1; RUN: llc -mcpu=pwr7 -O0 -disable-fp-elim < %s | FileCheck %s
2
3; FIXME: The code generation for packed structs is very poor because the
4; PowerPC target wrongly rejects all unaligned loads.  This test case will
5; need to be revised when that is fixed.
6
7target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v128:128:128-n32:64"
8target triple = "powerpc64-unknown-linux-gnu"
9
10%struct.s1 = type { i8 }
11%struct.s2 = type { i16 }
12%struct.s4 = type { i32 }
13%struct.t1 = type { i8 }
14%struct.t3 = type <{ i16, i8 }>
15%struct.t5 = type <{ i32, i8 }>
16%struct.t6 = type <{ i32, i16 }>
17%struct.t7 = type <{ i32, i16, i8 }>
18%struct.s3 = type { i16, i8 }
19%struct.s5 = type { i32, i8 }
20%struct.s6 = type { i32, i16 }
21%struct.s7 = type { i32, i16, i8 }
22%struct.t2 = type <{ i16 }>
23%struct.t4 = type <{ i32 }>
24
25@caller1.p1 = private unnamed_addr constant %struct.s1 { i8 1 }, align 1
26@caller1.p2 = private unnamed_addr constant %struct.s2 { i16 2 }, align 2
27@caller1.p3 = private unnamed_addr constant { i16, i8, i8 } { i16 4, i8 8, i8 undef }, align 2
28@caller1.p4 = private unnamed_addr constant %struct.s4 { i32 16 }, align 4
29@caller1.p5 = private unnamed_addr constant { i32, i8, [3 x i8] } { i32 32, i8 64, [3 x i8] undef }, align 4
30@caller1.p6 = private unnamed_addr constant { i32, i16, [2 x i8] } { i32 128, i16 256, [2 x i8] undef }, align 4
31@caller1.p7 = private unnamed_addr constant { i32, i16, i8, i8 } { i32 512, i16 1024, i8 -3, i8 undef }, align 4
32@caller2.p1 = private unnamed_addr constant %struct.t1 { i8 1 }, align 1
33@caller2.p2 = private unnamed_addr constant { i16 } { i16 2 }, align 1
34@caller2.p3 = private unnamed_addr constant %struct.t3 <{ i16 4, i8 8 }>, align 1
35@caller2.p4 = private unnamed_addr constant { i32 } { i32 16 }, align 1
36@caller2.p5 = private unnamed_addr constant %struct.t5 <{ i32 32, i8 64 }>, align 1
37@caller2.p6 = private unnamed_addr constant %struct.t6 <{ i32 128, i16 256 }>, align 1
38@caller2.p7 = private unnamed_addr constant %struct.t7 <{ i32 512, i16 1024, i8 -3 }>, align 1
39
40define i32 @caller1() nounwind {
41entry:
42  %p1 = alloca %struct.s1, align 1
43  %p2 = alloca %struct.s2, align 2
44  %p3 = alloca %struct.s3, align 2
45  %p4 = alloca %struct.s4, align 4
46  %p5 = alloca %struct.s5, align 4
47  %p6 = alloca %struct.s6, align 4
48  %p7 = alloca %struct.s7, align 4
49  %0 = bitcast %struct.s1* %p1 to i8*
50  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %0, i8* getelementptr inbounds (%struct.s1* @caller1.p1, i32 0, i32 0), i64 1, i32 1, i1 false)
51  %1 = bitcast %struct.s2* %p2 to i8*
52  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* bitcast (%struct.s2* @caller1.p2 to i8*), i64 2, i32 2, i1 false)
53  %2 = bitcast %struct.s3* %p3 to i8*
54  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %2, i8* bitcast ({ i16, i8, i8 }* @caller1.p3 to i8*), i64 4, i32 2, i1 false)
55  %3 = bitcast %struct.s4* %p4 to i8*
56  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* bitcast (%struct.s4* @caller1.p4 to i8*), i64 4, i32 4, i1 false)
57  %4 = bitcast %struct.s5* %p5 to i8*
58  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* bitcast ({ i32, i8, [3 x i8] }* @caller1.p5 to i8*), i64 8, i32 4, i1 false)
59  %5 = bitcast %struct.s6* %p6 to i8*
