1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,SI %s 2 ; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89 %s 3 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX89,GFX9 %s 4 5 ; GCN-LABEL: {{^}}v_clamp_add_src_f32: 6 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 7 ; GCN-NOT: [[A]] 8 ; GCN: v_add_f32_e64 v{{[0-9]+}}, [[A]], 1.0 clamp{{$}} 9 define amdgpu_kernel void @v_clamp_add_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 { 10 %tid = call i32 @llvm.amdgcn.workitem.id.x() 11 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid 12 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 13 %a = load float, float addrspace(1)* %gep0 14 %add = fadd float %a, 1.0 15 %max = call float @llvm.maxnum.f32(float %add, float 0.0) 16 %clamp = call float @llvm.minnum.f32(float %max, float 1.0) 17 store float %clamp, float addrspace(1)* %out.gep 18 ret void 19 } 20 21 ; GCN-LABEL: {{^}}v_clamp_multi_use_src_f32: 22 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 23 ; GCN: v_add_f32_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}} 24 ; GCN: v_max_f32_e64 v{{[0-9]+}}, [[ADD]], [[ADD]] clamp{{$}} 25 define amdgpu_kernel void @v_clamp_multi_use_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 { 26 %tid = call i32 @llvm.amdgcn.workitem.id.x() 27 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid 28 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 29 %a = load float, float addrspace(1)* %gep0 30 %add = fadd float %a, 1.0 31 %max = call float @llvm.maxnum.f32(float %add, float 0.0) 32 %clamp = call float @llvm.minnum.f32(float %max, float 1.0) 33 store float %clamp, float addrspace(1)* %out.gep 34 store volatile float %add, float addrspace(1)* undef 35 ret void 36 } 37 38 ; GCN-LABEL: {{^}}v_clamp_dbg_use_src_f32: 39 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 40 ; GCN-NOT: [[A]] 41 ; GCN: v_add_f32_e64 v{{[0-9]+}}, [[A]], 1.0 clamp{{$}} 42 define amdgpu_kernel void @v_clamp_dbg_use_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 { 43 %tid = call i32 @llvm.amdgcn.workitem.id.x() 44 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid 45 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 46 %a = load float, float addrspace(1)* %gep0 47 %add = fadd float %a, 1.0 48 call void @llvm.dbg.value(metadata float %add, i64 0, metadata !4, metadata !9), !dbg !10 49 %max = call float @llvm.maxnum.f32(float %add, float 0.0) 50 %clamp = call float @llvm.minnum.f32(float %max, float 1.0) 51 store float %clamp, float addrspace(1)* %out.gep 52 ret void 53 } 54 55 ; GCN-LABEL: {{^}}v_clamp_add_neg_src_f32: 56 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 57 ; GCN: v_floor_f32_e32 [[FLOOR:v[0-9]+]], [[A]] 58 ; GCN: v_max_f32_e64 v{{[0-9]+}}, -[[FLOOR]], -[[FLOOR]] clamp{{$}} 59 define amdgpu_kernel void @v_clamp_add_neg_src_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 { 60 %tid = call i32 @llvm.amdgcn.workitem.id.x() 61 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid 62 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 63 %a = load float, float addrspace(1)* %gep0 64 %floor = call float @llvm.floor.f32(float %a) 65 %neg.floor = fsub float -0.0, %floor 66 %max = call float @llvm.maxnum.f32(float %neg.floor, float 0.0) 67 %clamp = call float @llvm.minnum.f32(float %max, float 1.0) 68 store float %clamp, float addrspace(1)* %out.gep 69 ret void 70 } 71 72 ; GCN-LABEL: {{^}}v_non_clamp_max_f32: 73 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 74 ; GCN: v_add_f32_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}} 75 ; GCN: v_max_f32_e32 v{{[0-9]+}}, 0, [[ADD]]{{$}} 76 define amdgpu_kernel void @v_non_clamp_max_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 { 77 %tid = call i32 @llvm.amdgcn.workitem.id.x() 78 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid 79 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 80 %a = load float, float addrspace(1)* %gep0 81 %add = fadd float %a, 1.0 82 %max = call float @llvm.maxnum.f32(float %add, float 