1 /* 2 * Copyright (C) 2017 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 /** 18 * Tests for halving-add idiomatic vectorization. 19 */ 20 public class Main { 21 22 private static final int N = 64 * 1024; 23 private static final int M = N + 31; 24 25 static short[] sB1 = new short[M]; 26 static short[] sB2 = new short[M]; 27 static short[] sBo = new short[M]; 28 29 /// CHECK-START: void Main.halving_add_signed(short[], short[], short[]) loop_optimization (before) 30 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 31 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 32 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 33 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 34 /// CHECK-DAG: <<Add:i\d+>> Add [<<Get1>>,<<Get2>>] loop:<<Loop>> outer_loop:none 35 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 36 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 37 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 38 // 39 /// CHECK-START-ARM: void Main.halving_add_signed(short[], short[], short[]) loop_optimization (after) 40 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 41 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 42 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 43 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 44 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 45 // 46 /// CHECK-START-ARM64: void Main.halving_add_signed(short[], short[], short[]) loop_optimization (after) 47 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 48 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 49 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 50 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 51 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 52 // 53 /// CHECK-START-MIPS64: void Main.halving_add_signed(short[], short[], short[]) loop_optimization (after) 54 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 55 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 56 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 57 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 58 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 59 private static void halving_add_signed(short[] b1, short[] b2, short[] bo) { 60 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 61 for (int i = 0; i < min_length; i++) { 62 bo[i] = (short) ((b1[i] + b2[i]) >> 1); 63 } 64 } 65 66 /// CHECK-START: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (before) 67 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 68 /// CHECK-DAG: <<I10:i\d+>> IntConstant 10 loop:none 69 /// CHECK-DAG: <<M10:i\d+>> IntConstant -10 loop:none 70 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 71 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 72 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 73 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I10>>] loop:<<Loop>> outer_loop:none 74 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Get2>>,<<M10>>] loop:<<Loop>> outer_loop:none 75 /// CHECK-DAG: <<Add3:i\d+>> Add [<<Add1>>,<<Add2>>] loop:<<Loop>> outer_loop:none 76 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add3>>,<<I1>>] loop:<<Loop>> outer_loop:none 77 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 78 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 79 // 80 /// CHECK-START-ARM: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 81 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 82 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 83 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 84 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 85 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 86 // 87 /// CHECK-START-ARM64: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 88 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 89 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 90 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 91 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 92 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 93 // 94 /// CHECK-START-MIPS64: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 95 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 96 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 97 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 98 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 99 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 100 private static void halving_add_signed_alt(short[] b1, short[] b2, short[] bo) { 101 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 102 for (int i = 0; i < min_length; i++) { 103 // Cancelling constant computations do not confuse recognition. 104 bo[i] = (short) (((b1[i] + 10) + (b2[i] - 10)) >> 1); 105 } 106 } 107 108 /// CHECK-START: void Main.halving_add_unsigned(short[], short[], short[]) loop_optimization (before) 109 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 110 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 111 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 112 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 113 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 114 /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 115 /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 116 /// CHECK-DAG: <<Add:i\d+>> Add [<<And1>>,<<And2>>] loop:<<Loop>> outer_loop:none 117 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 118 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 119 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 120 // 121 /// CHECK-START-ARM: void Main.halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 122 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 123 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 124 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 125 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none 126 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 127 // 128 /// CHECK-START-ARM64: void Main.halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 129 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 130 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 131 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 132 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none 133 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 134 // 135 /// CHECK-START-MIPS64: void Main.halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 136 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 137 