1; RUN: llc -verify-machineinstrs -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -enable-amdgpu-aa=0 -mattr=+flat-for-global,-fp64-fp16-denormals < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9,GFX89 %s 2; RUN: llc -verify-machineinstrs -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -enable-amdgpu-aa=0 -mattr=+flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CIVI,VI,GFX89 %s 3; RUN: llc -verify-machineinstrs -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -enable-amdgpu-aa=0 -mattr=+flat-for-global < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CIVI,CI %s 4 5; GCN-LABEL: {{^}}s_insertelement_v2i16_0: 6; GCN: s_load_dword [[VEC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x0 7 8; CIVI: s_and_b32 [[ELT1:s[0-9]+]], [[VEC]], 0xffff0000{{$}} 9; CIVI: s_or_b32 s{{[0-9]+}}, [[ELT1]], 0x3e7{{$}} 10 11; GFX9-NOT: lshr 12; GFX9: s_pack_lh_b32_b16 s{{[0-9]+}}, 0x3e7, [[VEC]] 13define amdgpu_kernel void @s_insertelement_v2i16_0(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr) #0 { 14 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 15 %vecins = insertelement <2 x i16> %vec, i16 999, i32 0 16 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 17 ret void 18} 19 20; GCN-LABEL: {{^}}s_insertelement_v2i16_0_reg: 21; GCN-DAG: s_load_dword [[ELT_LOAD:s[0-9]+]], s[4:5], 22; GCN-DAG: s_load_dword [[VEC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x0 23 24; CIVI-DAG: s_and_b32 [[ELT0:s[0-9]+]], [[ELT_LOAD]], 0xffff{{$}} 25; CIVI-DAG: s_and_b32 [[ELT1:s[0-9]+]], [[VEC]], 0xffff0000{{$}} 26; CIVI: s_or_b32 s{{[0-9]+}}, [[ELT0]], [[ELT1]] 27 28; GFX9-NOT: [[ELT0]] 29; GFX9-NOT: [[VEC]] 30; GFX9: s_pack_lh_b32_b16 s{{[0-9]+}}, [[ELT_LOAD]], [[VEC]] 31define amdgpu_kernel void @s_insertelement_v2i16_0_reg(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr, [8 x i32], i16 %elt) #0 { 32 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 33 %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0 34 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 35 ret void 36} 37 38; GCN-LABEL: {{^}}s_insertelement_v2i16_0_multi_use_hi_reg: 39; GCN-DAG: s_load_dword [[ELT_LOAD:s[0-9]+]], s[4:5], 40; GCN-DAG: s_load_dword [[VEC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x0 41 42; CI-DAG: s_and_b32 [[ELT0_MASKED:s[0-9]+]], [[ELT_LOAD]], 0xffff{{$}} 43; CI: s_lshr_b32 [[SHR:s[0-9]+]], [[VEC]], 16 44; CI: s_lshl_b32 [[ELT1:s[0-9]+]], [[SHR]], 16 45; CI-DAG: s_or_b32 s{{[0-9]+}}, [[ELT0_MASKED]], [[ELT1]] 46; CI-DAG: ; use [[SHR]] 47 48 49; FIXME: Should be able to void mask of upper bits 50; VI-DAG: s_and_b32 [[ELT_MASKED:s[0-9]+]], [[ELT_LOAD]], 0xffff{{$}} 51; VI-DAG: s_and_b32 [[VEC_HIMASK:s[0-9]+]], [[VEC]], 0xffff0000{{$}} 52; VI-DAG: s_or_b32 [[OR:s[0-9]+]], [[ELT_MASKED]], [[VEC_HIMASK]] 53; VI-DAG: s_lshr_b32 [[SHR:s[0-9]+]], [[VEC]], 16 54 55; VI-DAG: ; use [[SHR]] 56 57 58; GFX9: s_lshr_b32 [[ELT1:s[0-9]+]], [[VEC]], 16 59; GFX9-DAG: s_pack_ll_b32_b16 s{{[0-9]+}}, [[ELT_LOAD]], [[ELT1]] 60; GFX9-DAG: ; use [[ELT1]] 61define amdgpu_kernel void @s_insertelement_v2i16_0_multi_use_hi_reg(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr, [8 x i32], i16 %elt) #0 { 62 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 63 %elt1 = extractelement <2 x i16> %vec, i32 1 64 %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0 65 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 66 %use1 = zext i16 %elt1 to i32 67 call void asm sideeffect "; use $0", "s"(i32 %use1) #0 68 ret void 69} 70 71; GCN-LABEL: {{^}}s_insertelement_v2i16_0_reghi: 72; GCN-DAG: s_load_dword [[ELT_ARG:s[0-9]+]], s[4:5], 73; GCN-DAG: s_load_dword [[VEC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x0 74 75; CIVI: s_lshr_b32 [[ELT_HI:s[0-9]+]], [[ELT_ARG]], 16 76; CIVI-DAG: s_and_b32 [[ELT1:s[0-9]+]], [[VEC]], 0xffff0000{{$}} 77; CIVI: s_or_b32 s{{[0-9]+}}, [[ELT_HI]], [[ELT1]] 78 79; GFX9-NOT: [[ELT0]] 80; GFX9-NOT: [[VEC]] 81; GFX9: s_pack_hh_b32_b16 s{{[0-9]+}}, [[ELT_ARG]], [[VEC]] 82define amdgpu_kernel void @s_insertelement_v2i16_0_reghi(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr, [8 x i32], i32 %elt.arg) #0 { 83 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 84 %elt.hi = lshr i32 %elt.arg, 16 85 %elt = trunc i32 %elt.hi to i16 86 %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0 87 