1; RUN: opt -arm-parallel-dsp -dce -mtriple=armv7-a -S %s -o - | FileCheck %s
2
3; CHECK-LABEL: single_block
4; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
5; CHECK: [[A:%[^ ]+]] = load i32, i32* [[CAST_A]]
6; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
7; CHECK: [[B:%[^ ]+]] = load i32, i32* [[CAST_B]]
8; CHECK: call i32 @llvm.arm.smlad(i32 [[A]], i32 [[B]], i32 %acc)
9define i32 @single_block(i16* %a, i16* %b, i32 %acc) {
10entry:
11  %ld.a.0 = load i16, i16* %a
12  %sext.a.0 = sext i16 %ld.a.0 to i32
13  %ld.b.0 = load i16, i16* %b
14  %sext.b.0 = sext i16 %ld.b.0 to i32
15  %mul.0 = mul i32 %sext.a.0, %sext.b.0
16  %addr.a.1 = getelementptr i16, i16* %a, i32 1
17  %addr.b.1 = getelementptr i16, i16* %b, i32 1
18  %ld.a.1 = load i16, i16* %addr.a.1
19  %sext.a.1 = sext i16 %ld.a.1 to i32
20  %ld.b.1 = load i16, i16* %addr.b.1
21  %sext.b.1 = sext i16 %ld.b.1 to i32
22  %mul.1 = mul i32 %sext.a.1, %sext.b.1
23  %add = add i32 %mul.0, %mul.1
24  %res = add i32 %add, %acc
25  ret i32 %res
26}
27
28; CHECK-LABEL: single_block_64
29; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
30; CHECK: [[A:%[^ ]+]] = load i32, i32* [[CAST_A]]
31; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
32; CHECK: [[B:%[^ ]+]] = load i32, i32* [[CAST_B]]
33; CHECK: call i64 @llvm.arm.smlald(i32 [[A]], i32 [[B]], i64 %acc)
34define i64 @single_block_64(i16* %a, i16* %b, i64 %acc) {
35entry:
36  %ld.a.0 = load i16, i16* %a
37  %sext.a.0 = sext i16 %ld.a.0 to i32
38  %ld.b.0 = load i16, i16* %b
39  %sext.b.0 = sext i16 %ld.b.0 to i32
40  %mul.0 = mul i32 %sext.a.0, %sext.b.0
41  %addr.a.1 = getelementptr i16, i16* %a, i32 1
42  %addr.b.1 = getelementptr i16, i16* %b, i32 1
43  %ld.a.1 = load i16, i16* %addr.a.1
44  %sext.a.1 = sext i16 %ld.a.1 to i32
45  %ld.b.1 = load i16, i16* %addr.b.1
46  %sext.b.1 = sext i16 %ld.b.1 to i32
47  %mul.1 = mul i32 %sext.a.1, %sext.b.1
48  %sext.mul.0 = sext i32 %mul.0 to i64
49  %sext.mul.1 = sext i32 %mul.1 to i64
50  %add = add i64 %sext.mul.0, %sext.mul.1
51  %res = add i64 %add, %acc
52  ret i64 %res
53}
54
55; CHECK-LABEL: multi_block
56; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
57; CHECK: [[A:%[^ ]+]] = load i32, i32* [[CAST_A]]
58; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
59; CHECK: [[B:%[^ ]+]] = load i32, i32* [[CAST_B]]
60; CHECK:  call i32 @llvm.arm.smlad(i32 [[A]], i32 [[B]], i32 0)
61define i32 @multi_block(i16* %a, i16* %b, i32 %acc) {
62entry:
63  %ld.a.0 = load i16, i16* %a
