1; RUN: llc -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=SI -check-prefix=FUNC %s 2; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=SI -check-prefix=FUNC %s 3; RUN: llc -mtriple=r600 -mcpu=redwood < %s | FileCheck -allow-deprecated-dag-overlap -check-prefix=R600 -check-prefix=FUNC %s 4 5; FUNC-LABEL: {{^}}s_uint_to_fp_i32_to_f32: 6; SI: v_cvt_f32_u32_e32 7 8; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW]}}, KC0[2].Z 9define amdgpu_kernel void @s_uint_to_fp_i32_to_f32(ptr addrspace(1) %out, i32 %in) #0 { 10 %result = uitofp i32 %in to float 11 store float %result, ptr addrspace(1) %out 12 ret void 13} 14 15; FUNC-LABEL: {{^}}v_uint_to_fp_i32_to_f32: 16; SI: v_cvt_f32_u32_e32 {{v[0-9]+}}, {{v[0-9]+$}} 17 18; R600: INT_TO_FLT 19define amdgpu_kernel void @v_uint_to_fp_i32_to_f32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { 20 %tid = call i32 @llvm.amdgcn.workitem.id.x() 21 %in.gep = getelementptr i32, ptr addrspace(1) %in, i32 %tid 22 %out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid 23 %val = load i32, ptr addrspace(1) %in.gep 24 %result = uitofp i32 %val to float 25 store float %result, ptr addrspace(1) %out.gep 26 ret void 27} 28 29; FUNC-LABEL: {{^}}s_uint_to_fp_v2i32_to_v2f32: 30; SI: v_cvt_f32_u32_e32 31; SI: v_cvt_f32_u32_e32 32 33; R600-DAG: UINT_TO_FLT * T{{[0-9]+\.[XYZW]}}, KC0[2].W 34; R600-DAG: UINT_TO_FLT * T{{[0-9]+\.[XYZW]}}, KC0[3].X 35define amdgpu_kernel void @s_uint_to_fp_v2i32_to_v2f32(ptr addrspace(1) %out, <2 x i32> %in) #0 { 36 %result = uitofp <2 x i32> %in to <2 x float> 37 store <2 x float> %result, ptr addrspace(1) %out 38 ret void 39} 40 41; FUNC-LABEL: {{^}}s_uint_to_fp_v4i32_to_v4f32: 42; SI: v_cvt_f32_u32_e32 43; SI: v_cvt_f32_u32_e32 44; SI: v_cvt_f32_u32_e32 45; SI: v_cvt_f32_u32_e32 46; SI: s_endpgm 47 48; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 49; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 50; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 51; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 52define amdgpu_kernel void @s_uint_to_fp_v4i32_to_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { 53 %value = load <4 x i32>, ptr addrspace(1) %in 54 %result = uitofp <4 x i32> %value to <4 x float> 55 store <4 x float> %result, ptr addrspace(1) %out 56 ret void 57} 58 59; FUNC-LABEL: {{^}}v_uint_to_fp_v4i32: 60; SI: v_cvt_f32_u32_e32 61; SI: v_cvt_f32_u32_e32 62; SI: v_cvt_f32_u32_e32 63; SI: v_cvt_f32_u32_e32 64 65; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 66; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 67; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 68; R600: UINT_TO_FLT * T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}} 69define amdgpu_kernel void @v_uint_to_fp_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { 70 %tid = call i32 @llvm.amdgcn.workitem.id.x() 71 %in.gep = getelementptr <4 x i32>, ptr addrspace(1) %in, i32 %tid 72 %out.gep = getelementptr <4 x float>, ptr addrspace(1) %out, i32 %tid 73 %value = load <4 x i32>, ptr addrspace(1) %in.gep 74 %result = uitofp <4 x i32> %value to <4 x float> 75 store <4 x float> %result, ptr addrspace(1) %out.gep 76 ret void 77} 78 79; FUNC-LABEL: {{^}}s_uint_to_fp_i1_to_f32: 80; SI: s_cmp_eq_u32 81; SI: s_cselect_b64 [[CMP:s\[[0-9]+:[0-9]\]]], -1, 0 82; SI: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1.0, [[CMP]] 83; SI: buffer_store_dword [[RESULT]], 84; SI: s_endpgm 85define amdgpu_kernel void @s_uint_to_fp_i1_to_f32(ptr addrspace(1) %out, i32 %in) #0 { 86 %cmp = icmp eq i32 %in, 0 87 %fp = uitofp i1 %cmp to float 88 store float %fp, ptr addrspace(1) %out 89 ret void 90} 91 92; FUNC-LABEL: {{^}}s_uint_to_fp_i1_to_f32_load: 93; SI: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1.0 94; SI: buffer_store_dword [[RESULT]], 95; SI: s_endpgm 96define amdgpu_kernel void @s_uint_to_fp_i1_to_f32_load(ptr addrspace(1) %out, i1 %in) #0 { 97 %fp = uitofp i1 %in to float 98 store float %fp, ptr addrspace(1) %out 99 ret void 100} 101 102; FUNC-LABEL: {{^}}v_uint_to_fp_i1_f32_load: 103; SI: {{buffer|flat}}_load_ubyte 104; SI: v_and_b32_e32 {{v[0-9]+}}, 1, {{v[0-9]+}} 105; SI: v_cmp_eq_u32 106; SI: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1.0 107; SI: {{buffer|flat}}_store_dword {{.*}}[[RESULT]] 108; SI: s_endpgm 109define amdgpu_kernel void @v_uint_to_fp_i1_f32_load(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 { 110 %tid = call i32 @llvm.amdgcn.workitem.id.x() 111 %in.gep = getelementptr i1, ptr addrspace(1) %in, i32 %tid 112 %out.gep = getelementptr float, ptr addrspace(1) %out, i32 %tid 113 %val = load i1, ptr addrspace(1) %in.gep 114 %fp = uitofp i1 %val to float 115 store float %fp, ptr addrspace(1) %out.gep 116 ret void 117} 118 119; FIXME: Repeated here to test r600 120; FUNC-LABEL: {{^}}s_uint_to_fp_i64_to_f32: 121; R600: FFBH_UINT 122; R600: CNDE_INT 123; R600: UINT_TO_FLT 124 125define amdgpu_kernel void @s_uint_to_fp_i64_to_f32(ptr addrspace(1) %out, i64 %in) #0 { 126entry: 127 %cvt = uitofp i64 %in to float 128 store float %cvt, ptr addrspace(1) %out 129 ret void 130} 131 132declare i32 @llvm.amdgcn.workitem.id.x() #1 133 134attributes #0 = { nounwind } 135attributes #1 = { nounwind readnone } 136