68 lines · plain
1; RUN: llc -mtriple=amdgcn--amdpal < %s | FileCheck -check-prefixes=GCN,GCN-PAL %s2; RUN: llc -mtriple=amdgcn-- -mcpu=kaveri < %s | FileCheck -check-prefixes=GCN,GCN-DEFAULT %s3; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=kaveri < %s | FileCheck -check-prefixes=GCN,GCN-MESA %s4; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri < %s | FileCheck -check-prefixes=GCN,GCN-DEFAULT %s5; RUN: llc -mtriple=r600-- -mcpu=cypress < %s | FileCheck -check-prefix=R600 %s6 7@private1 = private unnamed_addr addrspace(4) constant [4 x float] [float 0.0, float 1.0, float 2.0, float 3.0]8@private2 = private unnamed_addr addrspace(4) constant [4 x float] [float 4.0, float 5.0, float 6.0, float 7.0]9@available_externally = available_externally addrspace(4) global [256 x i32] zeroinitializer10 11; GCN-LABEL: {{^}}private_test:12 13; Non-R600 OSes use relocations.14; GCN-DEFAULT: s_getpc_b64 s[[[PC0_LO:[0-9]+]]:[[PC0_HI:[0-9]+]]]15; GCN-DEFAULT: s_add_u32 s{{[0-9]+}}, s[[PC0_LO]], .Lprivate1@rel32@lo+416; GCN-DEFAULT: s_addc_u32 s{{[0-9]+}}, s[[PC0_HI]], .Lprivate1@rel32@hi+1217; GCN-DEFAULT: s_getpc_b64 s[[[PC1_LO:[0-9]+]]:[[PC1_HI:[0-9]+]]]18; GCN-DEFAULT: s_add_u32 s{{[0-9]+}}, s[[PC1_LO]], .Lprivate2@rel32@lo+419; GCN-DEFAULT: s_addc_u32 s{{[0-9]+}}, s[[PC1_HI]], .Lprivate2@rel32@hi+1220 21; MESA uses absolute relocations.22; GCN-MESA: s_add_u32 s2, .Lprivate1@abs32@lo, s423; GCN-MESA: s_addc_u32 s3, .Lprivate1@abs32@hi, s524 25; PAL uses absolute relocations.26; GCN-PAL: s_add_u32 s2, .Lprivate1@abs32@lo, s427; GCN-PAL: s_addc_u32 s3, .Lprivate1@abs32@hi, s528; GCN-PAL: s_add_u32 s4, .Lprivate2@abs32@lo, s429; GCN-PAL: s_addc_u32 s5, .Lprivate2@abs32@hi, s530 31; R600-LABEL: private_test32define amdgpu_kernel void @private_test(i32 %index, ptr addrspace(1) %out) {33 %ptr = getelementptr [4 x float], ptr addrspace(4) @private1, i32 0, i32 %index34 %val = load float, ptr addrspace(4) %ptr35 store volatile float %val, ptr addrspace(1) %out36 %ptr2 = getelementptr [4 x float], ptr addrspace(4) @private2, i32 0, i32 %index37 %val2 = load float, ptr addrspace(4) %ptr238 store volatile float %val2, ptr addrspace(1) %out39 ret void40}41 42; GCN-LABEL: {{^}}available_externally_test:43; GCN-DEFAULT: s_getpc_b64 s[[[PC0_LO:[0-9]+]]:[[PC0_HI:[0-9]+]]]44; GCN-DEFAULT: s_add_u32 s{{[0-9]+}}, s[[PC0_LO]], available_externally@gotpcrel32@lo+445; GCN-DEFAULT: s_addc_u32 s{{[0-9]+}}, s[[PC0_HI]], available_externally@gotpcrel32@hi+1246 47; GCN-MESA: s_mov_b32 s1, available_externally@abs32@hi48; GCN-MESA: s_mov_b32 s0, available_externally@abs32@lo49 50; R600-LABEL: available_externally_test51 52; GCN-PAL: s_mov_b32 s1, available_externally@abs32@hi53; GCN-PAL: s_mov_b32 s0, available_externally@abs32@lo54define amdgpu_kernel void @available_externally_test(ptr addrspace(1) %out) {55 %ptr = getelementptr [256 x i32], ptr addrspace(4) @available_externally, i32 0, i32 156 %val = load i32, ptr addrspace(4) %ptr57 store i32 %val, ptr addrspace(1) %out58 ret void59}60 61; GCN: .section .rodata62; R600: .text63 64; GCN: private1:65; GCN: private2:66; R600: private1:67; R600: private2:68