1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji < %s | FileCheck %s 2 3@private = private addrspace(1) global [256 x i32] zeroinitializer 4@internal = internal addrspace(1) global [256 x i32] zeroinitializer 5@available_externally = available_externally addrspace(1) global [256 x i32] zeroinitializer 6@linkonce = linkonce addrspace(1) global [256 x i32] zeroinitializer 7@weak= weak addrspace(1) global [256 x i32] zeroinitializer 8@common = common addrspace(1) global [256 x i32] zeroinitializer 9@extern_weak = extern_weak addrspace(1) global [256 x i32] 10@linkonce_odr = linkonce_odr addrspace(1) global [256 x i32] zeroinitializer 11@weak_odr = weak_odr addrspace(1) global [256 x i32] zeroinitializer 12@external = external addrspace(1) global [256 x i32] 13@external_w_init = addrspace(1) global [256 x i32] zeroinitializer 14 15; CHECK-LABEL: private_test: 16; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 17; CHECK: s_add_u32 s[[ADDR_LO:[0-9]+]], s[[PC_LO]], private@rel32@lo+8 18; CHECK: s_addc_u32 s[[ADDR_HI:[0-9]+]], s[[PC_HI]], private@rel32@hi+8 19; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[ADDR_LO]] 20; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[ADDR_HI]] 21; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 22define amdgpu_kernel void @private_test(i32 addrspace(1)* %out) { 23 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @private, i32 0, i32 1 24 %val = load i32, i32 addrspace(1)* %ptr 25 store i32 %val, i32 addrspace(1)* %out 26 ret void 27} 28 29; CHECK-LABEL: internal_test: 30; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 31; CHECK: s_add_u32 s[[ADDR_LO:[0-9]+]], s[[PC_LO]], internal@rel32@lo+8 32; CHECK: s_addc_u32 s[[ADDR_HI:[0-9]+]], s[[PC_HI]], internal@rel32@hi+8 33; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[ADDR_LO]] 34; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[ADDR_HI]] 35; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 36define amdgpu_kernel void @internal_test(i32 addrspace(1)* %out) { 37 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @internal, i32 0, i32 1 38 %val = load i32, i32 addrspace(1)* %ptr 39 store i32 %val, i32 addrspace(1)* %out 40 ret void 41} 42 43; CHECK-LABEL: available_externally_test: 44; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 45; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], available_externally@gotpcrel32@lo+4 46; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], available_externally@gotpcrel32@hi+4 47; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 48; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 49; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 50; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 51; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 52; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 53define amdgpu_kernel void @available_externally_test(i32 addrspace(1)* %out) { 54 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @available_externally, i32 0, i32 1 55 %val = load i32, i32 addrspace(1)* %ptr 56 store i32 %val, i32 addrspace(1)* %out 57 ret void 58} 59 60; CHECK-LABEL: linkonce_test: 61; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 62; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], linkonce@gotpcrel32@lo+4 63; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], linkonce@gotpcrel32@hi+4 64; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 65; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 66; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 67; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 68; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 69; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 70define amdgpu_kernel void @linkonce_test(i32 addrspace(1)* %out) { 71 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @linkonce, i32 0, i32 1 72 %val = load i32, i32 addrspace(1)* %ptr 73 store i32 %val, i32 addrspace(1)* %out 74 ret void 75} 76 77; CHECK-LABEL: weak_test: 78; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 79; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], weak@gotpcrel32@lo+4 80; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], weak@gotpcrel32@hi+4 81; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 82; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 83; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 84; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 85; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 86; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 87define amdgpu_kernel void @weak_test(i32 addrspace(1)* %out) { 88 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @weak, i32 0, i32 1 89 %val = load i32, i32 addrspace(1)* %ptr 90 store i32 %val, i32 addrspace(1)* %out 91 ret void 92} 93 94; CHECK-LABEL: common_test: 95; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 96; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], common@gotpcrel32@lo+4 97; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], common@gotpcrel32@hi+4 98; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 99; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 100; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 101; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 102; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 103; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 104define amdgpu_kernel void @common_test(i32 addrspace(1)* %out) { 105 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @common, i32 0, i32 1 106 %val = load i32, i32 addrspace(1)* %ptr 107 store i32 %val, i32 addrspace(1)* %out 108 ret void 109} 110 111; CHECK-LABEL: extern_weak_test: 112; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 113; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], extern_weak@gotpcrel32@lo+4 114; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], extern_weak@gotpcrel32@hi+4 115; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 116; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 117; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 118; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 119; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 120; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 121define amdgpu_kernel void @extern_weak_test(i32 addrspace(1)* %out) { 122 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @extern_weak, i32 0, i32 1 123 %val = load i32, i32 addrspace(1)* %ptr 124 store i32 %val, i32 addrspace(1)* %out 125 ret void 126} 127 128; CHECK-LABEL: linkonce_odr_test: 129; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 130; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], linkonce_odr@gotpcrel32@lo+4 131; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], linkonce_odr@gotpcrel32@hi+4 132; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 133; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 134; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 135; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 136; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 137; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 138define amdgpu_kernel void @linkonce_odr_test(i32 addrspace(1)* %out) { 139 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @linkonce_odr, i32 0, i32 1 140 %val = load i32, i32 addrspace(1)* %ptr 141 store i32 %val, i32 addrspace(1)* %out 142 ret void 143} 144 145; CHECK-LABEL: weak_odr_test: 146; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 147; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], weak_odr@gotpcrel32@lo+4 148; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], weak_odr@gotpcrel32@hi+4 149; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 150; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 151; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 152; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 153; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 154; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 155define amdgpu_kernel void @weak_odr_test(i32 addrspace(1)* %out) { 156 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @weak_odr, i32 0, i32 1 157 %val = load i32, i32 addrspace(1)* %ptr 158 store i32 %val, i32 addrspace(1)* %out 159 ret void 160} 161 162; CHECK-LABEL: external_test: 163; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 164; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], external@gotpcrel32@lo+4 165; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], external@gotpcrel32@hi+4 166; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 167; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 168; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 169; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 170; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 171; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 172define amdgpu_kernel void @external_test(i32 addrspace(1)* %out) { 173 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @external, i32 0, i32 1 174 %val = load i32, i32 addrspace(1)* %ptr 175 store i32 %val, i32 addrspace(1)* %out 176 ret void 177} 178 179; CHECK-LABEL: external_w_init_test: 180; CHECK: s_getpc_b64 s{{\[}}[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]{{\]}} 181; CHECK: s_add_u32 s[[GOTADDR_LO:[0-9]+]], s[[PC_LO]], external_w_init@gotpcrel32@lo+4 182; CHECK: s_addc_u32 s[[GOTADDR_HI:[0-9]+]], s[[PC_HI]], external_w_init@gotpcrel32@hi+4 183; CHECK: s_load_dwordx2 s{{\[}}[[ADDR_LO:[0-9]+]]:[[ADDR_HI:[0-9]+]]{{\]}}, s{{\[}}[[GOTADDR_LO]]:[[GOTADDR_HI]]{{\]}}, 0x0 184; CHECK: s_add_u32 s[[GEP_LO:[0-9]+]], s[[ADDR_LO]], 4 185; CHECK: s_addc_u32 s[[GEP_HI:[0-9]+]], s[[ADDR_HI]], 0 186; CHECK-DAG: v_mov_b32_e32 v[[V_LO:[0-9]+]], s[[GEP_LO]] 187; CHECK-DAG: v_mov_b32_e32 v[[V_HI:[0-9]+]], s[[GEP_HI]] 188; CHECK: flat_load_dword v{{[0-9]+}}, v{{\[}}[[V_LO]]:[[V_HI]]{{\]}} 189define amdgpu_kernel void @external_w_init_test(i32 addrspace(1)* %out) { 190 %ptr = getelementptr [256 x i32], [256 x i32] addrspace(1)* @external_w_init, i32 0, i32 1 191 %val = load i32, i32 addrspace(1)* %ptr 192 store i32 %val, i32 addrspace(1)* %out 193 ret void 194} 195 196; CHECK: .local private 197; CHECK: .local internal 198; CHECK: .weak linkonce 199; CHECK: .weak weak 200; CHECK: .weak linkonce_odr 201; CHECK: .weak weak_odr 202; CHECK-NOT: external{{$}} 203; CHECK: .globl external_w_init 204