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 88 ret void 89} 90 91; GCN-LABEL: {{^}}s_insertelement_v2i16_0_reghi_multi_use_1: 92; GCN: s_load_dword [[ELT_ARG:s[0-9]+]], 93; GCN: s_load_dword [[VEC:s[0-9]+]], 94 95; CIVI-DAG: s_lshr_b32 [[ELT1:s[0-9]+]], [[ELT_ARG]], 16 96; CIVI-DAG: s_and_b32 [[ELT0:s[0-9]+]], [[VEC]], 0xffff0000{{$}} 97; CIVI: s_or_b32 s{{[0-9]+}}, [[ELT1]], [[ELT0]] 98 99; GFX9: s_lshr_b32 [[ELT1:s[0-9]+]], [[ELT_ARG]], 16 100; GFX9: s_pack_lh_b32_b16 s{{[0-9]+}}, [[ELT1]], [[VEC]] 101; GFX9: ; use [[ELT1]] 102define amdgpu_kernel void @s_insertelement_v2i16_0_reghi_multi_use_1(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr, i32 %elt.arg) #0 { 103 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 104 %elt.hi = lshr i32 %elt.arg, 16 105 %elt = trunc i32 %elt.hi to i16 106 %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0 107 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 108 %use1 = zext i16 %elt to i32 109 call void asm sideeffect "; use $0", "s"(i32 %use1) #0 110 ret void 111} 112 113; GCN-LABEL: {{^}}s_insertelement_v2i16_0_reghi_both_multi_use_1: 114; GCN: s_load_dword [[ELT_ARG:s[0-9]+]], 115; GCN: s_load_dword [[VEC:s[0-9]+]], 116 117; CI-DAG: s_lshr_b32 [[ELT_HI:s[0-9]+]], [[ELT_ARG]], 16 118; CI-DAG: s_lshr_b32 [[SHR:s[0-9]+]], [[VEC]], 16 119; CI-DAG: s_lshl_b32 [[VEC_HI:s[0-9]+]], [[SHR]], 16 120; CI: s_or_b32 s{{[0-9]+}}, [[ELT_HI]], [[VEC_HI]] 121 122 123; VI-DAG: s_lshr_b32 [[ELT_HI:s[0-9]+]], [[ELT_ARG]], 16 124; VI-DAG: s_lshr_b32 [[VEC_HI:s[0-9]+]], [[VEC]], 16 125; VI: s_and_b32 [[MASK_HI:s[0-9]+]], [[VEC]], 0xffff0000 126; VI: s_or_b32 s{{[0-9]+}}, [[ELT_HI]], [[MASK_HI]] 127 128; GFX9-DAG: s_lshr_b32 [[ELT_HI:s[0-9]+]], [[ELT_ARG]], 16 129; GFX9-DAG: s_lshr_b32 [[VEC_HI:s[0-9]+]], [[VEC]], 16 130; GFX9: s_pack_ll_b32_b16 s{{[0-9]+}}, [[ELT_HI]], [[VEC_HI]] 131; GFX9: ; use [[ELT_HI]] 132; GFX9: ; use [[VEC_HI]] 133define amdgpu_kernel void @s_insertelement_v2i16_0_reghi_both_multi_use_1(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr, i32 %elt.arg) #0 { 134 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 135 %elt.hi = lshr i32 %elt.arg, 16 136 %elt = trunc i32 %elt.hi to i16 137 %vec.hi = extractelement <2 x i16> %vec, i32 1 138 %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0 139 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 140 %use1 = zext i16 %elt to i32 141 %vec.hi.use1 = zext i16 %vec.hi to i32 142 143 call void asm sideeffect "; use $0", "s"(i32 %use1) #0 144 call void asm sideeffect "; use $0", "s"(i32 %vec.hi.use1) #0 145 ret void 146} 147 148; GCN-LABEL: {{^}}s_insertelement_v2i16_1: 149; GCN: s_load_dword [[VEC:s[0-9]+]] 150 151; GCN-NOT: s_lshr 152 153; CIVI: s_and_b32 [[ELT0:s[0-9]+]], [[VEC]], 0xffff{{$}} 154; CIVI: s_or_b32 [[INS:s[0-9]+]], [[ELT0]], 0x3e70000 155 156; GFX9: s_pack_ll_b32_b16 s{{[0-9]+}}, [[VEC]], 0x3e7 157define amdgpu_kernel void @s_insertelement_v2i16_1(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr) #0 { 158 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 159 %vecins = insertelement <2 x i16> %vec, i16 999, i32 1 160 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 161 ret void 162} 163 164; GCN-LABEL: {{^}}s_insertelement_v2i16_1_reg: 165; GCN-DAG: s_load_dword [[ELT1_LOAD:s[0-9]+]], s[4:5], 166; GCN-DAG: s_load_dword [[VEC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0x0 167 168; CIVI: s_lshl_b32 [[ELT1:s[0-9]+]], [[ELT1_LOAD]], 16 169; CIVI: s_and_b32 [[ELT0:s[0-9]+]], [[VEC]], 0xffff{{$}} 170; CIVI: s_or_b32 s{{[0-9]+}}, [[ELT0]], [[ELT1]] 171 172; GCN-NOT: shlr 173; GFX9: s_pack_ll_b32_b16 s{{[0-9]+}}, [[VEC]], [[ELT1_LOAD]] 174define amdgpu_kernel void @s_insertelement_v2i16_1_reg(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr, [8 x i32], i16 %elt) #0 { 175 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 176 %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 1 177 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 178 ret void 179} 180 181; GCN-LABEL: {{^}}s_insertelement_v2f16_0: 182; GCN: s_load_dword [[VEC:s[0-9]+]] 183; CIVI: s_and_b32 [[ELT1:s[0-9]+]], [[VEC:s[0-9]+]], 0xffff0000 184; CIVI: s_or_b32 s{{[0-9]+}}, [[ELT1]], 0x4500 185 186; GFX9: s_lshr_b32 [[ELT1:s[0-9]+]], [[VEC]], 16 187; GFX9: s_pack_ll_b32_b16 s{{[0-9]+}}, 0x4500, [[ELT1]] 188define amdgpu_kernel void @s_insertelement_v2f16_0(<2 