xref: /llvm-project/llvm/test/CodeGen/SystemZ/vec-bswap-07.ll (revision 0a76f7d9d8c1fc693568ed26420c47d92a6ba0e7) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
2; Test replications of a byte-swapped scalar memory value.
3;
4; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z15 | FileCheck %s
5
6declare i16 @llvm.bswap.i16(i16)
7declare i32 @llvm.bswap.i32(i32)
8declare i64 @llvm.bswap.i64(i64)
9declare <8 x i16> @llvm.bswap.v8i16(<8 x i16>)
10declare <4 x i32> @llvm.bswap.v4i32(<4 x i32>)
11declare <2 x i64> @llvm.bswap.v2i64(<2 x i64>)
12
13; Test a v8i16 replicating load with no offset.
14define <8 x i16> @f1(ptr %ptr) {
15; CHECK-LABEL: f1:
16; CHECK:       # %bb.0:
17; CHECK-NEXT:    vlbrreph %v24, 0(%r2)
18; CHECK-NEXT:    br %r14
19  %scalar = load i16, ptr %ptr
20  %swap = call i16 @llvm.bswap.i16(i16 %scalar)
21  %val = insertelement <8 x i16> undef, i16 %swap, i32 0
22  %ret = shufflevector <8 x i16> %val, <8 x i16> undef,
23                       <8 x i32> zeroinitializer
24  ret <8 x i16> %ret
25}
26
27; Test a v8i16 replicating load with the maximum in-range offset.
28define <8 x i16> @f2(ptr %base) {
29; CHECK-LABEL: f2:
30; CHECK:       # %bb.0:
31; CHECK-NEXT:    vlbrreph %v24, 4094(%r2)
32; CHECK-NEXT:    br %r14
33  %ptr = getelementptr i16, ptr %base, i64 2047
34  %scalar = load i16, ptr %ptr
35  %swap = call i16 @llvm.bswap.i16(i16 %scalar)
36  %val = insertelement <8 x i16> undef, i16 %swap, i32 0
37  %ret = shufflevector <8 x i16> %val, <8 x i16> undef,
38                       <8 x i32> zeroinitializer
39  ret <8 x i16> %ret
40}
41
42; Test a v8i16 replicating load with the first out-of-range offset.
43define <8 x i16> @f3(ptr %base) {
44; CHECK-LABEL: f3:
45; CHECK:       # %bb.0:
46; CHECK-NEXT:    aghi %r2, 4096
47; CHECK-NEXT:    vlbrreph %v24, 0(%r2)
48; CHECK-NEXT:    br %r14
49  %ptr = getelementptr i16, ptr %base, i64 2048
50  %scalar = load i16, ptr %ptr
51  %swap = call i16 @llvm.bswap.i16(i16 %scalar)
52  %val = insertelement <8 x i16> undef, i16 %swap, i32 0
53  %ret = shufflevector <8 x i16> %val, <8 x i16> undef,
54                       <8 x i32> zeroinitializer
55  ret <8 x i16> %ret
56}
57
58; Test a v8i16 replicating load using a vector bswap.
59define <8 x i16> @f4(ptr %ptr) {
60; CHECK-LABEL: f4:
61; CHECK:       # %bb.0:
62; CHECK-NEXT:    vlbrreph %v24, 0(%r2)
63; CHECK-NEXT:    br %r14
64  %scalar = load i16, ptr %ptr
65  %val = insertelement <8 x i16> undef, i16 %scalar, i32 0
66  %rep = shufflevector <8 x i16> %val, <8 x i16> undef,
67                       <8 x i32> zeroinitializer
68  %ret = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %rep)
69  ret <8 x i16> %ret
70}
71
72; Test a v4i32 replicating load with no offset.
73define <4 x i32> @f5(ptr %ptr) {
74; CHECK-LABEL: f5:
75; CHECK:       # %bb.0:
76; CHECK-NEXT:    vlbrrepf %v24, 0(%r2)
77; CHECK-NEXT:    br %r14
78  %scalar = load i32, ptr %ptr
79  %swap = call i32 @llvm.bswap.i32(i32 %scalar)
80  %val = insertelement <4 x i32> undef, i32 %swap, i32 0
81  %ret = shufflevector <4 x i32> %val, <4 x i32> undef,
82                       <4 x i32> zeroinitializer
83  ret <4 x i32> %ret
84}
85
86; Test a v4i32 replicating load with the maximum in-range offset.
87define <4 x i32> @f6(ptr %base) {
88; CHECK-LABEL: f6:
89; CHECK:       # %bb.0:
90; CHECK-NEXT:    vlbrrepf %v24, 4092(%r2)
91; CHECK-NEXT:    br %r14
92  %ptr = getelementptr i32, ptr %base, i64 1023
93  %scalar = load i32, ptr %ptr
94  %swap = call i32 @llvm.bswap.i32(i32 %scalar)
95  %val = insertelement <4 x i32> undef, i32 %swap, i32 0
96  %ret = shufflevector <4 x i32> %val, <4 x i32> undef,
97                       <4 x i32> zeroinitializer
98  ret <4 x i32> %ret
99}
100
101; Test a v4i32 replicating load with the first out-of-range offset.