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 138 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 139 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none 140 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 141 private static void halving_add_unsigned(short[] b1, short[] b2, short[] bo) { 142 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 143 for (int i = 0; i < min_length; i++) { 144 bo[i] = (short) (((b1[i] & 0xffff) + (b2[i] & 0xffff)) >> 1); 145 } 146 } 147 148 /// CHECK-START: void Main.rounding_halving_add_signed(short[], short[], short[]) loop_optimization (before) 149 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 150 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 151 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 152 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 153 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<Get2>>] loop:<<Loop>> outer_loop:none 154 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<I1>>] loop:<<Loop>> outer_loop:none 155 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 156 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 157 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 158 // 159 /// CHECK-START-ARM: void Main.rounding_halving_add_signed(short[], short[], short[]) loop_optimization (after) 160 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 161 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 162 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 163 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 164 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 165 // 166 /// CHECK-START-ARM64: void Main.rounding_halving_add_signed(short[], short[], short[]) loop_optimization (after) 167 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 168 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 169 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 170 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 171 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 172 // 173 /// CHECK-START-MIPS64: void Main.rounding_halving_add_signed(short[], short[], short[]) loop_optimization (after) 174 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 175 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 176 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 177 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 178 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 179 private static void rounding_halving_add_signed(short[] b1, short[] b2, short[] bo) { 180 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 181 for (int i = 0; i < min_length; i++) { 182 bo[i] = (short) ((b1[i] + b2[i] + 1) >> 1); 183 } 184 } 185 186 /// CHECK-START: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (before) 187 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 188 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 189 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 190 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 191 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I1>>] loop:<<Loop>> outer_loop:none 192 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<Get2>>] loop:<<Loop>> outer_loop:none 193 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 194 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 195 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 196 // 197 /// CHECK-START-ARM: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 198 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 199 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 200 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 201 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 202 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 203 // 204 /// CHECK-START-ARM64: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 205 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 206 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 207 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 208 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 209 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 210 // 211 /// CHECK-START-MIPS64: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 212 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 213 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 214 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 215 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 216 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 217 private static void rounding_halving_add_signed_alt(short[] b1, short[] b2, short[] bo) { 218 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 219 for (int i = 0; i < min_length; i++) { 220 // Slightly different order in idiom does not confuse recognition. 221 bo[i] = (short) (((1 + b1[i]) + b2[i]) >> 1); 222 } 223 } 224 225 /// CHECK-START: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (before) 226 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 227 /// CHECK-DAG: <<I10:i\d+>> IntConstant 10 loop:none 228 /// CHECK-DAG: <<M9:i\d+>> IntConstant -9 loop:none 229 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 230 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 231 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 232 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I10>>] loop:<<Loop>> outer_loop:none 233 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Get2>>,<<M9>>] loop:<<Loop>> outer_loop:none 234 /// CHECK-DAG: <<Add3:i\d+>> Add [<<Add1>>,<<Add2>>] loop:<<Loop>> outer_loop:none 235 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add3>>,<<I1>>] loop:<<Loop>> outer_loop:none 236 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 237 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 238 // 239 /// CHECK-START-ARM: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (after) 240 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 241 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 242 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 243 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 244 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 245 // 246 /// CHECK-START-ARM64: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (after) 247 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 248 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 249 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 250 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 251 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 252 // 253 /// CHECK-START-MIPS64: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (after) 254 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 255 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 256 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 257 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none 258 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 259 private static void rounding_halving_add_signed_alt2(short[] b1, short[] b2, short[] bo) { 260 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 261 for (int i = 0; i < min_length; i++) { 262 // Computations that cancel to adding 1 also do not confuse recognition. 