1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
| ; RUN: llc -march=hexagon -O0 < %s | FileCheck --check-prefix=CHECKO0 %s
; KP: Removed -O2 check. The code has become more aggressively optimized
; (some loads were found to be redundant and have been removed completely),
; and verifying correct code generation has become more difficult than
; its worth.
; CHECK: v{{[0-9]*}} = vsplat(r{{[0-9]*}})
; CHECK: v{{[0-9]*}} = vsplat(r{{[0-9]*}})
; CHECKO0: vmem(r{{[0-9]*}}+#0) = v{{[0-9]*}}
; CHECKO0: v{{[0-9]*}} = vmem(r{{[0-9]*}}+#0)
; CHECKO0: v{{[0-9]*}} = vmem(r{{[0-9]*}}+#0)
; Allow .cur loads.
; CHECKO2: v{{[0-9].*}} = vmem(r{{[0-9]*}}+#0)
; CHECKO2: vmem(r{{[0-9]*}}+#0) = v{{[0-9]*}}
; CHECKO2: v{{[0-9].*}} = vmem(r{{[0-9]*}}+#0)
; CHECK: v{{[0-9]*}}:{{[0-9]*}} = vcombine(v{{[0-9]*}},v{{[0-9]*}})
; CHECK: vmem(r{{[0-9]*}}+#0) = v{{[0-9]*}}
; CHECK: vmem(r{{[0-9]*}}+#32) = v{{[0-9]*}}
; CHECK: v{{[0-9]*}} = vmem(r{{[0-9]*}}+#0)
; CHECK: v{{[0-9]*}} = vmem(r{{[0-9]*}}+#32)
; CHECK: vmem(r{{[0-9]*}}+#0) = v{{[0-9]*}}
; CHECK: vmem(r{{[0-9]*}}+#32) = v{{[0-9]*}}
target triple = "hexagon"
@g0 = common global [10 x <32 x i32>] zeroinitializer, align 64
@g1 = private unnamed_addr constant [11 x i8] c"c[%d]= %x\0A\00", align 8
@g2 = common global [10 x <16 x i32>] zeroinitializer, align 64
@g3 = common global [10 x <16 x i32>] zeroinitializer, align 64
@g4 = common global [10 x <32 x i32>] zeroinitializer, align 64
declare i32 @f0(i8*, ...)
; Function Attrs: nounwind
define void @f1(i32 %a0) #0 {
b0:
%v0 = alloca i32, align 4
%v1 = alloca i32*, align 4
%v2 = alloca i32, align 4
store i32 %a0, i32* %v0, align 4
store i32* getelementptr inbounds ([10 x <32 x i32>], [10 x <32 x i32>]* @g0, i32 0, i32 0, i32 0), i32** %v1, align 4
%v3 = load i32, i32* %v0, align 4
%v4 = load i32*, i32** %v1, align 4
%v5 = getelementptr inbounds i32, i32* %v4, i32 %v3
store i32* %v5, i32** %v1, align 4
store i32 0, i32* %v2, align 4
br label %b1
b1: ; preds = %b3, %b0
%v6 = load i32, i32* %v2, align 4
%v7 = icmp slt i32 %v6, 16
br i1 %v7, label %b2, label %b4
b2: ; preds = %b1
%v8 = load i32, i32* %v2, align 4
%v9 = load i32*, i32** %v1, align 4
%v10 = getelementptr inbounds i32, i32* %v9, i32 1
store i32* %v10, i32** %v1, align 4
%v11 = load i32, i32* %v9, align 4
%v12 = call i32 (i8*, ...) @f0(i8* getelementptr inbounds ([11 x i8], [11 x i8]* @g1, i32 0, i32 0), i32 %v8, i32 %v11)
br label %b3
b3: ; preds = %b2
%v13 = load i32, i32* %v2, align 4
%v14 = add nsw i32 %v13, 1
store i32 %v14, i32* %v2, align 4
br label %b1
b4: ; preds = %b1
ret void
}
; Function Attrs: nounwind
define i32 @f2() #0 {
b0:
%v0 = alloca i32, align 4
%v1 = alloca i32, align 4
store i32 0, i32* %v0
store i32 0, i32* %v1, align 4
br label %b1
b1: ; preds = %b3, %b0
%v2 = load i32, i32* %v1, align 4
%v3 = icmp slt i32 %v2, 3
br i1 %v3, label %b2, label %b4