x half> addrspace(1)* %out, <2 x half> addrspace(4)* %vec.ptr) #0 { 189 %vec = load <2 x half>, <2 x half> addrspace(4)* %vec.ptr 190 %vecins = insertelement <2 x half> %vec, half 5.000000e+00, i32 0 191 store <2 x half> %vecins, <2 x half> addrspace(1)* %out 192 ret void 193} 194 195; GCN-LABEL: {{^}}s_insertelement_v2f16_1: 196; GCN: s_load_dword [[VEC:s[0-9]+]] 197; GCN-NOT: s_lshr 198 199; CIVI: s_and_b32 [[ELT0:s[0-9]+]], [[VEC]], 0xffff{{$}} 200; CIVI: s_or_b32 [[INS:s[0-9]+]], [[ELT0]], 0x45000000 201 202; GFX9: s_pack_ll_b32_b16 s{{[0-9]+}}, [[VEC]], 0x4500 203define amdgpu_kernel void @s_insertelement_v2f16_1(<2 x half> addrspace(1)* %out, <2 x half> addrspace(4)* %vec.ptr) #0 { 204 %vec = load <2 x half>, <2 x half> addrspace(4)* %vec.ptr 205 %vecins = insertelement <2 x half> %vec, half 5.000000e+00, i32 1 206 store <2 x half> %vecins, <2 x half> addrspace(1)* %out 207 ret void 208} 209 210; GCN-LABEL: {{^}}v_insertelement_v2i16_0: 211; GCN-DAG: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 212; CIVI: v_and_b32_e32 [[ELT1:v[0-9]+]], 0xffff0000, [[VEC]] 213; CIVI: v_or_b32_e32 [[RES:v[0-9]+]], 0x3e7, [[ELT1]] 214 215; GFX9-DAG: s_movk_i32 [[ELT0:s[0-9]+]], 0x3e7{{$}} 216; GFX9-DAG: v_mov_b32_e32 [[MASK:v[0-9]+]], 0xffff{{$}} 217; GFX9: v_bfi_b32 [[RES:v[0-9]+]], [[MASK]], [[ELT0]], [[VEC]] 218; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 219define amdgpu_kernel void @v_insertelement_v2i16_0(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %in) #0 { 220 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 221 %tid.ext = sext i32 %tid to i64 222 %in.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %in, i64 %tid.ext 223 %out.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 %tid.ext 224 %vec = load <2 x i16>, <2 x i16> addrspace(1)* %in.gep 225 %vecins = insertelement <2 x i16> %vec, i16 999, i32 0 226 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out.gep 227 ret void 228} 229 230; GCN-LABEL: {{^}}v_insertelement_v2i16_0_reghi: 231; GCN-DAG: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 232; GCN-DAG: s_load_dword [[ELT0:s[0-9]+]] 233 234; CIVI-DAG: s_lshr_b32 [[ELT0_SHIFT:s[0-9]+]], [[ELT0]], 16 235; CIVI-DAG: v_and_b32_e32 [[ELT1:v[0-9]+]], 0xffff0000, [[VEC]] 236; CIVI: v_or_b32_e32 [[RES:v[0-9]+]], [[ELT0_SHIFT]], [[ELT1]] 237 238; GFX9-DAG: v_mov_b32_e32 [[MASK:v[0-9]+]], 0xffff0000{{$}} 239; GFX9-DAG: v_lshrrev_b32_e64 [[ELT0_SHIFT:v[0-9]+]], 16, [[ELT0]] 240; GFX9: v_and_or_b32 [[RES:v[0-9]+]], [[VEC]], [[MASK]], [[ELT0_SHIFT]] 241 242; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 243define amdgpu_kernel void @v_insertelement_v2i16_0_reghi(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %in, i32 %elt.arg) #0 { 244 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 245 %tid.ext = sext i32 %tid to i64 246 %in.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %in, i64 %tid.ext 247 %out.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 %tid.ext 248 %vec = load <2 x i16>, <2 x i16> addrspace(1)* %in.gep 249 %elt.hi = lshr i32 %elt.arg, 16 250 %elt = trunc i32 %elt.hi to i16 251 %vecins = insertelement <2 x i16> %vec, i16 %elt, i32 0 252 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out.gep 253 ret void 254} 255 256; GCN-LABEL: {{^}}v_insertelement_v2i16_0_inlineimm: 257; GCN-DAG: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 258 259; CIVI: v_and_b32_e32 [[ELT1:v[0-9]+]], 0xffff0000, [[VEC]] 260; CIVI: v_or_b32_e32 [[RES:v[0-9]+]], 53, [[ELT1]] 261 262; GFX9-DAG: v_mov_b32_e32 [[MASK:v[0-9]+]], 0xffff{{$}} 263; GFX9: v_bfi_b32 [[RES:v[0-9]+]], [[MASK]], 53, [[VEC]] 264 265; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 266define amdgpu_kernel void @v_insertelement_v2i16_0_inlineimm(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %in) #0 { 267 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 268 %tid.ext = sext i32 %tid to i64 269 %in.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %in, i64 %tid.ext 270 %out.