1; RUN: llc -verify-machineinstrs -o - %s -mtriple=aarch64-linux-gnu | FileCheck %s
2; RUN: llc -verify-machineinstrs < %s -mtriple=aarch64-none-linux-gnu -mattr=-fp-armv8 | FileCheck --check-prefix=CHECK-NOFP %s
3
4@var_8bit = global i8 0
5@var_16bit = global i16 0
6@var_32bit = global i32 0
7@var_64bit = global i64 0
8
9@var_float = global float 0.0
10@var_double = global double 0.0
11
12@varptr = global i8* null
13
14define void @ldst_8bit() {
15; CHECK-LABEL: ldst_8bit:
16
17; No architectural support for loads to 16-bit or 8-bit since we
18; promote i8 during lowering.
19  %addr_8bit = load i8*, i8** @varptr
20
21; match a sign-extending load 8-bit -> 32-bit
22   %addr_sext32 = getelementptr i8, i8* %addr_8bit, i64 -256
23   %val8_sext32 = load volatile i8, i8* %addr_sext32
24   %val32_signed = sext i8 %val8_sext32 to i32
25   store volatile i32 %val32_signed, i32* @var_32bit
26; CHECK: ldursb {{w[0-9]+}}, [{{x[0-9]+}}, #-256]
27
28; match a zero-extending load volatile 8-bit -> 32-bit
29  %addr_zext32 = getelementptr i8, i8* %addr_8bit, i64 -12
30  %val8_zext32 = load volatile i8, i8* %addr_zext32
31  %val32_unsigned = zext i8 %val8_zext32 to i32
32  store volatile i32 %val32_unsigned, i32* @var_32bit
33; CHECK: ldurb {{w[0-9]+}}, [{{x[0-9]+}}, #-12]
34
35; match an any-extending load volatile 8-bit -> 32-bit
36  %addr_anyext = getelementptr i8, i8* %addr_8bit, i64 -1
37  %val8_anyext = load volatile i8, i8* %addr_anyext
38  %newval8 = add i8 %val8_anyext, 1
39  store volatile i8 %newval8, i8* @var_8bit
40; CHECK: ldurb {{w[0-9]+}}, [{{x[0-9]+}}, #-1]
41
42; match a sign-extending load volatile 8-bit -> 64-bit
43  %addr_sext64 = getelementptr i8, i8* %addr_8bit, i64 -5
44  %val8_sext64 = load volatile i8, i8* %addr_sext64
45  %val64_signed = sext i8 %val8_sext64 to i64
46  store volatile i64 %val64_signed, i64* @var_64bit
47; CHECK: ldursb {{x[0-9]+}}, [{{x[0-9]+}}, #-5]
48
49; match a zero-extending load volatile 8-bit -> 64-bit.
50; This uses the fact that ldrb w0, [x0] will zero out the high 32-bits
51; of x0 so it's identical to load volatileing to 32-bits.
52  %addr_zext64 = getelementptr i8, i8* %addr_8bit, i64 -9
53  %val8_zext64 = load volatile i8, i8* %addr_zext64
54  %val64_unsigned = zext i8 %val8_zext64 to i64
55  store volatile i64 %val64_unsigned, i64* @var_64bit
56; CHECK: ldurb {{w[0-9]+}}, [{{x[0-9]+}}, #-9]
57
58; truncating store volatile 32-bits to 8-bits
59  %addr_trunc32 = getelementptr i8, i8* %addr_8bit, i64 -256
60  %val32 = load volatile i32, i32* @var_32bit
61  %val8_trunc32 = trunc i32 %val32 to i8
62  store volatile i8 %val8_trunc32, i8* %addr_trunc32
63; CHECK: sturb {{w[0-9]+}}, [{{x[0-9]+}}, #-256]
64
65; truncating store volatile 64-bits to 8-bits
66  %addr_trunc64 = getelementptr i8, i8* %addr_8bit, i64 -1
67  %val64 = load volatile i64, i64* @var_64bit
68  %val8_trunc64 = trunc i64 %val64 to i8
69  store volatile i8 %val8_trunc64, i8* %addr_trunc64
70; CHECK: sturb {{w[0-9]+}}, [{{x[0-9]+}}, #-1]
71
72   ret void
73}
74
75define void @ldst_16bit() {
76; CHECK-LABEL: ldst_16bit:
77
78; No architectural support for loads to 16-bit or 16-bit since we
79; promote i16 during lowering.