64  %sext.a.0 = sext i16 %ld.a.0 to i32
65  %ld.b.0 = load i16, i16* %b
66  %sext.b.0 = sext i16 %ld.b.0 to i32
67  %mul.0 = mul i32 %sext.a.0, %sext.b.0
68  %addr.a.1 = getelementptr i16, i16* %a, i32 1
69  %addr.b.1 = getelementptr i16, i16* %b, i32 1
70  %ld.a.1 = load i16, i16* %addr.a.1
71  %sext.a.1 = sext i16 %ld.a.1 to i32
72  %ld.b.1 = load i16, i16* %addr.b.1
73  %sext.b.1 = sext i16 %ld.b.1 to i32
74  %mul.1 = mul i32 %sext.a.1, %sext.b.1
75  %add = add i32 %mul.0, %mul.1
76  br label %bb.1
77
78bb.1:
79  %res = add i32 %add, %acc
80  ret i32 %res
81}
82
83; CHECK-LABEL: multi_block_64
84; CHECK: [[CAST_A:%[^ ]+]] = bitcast i16* %a to i32*
85; CHECK: [[A:%[^ ]+]] = load i32, i32* [[CAST_A]]
86; CHECK: [[CAST_B:%[^ ]+]] = bitcast i16* %b to i32*
87; CHECK: [[B:%[^ ]+]] = load i32, i32* [[CAST_B]]
88; CHECK:  call i64 @llvm.arm.smlald(i32 [[A]], i32 [[B]], i64 0)
89define i64 @multi_block_64(i16* %a, i16* %b, i64 %acc) {
90entry:
91  %ld.a.0 = load i16, i16* %a
92  %sext.a.0 = sext i16 %ld.a.0 to i32
93  %ld.b.0 = load i16, i16* %b
94  %sext.b.0 = sext i16 %ld.b.0 to i32
95  %mul.0 = mul i32 %sext.a.0, %sext.b.0
96  %addr.a.1 = getelementptr i16, i16* %a, i32 1
97  %addr.b.1 = getelementptr i16, i16* %b, i32 1
98  %ld.a.1 = load i16, i16* %addr.a.1
99  %sext.a.1 = sext i16 %ld.a.1 to i32
100  %ld.b.1 = load i16, i16* %addr.b.1
101  %sext.b.1 = sext i16 %ld.b.1 to i32
102  %mul.1 = mul i32 %sext.a.1, %sext.b.1
103  %sext.mul.0 = sext i32 %mul.0 to i64
104  %sext.mul.1 = sext i32 %mul.1 to i64
105  %add = add i64 %sext.mul.0, %sext.mul.1
106  br label %bb.1
107
108bb.1:
109  %res = add i64 %add, %acc
110  ret i64 %res
111}
112
113; CHECK-LABEL: multi_block_1
114; CHECK-NOT: call i32 @llvm.arm.smlad
115define i32 @multi_block_1(i16* %a, i16* %b, i32 %acc) {
116entry:
117  %ld.a.0 = load i16, i16* %a
118  %sext.a.0 = sext i16 %ld.a.0 to i32
119  %ld.b.0 = load i16, i16* %b
120  %sext.b.0 = sext i16 %ld.b.0 to i32
121  %mul.0 = mul i32 %sext.a.0, %sext.b.0
122  br label %bb.1
123
124bb.1:
125  %addr.a.1 = getelementptr i16, i16* %a, i32 1
126  %addr.b.1 = getelementptr i16, i16* %b, i32 1
127  %ld.a.1 = load i16, i16* %addr.a.1
128  %sext.a.1 = sext i16 %ld.a.1 to i32
129  %ld.b.1 = load i16, i16* %addr.b.1
130  %sext.b.1 = sext i16 %ld.b.1 to i32
131  %mul.1 = mul i32 %sext.a.1, %sext.b.1
132  %add = add i32 %mul.0, %mul.1
133  %res = add i32 %add, %acc
134  ret i32 %res
135}
136
137; TODO: Four smlads should be generated here, but mul.0 and mul.3 remain as
138; scalars.