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 %tid.ext 271 %vec = load <2 x i16>, <2 x i16> addrspace(1)* %in.gep 272 %vecins = insertelement <2 x i16> %vec, i16 53, i32 0 273 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out.gep 274 ret void 275} 276 277; FIXME: fold lshl_or c0, c1, v0 -> or (c0 << c1), v0 278 279; GCN-LABEL: {{^}}v_insertelement_v2i16_1: 280; VI: v_mov_b32_e32 [[K:v[0-9]+]], 0x3e70000 281; GCN-DAG: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 282 283; GFX9-DAG: s_movk_i32 [[K:s[0-9]+]], 0x3e7 284; GFX9-DAG: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VEC]] 285; GFX9: v_lshl_or_b32 [[RES:v[0-9]+]], [[K]], 16, [[ELT0]] 286 287; CI: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff, [[VEC]] 288; CI: v_or_b32_e32 [[RES:v[0-9]+]], 0x3e70000, [[AND]] 289; VI: v_or_b32_sdwa [[RES:v[0-9]+]], [[K]], [[VEC]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 290 291; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 292define amdgpu_kernel void @v_insertelement_v2i16_1(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %in) #0 { 293 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 294 %tid.ext = sext i32 %tid to i64 295 %in.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %in, i64 %tid.ext 296 %out.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 %tid.ext 297 %vec = load <2 x i16>, <2 x i16> addrspace(1)* %in.gep 298 %vecins = insertelement <2 x i16> %vec, i16 999, i32 1 299 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out.gep 300 ret void 301} 302 303; GCN-LABEL: {{^}}v_insertelement_v2i16_1_inlineimm: 304; VI: v_mov_b32_e32 [[K:v[0-9]+]], 0xfff10000 305; GCN: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 306; CI: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VEC]] 307; GFX9: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VEC]] 308; CI: v_or_b32_e32 [[RES:v[0-9]+]], 0xfff10000, [[ELT0]] 309; VI: v_or_b32_sdwa [[RES:v[0-9]+]], [[K]], [[VEC]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 310; GFX9: v_lshl_or_b32 [[RES:v[0-9]+]], -15, 16, [[ELT0]] 311; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 312define amdgpu_kernel void @v_insertelement_v2i16_1_inlineimm(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %in) #0 { 313 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 314 %tid.ext = sext i32 %tid to i64 315 %in.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %in, i64 %tid.ext 316 %out.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 %tid.ext 317 %vec = load <2 x i16>, <2 x i16> addrspace(1)* %in.gep 318 %vecins = insertelement <2 x i16> %vec, i16 -15, i32 1 319 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out.gep 320 ret void 321} 322 323; GCN-LABEL: {{^}}v_insertelement_v2f16_0: 324; GCN-DAG: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 325 326; CIVI: v_and_b32_e32 [[ELT1:v[0-9]+]], 0xffff0000, [[VEC]] 327; CIVI: v_or_b32_e32 [[RES:v[0-9]+]], 0x4500, [[ELT1]] 328 329; GFX9-DAG: v_mov_b32_e32 [[ELT0:v[0-9]+]], 0x4500{{$}} 330; GFX9-DAG: v_lshrrev_b32_e32 [[ELT1:v[0-9]+]], 16, [[VEC]] 331; GFX9: v_lshl_or_b32 [[RES:v[0-9]+]], [[ELT1]], 16, [[ELT0]] 332 333; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 334define amdgpu_kernel void @v_insertelement_v2f16_0(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 { 335 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 336 %tid.ext = sext i32 %tid to i64 337 %in.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %in, i64 %tid.ext 338 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext 339 %vec = load <2 x half>, <2 x half> addrspace(1)* %in.gep 340 %vecins = insertelement <2 x half> %vec, half 5.000000e+00, i32 0 341 store <2 x half> %vecins, <2 x half> addrspace(1)* %out.gep 342 ret void 343} 344 345; GCN-LABEL: {{^}}v_insertelement_v2f16_0_inlineimm: 346; GCN: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 347 348; CIVI: v_and_b32_e32 [[ELT1:v[0-9]+]], 0xffff0000, [[VEC]] 349; CIVI: v_or_b32_e32 [[RES:v[0-9]+]], 53, [[ELT1]] 350 351; GFX9: v_lshrrev_b32_e32 [[ELT1:v[0-9]+]], 16, [[VEC]] 352; GFX9: v_lshl_or_b32 [[RES:v[0-9]+]], [[ELT1]], 16, 53 353; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 354define amdgpu_kernel void @v_insertelement_v2f16_0_inlineimm(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 { 355 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 356 %tid.ext = sext i32 %tid to i64 357 %in.