263 bo[i] = (short) (((b1[i] + 10) + (b2[i] - 9)) >> 1); 264 } 265 } 266 267 /// CHECK-START: void Main.rounding_halving_add_unsigned(short[], short[], short[]) loop_optimization (before) 268 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 269 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 270 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 271 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 272 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 273 /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 274 /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 275 /// CHECK-DAG: <<Add1:i\d+>> Add [<<And1>>,<<And2>>] loop:<<Loop>> outer_loop:none 276 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<I1>>] loop:<<Loop>> outer_loop:none 277 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 278 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 279 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 280 // 281 /// CHECK-START-ARM: void Main.rounding_halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 282 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 283 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 284 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 285 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:true loop:<<Loop>> outer_loop:none 286 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 287 // 288 /// CHECK-START-ARM64: void Main.rounding_halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 289 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 290 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 291 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 292 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:true loop:<<Loop>> outer_loop:none 293 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 294 // 295 /// CHECK-START-MIPS64: void Main.rounding_halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 296 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 297 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 298 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 299 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:true loop:<<Loop>> outer_loop:none 300 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 301 private static void rounding_halving_add_unsigned(short[] b1, short[] b2, short[] bo) { 302 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 303 for (int i = 0; i < min_length; i++) { 304 bo[i] = (short) (((b1[i] & 0xffff) + (b2[i] & 0xffff) + 1) >> 1); 305 } 306 } 307 308 /// CHECK-START: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (before) 309 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 310 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 311 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 312 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 313 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 314 /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 315 /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 316 /// CHECK-DAG: <<Add1:i\d+>> Add [<<And2>>,<<I1>>] loop:<<Loop>> outer_loop:none 317 /// CHECK-DAG: <<Add2:i\d+>> Add [<<And1>>,<<Add1>>] loop:<<Loop>> outer_loop:none 318 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 319 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 320 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 321 // 322 /// CHECK-START-ARM: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (after) 323 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 324 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 325 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 326 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:true loop:<<Loop>> outer_loop:none 327 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 328 // 329 /// CHECK-START-ARM64: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (after) 330 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 331 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 332 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 333 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:true loop:<<Loop>> outer_loop:none 334 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 335 // 336 /// CHECK-START-MIPS64: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (after) 337 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 338 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 339 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 340 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:true loop:<<Loop>> outer_loop:none 341 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 342 private static void rounding_halving_add_unsigned_alt(short[] b1, short[] b2, short[] bo) { 343 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 344 for (int i = 0; i < min_length; i++) { 345 // Slightly different order in idiom does not confuse recognition. 