1 ; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s 2 3 declare i32 @llvm.r600.read.tidig.x() #0 4 declare float @llvm.fabs.f32(float) #0 5 6 ; FUNC-LABEL: {{^}}mad_sub_f32: 7 ; SI: buffer_load_dword [[REGA:v[0-9]+]] 8 ; SI: buffer_load_dword [[REGB:v[0-9]+]] 9 ; SI: buffer_load_dword [[REGC:v[0-9]+]] 10 ; SI: v_mad_f32 [[RESULT:v[0-9]+]], [[REGA]], [[REGB]], -[[REGC]] 11 ; SI: buffer_store_dword [[RESULT]] 12 define void @mad_sub_f32(float addrspace(1)* noalias nocapture %out, float addrspace(1)* noalias nocapture readonly %ptr) #1 { 13 %tid = tail call i32 @llvm.r600.read.tidig.x() #0 14 %tid.ext = sext i32 %tid to i64 15 %gep0 = getelementptr float, float addrspace(1)* %ptr, i64 %tid.ext 16 %add1 = add i64 %tid.ext, 1 17 %gep1 = getelementptr float, float addrspace(1)* %ptr, i64 %add1 18 %add2 = add i64 %tid.ext, 2 19 %gep2 = getelementptr float, float addrspace(1)* %ptr, i64 %add2 20 %outgep = getelementptr float, float addrspace(1)* %out, i64 %tid.ext 21 %a = load float, float addrspace(1)* %gep0, align 4 22 %b = load float, float addrspace(1)* %gep1, align 4 23 %c = load float, float addrspace(1)* %gep2, align 4 24 %mul = fmul float %a, %b 25 %sub = fsub float %mul, %c 26 store float %sub, float addrspace(1)* %outgep, align 4 27 ret void 28 } 29 30 ; FUNC-LABEL: {{^}}mad_sub_inv_f32: 31 ; SI: buffer_load_dword [[REGA:v[0-9]+]] 32 ; SI: buffer_load_dword [[REGB:v[0-9]+]] 33 ; SI: buffer_load_dword [[REGC:v[0-9]+]] 34 ; SI: v_mad_f32 [[RESULT:v[0-9]+]], -[[REGA]], [[REGB]], [[REGC]] 35 ; SI: buffer_store_dword [[RESULT]] 36 define void @mad_sub_inv_f32(float addrspace(1)* noalias nocapture %out, float addrspace(1)* noalias nocapture readonly %ptr) #1 { 37 %tid = tail call i32 @llvm.r600.read.tidig.x() #0 38 %tid.ext = sext i32 %tid to i64 39 %gep0 = getelementptr float, float addrspace(1)* %ptr, i64 %tid.ext 40 %add1 = add i64 %tid.ext, 1 41 %gep1 = getelementptr float, float addrspace(1)* %ptr, i64 %add1 42 %add2 = add i64 %tid.ext, 2 43 %gep2 = getelementptr float, float addrspace(1)* %ptr, i64 %add2 44 %outgep = getelementptr float, float addrspace(1)* %out, i64 %tid.ext 45 %a = load float, float addrspace(1)* %gep0, align 4 46 %b = load float, float addrspace(1)* %gep1, align 4 47 %c = load float, float addrspace(1)* %gep2, align 4 48 %mul = fmul float %a, %b 49 %sub = fsub float %c, %mul 50 store float %sub, float addrspace(1)* %outgep, align 4 51 ret void 52 } 53 54 ; FUNC-LABEL: {{^}}mad_sub_f64: 55 ; SI: v_mul_f64 56 ; SI: v_add_f64 57 define void @mad_sub_f64(double addrspace(1)* noalias nocapture %out, double addrspace(1)* noalias nocapture readonly %ptr) #1 { 58 %tid = tail call i32 @llvm.r600.read.tidig.x() #0 59 %tid.ext = sext i32 %tid to i64 60 %gep0 = getelementptr double, double addrspace(1)* %ptr, i64 %tid.ext 61 %add1 = add i64 %tid.ext, 1 62 %gep1 = getelementptr double, double addrspace(1)* %ptr, i64 %add1 63 %add2 = add i64 %tid.ext, 2 64 %gep2 = getelementptr double, double addrspace(1)* %ptr, i64 %add2 65 %outgep = getelementptr double, double addrspace(1)* %out, i64 %tid.ext 66 %a = load double, double addrspace(1)* %gep0, align 8 67 %b = load double, double addrspace(1)* %gep1, align 8 68 %c = load double, double addrspace(1)* %gep2, align 8 69 %mul = fmul double %a, %b 70 %sub = fsub double %mul, %c 71 store double %sub, double addrspace(1)* %outgep, align 