• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc -verify-machineinstrs < %s -mtriple=aarch64-none-linux-gnu | FileCheck %s
2
3%myStruct = type { i64 , i8, i32 }
4
5@var8 = global i8 0
6@var32 = global i32 0
7@var64 = global i64 0
8@var128 = global i128 0
9@varfloat = global float 0.0
10@vardouble = global double 0.0
11@varstruct = global %myStruct zeroinitializer
12
13define void @take_i8s(i8 %val1, i8 %val2) {
14; CHECK: take_i8s:
15    store i8 %val2, i8* @var8
16    ; Not using w1 may be technically allowed, but it would indicate a
17    ; problem in itself.
18;  CHECK: strb w1, [{{x[0-9]+}}, #:lo12:var8]
19    ret void
20}
21
22define void @add_floats(float %val1, float %val2) {
23; CHECK: add_floats:
24    %newval = fadd float %val1, %val2
25; CHECK: fadd [[ADDRES:s[0-9]+]], s0, s1
26    store float %newval, float* @varfloat
27; CHECK: str [[ADDRES]], [{{x[0-9]+}}, #:lo12:varfloat]
28    ret void
29}
30
31; byval pointers should be allocated to the stack and copied as if
32; with memcpy.
33define void @take_struct(%myStruct* byval %structval) {
34; CHECK: take_struct:
35    %addr0 = getelementptr %myStruct* %structval, i64 0, i32 2
36    %addr1 = getelementptr %myStruct* %structval, i64 0, i32 0
37
38    %val0 = load i32* %addr0
39    ; Some weird move means x0 is used for one access
40; CHECK: ldr [[REG32:w[0-9]+]], [{{x[0-9]+|sp}}, #12]
41    store i32 %val0, i32* @var32
42; CHECK: str [[REG32]], [{{x[0-9]+}}, #:lo12:var32]
43
44    %val1 = load i64* %addr1
45; CHECK: ldr [[REG64:x[0-9]+]], [{{x[0-9]+|sp}}]
46    store i64 %val1, i64* @var64
47; CHECK: str [[REG64]], [{{x[0-9]+}}, #:lo12:var64]
48
49    ret void
50}
51
52; %structval should be at sp + 16
53define void @check_byval_align(i32* byval %ignore, %myStruct* byval align 16 %structval) {
54; CHECK: check_byval_align:
55
56    %addr0 = getelementptr %myStruct* %structval, i64 0, i32 2
57    %addr1 = getelementptr %myStruct* %structval, i64 0, i32 0
58
59    %val0 = load i32* %addr0
60    ; Some weird move means x0 is used for one access
61; CHECK: add x[[STRUCTVAL_ADDR:[0-9]+]], sp, #16
62; CHECK: ldr [[REG32:w[0-9]+]], [x[[STRUCTVAL_ADDR]], #12]
63    store i32 %val0, i32* @var32
64; CHECK: str [[REG32]], [{{x[0-9]+}}, #:lo12:var32]
65
66    %val1 = load i64* %addr1
67; CHECK: ldr [[REG64:x[0-9]+]], [sp, #16]
68    store i64 %val1, i64* @var64
69; CHECK: str [[REG64]], [{{x[0-9]+}}, #:lo12:var64]
70
71    ret void
72}
73
74define i32 @return_int() {
75; CHECK: return_int:
76    %val = load i32* @var32
77    ret i32 %val
78; CHECK: ldr w0, [{{x[0-9]+}}, #:lo12:var32]
79    ; Make sure epilogue follows
80; CHECK-NEXT: ret
81}
82
83define double @return_double() {
84; CHECK: return_double:
85    ret double 3.14
86; CHECK: ldr d0, [{{x[0-9]+}}, #:lo12:.LCPI
87}
88
89; This is the kind of IR clang will produce for returning a struct
90; small enough to go into registers. Not all that pretty, but it
91; works.
92define [2 x i64] @return_struct() {
93; CHECK: return_struct:
94    %addr = bitcast %myStruct* @varstruct to [2 x i64]*
95    %val = load [2 x i64]* %addr
96    ret [2 x i64] %val
97; CHECK: ldr x0, [{{x[0-9]+}}, #:lo12:varstruct]
98    ; Odd register regex below disallows x0 which we want to be live now.