0.0) 83 store float %max, float addrspace(1)* %out.gep 84 ret void 85 } 86 87 ; GCN-LABEL: {{^}}v_clamp_add_src_f32_denormals: 88 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 89 ; GCN: v_add_f32_e64 [[ADD:v[0-9]+]], [[A]], 1.0 clamp{{$}} 90 define amdgpu_kernel void @v_clamp_add_src_f32_denormals(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 { 91 %tid = call i32 @llvm.amdgcn.workitem.id.x() 92 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid 93 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 94 %a = load float, float addrspace(1)* %gep0 95 %add = fadd float %a, 1.0 96 %max = call float @llvm.maxnum.f32(float %add, float 0.0) 97 %clamp = call float @llvm.minnum.f32(float %max, float 1.0) 98 store float %clamp, float addrspace(1)* %out.gep 99 ret void 100 } 101 102 ; GCN-LABEL: {{^}}v_clamp_add_src_f16_denorm: 103 ; GCN: {{buffer|flat|global}}_load_ushort [[A:v[0-9]+]] 104 ; GFX89: v_add_f16_e64 [[ADD:v[0-9]+]], [[A]], 1.0 clamp{{$}} 105 106 ; SI: v_cvt_f32_f16_e32 [[CVT:v[0-9]+]], [[A]] 107 ; SI: v_add_f32_e64 [[ADD:v[0-9]+]], [[CVT]], 1.0 clamp{{$}} 108 ; SI: v_cvt_f16_f32_e32 v{{[0-9]+}}, [[ADD]] 109 define amdgpu_kernel void @v_clamp_add_src_f16_denorm(half addrspace(1)* %out, half addrspace(1)* %aptr) #0 { 110 %tid = call i32 @llvm.amdgcn.workitem.id.x() 111 %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid 112 %out.gep = getelementptr half, half addrspace(1)* %out, i32 %tid 113 %a = load half, half addrspace(1)* %gep0 114 %add = fadd half %a, 1.0 115 %max = call half @llvm.maxnum.f16(half %add, half 0.0) 116 %clamp = call half @llvm.minnum.f16(half %max, half 1.0) 117 store half %clamp, half addrspace(1)* %out.gep 118 ret void 119 } 120 121 ; GCN-LABEL: {{^}}v_clamp_add_src_f16_no_denormals: 122 ; GCN: {{buffer|flat|global}}_load_ushort [[A:v[0-9]+]] 123 ; GFX89-NOT: [[A]] 124 ; GFX89: v_add_f16_e64 v{{[0-9]+}}, [[A]], 1.0 clamp{{$}} 125 126 ; SI: v_cvt_f32_f16_e32 [[CVT:v[0-9]+]], [[A]] 127 ; SI: v_add_f32_e64 [[ADD:v[0-9]+]], [[CVT]], 1.0 clamp{{$}} 128 ; SI: v_cvt_f16_f32_e32 v{{[0-9]+}}, [[ADD]] 129 define amdgpu_kernel void @v_clamp_add_src_f16_no_denormals(half addrspace(1)* %out, half addrspace(1)* %aptr) #3 { 130 %tid = call i32 @llvm.amdgcn.workitem.id.x() 131 %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid 132 %out.gep = getelementptr half, half addrspace(1)* %out, i32 %tid 133 %a = load half, half addrspace(1)* %gep0 134 %add = fadd half %a, 1.0 135 %max = call half @llvm.maxnum.f16(half %add, half 0.0) 136 %clamp = call half @llvm.minnum.f16(half %max, half 1.0) 137 store half %clamp, half addrspace(1)* %out.gep 138 ret void 139 } 140 141 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f32: 142 ; GCN: {{buffer|flat|global}}_load_dwordx2 v{{\[}}[[A:[0-9]+]]:[[B:[0-9]+]]{{\]}} 143 ; GCN-DAG: v_add_f32_e64 v{{[0-9]+}}, v[[A]], 1.0 clamp{{$}} 144 ; GCN-DAG: v_add_f32_e64 v{{[0-9]+}}, v[[B]], 1.0 clamp{{$}} 145 define amdgpu_kernel void @v_clamp_add_src_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %aptr) #0 { 146 %tid = call i32 @llvm.amdgcn.workitem.id.x() 147 %gep0 = getelementptr <2 x float>, <2 x float> addrspace(1)* %aptr, i32 %tid 148 %out.gep = getelementptr <2 x float>, <2 x float> addrspace(1)* %out, i32 %tid 149 %a = load <2 x float>, <2 x float> addrspace(1)* %gep0 150 %add = fadd <2 x float> %a, <float 1.0, float 1.0> 151 %max = call <2 x float> @llvm.maxnum.v2f32(<2 x float> %add, <2 x float> zeroinitializer) 152 %clamp = call <2 x float> @llvm.minnum.v2f32(<2 x float> %max, <2 x float> <float 1.0, float 1.0>) 153 store <2 x float> %clamp, <2 x float> addrspace(1)* %out.gep 154 ret void 155 } 156 157 ; GCN-LABEL: {{^}}v_clamp_add_src_f64: 158 ; GCN: {{buffer|flat|global}}_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]] 159 ; GCN: v_add_f64 v{{\[[0-9]+:[0-9]+\]}}, [[A]], 1.0 clamp{{$}} 160 define amdgpu_kernel void @v_clamp_add_src_f64(double addrspace(1)* %out, double addrspace(1)* %aptr) #0 { 161 %tid = call i32 @llvm.amdgcn.workitem.id.x() 162 %gep0 = getelementptr double, double addrspace(1)* %aptr, i32 %tid 163 %out.gep = getelementptr double, double addrspace(1)* %out, i32 %tid 164 %a = load double, double addrspace(1)* %gep0 165 %add = fadd double %a, 1.0 166 %max = call double @llvm.maxnum.f64(double %add, double 0.0) 167 %clamp = call double @llvm.minnum.f64(double %max, double 1.0) 168 store double %clamp, double addrspace(1)* %out.gep 169 ret void 170 } 171 172 ; GCN-LABEL: {{^}}v_clamp_mac_to_mad: 173 ; GCN: v_mad_f32 v{{[0-9]+}}, s{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]}} clamp{{$}} 174 define amdgpu_kernel void @v_clamp_mac_to_mad(float addrspace(1)* %out, float addrspace(1)* %aptr, float %a) #0 { 175 %tid = call i32 @llvm.amdgcn.workitem.id.x() 176 %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid 177 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 178 %b = load float, float addrspace(1)* %gep0 179 180 %mul = fmul float %a, %a 181 %add = fadd float %mul, %b 182 %max = call float @llvm.maxnum.f32(float %add, float 0.0) 183 %clamp = call float @llvm.minnum.f32(float %max, float 1.0) 184 %res = fadd float %clamp, %b 185 store float %res, float addrspace(1)* %out.gep 186 ret void 187 } 188 189 190 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm: 191 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 192 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], 1.0 op_sel_hi:[1,0] clamp{{$}} 193 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 { 194 %tid = call i32 @llvm.amdgcn.workitem.id.x() 195 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 196 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 197 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 198 %add = fadd <2 x half> %a, <half 1.0, half 1.0> 199 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %add, <2 x half> zeroinitializer) 200 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 201 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 202 ret void 203 } 204 205 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_no_denormals: 206 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 207 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], 1.0 op_sel_hi:[1,0] clamp{{$}} 208 define amdgpu_kernel void @v_clamp_add_src_v2f16_no_denormals(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #3 { 209 %tid = call i32 @llvm.amdgcn.workitem.id.x() 210 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 211 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 212 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 213 %add = fadd <2 x half> %a, <half 1.0, half 1.0> 214 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %add, <2 x half> zeroinitializer) 215 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 216 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 217 ret void 218 } 219 220 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_neg: 221 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 222 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], 1.0 op_sel_hi:[1,0]{{$}} 223 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] neg_lo:[1,1] neg_hi:[1,1] clamp{{$}} 224 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_neg(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 { 225 %tid = call i32 @llvm.amdgcn.workitem.id.x() 226 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 227 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 228 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 229 %add = fadd <2 x half> %a, <half 1.0, half 1.0> 230 %neg.add = fsub <2 x half> <half -0.0, half -0.0>, %add 231 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %neg.add, <2 x half> zeroinitializer) 232 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 233 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 234 ret void 235 } 236 237 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_neg_lo: 238 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 239 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], 1.0 