1 ; RUN: llc -verify-machineinstrs < %s -mtriple=aarch64-none-linux-gnu | FileCheck --check-prefix=CHECK %s 2 ; RUN: llc -verify-machineinstrs < %s -mtriple=aarch64-none-linux-gnu -mattr=-fp-armv8 | FileCheck --check-prefix=CHECK-NOFP %s 3 4 %myStruct = type { i64 , i8, i32 } 5 6 @var8 = global i8 0 7 @var32 = global i32 0 8 @var64 = global i64 0 9 @var128 = global i128 0 10 @varfloat = global float 0.0 11 @vardouble = global double 0.0 12 @varstruct = global %myStruct zeroinitializer 13 14 define void @take_i8s(i8 %val1, i8 %val2) { 15 ; CHECK-LABEL: take_i8s: 16 store i8 %val2, i8* @var8 17 ; Not using w1 may be technically allowed, but it would indicate a 18 ; problem in itself. 19 ; CHECK: strb w1, [{{x[0-9]+}}, {{#?}}:lo12:var8] 20 ret void 21 } 22 23 define void @add_floats(float %val1, float %val2) { 24 ; CHECK-LABEL: add_floats: 25 %newval = fadd float %val1, %val2 26 ; CHECK: fadd [[ADDRES:s[0-9]+]], s0, s1 27 ; CHECK-NOFP-NOT: fadd 28 store float %newval, float* @varfloat 29 ; CHECK: str [[ADDRES]], [{{x[0-9]+}}, {{#?}}:lo12:varfloat] 30 ret void 31 } 32 33 ; byval pointers should be allocated to the stack and copied as if 34 ; with memcpy. 35 define void @take_struct(%myStruct* byval %structval) { 36 ; CHECK-LABEL: take_struct: 37 %addr0 = getelementptr %myStruct* %structval, i64 0, i32 2 38 %addr1 = getelementptr %myStruct* %structval, i64 0, i32 0 39 40 %val0 = load volatile i32* %addr0 41 ; Some weird move means x0 is used for one access 42 ; CHECK: ldr [[REG32:w[0-9]+]], [{{x[0-9]+|sp}}, #12] 43 store volatile i32 %val0, i32* @var32 44 ; CHECK: str [[REG32]], [{{x[0-9]+}}, {{#?}}:lo12:var32] 45 46 %val1 = load volatile i64* %addr1 47 ; CHECK: ldr [[REG64:x[0-9]+]], [{{x[0-9]+|sp}}] 48 store volatile i64 %val1, i64* @var64 49 ; CHECK: str [[REG64]], [{{x[0-9]+}}, {{#?}}:lo12:var64] 50 51 ret void 52 } 53 54 ; %structval should be at sp + 16 55 define void @check_byval_align(i32* byval %ignore, %myStruct* byval align 16 %structval) { 56 ; CHECK-LABEL: check_byval_align: 57 58 %addr0 = getelementptr %myStruct* %structval, i64 0, i32 2 59 %addr1 = getelementptr %myStruct* %structval, i64 0, i32 0 60 61 %val0 = load volatile i32* %addr0 62 ; Some weird move means x0 is used for one access 63 ; CHECK: ldr [[REG32:w[0-9]+]], [sp, #28] 64 store i32 %val0, i32* @var32 65 ; CHECK: str [[REG32]], [{{x[0-9]+}}, {{#?}}:lo12:var32] 66 67 %val1 = load volatile i64* %addr1 68 ; CHECK: ldr [[REG64:x[0-9]+]], [sp, #16] 69 store i64 %val1, i64* @var64 70 ; CHECK: str [[REG64]], [{{x[0-9]+}}, {{#?}}:lo12:var64] 71 72 ret void 73 } 74 75 define i32 @return_int() { 76 ; CHECK-LABEL: return_int: 77 %val = load i32* @var32 78 ret i32 %val 79 ; CHECK: ldr w0, [{{x[0-9]+}}, {{#?}}:lo12:var32] 80 ; Make sure epilogue follows 81 ; CHECK-NEXT: ret 82 } 83 84 define double @return_double() { 85 ; CHECK-LABEL: return_double: 86 ret double 3.14 87 ; CHECK: ldr d0, [{{x[0-9]+}}, {{#?}}:lo12:.LCPI 88 ; CHECK-NOFP-NOT: ldr d0, 89 } 90 91 ; This is the kind of IR clang will produce for returning a struct 92 ; small enough to go into registers. Not all that pretty, but it 93 ; works. 94 define [2 x i64] @return_struct() { 95 ; CHECK-LABEL: return_struct: 96 %addr = bitcast %myStruct* @varstruct to [2 x i64]* 97 %val = load [2 x i64]* %addr 98 ret [2 x i64] %val 99 ; CHECK-DAG: ldr x0, [{{x[0-9]+}}, {{#?}}:lo12:varstruct] 100 ; Odd register regex below disallows x0 which we want to be live now. 101 ; CHECK-DAG: add {{x[1-9][0-9]*}}, {{x[1-9][0-9]*}}, {{#?