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %in, i64 %tid.ext 358 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext 359 %vec = load <2 x half>, <2 x half> addrspace(1)* %in.gep 360 %vecins = insertelement <2 x half> %vec, half 0xH0035, i32 0 361 store <2 x half> %vecins, <2 x half> addrspace(1)* %out.gep 362 ret void 363} 364 365; GCN-LABEL: {{^}}v_insertelement_v2f16_1: 366; VI: v_mov_b32_e32 [[K:v[0-9]+]], 0x45000000 367; GCN-DAG: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 368 369; GFX9-DAG: s_movk_i32 [[K:s[0-9]+]], 0x4500 370; GFX9-DAG: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VEC]] 371; GFX9: v_lshl_or_b32 [[RES:v[0-9]+]], [[K]], 16, [[ELT0]] 372 373; CI: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff, [[VEC]] 374; CI: v_or_b32_e32 [[RES:v[0-9]+]], 0x45000000, [[AND]] 375 376; VI: v_or_b32_sdwa [[RES:v[0-9]+]], [[K]], [[VEC]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 377 378; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 379define amdgpu_kernel void @v_insertelement_v2f16_1(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 { 380 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 381 %tid.ext = sext i32 %tid to i64 382 %in.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %in, i64 %tid.ext 383 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext 384 %vec = load <2 x half>, <2 x half> addrspace(1)* %in.gep 385 %vecins = insertelement <2 x half> %vec, half 5.000000e+00, i32 1 386 store <2 x half> %vecins, <2 x half> addrspace(1)* %out.gep 387 ret void 388} 389 390; GCN-LABEL: {{^}}v_insertelement_v2f16_1_inlineimm: 391; VI: v_mov_b32_e32 [[K:v[0-9]+]], 0x230000 392; GCN: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 393; CI: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VEC]] 394; GFX9: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VEC]] 395; CI: v_or_b32_e32 [[RES:v[0-9]+]], 0x230000, [[ELT0]] 396; VI: v_or_b32_sdwa [[RES:v[0-9]+]], [[K]], [[VEC]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 397; GFX9: v_lshl_or_b32 [[RES:v[0-9]+]], 35, 16, [[ELT0]] 398; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RES]] 399define amdgpu_kernel void @v_insertelement_v2f16_1_inlineimm(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 { 400 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 401 %tid.ext = sext i32 %tid to i64 402 %in.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %in, i64 %tid.ext 403 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext 404 %vec = load <2 x half>, <2 x half> addrspace(1)* %in.gep 405 %vecins = insertelement <2 x half> %vec, half 0xH0023, i32 1 406 store <2 x half> %vecins, <2 x half> addrspace(1)* %out.gep 407 ret void 408} 409 410; FIXME: Enable for others when argument load not split 411; GCN-LABEL: {{^}}s_insertelement_v2i16_dynamic: 412; GCN-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0x3e7 413; GCN: s_load_dword [[IDX:s[0-9]+]] 414; GCN: s_load_dword [[VEC:s[0-9]+]] 415; GCN-DAG: v_mov_b32_e32 [[VVEC:v[0-9]+]], [[VEC]] 416; GCN-DAG: s_lshl_b32 [[SCALED_IDX:s[0-9]+]], [[IDX]], 4 417; GCN-DAG: s_lshl_b32 [[MASK:s[0-9]+]], 0xffff, [[SCALED_IDX]] 418; GCN: v_bfi_b32 [[RESULT:v[0-9]+]], [[MASK]], [[K]], [[VVEC]] 419; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]] 420define amdgpu_kernel void @s_insertelement_v2i16_dynamic(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(4)* %vec.ptr, i32 addrspace(4)* %idx.ptr) #0 { 421 %idx = load volatile i32, i32 addrspace(4)* %idx.ptr 422 %vec = load <2 x i16>, <2 x i16> addrspace(4)* %vec.ptr 423 %vecins = insertelement <2 x i16> %vec, i16 999, i32 %idx 424 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out 425 ret void 426} 427 428; GCN-LABEL: {{^}}v_insertelement_v2i16_dynamic_sgpr: 429; GCN-DAG: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 430; GCN-DAG: s_load_dword [[IDX:s[0-9]+]] 431; GCN-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0x3e7 432; GCN-DAG: s_lshl_b32 [[SCALED_IDX:s[0-9]+]], [[IDX]], 4 433; GCN-DAG: s_lshl_b32 [[MASK:s[0-9]+]], 0xffff, [[SCALED_IDX]] 434; GCN: v_bfi_b32 [[RESULT:v[0-9]+]], [[MASK]], [[K]], [[VEC]] 435; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]] 436define amdgpu_kernel void @v_insertelement_v2i16_dynamic_sgpr(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %in, i32 %idx) #0 { 437 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 438 %tid.ext = sext i32 %tid to i64 439 %in.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %in, i64 %tid.ext 440 %out.