80  %addr_8bit = load i8*, i8** @varptr
81
82; match a sign-extending load 16-bit -> 32-bit
83   %addr8_sext32 = getelementptr i8, i8* %addr_8bit, i64 -256
84   %addr_sext32 = bitcast i8* %addr8_sext32 to i16*
85   %val16_sext32 = load volatile i16, i16* %addr_sext32
86   %val32_signed = sext i16 %val16_sext32 to i32
87   store volatile i32 %val32_signed, i32* @var_32bit
88; CHECK: ldursh {{w[0-9]+}}, [{{x[0-9]+}}, #-256]
89
90; match a zero-extending load volatile 16-bit -> 32-bit. With offset that would be unaligned.
91  %addr8_zext32 = getelementptr i8, i8* %addr_8bit, i64 15
92  %addr_zext32 = bitcast i8* %addr8_zext32 to i16*
93  %val16_zext32 = load volatile i16, i16* %addr_zext32
94  %val32_unsigned = zext i16 %val16_zext32 to i32
95  store volatile i32 %val32_unsigned, i32* @var_32bit
96; CHECK: ldurh {{w[0-9]+}}, [{{x[0-9]+}}, #15]
97
98; match an any-extending load volatile 16-bit -> 32-bit
99  %addr8_anyext = getelementptr i8, i8* %addr_8bit, i64 -1
100  %addr_anyext = bitcast i8* %addr8_anyext to i16*
101  %val16_anyext = load volatile i16, i16* %addr_anyext
102  %newval16 = add i16 %val16_anyext, 1
103  store volatile i16 %newval16, i16* @var_16bit
104; CHECK: ldurh {{w[0-9]+}}, [{{x[0-9]+}}, #-1]
105
106; match a sign-extending load volatile 16-bit -> 64-bit
107  %addr8_sext64 = getelementptr i8, i8* %addr_8bit, i64 -5
108  %addr_sext64 = bitcast i8* %addr8_sext64 to i16*
109  %val16_sext64 = load volatile i16, i16* %addr_sext64
110  %val64_signed = sext i16 %val16_sext64 to i64
111  store volatile i64 %val64_signed, i64* @var_64bit
112; CHECK: ldursh {{x[0-9]+}}, [{{x[0-9]+}}, #-5]
113
114; match a zero-extending load volatile 16-bit -> 64-bit.
115; This uses the fact that ldrb w0, [x0] will zero out the high 32-bits
116; of x0 so it's identical to load volatileing to 32-bits.
117  %addr8_zext64 = getelementptr i8, i8* %addr_8bit, i64 9
118  %addr_zext64 = bitcast i8* %addr8_zext64 to i16*
119  %val16_zext64 = load volatile i16, i16* %addr_zext64
120  %val64_unsigned = zext i16 %val16_zext64 to i64
121  store volatile i64 %val64_unsigned, i64* @var_64bit
122; CHECK: ldurh {{w[0-9]+}}, [{{x[0-9]+}}, #9]
123
124; truncating store volatile 32-bits to 16-bits
125  %addr8_trunc32 = getelementptr i8, i8* %addr_8bit, i64 -256
126  %addr_trunc32 = bitcast i8* %addr8_trunc32 to i16*
127  %val32 = load volatile i32, i32* @var_32bit
128  %val16_trunc32 = trunc i32 %val32 to i16
129  store volatile i16 %val16_trunc32, i16* %addr_trunc32
130; CHECK: sturh {{w[0-9]+}}, [{{x[0-9]+}}, #-256]
131
132; truncating store volatile 64-bits to 16-bits
133  %addr8_trunc64 = getelementptr i8, i8* %addr_8bit, i64 -1
134  %addr_trunc64 = bitcast i8* %addr8_trunc64 to i16*
135  %val64 = load volatile i64, i64* @var_64bit
136  %val16_trunc64 = trunc i64 %val64 to i16
137  store volatile i16 %val16_trunc64, i16* %addr_trunc64
138; CHECK: sturh {{w[0-9]+}}, [{{x[0-9]+}}, #-1]
139
140   ret void
141}
142
