70 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GCN %s2 3; GCN-LABEL: {{^}}const_load_no_shrink_dword_to_unaligned_byte:4; GCN: s_load_dword s{{[0-9]+}}5; GCN: s_load_dword [[LD:s[0-9]+]],6; GCN: s_bfe_i32 s{{[0-9]+}}, [[LD]], 0x100137define amdgpu_kernel void @const_load_no_shrink_dword_to_unaligned_byte(ptr addrspace(1) %out, ptr addrspace(4) %in, i32 %x) {8 %ptr = getelementptr i32, ptr addrspace(4) %in, i32 %x9 %load = load i32, ptr addrspace(4) %ptr, align 410 %and = and i32 %load, 52428811 %cmp = icmp eq i32 %and, 012 %sel = select i1 %cmp, i32 0, i32 -113 store i32 %sel, ptr addrspace(1) %out14 ret void15}16 17; GCN-LABEL: const_load_no_shrink_dword_to_aligned_byte:18; GCN: s_load_dword s{{[0-9]+}}19; GCN: s_load_dword [[LD:s[0-9]+]],20; GCN: s_bfe_i32 s{{[0-9]+}}, [[LD]], 0x1000321define amdgpu_kernel void @const_load_no_shrink_dword_to_aligned_byte(ptr addrspace(1) %out, ptr addrspace(4) %in, i32 %x) {22 %ptr = getelementptr i32, ptr addrspace(4) %in, i32 %x23 %load = load i32, ptr addrspace(4) %ptr, align 424 %and = and i32 %load, 825 %cmp = icmp eq i32 %and, 026 %sel = select i1 %cmp, i32 0, i32 -127 store i32 %sel, ptr addrspace(1) %out28 ret void29}30 31; GCN-LABEL: global_load_no_shrink_dword_to_unaligned_byte:32; GCN: s_load_dword s{{[0-9]+}}33; GCN: s_load_dword [[LD:s[0-9]+]],34; GCN: s_bfe_i32 s{{[0-9]+}}, [[LD]], 0x1001335define amdgpu_kernel void @global_load_no_shrink_dword_to_unaligned_byte(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in, i32 %x) {36 %ptr = getelementptr i32, ptr addrspace(1) %in, i32 %x37 %load = load i32, ptr addrspace(1) %ptr, align 438 %and = and i32 %load, 52428839 %cmp = icmp eq i32 %and, 040 %sel = select i1 %cmp, i32 0, i32 -141 store i32 %sel, ptr addrspace(1) %out42 ret void43}44 45; GCN-LABEL: global_load_no_shrink_dword_to_aligned_byte:46; GCN: s_load_dword s{{[0-9]+}}47; GCN: s_load_dword [[LD:s[0-9]+]],48; GCN: s_bfe_i32 s{{[0-9]+}}, [[LD]], 0x1000349define amdgpu_kernel void @global_load_no_shrink_dword_to_aligned_byte(ptr addrspace(1) %out, ptr addrspace(1) %in, i32 %x) {50 %ptr = getelementptr i32, ptr addrspace(1) %in, i32 %x51 %load = load i32, ptr addrspace(1) %ptr, align 452 %and = and i32 %load, 853 %cmp = icmp eq i32 %and, 054 %sel = select i1 %cmp, i32 0, i32 -155 store i32 %sel, ptr addrspace(1) %out56 ret void57}58 59; GCN-LABEL: const_load_shrink_dword_to_unaligned_byte:60; GCN: global_load_ushort61define amdgpu_kernel void @const_load_shrink_dword_to_unaligned_byte(ptr addrspace(1) %out, ptr addrspace(4) %in, i32 %x) {62 %ptr = getelementptr i32, ptr addrspace(4) %in, i32 %x63 %load = load i32, ptr addrspace(4) %ptr, align 264 %and = and i32 %load, 52428865 %cmp = icmp eq i32 %and, 066 %sel = select i1 %cmp, i32 0, i32 -167 store i32 %sel, ptr addrspace(1) %out68 ret void69}70