1; RUN: llc -mtriple=amdgcn-amd-amdhsa --amdhsa-code-object-version=2 -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,HSA %s 2; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,MESA %s 3 4; GCN-LABEL: {{^}}kernel_implicitarg_ptr_empty: 5; GCN: enable_sgpr_kernarg_segment_ptr = 1 6 7; HSA: kernarg_segment_byte_size = 0 8; MESA: kernarg_segment_byte_size = 16 9 10; HSA: s_load_dword s0, s[4:5], 0x0 11define amdgpu_kernel void @kernel_implicitarg_ptr_empty() #0 { 12 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 13 %cast = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 14 %load = load volatile i32, i32 addrspace(4)* %cast 15 ret void 16} 17 18; GCN-LABEL: {{^}}opencl_kernel_implicitarg_ptr_empty: 19; GCN: enable_sgpr_kernarg_segment_ptr = 1 20 21; HSA: kernarg_segment_byte_size = 48 22; MESA: kernarg_segment_byte_size = 16 23 24; HSA: s_load_dword s0, s[4:5], 0x0 25define amdgpu_kernel void @opencl_kernel_implicitarg_ptr_empty() #1 { 26 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 27 %cast = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 28 %load = load volatile i32, i32 addrspace(4)* %cast 29 ret void 30} 31 32; GCN-LABEL: {{^}}kernel_implicitarg_ptr: 33; GCN: enable_sgpr_kernarg_segment_ptr = 1 34 35; HSA: kernarg_segment_byte_size = 112 36; MESA: kernarg_segment_byte_size = 128 37 38; HSA: s_load_dword s0, s[4:5], 0x1c 39define amdgpu_kernel void @kernel_implicitarg_ptr([112 x i8]) #0 { 40 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 41 %cast = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 42 %load = load volatile i32, i32 addrspace(4)* %cast 43 ret void 44} 45 46; GCN-LABEL: {{^}}opencl_kernel_implicitarg_ptr: 47; GCN: enable_sgpr_kernarg_segment_ptr = 1 48 49; HSA: kernarg_segment_byte_size = 160 50; MESA: kernarg_segment_byte_size = 128 51 52; HSA: s_load_dword s0, s[4:5], 0x1c 53define amdgpu_kernel void @opencl_kernel_implicitarg_ptr([112 x i8]) #1 { 54 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 55 %cast = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 56 %load = load volatile i32, i32 addrspace(4)* %cast 57 ret void 58} 59 60; GCN-LABEL: {{^}}func_implicitarg_ptr: 61; GCN: s_waitcnt 62; HSA: s_load_dword s{{[0-9]+}}, s[8:9], 0x0 63; MESA: s_load_dword s{{[0-9]+}}, s[4:5], 0x0 64; GCN-NEXT: s_waitcnt 65; GCN-NEXT: s_setpc_b64 66define void @func_implicitarg_ptr() #0 { 67 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 68 %cast = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 69 %load = load volatile i32, i32 addrspace(4)* %cast 70 ret void 71} 72 73; GCN-LABEL: {{^}}opencl_func_implicitarg_ptr: 74; GCN: s_waitcnt 75; HSA: s_load_dword s{{[0-9]+}}, s[8:9], 0x0 76; MESA: s_load_dword s{{[0-9]+}}, s[4:5], 0x0 77; GCN-NEXT: s_waitcnt 78; GCN-NEXT: s_setpc_b64 79define void @opencl_func_implicitarg_ptr() #0 { 80 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 81 %cast = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 82 %load = load volatile i32, i32 addrspace(4)* %cast 83 ret void 84} 85 86; GCN-LABEL: {{^}}kernel_call_implicitarg_ptr_func_empty: 87; GCN: enable_sgpr_kernarg_segment_ptr = 1 88; HSA: kernarg_segment_byte_size = 0 89; MESA: kernarg_segment_byte_size = 16 90; GCN-NOT: s[4:5] 91; GCN-NOT: s4 92; GCN-NOT: s5 93; GCN: s_swappc_b64 94define amdgpu_kernel void @kernel_call_implicitarg_ptr_func_empty() #0 { 95 call void @func_implicitarg_ptr() 96 ret void 97} 98 99; GCN-LABEL: {{^}}opencl_kernel_call_implicitarg_ptr_func_empty: 100; GCN: enable_sgpr_kernarg_segment_ptr = 1 101; HSA: kernarg_segment_byte_size = 48 102; MESA: kernarg_segment_byte_size = 16 103; GCN-NOT: s[4:5] 104; GCN-NOT: s4 105; GCN-NOT: s5 106; GCN: s_swappc_b64 107define amdgpu_kernel void @opencl_kernel_call_implicitarg_ptr_func_empty() #1 { 108 call void @func_implicitarg_ptr() 109 ret void 110} 111 112; GCN-LABEL: {{^}}kernel_call_implicitarg_ptr_func: 113; GCN: enable_sgpr_kernarg_segment_ptr = 1 114; HSA: kernarg_segment_byte_size = 112 115; MESA: kernarg_segment_byte_size = 128 116 117; HSA: s_add_u32 s8, s8, 0x70 118; MESA: s_add_u32 s4, s4, 0x70 119 120; HSA: s_addc_u32 s9, s9, 0{{$}} 121; MESA: s_addc_u32 