99; CHECK: add {{x[1-9][0-9]*}}, {{x[1-9][0-9]*}}, #:lo12:varstruct
100; CHECK-NEXT: ldr x1, [{{x[1-9][0-9]*}}, #8]
101    ; Make sure epilogue immediately follows
102; CHECK-NEXT: ret
103}
104
105; Large structs are passed by reference (storage allocated by caller
106; to preserve value semantics) in x8. Strictly this only applies to
107; structs larger than 16 bytes, but C semantics can still be provided
108; if LLVM does it to %myStruct too. So this is the simplest check
109define void @return_large_struct(%myStruct* sret %retval) {
110; CHECK: return_large_struct:
111    %addr0 = getelementptr %myStruct* %retval, i64 0, i32 0
112    %addr1 = getelementptr %myStruct* %retval, i64 0, i32 1
113    %addr2 = getelementptr %myStruct* %retval, i64 0, i32 2
114
115    store i64 42, i64* %addr0
116    store i8 2, i8* %addr1
117    store i32 9, i32* %addr2
118; CHECK: str {{x[0-9]+}}, [x8]
119; CHECK: strb {{w[0-9]+}}, [x8, #8]
120; CHECK: str {{w[0-9]+}}, [x8, #12]
121
122    ret void
123}
124
125; This struct is just too far along to go into registers: (only x7 is
126; available, but it needs two). Also make sure that %stacked doesn't
127; sneak into x7 behind.
128define i32 @struct_on_stack(i8 %var0, i16 %var1, i32 %var2, i64 %var3, i128 %var45,
129                          i32* %var6, %myStruct* byval %struct, i32* byval %stacked,
130                          double %notstacked) {
131; CHECK: struct_on_stack:
132    %addr = getelementptr %myStruct* %struct, i64 0, i32 0
133    %val64 = load i64* %addr
134    store i64 %val64, i64* @var64
135    ; Currently nothing on local stack, so struct should be at sp
136; CHECK: ldr [[VAL64:x[0-9]+]], [sp]
137; CHECK: str [[VAL64]], [{{x[0-9]+}}, #:lo12:var64]
138
139    store double %notstacked, double* @vardouble
140; CHECK-NOT: ldr d0
141; CHECK: str d0, [{{x[0-9]+}}, #:lo12:vardouble
142
143    %retval = load i32* %stacked
144    ret i32 %retval
145; CHECK: ldr w0, [sp, #16]
146}
147
148define void @stacked_fpu(float %var0, double %var1, float %var2, float %var3,
149                         float %var4, float %var5, float %var6, float %var7,
150                         float %var8) {
151; CHECK: stacked_fpu:
152    store float %var8, float* @varfloat
153    ; Beware as above: the offset would be different on big-endian
154    ; machines if the first ldr were changed to use s-registers.
155; CHECK: ldr d[[VALFLOAT:[0-9]+]], [sp]
156; CHECK: str s[[VALFLOAT]], [{{x[0-9]+}}, #:lo12:varfloat]
157
158    ret void
159}
160
161; 128-bit integer types should be passed in xEVEN, xODD rather than
162; the reverse. In this case x2 and x3. Nothing should use x1.
163define i32 @check_i128_regalign(i32 %val0, i128 %val1, i32 %val2) {
164; CHECK: check_i128_regalign
165    store i128 %val1, i128* @var128
166; CHECK: str x2, [{{x[0-9]+}}, #:lo12:var128]
167; CHECK: str x3, [{{x[0-9]+}}, #8]
168
169    ret i32 %val2
170; CHECK: mov x0, x4
171}
172
173define void @check_i128_stackalign(i32 %val0, i32 %val1, i32 %val2, i32 %val3,
174                                   i32 %val4, i32 %val5, i32 %val6, i32 %val7,
175                                   i32 %stack1, i128 %stack2) {
176; CHECK: check_i128_stackalign
177    store i128 %stack2, i128* @var128
178    ; Nothing local on stack in current codegen, so first stack is 16 away
179; CHECK: ldr {{x[0-9]+}}, [sp, #16]
180    ; Important point is that we address sp+24 for second dword
181; CHECK: add     [[REG:x[0-9]+]], sp, #16
182; CHECK: ldr     {{x[0-9]+}}, {{\[}}[[REG]], #8]
183    ret void
184}
185
186declare void @llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
187
188define i32 @test_extern() {
189; CHECK: test_extern:
190  call void @llvm.memcpy.p0i8.p0i8.i32(i8* undef, i8* undef, i32 undef, i32 4, i1 0)
191; CHECK: bl memcpy
192  ret i32 0
193}
194