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 %tid.ext 441 %vec = load <2 x i16>, <2 x i16> addrspace(1)* %in.gep 442 %vecins = insertelement <2 x i16> %vec, i16 999, i32 %idx 443 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out.gep 444 ret void 445} 446 447; GCN-LABEL: {{^}}v_insertelement_v2i16_dynamic_vgpr: 448; GFX89-DAG: s_mov_b32 [[MASKK:s[0-9]+]], 0xffff{{$}} 449; GCN-DAG: s_movk_i32 [[K:s[0-9]+]], 0x3e7 450 451; GCN: {{flat|global}}_load_dword [[IDX:v[0-9]+]] 452; GCN: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 453 454; GFX89-DAG: v_lshlrev_b32_e32 [[SCALED_IDX:v[0-9]+]], 4, [[IDX]] 455; GFX89-DAG: v_lshlrev_b32_e64 [[MASK:v[0-9]+]], [[SCALED_IDX]], [[MASKK]] 456 457; CI-DAG: v_lshlrev_b32_e32 [[SCALED_IDX:v[0-9]+]], 4, [[IDX]] 458; CI-DAG: v_lshl_b32_e32 [[MASK:v[0-9]+]], 0xffff, [[SCALED_IDX]] 459 460; GCN: v_bfi_b32 [[RESULT:v[0-9]+]], [[MASK]], [[K]], [[VEC]] 461; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]] 462define amdgpu_kernel void @v_insertelement_v2i16_dynamic_vgpr(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %in, i32 addrspace(1)* %idx.ptr) #0 { 463 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 464 %tid.ext = sext i32 %tid to i64 465 %in.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %in, i64 %tid.ext 466 %idx.gep = getelementptr inbounds i32, i32 addrspace(1)* %idx.ptr, i64 %tid.ext 467 %out.gep = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 %tid.ext 468 %idx = load i32, i32 addrspace(1)* %idx.gep 469 %vec = load <2 x i16>, <2 x i16> addrspace(1)* %in.gep 470 %vecins = insertelement <2 x i16> %vec, i16 999, i32 %idx 471 store <2 x i16> %vecins, <2 x i16> addrspace(1)* %out.gep 472 ret void 473} 474 475; GCN-LABEL: {{^}}v_insertelement_v2f16_dynamic_vgpr: 476; GFX89-DAG: s_mov_b32 [[MASKK:s[0-9]+]], 0xffff{{$}} 477; GCN-DAG: s_movk_i32 [[K:s[0-9]+]], 0x1234 478 479; GCN: {{flat|global}}_load_dword [[IDX:v[0-9]+]] 480; GCN: {{flat|global}}_load_dword [[VEC:v[0-9]+]] 481 482; GFX89-DAG: v_lshlrev_b32_e32 [[SCALED_IDX:v[0-9]+]], 4, [[IDX]] 483; GFX89-DAG: v_lshlrev_b32_e64 [[MASK:v[0-9]+]], [[SCALED_IDX]], [[MASKK]] 484 485; CI-DAG: v_lshlrev_b32_e32 [[SCALED_IDX:v[0-9]+]], 4, [[IDX]] 486; CI-DAG: v_lshl_b32_e32 [[MASK:v[0-9]+]], 0xffff, [[SCALED_IDX]] 487 488; GCN: v_bfi_b32 [[RESULT:v[0-9]+]], [[MASK]], [[K]], [[VEC]] 489; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]] 490define amdgpu_kernel void @v_insertelement_v2f16_dynamic_vgpr(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in, i32 addrspace(1)* %idx.ptr) #0 { 491 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 492 %tid.ext = sext i32 %tid to i64 493 %in.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %in, i64 %tid.ext 494 %idx.gep = getelementptr inbounds i32, i32 addrspace(1)* %idx.ptr, i64 %tid.ext 495 %out.gep = getelementptr inbounds <2 x half>, <2 x half> addrspace(1)* %out, i64 %tid.ext 496 %idx = load i32, i32 addrspace(1)* %idx.gep 497 %vec = load <2 x half>, <2 x half> addrspace(1)* %in.gep 498 %vecins = insertelement <2 x half> %vec, half 0xH1234, i32 %idx 499 store <2 x half> %vecins, <2 x half> addrspace(1)* %out.gep 500 ret void 501} 502 503; GCN-LABEL: {{^}}v_insertelement_v4f16_0: 504; GCN-DAG: s_load_dword [[VAL:s[0-9]+]], s[4:5], 505; GCN-DAG: {{flat|global}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 506 507; GFX9-DAG: v_mov_b32_e32 [[BFI_MASK:v[0-9]+]], 0xffff{{$}} 508; GFX9: v_bfi_b32 v[[INS_LO:[0-9]+]], [[BFI_MASK]], [[VAL]], v[[LO]] 509 510; CIVI: s_and_b32 [[VAL_MASKED:s[0-9]+]], [[VAL]], 0xffff{{$}} 511; CIVI: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff0000, v[[LO]] 512; CIVI: v_or_b32_e32 v[[INS_LO:[0-9]+]], [[VAL_MASKED]], [[AND]] 513 514; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[INS_LO]]:[[HI]]{{\]}} 515define amdgpu_kernel void @v_insertelement_v4f16_0(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %in, [8 x i32], i32 %val) #0 { 516 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 517 %tid.ext = sext i32 %tid to i64 518 %in.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %in, i64 %tid.ext 519 %out.