139; CHECK-LABEL: num_load_limit
140; CHECK: call i32 @llvm.arm.smlad
141; CHECK: call i32 @llvm.arm.smlad
142; CHECK: call i32 @llvm.arm.smlad
143; CHECK-NOT: call i32 @llvm.arm.smlad
144define i32 @num_load_limit(i16* %a, i16* %b, i32 %acc) {
145entry:
146  %ld.a.0 = load i16, i16* %a
147  %sext.a.0 = sext i16 %ld.a.0 to i32
148  %ld.b.0 = load i16, i16* %b
149  %sext.b.0 = sext i16 %ld.b.0 to i32
150  %mul.0 = mul i32 %sext.a.0, %sext.b.0
151  %addr.a.1 = getelementptr i16, i16* %a, i32 1
152  %addr.b.1 = getelementptr i16, i16* %b, i32 1
153  %ld.a.1 = load i16, i16* %addr.a.1
154  %sext.a.1 = sext i16 %ld.a.1 to i32
155  %ld.b.1 = load i16, i16* %addr.b.1
156  %sext.b.1 = sext i16 %ld.b.1 to i32
157  %mul.1 = mul i32 %sext.a.1, %sext.b.1
158  %add.0 = add i32 %mul.0, %mul.1
159
160  %addr.a.2 = getelementptr i16, i16* %a, i32 2
161  %addr.b.2 = getelementptr i16, i16* %b, i32 2
162  %ld.a.2 = load i16, i16* %addr.a.2
163  %sext.a.2 = sext i16 %ld.a.2 to i32
164  %ld.b.2 = load i16, i16* %addr.b.2
165  %sext.b.2 = sext i16 %ld.b.2 to i32
166  %mul.2 = mul i32 %sext.a.0, %sext.b.0
167  %addr.a.3 = getelementptr i16, i16* %a, i32 3
168  %addr.b.3 = getelementptr i16, i16* %b, i32 3
169  %ld.a.3 = load i16, i16* %addr.a.3
170  %sext.a.3 = sext i16 %ld.a.3 to i32
171  %ld.b.3 = load i16, i16* %addr.b.3
172  %sext.b.3 = sext i16 %ld.b.3 to i32
173  %mul.3 = mul i32 %sext.a.1, %sext.b.3
174  %add.3 = add i32 %mul.2, %mul.3
175
176  %addr.a.4 = getelementptr i16, i16* %a, i32 4
177  %addr.b.4 = getelementptr i16, i16* %b, i32 4
178  %ld.a.4 = load i16, i16* %addr.a.4
179  %sext.a.4 = sext i16 %ld.a.4 to i32
180  %ld.b.4 = load i16, i16* %addr.b.4
181  %sext.b.4 = sext i16 %ld.b.4 to i32
182  %mul.4 = mul i32 %sext.a.4, %sext.b.4
183  %addr.a.5 = getelementptr i16, i16* %a, i32 5
184  %addr.b.5 = getelementptr i16, i16* %b, i32 5
185  %ld.a.5 = load i16, i16* %addr.a.5
186  %sext.a.5 = sext i16 %ld.a.5 to i32
187  %ld.b.5 = load i16, i16* %addr.b.5
188  %sext.b.5 = sext i16 %ld.b.5 to i32
189  %mul.5 = mul i32 %sext.a.5, %sext.b.5
190  %add.5 = add i32 %mul.4, %mul.5
191
192  %addr.a.6 = getelementptr i16, i16* %a, i32 6
193  %addr.b.6 = getelementptr i16, i16* %b, i32 6
194  %ld.a.6 = load i16, i16* %addr.a.6
195  %sext.a.6 = sext i16 %ld.a.6 to i32
196  %ld.b.6 = load i16, i16* %addr.b.6
197  %sext.b.6 = sext i16 %ld.b.6 to i32
198  %mul.6 = mul i32 %sext.a.6, %sext.b.6
199  %addr.a.7 = getelementptr i16, i16* %a, i32 7
200  %addr.b.7 = getelementptr i16, i16* %b, i32 7
201  %ld.a.7 = load i16, i16* %addr.a.7
202  %sext.a.7 = sext i16 %ld.a.7 to i32
203  %ld.b.7 = load i16, i16* %addr.b.7
204  %sext.b.7 = sext i16 %ld.b.7 to i32
205  %mul.7 = mul i32 %sext.a.7, %sext.b.7
206  %add.7 = add i32 %mul.6, %mul.7
207
208  %add.10 = add i32 %add.7, %add.5
209  %add.11 = add i32 %add.3, %add.0
210  %add.12 = add i32 %add.10, %add.11
211  %res = add i32 %add.12, %acc
212  ret i32 %res
213}
214
215; CHECK-LABEL: too_many_loads
216; CHECK-NOT: call i32 @llvm.arm.smlad
217define i32 @too_many_loads(i16* %a, i16* %b, i32 %acc) {