8 72 ret void 73 } 74 75 ; FUNC-LABEL: {{^}}mad_sub_fabs_f32: 76 ; SI: buffer_load_dword [[REGA:v[0-9]+]] 77 ; SI: buffer_load_dword [[REGB:v[0-9]+]] 78 ; SI: buffer_load_dword [[REGC:v[0-9]+]] 79 ; SI: v_mad_f32 [[RESULT:v[0-9]+]], [[REGA]], [[REGB]], -|[[REGC]]| 80 ; SI: buffer_store_dword [[RESULT]] 81 define void @mad_sub_fabs_f32(float addrspace(1)* noalias nocapture %out, float addrspace(1)* noalias nocapture readonly %ptr) #1 { 82 %tid = tail call i32 @llvm.r600.read.tidig.x() #0 83 %tid.ext = sext i32 %tid to i64 84 %gep0 = getelementptr float, float addrspace(1)* %ptr, i64 %tid.ext 85 %add1 = add i64 %tid.ext, 1 86 %gep1 = getelementptr float, float addrspace(1)* %ptr, i64 %add1 87 %add2 = add i64 %tid.ext, 2 88 %gep2 = getelementptr float, float addrspace(1)* %ptr, i64 %add2 89 %outgep = getelementptr float, float addrspace(1)* %out, i64 %tid.ext 90 %a = load float, float addrspace(1)* %gep0, align 4 91 %b = load float, float addrspace(1)* %gep1, align 4 92 %c = load float, float addrspace(1)* %gep2, align 4 93 %c.abs = call float @llvm.fabs.f32(float %c) #0 94 %mul = fmul float %a, %b 95 %sub = fsub float %mul, %c.abs 96 store float %sub, float addrspace(1)* %outgep, align 4 97 ret void 98 } 99 100 ; FUNC-LABEL: {{^}}mad_sub_fabs_inv_f32: 101 ; SI: buffer_load_dword [[REGA:v[0-9]+]] 102 ; SI: buffer_load_dword [[REGB:v[0-9]+]] 103 ; SI: buffer_load_dword [[REGC:v[0-9]+]] 104 ; SI: v_mad_f32 [[RESULT:v[0-9]+]], -[[REGA]], [[REGB]], |[[REGC]]| 105 ; SI: buffer_store_dword [[RESULT]] 106 define void @mad_sub_fabs_inv_f32(float addrspace(1)* noalias nocapture %out, float addrspace(1)* noalias nocapture readonly %ptr) #1 { 107 %tid = tail call i32 @llvm.r600.read.tidig.x() #0 108 %tid.ext = sext i32 %tid to i64 109 %gep0 = getelementptr float, float addrspace(1)* %ptr, i64 %tid.ext 110 %add1 = add i64 %tid.ext, 1 111 %gep1 = getelementptr float, float addrspace(1)* %ptr, i64 %add1 112 %add2 = add i64 %tid.ext, 2 113 %gep2 = getelementptr float, float addrspace(1)* %ptr, i64 %add2 114 %outgep = getelementptr float, float addrspace(1)* %out, i64 %tid.ext 115 %a = load float, float addrspace(1)* %gep0, align 4 116 %b = load float, float addrspace(1)* %gep1, align 4 117 %c = load float, float addrspace(1)* %gep2, align 4 118 %c.abs = call float @llvm.fabs.f32(float %c) #0 119 %mul = fmul float %a, %b 120 %sub = fsub float %c.abs, %mul 121 store float %sub, float addrspace(1)* %outgep, align 4 122 ret void 123 } 124 125 ; FUNC-LABEL: {{^}}neg_neg_mad_f32: 126 ; SI: v_mac_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}} 127 define void @neg_neg_mad_f32(float addrspace(1)* noalias nocapture %out, float addrspace(1)* noalias nocapture readonly %ptr) #1 { 128 %tid = tail call i32 @llvm.r600.read.tidig.x() #0 129 %tid.ext = sext i32 %tid to i64 130 %gep0 = getelementptr float, float addrspace(1)* %ptr, i64 %tid.ext 131 %add1 = add i64 %tid.ext, 1 132 %gep1 = getelementptr float, float addrspace(1)* %ptr, i64 %add1 133 %add2 = add i64 %tid.ext, 2 134 %gep2 = getelementptr float, float addrspace(1)* %ptr, i64 %add2 135 %outgep = getelementptr float, float addrspace(1)* %out, i64 %tid.ext 136 %a = load float, float addrspace(1)* %gep0, align 4 137 %b = load float, float addrspace(1)* %gep1, align 4 138 %c = load float, float addrspace(1)* %gep2, align 4 139 %nega = fsub float -0.000000e+00, %a 140 %negb = fsub float -0.000000e+00, %b 141 %mul = fmul float %nega, %negb 142 %sub = fadd float %mul, %c 143 store float %sub, float addrspace(1)* %outgep, align 4 144 ret void 145 } 146 147 ; FUNC-LABEL: {{^}}mad_fabs_sub_f32: 148 ; SI: buffer_load_dword [[REGA:v[0-9]+]] 149 ; SI: buffer_load_dword [[REGB:v[0-9]+]] 150 ; SI: buffer_load_dword [[REGC:v[0-9]+]] 151 ; SI: v_mad_f32 [[RESULT:v[0-9]+]], [[REGA]], |[[REGB]]|, -[[REGC]] 152 ; SI: buffer_store_dword [[RESULT]] 153 define void @mad_fabs_sub_f32(float addrspace(1)* noalias nocapture %out, float addrspace(1)* noalias nocapture readonly %ptr) #1 { 154 %tid = tail call i32 @llvm.r600.read.tidig.x() #0 155 %tid.ext = sext i32 %tid to i64 156 %gep0 = getelementptr float, float addrspace(1)* %ptr, i64 %tid.ext 157 %add1 = add i64 %tid.ext, 1 158 %gep1 = getelementptr float, float addrspace(1)* %ptr, i64 %add1 159 %add2 = add i64 %tid.ext, 2 160 %gep2 = getelementptr float, float addrspace(1)* %ptr, i64 %add2 161 %outgep = getelementptr float, float addrspace(1)* %out, i64 %tid.ext 162 %a = load float, float addrspace(1)* %gep0, align 4 163 %b = load float, float addrspace(1)* %gep1, align 4 164 %c = load float, float addrspace(1)* %gep2, align 4 165 %b.abs = call float @llvm.fabs.f32(float %b) #0 166 %mul = fmul float %a, %b.abs 167 %sub = fsub float %mul, %c 168 store float %sub, float addrspace(1)* %outgep, align 4 169 ret void 170 } 171 172 ; FUNC-LABEL: {{^}}fsub_c_fadd_a_a: 173 ; SI-DAG: buffer_load_dword [[R1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}} 174 ; SI-DAG: buffer_load_dword [[R2:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4 175 ; SI: v_mac_f32_e32 [[R2]], -2.0, [[R1]] 176 ; SI: buffer_store_dword [[R2]] 177 define void @fsub_c_fadd_a_a(float addrspace(1)* %out, float addrspace(1)* %in) { 178 %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone 179 %gep.0 = getelementptr float, float addrspace(1)* %out, i32 %tid 180 %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1 181 %gep.out = getelementptr float, float addrspace(1)* %out, i32 %tid 182 183 %r1 = load float, float addrspace(1)* %gep.0 184 %r2 = load float, float addrspace(1)* %gep.1 185 186 %add = fadd float %r1, %r1 187 %r3 = fsub float %r2, %add 188 189 store float %r3, float addrspace(1)* %gep.out 190 ret void 191 } 192 193 ; FUNC-LABEL: {{^}}fsub_fadd_a_a_c: 194 ; SI-DAG: buffer_load_dword [[R1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}} 195 ; SI-DAG: buffer_load_dword [[R2:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4 196 ; SI: v_mad_f32 [[RESULT:v[0-9]+]], 2.0, [[R1]], -[[R2]] 197 ; SI: buffer_store_dword [[RESULT]] 198 define void @fsub_fadd_a_a_c(float addrspace(1)* %out, float addrspace(1)* %in) { 199 %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone 200 %gep.0 = getelementptr float, float addrspace(1)* %out, i32 %tid 201 %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1 202 %gep.out = getelementptr float, float addrspace(1)* %out, i32 %tid 203 204 %r1 = load float, float addrspace(1)* %gep.0 205 %r2 = load float, float addrspace(1)* %gep.1 206 207 %add = fadd float %r1, %r1 208 %r3 = fsub float %add, %r2 209 210 store float %r3, float addrspace(1)* %gep.out 211 ret void 212 } 213 214 attributes #0 = { nounwind readnone } 215 attributes #1 = { nounwind } 216