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %out, i64 %tid.ext 520 %vec = load <4 x half>, <4 x half> addrspace(1)* %in.gep 521 %val.trunc = trunc i32 %val to i16 522 %val.cvt = bitcast i16 %val.trunc to half 523 %vecins = insertelement <4 x half> %vec, half %val.cvt, i32 0 524 store <4 x half> %vecins, <4 x half> addrspace(1)* %out.gep 525 ret void 526} 527 528; GCN-LABEL: {{^}}v_insertelement_v4f16_1: 529; GCN-DAG: s_load_dword [[VAL:s[0-9]+]] 530; GCN-DAG: {{flat|global}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 531 532; GFX9: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff, v[[LO]] 533; GFX9: v_lshl_or_b32 v[[INS_HALF:[0-9]+]], [[VAL]], 16, [[AND]] 534 535; VI: s_lshl_b32 [[VAL_HI:s[0-9]+]], [[VAL]], 16 536; VI-DAG: v_mov_b32_e32 [[COPY_VAL:v[0-9]+]], [[VAL_HI]] 537; VI: v_or_b32_sdwa v[[INS_HALF:[0-9]+]], [[COPY_VAL]], v[[LO]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 538 539; CI: s_lshl_b32 [[VAL_HI:s[0-9]+]], [[VAL]], 16 540; CI: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff, v[[LO]] 541; CI: v_or_b32_e32 v[[INS_HALF:[0-9]+]], [[VAL_HI]], [[AND]] 542 543; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[INS_HALF]]:[[HI]]{{\]}} 544define amdgpu_kernel void @v_insertelement_v4f16_1(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %in, i32 %val) #0 { 545 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 546 %tid.ext = sext i32 %tid to i64 547 %in.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %in, i64 %tid.ext 548 %out.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %out, i64 %tid.ext 549 %vec = load <4 x half>, <4 x half> addrspace(1)* %in.gep 550 %val.trunc = trunc i32 %val to i16 551 %val.cvt = bitcast i16 %val.trunc to half 552 %vecins = insertelement <4 x half> %vec, half %val.cvt, i32 1 553 store <4 x half> %vecins, <4 x half> addrspace(1)* %out.gep 554 ret void 555} 556 557; GCN-LABEL: {{^}}v_insertelement_v4f16_2: 558; GCN-DAG: s_load_dword [[VAL:s[0-9]+]], s[4:5], 559; GCN-DAG: {{flat|global}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 560 561; GFX9-DAG: v_mov_b32_e32 [[BFI_MASK:v[0-9]+]], 0xffff{{$}} 562; GFX9: v_bfi_b32 v[[INS_HI:[0-9]+]], [[BFI_MASK]], [[VAL]], v[[HI]] 563 564; CIVI: s_and_b32 [[VAL_MASKED:s[0-9]+]], [[VAL]], 0xffff{{$}} 565; CIVI: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff0000, v[[HI]] 566; CIVI: v_or_b32_e32 v[[INS_HI:[0-9]+]], [[VAL_MASKED]], [[AND]] 567 568; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[INS_HI]]{{\]}} 569define amdgpu_kernel void @v_insertelement_v4f16_2(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %in, [8 x i32], i32 %val) #0 { 570 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 571 %tid.ext = sext i32 %tid to i64 572 %in.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %in, i64 %tid.ext 573 %out.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %out, i64 %tid.ext 574 %vec = load <4 x half>, <4 x half> addrspace(1)* %in.gep 575 %val.trunc = trunc i32 %val to i16 576 %val.cvt = bitcast i16 %val.trunc to half 577 %vecins = insertelement <4 x half> %vec, half %val.cvt, i32 2 578 store <4 x half> %vecins, <4 x half> addrspace(1)* %out.gep 579 ret void 580} 581 582; GCN-LABEL: {{^}}v_insertelement_v4f16_3: 583; GCN-DAG: s_load_dword [[VAL:s[0-9]+]] 584; GCN-DAG: {{flat|global}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 585 586; GFX9: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff, v[[HI]] 587; GFX9: v_lshl_or_b32 v[[INS_HI:[0-9]+]], [[VAL]], 16, [[AND]] 588 589; VI: s_lshl_b32 [[VAL_HI:s[0-9]+]], [[VAL]], 16 590; VI-DAG: v_mov_b32_e32 [[COPY_VAL:v[0-9]+]], [[VAL_HI]] 591; VI: v_or_b32_sdwa v[[INS_HI:[0-9]+]], [[COPY_VAL]], v[[HI]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 592 593; CI: s_lshl_b32 [[VAL_HI:s[0-9]+]], [[VAL]], 16 594; CI: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff, v[[HI]] 595; CI: v_or_b32_e32 v[[INS_HI:[0-9]+]], [[VAL_HI]], [[AND]] 596 597; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[INS_HI]]{{\]}} 598define amdgpu_kernel void @v_insertelement_v4f16_3(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %in, i32 %val) #0 { 599 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 600 %tid.ext = sext i32 %tid to i64 601 %in.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %in, i64 %tid.ext 602 %out.