218entry:
219  %ld.a.0 = load i16, i16* %a
220  %sext.a.0 = sext i16 %ld.a.0 to i32
221  %ld.b.0 = load i16, i16* %b
222  %sext.b.0 = sext i16 %ld.b.0 to i32
223  %mul.0 = mul i32 %sext.a.0, %sext.b.0
224  %addr.a.1 = getelementptr i16, i16* %a, i32 1
225  %addr.b.1 = getelementptr i16, i16* %b, i32 1
226  %ld.a.1 = load i16, i16* %addr.a.1
227  %sext.a.1 = sext i16 %ld.a.1 to i32
228  %ld.b.1 = load i16, i16* %addr.b.1
229  %sext.b.1 = sext i16 %ld.b.1 to i32
230  %mul.1 = mul i32 %sext.a.1, %sext.b.1
231  %add.0 = add i32 %mul.0, %mul.1
232
233  %addr.a.2 = getelementptr i16, i16* %a, i32 2
234  %addr.b.2 = getelementptr i16, i16* %b, i32 2
235  %ld.a.2 = load i16, i16* %addr.a.2
236  %sext.a.2 = sext i16 %ld.a.2 to i32
237  %ld.b.2 = load i16, i16* %addr.b.2
238  %sext.b.2 = sext i16 %ld.b.2 to i32
239  %mul.2 = mul i32 %sext.a.0, %sext.b.0
240  %addr.a.3 = getelementptr i16, i16* %a, i32 3
241  %addr.b.3 = getelementptr i16, i16* %b, i32 3
242  %ld.a.3 = load i16, i16* %addr.a.3
243  %sext.a.3 = sext i16 %ld.a.3 to i32
244  %ld.b.3 = load i16, i16* %addr.b.3
245  %sext.b.3 = sext i16 %ld.b.3 to i32
246  %mul.3 = mul i32 %sext.a.1, %sext.b.3
247  %add.3 = add i32 %mul.2, %mul.3
248
249  %addr.a.4 = getelementptr i16, i16* %a, i32 4
250  %addr.b.4 = getelementptr i16, i16* %b, i32 4
251  %ld.a.4 = load i16, i16* %addr.a.4
252  %sext.a.4 = sext i16 %ld.a.4 to i32
253  %ld.b.4 = load i16, i16* %addr.b.4
254  %sext.b.4 = sext i16 %ld.b.4 to i32
255  %mul.4 = mul i32 %sext.a.4, %sext.b.4
256  %addr.a.5 = getelementptr i16, i16* %a, i32 5
257  %addr.b.5 = getelementptr i16, i16* %b, i32 5
258  %ld.a.5 = load i16, i16* %addr.a.5
259  %sext.a.5 = sext i16 %ld.a.5 to i32
260  %ld.b.5 = load i16, i16* %addr.b.5
261  %sext.b.5 = sext i16 %ld.b.5 to i32
262  %mul.5 = mul i32 %sext.a.5, %sext.b.5
263  %add.5 = add i32 %mul.4, %mul.5
264
265  %addr.a.6 = getelementptr i16, i16* %a, i32 6
266  %addr.b.6 = getelementptr i16, i16* %b, i32 6
267  %ld.a.6 = load i16, i16* %addr.a.6
268  %sext.a.6 = sext i16 %ld.a.6 to i32
269  %ld.b.6 = load i16, i16* %addr.b.6
270  %sext.b.6 = sext i16 %ld.b.6 to i32
271  %mul.6 = mul i32 %sext.a.6, %sext.b.6
272  %addr.a.7 = getelementptr i16, i16* %a, i32 7
273  %addr.b.7 = getelementptr i16, i16* %b, i32 7
274  %ld.a.7 = load i16, i16* %addr.a.7
275  %sext.a.7 = sext i16 %ld.a.7 to i32
276  %ld.b.7 = load i16, i16* %addr.b.7
277  %sext.b.7 = sext i16 %ld.b.7 to i32
278  %mul.7 = mul i32 %sext.a.7, %sext.b.7
279  %add.7 = add i32 %mul.6, %mul.7
280
281  %addr.a.8 = getelementptr i16, i16* %a, i32 7
282  %addr.b.8 = getelementptr i16, i16* %b, i32 7
283  %ld.a.8 = load i16, i16* %addr.a.8
284  %sext.a.8 = sext i16 %ld.a.8 to i32
285  %ld.b.8 = load i16, i16* %addr.b.8
286  %sext.b.8 = sext i16 %ld.b.8 to i32
287  %mul.8 = mul i32 %sext.a.8, %sext.b.8
288
289  %add.10 = add i32 %add.7, %add.5
290  %add.11 = add i32 %add.3, %add.0
291  %add.12 = add i32 %add.10, %add.11
292  %add.13 = add i32 %add.12, %acc
293  %res = add i32 %add.13, %mul.8
294  ret i32 %res
295}
296