60  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %5, i8* bitcast ({ i32, i16, [2 x i8] }* @caller1.p6 to i8*), i64 8, i32 4, i1 false)
61  %6 = bitcast %struct.s7* %p7 to i8*
62  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %6, i8* bitcast ({ i32, i16, i8, i8 }* @caller1.p7 to i8*), i64 8, i32 4, i1 false)
63  %call = call i32 @callee1(%struct.s1* byval %p1, %struct.s2* byval %p2, %struct.s3* byval %p3, %struct.s4* byval %p4, %struct.s5* byval %p5, %struct.s6* byval %p6, %struct.s7* byval %p7)
64  ret i32 %call
65
66; CHECK: ld 9, 128(31)
67; CHECK: ld 8, 136(31)
68; CHECK: ld 7, 144(31)
69; CHECK: lwz 6, 152(31)
70; CHECK: lwz 5, 160(31)
71; CHECK: lhz 4, 168(31)
72; CHECK: lbz 3, 176(31)
73}
74
75declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind
76
77define internal i32 @callee1(%struct.s1* byval %v1, %struct.s2* byval %v2, %struct.s3* byval %v3, %struct.s4* byval %v4, %struct.s5* byval %v5, %struct.s6* byval %v6, %struct.s7* byval %v7) nounwind {
78entry:
79  %a = getelementptr inbounds %struct.s1* %v1, i32 0, i32 0
80  %0 = load i8* %a, align 1
81  %conv = zext i8 %0 to i32
82  %a1 = getelementptr inbounds %struct.s2* %v2, i32 0, i32 0
83  %1 = load i16* %a1, align 2
84  %conv2 = sext i16 %1 to i32
85  %add = add nsw i32 %conv, %conv2
86  %a3 = getelementptr inbounds %struct.s3* %v3, i32 0, i32 0
87  %2 = load i16* %a3, align 2
88  %conv4 = sext i16 %2 to i32
89  %add5 = add nsw i32 %add, %conv4
90  %a6 = getelementptr inbounds %struct.s4* %v4, i32 0, i32 0
91  %3 = load i32* %a6, align 4
92  %add7 = add nsw i32 %add5, %3
93  %a8 = getelementptr inbounds %struct.s5* %v5, i32 0, i32 0
94  %4 = load i32* %a8, align 4
95  %add9 = add nsw i32 %add7, %4
96  %a10 = getelementptr inbounds %struct.s6* %v6, i32 0, i32 0
97  %5 = load i32* %a10, align 4
98  %add11 = add nsw i32 %add9, %5
99  %a12 = getelementptr inbounds %struct.s7* %v7, i32 0, i32 0
100  %6 = load i32* %a12, align 4
101  %add13 = add nsw i32 %add11, %6
102  ret i32 %add13
103
104; CHECK: std 9, 96(1)
105; CHECK: std 8, 88(1)
106; CHECK: std 7, 80(1)
107; CHECK: stw 6, 76(1)
108; CHECK: stw 5, 68(1)
109; CHECK: sth 4, 62(1)
110; CHECK: stb 3, 55(1)
111; CHECK: lha {{[0-9]+}}, 62(1)
112; CHECK: lbz {{[0-9]+}}, 55(1)
113; CHECK: lha {{[0-9]+}}, 68(1)
114; CHECK: lwz {{[0-9]+}}, 76(1)
115; CHECK: lwz {{[0-9]+}}, 80(1)
116; CHECK: lwz {{[0-9]+}}, 88(1)
117; CHECK: lwz {{[0-9]+}}, 96(1)
118}
119
120define i32 @caller2() nounwind {
121entry:
122  %p1 = alloca %struct.t1, align 1
123  %p2 = alloca %struct.t2, align 1
124  %p3 = alloca %struct.t3, align 1
125  %p4 = alloca %struct.t4, align 1
126  %p5 = alloca %struct.t5, align 1
127  %p6 = alloca %struct.t6, align 1
128  %p7 = alloca %struct.t7, align 1
129  %0 = bitcast %struct.t1* %p1 to i8*
130  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %0, i8* getelementptr inbounds (%struct.t1* @caller2.p1, i32 0, i32 0), i64 1, i32 1, i1 false)
131  %1 = bitcast %struct.t2* %p2 to i8*
132  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* bitcast ({ i16 }* @caller2.p2 to i8*), i64 2, i32 1, i1 false)
133  %2 = bitcast %struct.t3* %p3 to i8*
134  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %2, i8* bitcast (%struct.t3* @caller2.p3 to i8*), i64 3, i32 1, i1 false)