102define <4 x i32> @f7(ptr %base) {
103; CHECK-LABEL: f7:
104; CHECK:       # %bb.0:
105; CHECK-NEXT:    aghi %r2, 4096
106; CHECK-NEXT:    vlbrrepf %v24, 0(%r2)
107; CHECK-NEXT:    br %r14
108  %ptr = getelementptr i32, ptr %base, i64 1024
109  %scalar = load i32, ptr %ptr
110  %swap = call i32 @llvm.bswap.i32(i32 %scalar)
111  %val = insertelement <4 x i32> undef, i32 %swap, i32 0
112  %ret = shufflevector <4 x i32> %val, <4 x i32> undef,
113                       <4 x i32> zeroinitializer
114  ret <4 x i32> %ret
115}
116
117; Test a v4i32 replicating load using a vector bswap.
118define <4 x i32> @f8(ptr %ptr) {
119; CHECK-LABEL: f8:
120; CHECK:       # %bb.0:
121; CHECK-NEXT:    vlbrrepf %v24, 0(%r2)
122; CHECK-NEXT:    br %r14
123  %scalar = load i32, ptr %ptr
124  %val = insertelement <4 x i32> undef, i32 %scalar, i32 0
125  %rep = shufflevector <4 x i32> %val, <4 x i32> undef,
126                       <4 x i32> zeroinitializer
127  %ret = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %rep)
128  ret <4 x i32> %ret
129}
130
131; Test a v2i64 replicating load with no offset.
132define <2 x i64> @f9(ptr %ptr) {
133; CHECK-LABEL: f9:
134; CHECK:       # %bb.0:
135; CHECK-NEXT:    vlbrrepg %v24, 0(%r2)
136; CHECK-NEXT:    br %r14
137  %scalar = load i64, ptr %ptr
138  %swap = call i64 @llvm.bswap.i64(i64 %scalar)
139  %val = insertelement <2 x i64> undef, i64 %swap, i32 0
140  %ret = shufflevector <2 x i64> %val, <2 x i64> undef,
141                       <2 x i32> zeroinitializer
142  ret <2 x i64> %ret
143}
144
145; Test a v2i64 replicating load with the maximum in-range offset.
146define <2 x i64> @f10(ptr %base) {
147; CHECK-LABEL: f10:
148; CHECK:       # %bb.0:
149; CHECK-NEXT:    vlbrrepg %v24, 4088(%r2)
150; CHECK-NEXT:    br %r14
151  %ptr = getelementptr i64, ptr %base, i32 511
152  %scalar = load i64, ptr %ptr
153  %swap = call i64 @llvm.bswap.i64(i64 %scalar)
154  %val = insertelement <2 x i64> undef, i64 %swap, i32 0
155  %ret = shufflevector <2 x i64> %val, <2 x i64> undef,
156                       <2 x i32> zeroinitializer
157  ret <2 x i64> %ret
158}
159
160; Test a v2i64 replicating load with the first out-of-range offset.
161define <2 x i64> @f11(ptr %base) {
162; CHECK-LABEL: f11:
163; CHECK:       # %bb.0:
164; CHECK-NEXT:    aghi %r2, 4096
165; CHECK-NEXT:    vlbrrepg %v24, 0(%r2)
166; CHECK-NEXT:    br %r14
167  %ptr = getelementptr i64, ptr %base, i32 512
168  %scalar = load i64, ptr %ptr
169  %swap = call i64 @llvm.bswap.i64(i64 %scalar)
170  %val = insertelement <2 x i64> undef, i64 %swap, i32 0
171  %ret = shufflevector <2 x i64> %val, <2 x i64> undef,
172                       <2 x i32> zeroinitializer
173  ret <2 x i64> %ret
174}
175
176; Test a v2i64 replicating load using a vector bswap.
177define <2 x i64> @f12(ptr %ptr) {
178; CHECK-LABEL: f12:
179; CHECK:       # %bb.0:
180; CHECK-NEXT:    vlbrrepg %v24, 0(%r2)
181; CHECK-NEXT:    br %r14
182  %scalar = load i64, ptr %ptr
183  %val = insertelement <2 x i64> undef, i64 %scalar, i32 0
184  %rep = shufflevector <2 x i64> %val, <2 x i64> undef,
185                       <2 x i32> zeroinitializer
186  %ret = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %rep)
187  ret <2 x i64> %ret
188}
189
190; Test a v8i16 replicating load with an index.
191define <8 x i16> @f13(ptr %base, i64 %index) {
192; CHECK-LABEL: f13:
193; CHECK:       # %bb.0:
194; CHECK-NEXT:    sllg %r1, %r3, 1
195; CHECK-NEXT:    vlbrreph %v24, 2046(%r1,%r2)
196; CHECK-NEXT:    br %r14
197  %ptr1 = getelementptr i16, ptr %base, i64 %index
198  %ptr = getelementptr i16, ptr %ptr1, i64 1023
199  %scalar = load i16, ptr %ptr
200  %swap = call i16 @llvm.bswap.i16(i16 %scalar)
201  %val = insertelement <8 x i16> undef, i16 %swap, i32 0
202  %ret = shufflevector <8 x i16> %val, <8 x i16> undef,
203                       <8 x i32> zeroinitializer
204  ret <8 x i16> %ret
205}
206
207