}}:lo12:varstruct 102 ; CHECK: ldr x1, [{{x[1-9][0-9]*}}, #8] 103 ; Make sure epilogue immediately follows 104 ; CHECK-NEXT: ret 105 } 106 107 ; Large structs are passed by reference (storage allocated by caller 108 ; to preserve value semantics) in x8. Strictly this only applies to 109 ; structs larger than 16 bytes, but C semantics can still be provided 110 ; if LLVM does it to %myStruct too. So this is the simplest check 111 define void @return_large_struct(%myStruct* sret %retval) { 112 ; CHECK-LABEL: return_large_struct: 113 %addr0 = getelementptr %myStruct* %retval, i64 0, i32 0 114 %addr1 = getelementptr %myStruct* %retval, i64 0, i32 1 115 %addr2 = getelementptr %myStruct* %retval, i64 0, i32 2 116 117 store i64 42, i64* %addr0 118 store i8 2, i8* %addr1 119 store i32 9, i32* %addr2 120 ; CHECK: str {{x[0-9]+}}, [x8] 121 ; CHECK: strb {{w[0-9]+}}, [x8, #8] 122 ; CHECK: str {{w[0-9]+}}, [x8, #12] 123 124 ret void 125 } 126 127 ; This struct is just too far along to go into registers: (only x7 is 128 ; available, but it needs two). Also make sure that %stacked doesn't 129 ; sneak into x7 behind. 130 define i32 @struct_on_stack(i8 %var0, i16 %var1, i32 %var2, i64 %var3, i128 %var45, 131 i32* %var6, %myStruct* byval %struct, i32* byval %stacked, 132 double %notstacked) { 133 ; CHECK-LABEL: struct_on_stack: 134 %addr = getelementptr %myStruct* %struct, i64 0, i32 0 135 %val64 = load volatile i64* %addr 136 store volatile i64 %val64, i64* @var64 137 ; Currently nothing on local stack, so struct should be at sp 138 ; CHECK: ldr [[VAL64:x[0-9]+]], [sp] 139 ; CHECK: str [[VAL64]], [{{x[0-9]+}}, {{#?}}:lo12:var64] 140 141 store volatile double %notstacked, double* @vardouble 142 ; CHECK-NOT: ldr d0 143 ; CHECK: str d0, [{{x[0-9]+}}, {{#?}}:lo12:vardouble 144 ; CHECK-NOFP-NOT: str d0, 145 146 %retval = load volatile i32* %stacked 147 ret i32 %retval 148 ; CHECK-LE: ldr w0, [sp, #16] 149 } 150 151 define void @stacked_fpu(float %var0, double %var1, float %var2, float %var3, 152 float %var4, float %var5, float %var6, float %var7, 153 float %var8) { 154 ; CHECK-LABEL: stacked_fpu: 155 store float %var8, float* @varfloat 156 ; Beware as above: the offset would be different on big-endian 157 ; machines if the first ldr were changed to use s-registers. 158 ; CHECK: ldr {{[ds]}}[[VALFLOAT:[0-9]+]], [sp] 159 ; CHECK: str s[[VALFLOAT]], [{{x[0-9]+}}, {{#?}}:lo12:varfloat] 160 161 ret void 162 } 163 164 ; 128-bit integer types should be passed in xEVEN, xODD rather than 165 ; the reverse. In this case x2 and x3. Nothing should use x1. 166 define i64 @check_i128_regalign(i32 %val0, i128 %val1, i64 %val2) { 167 ; CHECK-LABEL: check_i128_regalign 168 store i128 %val1, i128* @var128 169 ; CHECK-DAG: str x2, [{{x[0-9]+}}, {{#?}}:lo12:var128] 170 ; CHECK-DAG: str x3, [{{x[0-9]+}}, #8] 171 172 ret i64 %val2 173 ; CHECK: mov x0, x4 174 } 175 176 define void @check_i128_stackalign(i32 %val0, i32 %val1, i32 %val2, i32 %val3, 177 i32 %val4, i32 %val5, i32 %val6, i32 %val7, 178 i32 %stack1, i128 %stack2) { 179 ; CHECK-LABEL: check_i128_stackalign 180 store i128 %stack2, i128* @var128 181 ; Nothing local on stack in current codegen, so first stack is 16 away 182 ; CHECK-LE: add x[[REG:[0-9]+]], sp, #16 183 ; CHECK-LE: ldr {{x[0-9]+}}, [x[[REG]], #8] 184 185 ; Important point is that we address sp+24 for second dword 186 187 ; CHECK: ldp {{x[0-9]+}}, {{x[0-9]+}}, [sp, #16] 188 ret void 189 } 190 191 declare void @llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1) 192 193 define i32 @test_extern() { 194 ; CHECK-LABEL: test_extern: 195 call void @llvm.memcpy.p0i8.p0i8.i32(i8* undef, i8* undef, i32 undef, i32 4, i1 0) 196 ; CHECK: bl memcpy 197 ret i32 0 198 } 199 200 201 ; A sub-i32 stack argument must be loaded on big endian with ldr{h,b}, not just 202 ; implicitly extended to a 32-bit load. 203 define i16 @stacked_i16(i32 %val0, i32 %val1, i32 %val2, i32 %val3, 204 i32 %val4, i32 %val5, i32 %val6, i32 %val7, 205 i16 %stack1) { 206 ; CHECK-LABEL: stacked_i16 207 ret i16 %stack1 208 } 209