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %out, i64 %tid.ext 603 %vec = load <4 x half>, <4 x half> addrspace(1)* %in.gep 604 %val.trunc = trunc i32 %val to i16 605 %val.cvt = bitcast i16 %val.trunc to half 606 %vecins = insertelement <4 x half> %vec, half %val.cvt, i32 3 607 store <4 x half> %vecins, <4 x half> addrspace(1)* %out.gep 608 ret void 609} 610 611; GCN-LABEL: {{^}}v_insertelement_v4i16_2: 612; GCN-DAG: s_load_dword [[VAL:s[0-9]+]] 613; GCN-DAG: {{flat|global}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 614 615; GFX9-DAG: v_mov_b32_e32 [[BFI_MASK:v[0-9]+]], 0xffff{{$}} 616; GFX9: v_bfi_b32 v[[INS_HI:[0-9]+]], [[BFI_MASK]], [[VAL]], v[[HI]] 617 618; CIVI: s_and_b32 [[VAL_MASKED:s[0-9]+]], [[VAL]], 0xffff{{$}} 619; CIVI: v_and_b32_e32 [[AND:v[0-9]+]], 0xffff0000, v[[HI]] 620; CIVI: v_or_b32_e32 v[[INS_HI:[0-9]+]], [[VAL_MASKED]], [[AND]] 621 622; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO]]:[[INS_HI]]{{\]}} 623define amdgpu_kernel void @v_insertelement_v4i16_2(<4 x i16> addrspace(1)* %out, <4 x i16> addrspace(1)* %in, i32 %val) #0 { 624 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 625 %tid.ext = sext i32 %tid to i64 626 %in.gep = getelementptr inbounds <4 x i16>, <4 x i16> addrspace(1)* %in, i64 %tid.ext 627 %out.gep = getelementptr inbounds <4 x i16>, <4 x i16> addrspace(1)* %out, i64 %tid.ext 628 %vec = load <4 x i16>, <4 x i16> addrspace(1)* %in.gep 629 %val.trunc = trunc i32 %val to i16 630 %val.cvt = bitcast i16 %val.trunc to i16 631 %vecins = insertelement <4 x i16> %vec, i16 %val.cvt, i32 2 632 store <4 x i16> %vecins, <4 x i16> addrspace(1)* %out.gep 633 ret void 634} 635 636; FIXME: Better code on CI? 637; GCN-LABEL: {{^}}v_insertelement_v4i16_dynamic_vgpr: 638; GCN-DAG: {{flat|global}}_load_dword [[IDX:v[0-9]+]], 639; GCN-DAG: s_load_dword [[VAL:s[0-9]+]] 640; GCN-DAG: {{flat|global}}_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}} 641 642; GCN-DAG: v_lshlrev_b32_e32 [[SCALED_IDX:v[0-9]+]], 4, [[IDX]] 643; GCN-DAG: s_mov_b32 s[[MASK_HI:[0-9]+]], 0 644; GCN-DAG: s_mov_b32 s[[MASK_LO:[0-9]+]], 0xffff{{$}} 645 646; GFX89: v_lshlrev_b64 v{{\[}}[[SHIFT_LO:[0-9]+]]:[[SHIFT_HI:[0-9]+]]{{\]}}, [[SCALED_IDX]], s{{\[}}[[MASK_LO]]:[[MASK_HI]]{{\]}} 647; GFX89-DAG: v_not_b32_e32 v[[NOT_SHIFT_LO:[0-9+]]], v[[SHIFT_LO]] 648; GFX89-DAG: v_not_b32_e32 v[[NOT_SHIFT_HI:[0-9+]]], v[[SHIFT_HI]] 649; GFX89-DAG: v_and_b32_e32 v[[MASK:[0-9]+]], [[VAL]], v[[SHIFT_LO]] 650 651; GFX89-DAG: v_and_b32_e32 v[[AND0:[0-9]+]], v[[NOT_SHIFT_LO]], v[[LO]] 652; GFX89-DAG: v_and_b32_e32 v[[AND1:[0-9]+]], v[[NOT_SHIFT_HI]], v[[HI]] 653; GFX89: v_or_b32_sdwa v[[OR_SDWA:[0-9]+]], v[[MASK]], v[[AND0]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD 654 655 656; CI: v_lshl_b64 v{{\[}}[[SHIFT_LO:[0-9]+]]:[[SHIFT_HI:[0-9]+]]{{\]}}, s{{\[}}[[MASK_LO]]:[[MASK_HI]]{{\]}}, [[SCALED_IDX]] 657; CI-DAG: v_bfi_b32 v[[OR_SDWA:[0-9]+]], v[[SHIFT_LO]], 658; CI-DAG: v_bfi_b32 v[[AND1:[0-9]+]], v[[SHIFT_HI]], 0, 659 660; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[OR_SDWA]]:[[AND1]]{{\]}} 661define amdgpu_kernel void @v_insertelement_v4i16_dynamic_vgpr(<4 x i16> addrspace(1)* %out, <4 x i16> addrspace(1)* %in, i32 %val) #0 { 662 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 663 %tid.ext = sext i32 %tid to i64 664 %in.gep = getelementptr inbounds <4 x i16>, <4 x i16> addrspace(1)* %in, i64 %tid.ext 665 %out.gep = getelementptr inbounds <4 x i16>, <4 x i16> addrspace(1)* %out, i64 %tid.ext 666 %idx.val = load volatile i32, i32 addrspace(1)* undef 667 %vec = load <4 x i16>, <4 x i16> addrspace(1)* %in.gep 668 %val.trunc = trunc i32 %val to i16 669 %val.cvt = bitcast i16 %val.trunc to i16 670 %vecins = insertelement <4 x i16> %vec, i16 %val.cvt, i32 %idx.val 671 store <4 x i16> %vecins, <4 x i16> addrspace(1)* %out.gep 672 ret void 673} 674 675; GCN-LABEL: {{^}}v_insertelement_v4f16_dynamic_sgpr: 676define amdgpu_kernel void @v_insertelement_v4f16_dynamic_sgpr(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %in, i32 %val, i32 %idxval) #0 { 677 %tid = call i32 @llvm.amdgcn.workitem.id.x() #1 678 %tid.ext = sext i32 %tid to i64 679 %in.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %in, i64 %tid.ext 680 %out.gep = getelementptr inbounds <4 x half>, <4 x half> addrspace(1)* %out, i64 %tid.ext 681 %vec = load <4 x half>, <4 x half> addrspace(1)* %in.gep 682 %val.trunc = trunc i32 %val to i16 683 %val.cvt = bitcast i16 %val.trunc to half 684 %vecins = insertelement <4 x half> %vec, half %val.cvt, i32 %idxval 685 store <4 x half> %vecins, <4 x half> addrspace(1)* %out.gep 686 ret void 687} 688 689declare i32 @llvm.amdgcn.workitem.id.x() #1 690 691attributes #0 = { nounwind } 692attributes #1 = { nounwind readnone } 693