1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx901 -mattr=-flat-for-global,+fp64-fp16-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -check-prefix=GFX9-DENORM %s 2; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx901 -mattr=-flat-for-global,-fp64-fp16-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -check-prefix=GFX9-FLUSH %s 3; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s 4; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=CI %s 5 6 7; GCN-LABEL: {{^}}s_pack_v2f16: 8; GFX9: s_load_dword [[VAL0:s[0-9]+]] 9; GFX9: s_load_dword [[VAL1:s[0-9]+]] 10; GFX9: s_pack_ll_b32_b16 [[PACKED:s[0-9]+]], [[VAL0]], [[VAL1]] 11; GFX9: ; use [[PACKED]] 12define void @s_pack_v2f16(i32 addrspace(2)* %in0, i32 addrspace(2)* %in1) #0 { 13 %val0 = load volatile i32, i32 addrspace(2)* %in0 14 %val1 = load volatile i32, i32 addrspace(2)* %in1 15 %lo.i = trunc i32 %val0 to i16 16 %hi.i = trunc i32 %val1 to i16 17 %lo = bitcast i16 %lo.i to half 18 %hi = bitcast i16 %hi.i to half 19 %vec.0 = insertelement <2 x half> undef, half %lo, i32 0 20 %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1 21 %vec.i32 = bitcast <2 x half> %vec.1 to i32 22 23 call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0 24 ret void 25} 26 27; GCN-LABEL: {{^}}s_pack_v2f16_imm_lo: 28; GFX9: s_load_dword [[VAL1:s[0-9]+]] 29; GFX9: s_pack_ll_b32_b16 [[PACKED:s[0-9]+]], 0x1234, [[VAL1]] 30; GFX9: ; use [[PACKED]] 31define void @s_pack_v2f16_imm_lo(i32 addrspace(2)* %in1) #0 { 32 %val1 = load i32, i32 addrspace(2)* %in1 33 %hi.i = trunc i32 %val1 to i16 34 %hi = bitcast i16 %hi.i to half 35 %vec.0 = insertelement <2 x half> undef, half 0xH1234, i32 0 36 %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1 37 %vec.i32 = bitcast <2 x half> %vec.1 to i32 38 39 call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0 40 ret void 41} 42 43; GCN-LABEL: {{^}}s_pack_v2f16_imm_hi: 44; GFX9: s_load_dword [[VAL0:s[0-9]+]] 45; GFX9: s_pack_ll_b32_b16 [[PACKED:s[0-9]+]], [[VAL0]], 0x1234 46; GFX9: ; use [[PACKED]] 47define void @s_pack_v2f16_imm_hi(i32 addrspace(2)* %in0) #0 { 48 %val0 = load i32, i32 addrspace(2)* %in0 49 %lo.i = trunc i32 %val0 to i16 50 %lo = bitcast i16 %lo.i to half 51 %vec.0 = insertelement <2 x half> undef, half %lo, i32 0 52 %vec.1 = insertelement <2 x half> %vec.0, half 0xH1234, i32 1 53 %vec.i32 = bitcast <2 x half> %vec.1 to i32 54 55 call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0 56 ret void 57} 58 59; GCN-LABEL: {{^}}v_pack_v2f16: 60; GFX9: flat_load_dword [[VAL0:v[0-9]+]] 61; GFX9: flat_load_dword [[VAL1:v[0-9]+]] 62; GFX9-DENORM: v_pack_b32_f16 [[PACKED:v[0-9]+]], [[VAL0]], [[VAL1]] 63 64; GFX9-FLUSH: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VAL0]] 65; GFX9-FLUSH: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[ELT0]] 66; GFX9: ; use [[PACKED]] 67define void @v_pack_v2f16(i32 addrspace(1)* %in0, i32 addrspace(1)* %in1) #0 { 68 %tid = call i32 @llvm.amdgcn.workitem.id.x() 69 %tid.ext = sext i32 %tid to i64 70 %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext 71 %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext 72 %val0 = load volatile i32, i32 addrspace(1)* %in0.gep 73 %val1 = load volatile i32, i32 addrspace(1)* %in1.gep 74 %lo.i = trunc i32 %val0 to i16 75 %hi.i = trunc i32 %val1 to i16 76 %lo = bitcast i16 %lo.i to half 77 %hi = bitcast i16 %hi.i to half 78 %vec.0 = insertelement <2 x half> undef, half %lo, i32 0 79 %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1 80 %vec.i32 = bitcast <2 x half> %vec.1 to i32 81 call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0 82 ret void 83} 84 85; GCN-LABEL: {{^}}v_pack_v2f16_user: 86; GFX9: flat_load_dword [[VAL0:v[0-9]+]] 87; GFX9: flat_load_dword [[VAL1:v[0-9]+]] 88; GFX9-DENORM: v_pack_b32_f16 [[PACKED:v[0-9]+]], [[VAL0]], [[VAL1]] 89 90; GFX9-FLUSH: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VAL0]] 91; GFX9-FLUSH: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[ELT0]] 92 93; GFX9: v_add_i32_e32 v{{[0-9]+}}, vcc, 9, [[PACKED]] 94define void @v_pack_v2f16_user(i32 addrspace(1)* %in0, i32 addrspace(1)* %in1) #0 { 95 %tid = call i32 @llvm.amdgcn.workitem.id.x() 96 %tid.ext = sext i32 %tid to i64 97 %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext 98 %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext 99 %val0 = load volatile i32, i32 addrspace(1)* %in0.gep 100 %val1 = load volatile i32, i32 addrspace(1)* %in1.gep 101 %lo.i = trunc i32 %val0 to i16 102 %hi.i = trunc i32 %val1 to i16 103 %lo = bitcast i16 %lo.i to half 104 %hi = bitcast i16 %hi.i to half 105 %vec.0 = insertelement <2 x half> undef, half %lo, i32 0 106 %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1 107 %vec.i32 = bitcast <2 x half> %vec.1 to i32 108 %foo = add i32 %vec.i32, 9 109 store volatile i32 %foo, i32 addrspace(1)* undef 110 ret void 111} 112 113; GCN-LABEL: {{^}}v_pack_v2f16_imm_lo: 114; GFX9: flat_load_dword [[VAL1:v[0-9]+]] 115; GFX9-DENORM: s_movk_i32 [[K:s[0-9]+]], 0x1234{{$}} 116; GFX9-DENORM: v_pack_b32_f16 [[PACKED:v[0-9]+]], [[K]], [[VAL1]] 117 118; GFX9-FLUSH: v_mov_b32_e32 [[K:v[0-9]+]], 0x1234{{$}} 119; GFX9-FLUSH: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[K]] 120; GFX9: ; use [[PACKED]] 121define void @v_pack_v2f16_imm_lo(i32 addrspace(1)* %in1) #0 { 122 %tid = call i32 @llvm.amdgcn.workitem.id.x() 123 %tid.ext = sext i32 %tid to i64 124 %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext 125 %val1 = load volatile i32, i32 addrspace(1)* %in1.gep 126 %hi.i = trunc i32 %val1 to i16 127 %hi = bitcast i16 %hi.i to half 128 %vec.0 = insertelement <2 x half> undef, half 0xH1234, i32 0 129 %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1 130 %vec.i32 = bitcast <2 x half> %vec.1 to i32 131 call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0 132 ret void 133} 134 135; GCN-LABEL: {{^}}v_pack_v2f16_inline_imm_lo: 136; GFX9: flat_load_dword [[VAL1:v[0-9]+]] 137; GFX9-DENORM: v_pack_b32_f16 [[PACKED:v[0-9]+]], 4.0, [[VAL1]] 138 139; GFX9-FLUSH: v_mov_b32_e32 [[K:v[0-9]+]], 0x4400{{$}} 140; GFX9-FLUSH: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[K]] 141 142; GFX9: ; use [[PACKED]] 143define void @v_pack_v2f16_inline_imm_lo(i32 addrspace(1)* %in1) #0 { 144 %tid = call i32 @llvm.amdgcn.workitem.id.x() 145 %tid.ext = sext i32 %tid to i64 146 %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext 147 %val1 = load volatile i32, i32 addrspace(1)* %in1.gep 148 %hi.i = trunc i32 %val1 to i16 149 %hi = bitcast i16 %hi.i to half 150 %vec.0 = insertelement <2 x half> undef, half 4.0, i32 0 151 %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1 152 %vec.i32 = bitcast <2 x half> %vec.1 to i32 153 call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0 154 ret void 155} 156 157; GCN-LABEL: {{^}}v_pack_v2f16_imm_hi: 158; GFX9: flat_load_dword [[VAL0:v[0-9]+]] 159; GFX9-DENORM: s_movk_i32 [[K:s[0-9]+]], 0x1234 160; GFX9-DENORM: v_pack_b32_f16 [[PACKED:v[0-9]+]], [[VAL0]], [[K]] 161 162; GFX9-FLUSH: s_movk_i32 [[K:s[0-9]+]], 0x1234 163; GFX9-FLUSH: v_and_b32_e32 [[MASKED:v[0-9]+]], 0xffff, [[VAL0]] 164; GFX9-FLUSH: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[K]], 16, [[MASKED]] 165 166; GFX9: ; use [[PACKED]] 167define void @v_pack_v2f16_imm_hi(i32 addrspace(1)* %in0) #0 { 168 %tid = call i32 @llvm.amdgcn.workitem.id.x() 169 %tid.ext = sext i32 %tid to i64 170 %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext 171 %val0 = load volatile i32, i32 addrspace(1)* %in0.gep 172 %lo.i = trunc i32 %val0 to i16 173 %lo = bitcast i16 %lo.i to half 174 %vec.0 = insertelement <2 x half> undef, half %lo, i32 0 175 %vec.1 = insertelement <2 x half> %vec.0, half 0xH1234, i32 1 176 %vec.i32 = bitcast <2 x half> %vec.1 to i32 177 call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0 178 ret void 179} 180 181; GCN-LABEL: {{^}}v_pack_v2f16_inline_f16imm_hi: 182; GFX9: flat_load_dword [[VAL:v[0-9]+]] 183; GFX9-DENORM: v_pack_b32_f16 [[PACKED:v[0-9]+]], [[VAL]], 1.0 184 185; GFX9-FLUSH: s_movk_i32 [[K:s[0-9]+]], 0x3c00 186; GFX9-FLUSH: v_and_b32_e32 [[MASKED:v[0-9]+]], 0xffff, [[VAL]] 187; GFX9-FLUSH: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[K]], 16, [[MASKED]] 188 189; GFX9: ; use [[PACKED]] 190define void @v_pack_v2f16_inline_f16imm_hi(i32 addrspace(1)* %in0) #0 { 191 %tid = call i32 @llvm.amdgcn.workitem.id.x() 192 %tid.ext = sext i32 %tid to i64 193 %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext 194 %val0 = load volatile i32, i32 addrspace(1)* %in0.gep 195 %lo.i = trunc i32 %val0 to i16 196 %lo = bitcast i16 %lo.i to half 197 %vec.0 = insertelement <2 x half> undef, half %lo, i32 0 198 %vec.1 = insertelement <2 x half> %vec.0, half 1.0, i32 1 199 %vec.i32 = bitcast <2 x half> %vec.1 to i32 200 call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0 201 ret void 202} 203 204; GCN-LABEL: {{^}}v_pack_v2f16_inline_imm_hi: 205; GFX9: flat_load_dword [[VAL:v[0-9]+]] 206; GFX9-DENORM: v_pack_b32_f16 [[PACKED:v[0-9]+]], [[VAL]], 64 207 208; GFX9-FLUSH: v_and_b32_e32 [[MASKED:v[0-9]+]], 0xffff, [[VAL]] 209; GFX9-FLUSH: v_lshl_or_b32 [[PACKED:v[0-9]+]], 64, 16, [[MASKED]] 210 211; GFX9: ; use [[PACKED]] 212define void @v_pack_v2f16_inline_imm_hi(i32 addrspace(1)* %in0) #0 { 213 %tid = call i32 @llvm.amdgcn.workitem.id.x() 214 %tid.ext = sext i32 %tid to i64 215 %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext 216 %val0 = load volatile i32, i32 addrspace(1)* %in0.gep 217 %lo.i = trunc i32 %val0 to i16 218 %lo = bitcast i16 %lo.i to half 219 %vec.0 = insertelement <2 x half> undef, half %lo, i32 0 220 %vec.1 = insertelement <2 x half> %vec.0, half 0xH0040, i32 1 221 %vec.i32 = bitcast <2 x half> %vec.1 to i32 222 call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0 223 ret void 224} 225 226declare i32 @llvm.amdgcn.workitem.id.x() #1 227 228attributes #0 = { nounwind } 229attributes #1 = { nounwind readnone } 230