143define void @ldst_32bit() {
144; CHECK-LABEL: ldst_32bit:
145
146  %addr_8bit = load i8*, i8** @varptr
147
148; Straight 32-bit load/store
149  %addr32_8_noext = getelementptr i8, i8* %addr_8bit, i64 1
150  %addr32_noext = bitcast i8* %addr32_8_noext to i32*
151  %val32_noext = load volatile i32, i32* %addr32_noext
152  store volatile i32 %val32_noext, i32* %addr32_noext
153; CHECK: ldur {{w[0-9]+}}, [{{x[0-9]+}}, #1]
154; CHECK: stur {{w[0-9]+}}, [{{x[0-9]+}}, #1]
155
156; Zero-extension to 64-bits
157  %addr32_8_zext = getelementptr i8, i8* %addr_8bit, i64 -256
158  %addr32_zext = bitcast i8* %addr32_8_zext to i32*
159  %val32_zext = load volatile i32, i32* %addr32_zext
160  %val64_unsigned = zext i32 %val32_zext to i64
161  store volatile i64 %val64_unsigned, i64* @var_64bit
162; CHECK: ldur {{w[0-9]+}}, [{{x[0-9]+}}, #-256]
163; CHECK: str {{x[0-9]+}}, [{{x[0-9]+}}, {{#?}}:lo12:var_64bit]
164
165; Sign-extension to 64-bits
166  %addr32_8_sext = getelementptr i8, i8* %addr_8bit, i64 -12
167  %addr32_sext = bitcast i8* %addr32_8_sext to i32*
168  %val32_sext = load volatile i32, i32* %addr32_sext
169  %val64_signed = sext i32 %val32_sext to i64
170  store volatile i64 %val64_signed, i64* @var_64bit
171; CHECK: ldursw {{x[0-9]+}}, [{{x[0-9]+}}, #-12]
172; CHECK: str {{x[0-9]+}}, [{{x[0-9]+}}, {{#?}}:lo12:var_64bit]
173
174; Truncation from 64-bits
175  %addr64_8_trunc = getelementptr i8, i8* %addr_8bit, i64 255
176  %addr64_trunc = bitcast i8* %addr64_8_trunc to i64*
177  %addr32_8_trunc = getelementptr i8, i8* %addr_8bit, i64 -20
178  %addr32_trunc = bitcast i8* %addr32_8_trunc to i32*
179
180  %val64_trunc = load volatile i64, i64* %addr64_trunc
181  %val32_trunc = trunc i64 %val64_trunc to i32
182  store volatile i32 %val32_trunc, i32* %addr32_trunc
183; CHECK: ldur {{x[0-9]+}}, [{{x[0-9]+}}, #255]
184; CHECK: stur {{w[0-9]+}}, [{{x[0-9]+}}, #-20]
185
186  ret void
187}
188
189define void @ldst_float() {
190; CHECK-LABEL: ldst_float:
191
192  %addr_8bit = load i8*, i8** @varptr
193  %addrfp_8 = getelementptr i8, i8* %addr_8bit, i64 -5
194  %addrfp = bitcast i8* %addrfp_8 to float*
195
196  %valfp = load volatile float, float* %addrfp
197; CHECK: ldur {{s[0-9]+}}, [{{x[0-9]+}}, #-5]
198; CHECK-NOFP-NOT: ldur {{s[0-9]+}},
199
200  store volatile float %valfp, float* %addrfp
201; CHECK: stur {{s[0-9]+}}, [{{x[0-9]+}}, #-5]
202; CHECK-NOFP-NOT: stur {{s[0-9]+}},
203
204  ret void
205}
206
207define void @ldst_double() {
208; CHECK-LABEL: ldst_double:
209
210  %addr_8bit = load i8*, i8** @varptr
211  %addrfp_8 = getelementptr i8, i8* %addr_8bit, i64 4
212  %addrfp = bitcast i8* %addrfp_8 to double*
213
214  %valfp = load volatile double, double* %addrfp
215; CHECK: ldur {{d[0-9]+}}, [{{x[0-9]+}}, #4]
216; CHECK-NOFP-NOT: ldur {{d[0-9]+}},
217
218  store volatile double %valfp, double* %addrfp
219; CHECK: stur {{d[0-9]+}}, [{{x[0-9]+}}, #4]
220; CHECK-NOFP-NOT: stur {{d[0-9]+}},
221
222   ret void
223}
224