op_sel_hi:[1,0]{{$}} 240 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] neg_lo:[1,1] clamp{{$}} 241 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_neg_lo(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 { 242 %tid = call i32 @llvm.amdgcn.workitem.id.x() 243 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 244 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 245 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 246 %add = fadd <2 x half> %a, <half 1.0, half 1.0> 247 %lo = extractelement <2 x half> %add, i32 0 248 %neg.lo = fsub half -0.0, %lo 249 %neg.lo.add = insertelement <2 x half> %add, half %neg.lo, i32 0 250 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %neg.lo.add, <2 x half> zeroinitializer) 251 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 252 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 253 ret void 254 } 255 256 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_neg_hi: 257 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 258 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], 1.0 op_sel_hi:[1,0]{{$}} 259 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] neg_hi:[1,1] clamp{{$}} 260 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_neg_hi(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 { 261 %tid = call i32 @llvm.amdgcn.workitem.id.x() 262 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 263 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 264 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 265 %add = fadd <2 x half> %a, <half 1.0, half 1.0> 266 %hi = extractelement <2 x half> %add, i32 1 267 %neg.hi = fsub half -0.0, %hi 268 %neg.hi.add = insertelement <2 x half> %add, half %neg.hi, i32 1 269 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %neg.hi.add, <2 x half> zeroinitializer) 270 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 271 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 272 ret void 273 } 274 275 ; GCN-LABEL: {{^}}v_clamp_add_src_v2f16_denorm_shuf: 276 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 277 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], 1.0 op_sel_hi:[1,0]{{$}} 278 ; GFX9: v_pk_max_f16 [[MAX:v[0-9]+]], [[ADD]], [[ADD]] op_sel:[1,1] op_sel_hi:[0,0] clamp{{$}} 279 define amdgpu_kernel void @v_clamp_add_src_v2f16_denorm_shuf(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 { 280 %tid = call i32 @llvm.amdgcn.workitem.id.x() 281 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 282 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 283 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 284 %add = fadd <2 x half> %a, <half 1.0, half 1.0> 285 %shuf = shufflevector <2 x half> %add, <2 x half> undef, <2 x i32> <i32 1, i32 0> 286 287 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %shuf, <2 x half> zeroinitializer) 288 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 289 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 290 ret void 291 } 292 293 ; GCN-LABEL: {{^}}v_no_clamp_add_src_v2f16_f32_src: 294 ; GCN-DAG: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 295 ; GFX9: v_add_f32_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}} 296 ; GFX9: v_pk_max_f16 [[CLAMP:v[0-9]+]], [[ADD]], [[ADD]] clamp{{$}} 297 define amdgpu_kernel void @v_no_clamp_add_src_v2f16_f32_src(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 { 298 %tid = call i32 @llvm.amdgcn.workitem.id.x() 299 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 300 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 301 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 302 %bc = bitcast <2 x half> %a to float 303 %f32.op = fadd float %bc, 1.0 304 %f32.op.cast = bitcast float %f32.op to <2 x half> 305 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %f32.op.cast, <2 x half> zeroinitializer) 306 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 307 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 308 ret void 309 } 310 311 ; GCN-LABEL: {{^}}v_no_clamp_add_packed_src_f32: 