s5, s5, 0{{$}} 122; GCN: s_swappc_b64 123define amdgpu_kernel void @kernel_call_implicitarg_ptr_func([112 x i8]) #0 { 124 call void @func_implicitarg_ptr() 125 ret void 126} 127 128; GCN-LABEL: {{^}}opencl_kernel_call_implicitarg_ptr_func: 129; GCN: enable_sgpr_kernarg_segment_ptr = 1 130; HSA: kernarg_segment_byte_size = 160 131; MESA: kernarg_segment_byte_size = 128 132 133; HSA: s_add_u32 s8, s8, 0x70 134; HSA: s_addc_u32 s9, s9, 0{{$}} 135; MESA: s_add_u32 s4, s4, 0x70 136; MESA: s_addc_u32 s5, s5, 0{{$}} 137; GCN: s_swappc_b64 138define amdgpu_kernel void @opencl_kernel_call_implicitarg_ptr_func([112 x i8]) #1 { 139 call void @func_implicitarg_ptr() 140 ret void 141} 142 143; GCN-LABEL: {{^}}func_call_implicitarg_ptr_func: 144; HSA-NOT: s8 145; HSA-NOT: s9 146; HSA-NOT: s[8:9] 147; MESA-NOT: s4 148; MESA-NOT: s5 149; MESA-NOT: s[4:5] 150define void @func_call_implicitarg_ptr_func() #0 { 151 call void @func_implicitarg_ptr() 152 ret void 153} 154 155; GCN-LABEL: {{^}}opencl_func_call_implicitarg_ptr_func: 156; HSA-NOT: s8 157; HSA-NOT: s9 158; HSA-NOT: s[8:9] 159; MESA-NOT: s4 160; MESA-NOT: s5 161; MESA-NOT: s[4:5] 162define void @opencl_func_call_implicitarg_ptr_func() #0 { 163 call void @func_implicitarg_ptr() 164 ret void 165} 166 167; GCN-LABEL: {{^}}func_kernarg_implicitarg_ptr: 168; GCN: s_waitcnt 169; GCN-DAG: s_mov_b64 [[NULL:s\[[0-9]+:[0-9]+\]]], 0 170; GCN-DAG: s_load_dword s{{[0-9]+}}, [[NULL]], 0x0 171; HSA: s_load_dword s{{[0-9]+}}, s[8:9], 0x0 172; MESA: s_load_dword s{{[0-9]+}}, s[4:5], 0x0 173; GCN: s_waitcnt lgkmcnt(0) 174define void @func_kernarg_implicitarg_ptr() #0 { 175 %kernarg.segment.ptr = call i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr() 176 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 177 %cast.kernarg.segment.ptr = bitcast i8 addrspace(4)* %kernarg.segment.ptr to i32 addrspace(4)* 178 %cast.implicitarg = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 179 %load0 = load volatile i32, i32 addrspace(4)* %cast.kernarg.segment.ptr 180 %load1 = load volatile i32, i32 addrspace(4)* %cast.implicitarg 181 ret void 182} 183 184; GCN-LABEL: {{^}}opencl_func_kernarg_implicitarg_ptr: 185; GCN: s_waitcnt 186; GCN-DAG: s_mov_b64 [[NULL:s\[[0-9]+:[0-9]+\]]], 0 187; GCN-DAG: s_load_dword s{{[0-9]+}}, [[NULL]], 0x0 188; HSA: s_load_dword s{{[0-9]+}}, s[8:9], 0x0 189; MESA: s_load_dword s{{[0-9]+}}, s[4:5], 0x0 190; GCN: s_waitcnt lgkmcnt(0) 191define void @opencl_func_kernarg_implicitarg_ptr() #0 { 192 %kernarg.segment.ptr = call i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr() 193 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 194 %cast.kernarg.segment.ptr = bitcast i8 addrspace(4)* %kernarg.segment.ptr to i32 addrspace(4)* 195 %cast.implicitarg = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 196 %load0 = load volatile i32, i32 addrspace(4)* %cast.kernarg.segment.ptr 197 %load1 = load volatile i32, i32 addrspace(4)* %cast.implicitarg 198 ret void 199} 200 201; GCN-LABEL: {{^}}kernel_call_kernarg_implicitarg_ptr_func: 202; HSA: s_add_u32 s8, s8, 0x70 203; HSA: s_addc_u32 s9, s9, 0 204; MESA: s_add_u32 s4, s4, 0x70 205; MESA: s_addc_u32 s5, s5, 0 206; GCN: s_swappc_b64 207define amdgpu_kernel void @kernel_call_kernarg_implicitarg_ptr_func([112 x i8]) #0 { 208 call void @func_kernarg_implicitarg_ptr() 209 ret void 210} 211 212; GCN-LABEL: {{^}}kernel_implicitarg_no_struct_align_padding: 213; HSA: kernarg_segment_byte_size = 120 214; MESA: kernarg_segment_byte_size = 84 215; GCN: kernarg_segment_alignment = 6 216define amdgpu_kernel void @kernel_implicitarg_no_struct_align_padding(<16 x i32>, i32) #1 { 217 %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() 218 %cast = bitcast i8 addrspace(4)* %implicitarg.ptr to i32 addrspace(4)* 219 %load = load volatile i32, i32 addrspace(4)* %cast 220 ret void 221} 222 223declare i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr() #2 224declare i8 addrspace(4)* @llvm.amdgcn.kernarg.segment.ptr() #2 225 226attributes #0 = { nounwind noinline } 227attributes #1 = { nounwind noinline "amdgpu-implicitarg-num-bytes"="48" } 228attributes #2 = { nounwind readnone speculatable } 229