b2: ; preds = %b1
%v4 = load i32, i32* %v1, align 4
%v5 = add nsw i32 %v4, 1
%v6 = call <16 x i32> @llvm.hexagon.V6.lvsplatw(i32 %v5)
%v7 = load i32, i32* %v1, align 4
%v8 = getelementptr inbounds [10 x <16 x i32>], [10 x <16 x i32>]* @g2, i32 0, i32 %v7
store <16 x i32> %v6, <16 x i32>* %v8, align 64
%v9 = load i32, i32* %v1, align 4
%v10 = mul nsw i32 %v9, 10
%v11 = add nsw i32 %v10, 1
%v12 = call <16 x i32> @llvm.hexagon.V6.lvsplatw(i32 %v11)
%v13 = load i32, i32* %v1, align 4
%v14 = getelementptr inbounds [10 x <16 x i32>], [10 x <16 x i32>]* @g3, i32 0, i32 %v13
store <16 x i32> %v12, <16 x i32>* %v14, align 64
%v15 = load i32, i32* %v1, align 4
%v16 = getelementptr inbounds [10 x <16 x i32>], [10 x <16 x i32>]* @g2, i32 0, i32 %v15
%v17 = load <16 x i32>, <16 x i32>* %v16, align 64
%v18 = load i32, i32* %v1, align 4
%v19 = getelementptr inbounds [10 x <16 x i32>], [10 x <16 x i32>]* @g3, i32 0, i32 %v18
%v20 = load <16 x i32>, <16 x i32>* %v19, align 64
%v21 = call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> %v17, <16 x i32> %v20)
%v22 = load i32, i32* %v1, align 4
%v23 = getelementptr inbounds [10 x <32 x i32>], [10 x <32 x i32>]* @g4, i32 0, i32 %v22
store <32 x i32> %v21, <32 x i32>* %v23, align 64
br label %b3
b3: ; preds = %b2
%v24 = load i32, i32* %v1, align 4
%v25 = add nsw i32 %v24, 1
store i32 %v25, i32* %v1, align 4
br label %b1
b4: ; preds = %b1
store i32 0, i32* %v1, align 4
br label %b5
b5: ; preds = %b7, %b4
%v26 = load i32, i32* %v1, align 4
%v27 = icmp slt i32 %v26, 3
br i1 %v27, label %b6, label %b8
b6: ; preds = %b5
%v28 = load i32, i32* %v1, align 4
%v29 = getelementptr inbounds [10 x <32 x i32>], [10 x <32 x i32>]* @g4, i32 0, i32 %v28
%v30 = load <32 x i32>, <32 x i32>* %v29, align 64
%v31 = load i32, i32* %v1, align 4
%v32 = getelementptr inbounds [10 x <32 x i32>], [10 x <32 x i32>]* @g0, i32 0, i32 %v31
store <32 x i32> %v30, <32 x i32>* %v32, align 64
br label %b7
b7: ; preds = %b6
%v33 = load i32, i32* %v1, align 4
%v34 = add nsw i32 %v33, 1
store i32 %v34, i32* %v1, align 4
br label %b5
b8: ; preds = %b5
store i32 0, i32* %v1, align 4
br label %b9
b9: ; preds = %b11, %b8
%v35 = load i32, i32* %v1, align 4
%v36 = icmp slt i32 %v35, 3
br i1 %v36, label %b10, label %b12
b10: ; preds = %b9
%v37 = load i32, i32* %v1, align 4
%v38 = mul nsw i32 %v37, 16
call void @f1(i32 %v38)
br label %b11
b11: ; preds = %b10
%v39 = load i32, i32* %v1, align 4
%v40 = add nsw i32 %v39, 1
store i32 %v40, i32* %v1, align 4
br label %b9
b12: ; preds = %b9
ret i32 0
}
; Function Attrs: nounwind readnone
declare <16 x i32> @llvm.hexagon.V6.lvsplatw(i32) #1
; Function Attrs: nounwind readnone
declare <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32>, <16 x i32>) #1
attributes #0 = { nounwind "target-cpu"="hexagonv60" "target-features"="+hvxv60,+hvx-length64b" }
attributes #1 = { nounwind readnone }
|