346 bo[i] = (short) ((b1[i] & 0xffff) + ((b2[i] & 0xffff) + 1) >> 1); 347 } 348 } 349 350 /// CHECK-START: void Main.halving_add_signed_constant(short[], short[]) loop_optimization (before) 351 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 352 /// CHECK-DAG: <<SMAX:i\d+>> IntConstant 32767 loop:none 353 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 354 /// CHECK-DAG: <<Get:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 355 /// CHECK-DAG: <<Add:i\d+>> Add [<<Get>>,<<SMAX>>] loop:<<Loop>> outer_loop:none 356 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 357 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 358 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 359 // 360 /// CHECK-START-ARM: void Main.halving_add_signed_constant(short[], short[]) loop_optimization (after) 361 /// CHECK-DAG: <<SMAX:i\d+>> IntConstant 32767 loop:none 362 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<SMAX>>] loop:none 363 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 364 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 365 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 366 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 367 // 368 /// CHECK-START-ARM64: void Main.halving_add_signed_constant(short[], short[]) loop_optimization (after) 369 /// CHECK-DAG: <<SMAX:i\d+>> IntConstant 32767 loop:none 370 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<SMAX>>] loop:none 371 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 372 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 373 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 374 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 375 // 376 /// CHECK-START-MIPS64: void Main.halving_add_signed_constant(short[], short[]) loop_optimization (after) 377 /// CHECK-DAG: <<SMAX:i\d+>> IntConstant 32767 loop:none 378 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<SMAX>>] loop:none 379 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 380 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 381 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none 382 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 383 private static void halving_add_signed_constant(short[] b1, short[] bo) { 384 int min_length = Math.min(bo.length, b1.length); 385 for (int i = 0; i < min_length; i++) { 386 bo[i] = (short) ((b1[i] + 0x7fff) >> 1); 387 } 388 } 389 390 /// CHECK-START: void Main.halving_add_unsigned_constant(short[], short[]) loop_optimization (before) 391 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 392 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 393 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 394 /// CHECK-DAG: <<Get:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 395 /// CHECK-DAG: <<And:i\d+>> And [<<Get>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 396 /// CHECK-DAG: <<Add:i\d+>> Add [<<And>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 397 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 398 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 399 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 400 // 401 /// CHECK-START-ARM: void Main.halving_add_unsigned_constant(short[], short[]) loop_optimization (after) 402 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 403 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<UMAX>>] loop:none 404 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 405 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 406 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none 407 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 408 // 409 /// CHECK-START-ARM64: void Main.halving_add_unsigned_constant(short[], short[]) loop_optimization (after) 410 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 411 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<UMAX>>] loop:none 412 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 413 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 414 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none 415 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 416 // 417 /// CHECK-START-MIPS64: void Main.halving_add_unsigned_constant(short[], short[]) loop_optimization (after) 418 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 419 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<UMAX>>] loop:none 420 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 421 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 422 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none 423 /// CHECK-DAG: VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>> outer_loop:none 424 private static void halving_add_unsigned_constant(short[] b1, short[] bo) { 425 int min_length = Math.min(bo.length, b1.length); 426 for (int i = 0; i < min_length; i++) { 427 bo[i] = (short) (((b1[i] & 0xffff) + 0xffff) >> 1); 428 } 429 } 430 431 public static void main(String[] args) { 432 // Some interesting values. 433 short[] interesting = { 434 (short) 0x0000, 435 (short) 0x0001, 436 (short) 0x0002, 437 (short) 0x1234, 438 (short) 0x8000, 439 (short) 0x8001, 440 (short) 0x7fff, 441 (short) 0xffff 442 }; 443 // Initialize cross-values to test all cases, and also 444 // set up some extra values to exercise the cleanup loop. 445 for (int i = 0; i < M; i++) { 446 sB1[i] = (short) i; 447 sB2[i] = interesting[i & 7]; 448 } 449 450 // Test halving add idioms. 451 halving_add_signed(sB1, sB2, sBo); 452 for (int i = 0; i < M; i++) { 453 short e = (short) ((sB1[i] + sB2[i]) >> 1); 454 expectEquals(e, sBo[i]); 455 } 456 halving_add_signed_alt(sB1, sB2, sBo); 457 for (int i = 0; i < M; i++) { 458 short e = (short) ((sB1[i] + sB2[i]) >> 1); 459 expectEquals(e, sBo[i]); 460 } 461 halving_add_unsigned(sB1, sB2, sBo); 462 for (int i = 0; i < M; i++) { 463 short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff)) >> 1); 464 expectEquals(e, sBo[i]); 465 } 466 rounding_halving_add_signed(sB1, sB2, sBo); 467 for (int i = 0; i < M; i++) { 468 short e = (short) ((sB1[i] + sB2[i] + 1) >> 1); 469 expectEquals(e, sBo[i]); 470 } 471 rounding_halving_add_signed_alt(sB1, sB2, sBo); 472 for (int i = 0; i < M; i++) { 473 short e = (short) ((sB1[i] + sB2[i] + 1) >> 1); 474 expectEquals(e, sBo[i]); 475 } 476 rounding_halving_add_signed_alt2(sB1, sB2, sBo); 477 for (int i = 0; i < M; i++) { 478 short e = (short) ((sB1[i] + sB2[i] + 1) >> 1); 479 expectEquals(e, sBo[i]); 480 } 481 rounding_halving_add_unsigned(sB1, sB2, sBo); 482 for (int i = 0; i < M; i++) { 483 short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff) + 1) >> 1); 484 expectEquals(e, sBo[i]); 485 } 486 rounding_halving_add_unsigned_alt(sB1, sB2, sBo); 487 for (int i = 0; i < M; i++) { 488 short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff) + 1) >> 1); 489 expectEquals(e, sBo[i]); 490 } 491 halving_add_signed_constant(sB1, sBo); 492 for (int i = 0; i < M; i++) { 493 short e = (short) ((sB1[i] + 0x7fff) >> 1); 494 expectEquals(e, sBo[i]); 495 } 496 halving_add_unsigned_constant(sB1, sBo); 497 for (int i = 0; i < M; i++) { 498 short e = (short) (((sB1[i] & 0xffff) + 0xffff) >> 1); 499 expectEquals(e, sBo[i]); 500 } 501 502 System.out.println("passed"); 503 } 504 505 private static void expectEquals(int expected, int result) { 506 if (expected != result) { 507 throw new Error("Expected: " + expected + ", found: " + result); 508 } 509 } 510 } 511