135  %3 = bitcast %struct.t4* %p4 to i8*
136  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* bitcast ({ i32 }* @caller2.p4 to i8*), i64 4, i32 1, i1 false)
137  %4 = bitcast %struct.t5* %p5 to i8*
138  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* bitcast (%struct.t5* @caller2.p5 to i8*), i64 5, i32 1, i1 false)
139  %5 = bitcast %struct.t6* %p6 to i8*
140  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %5, i8* bitcast (%struct.t6* @caller2.p6 to i8*), i64 6, i32 1, i1 false)
141  %6 = bitcast %struct.t7* %p7 to i8*
142  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %6, i8* bitcast (%struct.t7* @caller2.p7 to i8*), i64 7, i32 1, i1 false)
143  %call = call i32 @callee2(%struct.t1* byval %p1, %struct.t2* byval %p2, %struct.t3* byval %p3, %struct.t4* byval %p4, %struct.t5* byval %p5, %struct.t6* byval %p6, %struct.t7* byval %p7)
144  ret i32 %call
145
146; CHECK: stb {{[0-9]+}}, 71(1)
147; CHECK: sth {{[0-9]+}}, 69(1)
148; CHECK: stb {{[0-9]+}}, 87(1)
149; CHECK: stw {{[0-9]+}}, 83(1)
150; CHECK: sth {{[0-9]+}}, 94(1)
151; CHECK: stw {{[0-9]+}}, 90(1)
152; CHECK: stb {{[0-9]+}}, 103(1)
153; CHECK: sth {{[0-9]+}}, 101(1)
154; CHECK: stw {{[0-9]+}}, 97(1)
155; CHECK: ld 9, 96(1)
156; CHECK: ld 8, 88(1)
157; CHECK: ld 7, 80(1)
158; CHECK: lwz 6, 152(31)
159; CHECK: ld 5, 64(1)
160; CHECK: lhz 4, 168(31)
161; CHECK: lbz 3, 176(31)
162}
163
164define internal i32 @callee2(%struct.t1* byval %v1, %struct.t2* byval %v2, %struct.t3* byval %v3, %struct.t4* byval %v4, %struct.t5* byval %v5, %struct.t6* byval %v6, %struct.t7* byval %v7) nounwind {
165entry:
166  %a = getelementptr inbounds %struct.t1* %v1, i32 0, i32 0
167  %0 = load i8* %a, align 1
168  %conv = zext i8 %0 to i32
169  %a1 = getelementptr inbounds %struct.t2* %v2, i32 0, i32 0
170  %1 = load i16* %a1, align 1
171  %conv2 = sext i16 %1 to i32
172  %add = add nsw i32 %conv, %conv2
173  %a3 = getelementptr inbounds %struct.t3* %v3, i32 0, i32 0
174  %2 = load i16* %a3, align 1
175  %conv4 = sext i16 %2 to i32
176  %add5 = add nsw i32 %add, %conv4
177  %a6 = getelementptr inbounds %struct.t4* %v4, i32 0, i32 0
178  %3 = load i32* %a6, align 1
179  %add7 = add nsw i32 %add5, %3
180  %a8 = getelementptr inbounds %struct.t5* %v5, i32 0, i32 0
181  %4 = load i32* %a8, align 1
182  %add9 = add nsw i32 %add7, %4
183  %a10 = getelementptr inbounds %struct.t6* %v6, i32 0, i32 0
184  %5 = load i32* %a10, align 1
185  %add11 = add nsw i32 %add9, %5
186  %a12 = getelementptr inbounds %struct.t7* %v7, i32 0, i32 0
187  %6 = load i32* %a12, align 1
188  %add13 = add nsw i32 %add11, %6
189  ret i32 %add13
190
191; CHECK: std 9, 96(1)
192; CHECK: std 8, 88(1)
193; CHECK: std 7, 80(1)
194; CHECK: stw 6, 76(1)
195; CHECK: std 5, 64(1)
196; CHECK: sth 4, 62(1)
197; CHECK: stb 3, 55(1)
198; CHECK: lbz {{[0-9]+}}, 85(1)
199; CHECK: lbz {{[0-9]+}}, 86(1)
200; CHECK: lbz {{[0-9]+}}, 83(1)
201; CHECK: lbz {{[0-9]+}}, 84(1)
202; CHECK: lbz {{[0-9]+}}, 69(1)
203; CHECK: lbz {{[0-9]+}}, 70(1)
204; CHECK: lha {{[0-9]+}}, 62(1)
205; CHECK: lbz {{[0-9]+}}, 55(1)
206; CHECK: lwz {{[0-9]+}}, 76(1)
207; CHECK: lhz {{[0-9]+}}, 90(1)
208; CHECK: lhz {{[0-9]+}}, 92(1)
209; CHECK: lbz {{[0-9]+}}, 99(1)
210; CHECK: lbz {{[0-9]+}}, 100(1)
211; CHECK: lbz {{[0-9]+}}, 97(1)
212; CHECK: lbz {{[0-9]+}}, 98(1)
213}
214