312 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]] 313 ; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[A]], 1.0 op_sel_hi:[1,0]{{$}} 314 ; GFX9: v_max_f32_e64 [[CLAMP:v[0-9]+]], [[ADD]], [[ADD]] clamp{{$}} 315 define amdgpu_kernel void @v_no_clamp_add_packed_src_f32(float addrspace(1)* %out, <2 x half> addrspace(1)* %aptr) #0 { 316 %tid = call i32 @llvm.amdgcn.workitem.id.x() 317 %gep0 = getelementptr <2 x half>, <2 x half> addrspace(1)* %aptr, i32 %tid 318 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid 319 %a = load <2 x half>, <2 x half> addrspace(1)* %gep0 320 %add = fadd <2 x half> %a, <half 1.0, half 1.0> 321 %bc.add = bitcast <2 x half> %add to float 322 %max = call float @llvm.maxnum.f32(float %bc.add, float 0.0) 323 %clamp = call float @llvm.minnum.f32(float %max, float 1.0) 324 store float %clamp, float addrspace(1)* %out.gep 325 ret void 326 } 327 328 ; Since the high bits are zeroed, it probably would be OK in this case 329 ; to use clamp. 330 ; GCN-LABEL: {{^}}v_no_clamp_add_src_v2f16_f16_src: 331 ; GCN-DAG: {{buffer|flat|global}}_load_ushort [[A:v[0-9]+]] 332 ; GFX9: v_add_f16_e32 [[ADD:v[0-9]+]], 1.0, [[A]]{{$}} 333 ; GFX9: v_pk_max_f16 [[CLAMP:v[0-9]+]], [[ADD]], [[ADD]] clamp{{$}} 334 define amdgpu_kernel void @v_no_clamp_add_src_v2f16_f16_src(<2 x half> addrspace(1)* %out, half addrspace(1)* %aptr) #0 { 335 %tid = call i32 @llvm.amdgcn.workitem.id.x() 336 %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid 337 %out.gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid 338 %a = load half, half addrspace(1)* %gep0 339 %add = fadd half %a, 1.0 340 %bc = bitcast half %add to i16 341 %zext = zext i16 %bc to i32 342 %v2f16 = bitcast i32 %zext to <2 x half> 343 %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %v2f16, <2 x half> zeroinitializer) 344 %clamp = call <2 x half> @llvm.minnum.v2f16(<2 x half> %max, <2 x half> <half 1.0, half 1.0>) 345 store <2 x half> %clamp, <2 x half> addrspace(1)* %out.gep 346 ret void 347 } 348 349 declare i32 @llvm.amdgcn.workitem.id.x() #1 350 declare float @llvm.fabs.f32(float) #1 351 declare float @llvm.floor.f32(float) #1 352 declare float @llvm.minnum.f32(float, float) #1 353 declare float @llvm.maxnum.f32(float, float) #1 354 declare float @llvm.amdgcn.fmed3.f32(float, float, float) #1 355 declare double @llvm.fabs.f64(double) #1 356 declare double @llvm.minnum.f64(double, double) #1 357 declare double @llvm.maxnum.f64(double, double) #1 358 declare half @llvm.fabs.f16(half) #1 359 declare half @llvm.minnum.f16(half, half) #1 360 declare half @llvm.maxnum.f16(half, half) #1 361 declare <2 x half> @llvm.minnum.v2f16(<2 x half>, <2 x half>) #1 362 declare <2 x half> @llvm.maxnum.v2f16(<2 x half>, <2 x half>) #1 363 declare <2 x float> @llvm.minnum.v2f32(<2 x float>, <2 x float>) #1 364 declare <2 x float> @llvm.maxnum.v2f32(<2 x float>, <2 x float>) #1 365 366 367 declare void @llvm.dbg.value(metadata, i64, metadata, metadata) #1 368 369 attributes #0 = { nounwind } 370 attributes #1 = { nounwind readnone } 371 attributes #2 = { nounwind "target-features"="+fp32-denormals" } 372 attributes #3 = { nounwind "target-features"="-fp64-fp16-denormals" } 373 374 !llvm.dbg.cu = !{!0} 375 !llvm.module.flags = !{!2, !3} 376 377 !0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, isOptimized: true, runtimeVersion: 0, emissionKind: NoDebug) 378 !1 = !DIFile(filename: "/tmp/foo.cl", directory: "/dev/null") 379 !2 = !{i32 2, !"Dwarf Version", i32 4} 380 !3 = !{i32 2, !"Debug Info Version", i32 3} 381 !4 = !DILocalVariable(name: "add", arg: 1, scope: !5, file: !1, line: 1) 382 !5 = distinct !DISubprogram(name: "foo", scope: !1, file: !1, line: 1, type: !6, isLocal: false, isDefinition: true, scopeLine: 2, flags: DIFlagPrototyped, isOptimized: true, unit: !0) 383 !6 = !DISubroutineType(types: !7) 384 !7 = !{null, !8} 385 !8 = !DIBasicType(name: "float", size: 32, align: 32) 386 !9 = !DIExpression() 387 !10 = !DILocation(line: 1, column: 42, scope: !5) 388
