47, %for.end ], [ %Input, %for.body.lr.ph ]
173  %i.037 = phi i32 [ 0, %for.body.lr.ph ], [ %inc16, %for.end ]
174  %i1 = mul nsw i32 %i.037, -1
175  %i2 = add i32 %i, %i1
176  %i3 = lshr i32 %i2, 2
177  %i4 = shl nuw i32 %i3, 2
178  %i5 = add i32 %i4, -4
179  %i6 = lshr i32 %i5, 2
180  %i7 = add nuw nsw i32 %i6, 1
181  %i8 = sub i32 %conv2, %i.037
182  %cmp433 = icmp slt i32 %i.037, %conv2
183  br i1 %cmp433, label %vector.ph, label %for.end
184
185vector.ph:                                        ; preds = %for.body
186  %start = call i32 @llvm.start.loop.iterations.i32(i32 %i7)
187  br label %vector.body
188
189vector.body:                                      ; preds = %vector.body, %vector.ph
190  %lsr.iv48 = phi i16* [ %scevgep49, %vector.body ], [ %lsr.iv46, %vector.ph ]
191  %lsr.iv = phi i16* [ %scevgep, %vector.body ], [ %Input, %vector.ph ]
192  %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
193  %vec.phi = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %i16, %vector.body ]
194  %i9 = phi i32 [ %start, %vector.ph ], [ %i17, %vector.body ]
195  %lsr.iv4850 = bitcast i16* %lsr.iv48 to <4 x i16>*
196  %lsr.iv45 = bitcast i16* %lsr.iv to <4 x i16>*
197  %active.lane.mask = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %i8)
198  %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %lsr.iv45, i32 2, <4 x i1> %active.lane.mask, <4 x i16> undef)
199  %i10 = sext <4 x i16> %wide.masked.load to <4 x i32>
200  %wide.masked.load42 = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %lsr.iv4850, i32 2, <4 x i1> %active.lane.mask, <4 x i16> undef)
201  %i11 = sext <4 x i16> %wide.masked.load42 to <4 x i32>
202  %i12 = mul nsw <4 x i32> %i11, %i10
203  %i13 = insertelement <4 x i32> undef, i32 %conv1032, i32 0
204  %i14 = shufflevector <4 x i32> %i13, <4 x i32> undef, <4 x i32> zeroinitializer
205  %i15 = ashr <4 x i32> %i12, %i14
206  %i16 = add <4 x i32> %i15, %vec.phi
207  %index.next = add i32 %index, 4
208  %scevgep = getelementptr i16, i16* %lsr.iv, i32 4
209  %scevgep49 = getelementptr i16, i16* %lsr.iv48, i32 4
210  %i17 = call i32 @llvm.loop.decrement.reg.i32(i32 %i9, i32 1)
211  %i18 = icmp ne i32 %i17, 0
212  br i1 %i18, label %vector.body, label %middle.block
213
214middle.block:                                     ; preds = %vector.body
215  %i19 = select <4 x i1> %active.lane.mask, <4 x i32> %i16, <4 x i32> %vec.phi
216  %i20 = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %i19)
217  br label %for.end
218
219for.end:                                          ; preds = %middle.block, %for.body
220  %Sum.0.lcssa = phi i32 [ 0, %for.body ], [ %i20, %middle.block ]
221  %i21 = lshr i32 %Sum.0.lcssa, 16
222  %conv13 = trunc i32 %i21 to i16
223  %arrayidx14 = getelementptr inbounds i16, i16* %Output, i32 %i.037
224  store i16 %conv13, i16* %arrayidx14, align 2
225  %inc16 = add nuw nsw i32 %i.037, 1
226  %scevgep47 = getelementptr i16, i16* %lsr.iv46, i32 1
227  %lsr.iv.next = add i32 %lsr.iv51, -1
228  %exitcond39 = icmp eq i32 %inc16, %conv
229  br i1 %exitcond39, label %for.end17, label %for.body
230
231for.end17:                                        ; preds = %for.end, %entry
232  ret void
233}
234
235declare <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32, i32)
236declare <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>*, i32 immarg, <4 x i1>, <4 x i16>)
237declare i32 @llvm.vector.reduce.add.v4i32(<4 x i32>)
238declare i32 @llvm.loop.decrement.reg.i32(i32, i32)
239declare i32 @llvm.start.loop.iterations.i32(i32)
240