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,ARM64,MIPS64}: void Main.halving_add_signed(short[], short[], short[]) loop_optimization (after) 40 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 41 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 42 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Int16 rounded:false loop:<<Loop>> outer_loop:none 43 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 44 private static void halving_add_signed(short[] b1, short[] b2, short[] bo) { 45 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 46 for (int i = 0; i < min_length; i++) { 47 bo[i] = (short) ((b1[i] + b2[i]) >> 1); 48 } 49 } 50 51 /// CHECK-START: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (before) 52 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 53 /// CHECK-DAG: <<I10:i\d+>> IntConstant 10 loop:none 54 /// CHECK-DAG: <<M10:i\d+>> IntConstant -10 loop:none 55 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 56 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 57 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 58 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I10>>] loop:<<Loop>> outer_loop:none 59 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Get2>>,<<M10>>] loop:<<Loop>> outer_loop:none 60 /// CHECK-DAG: <<Add3:i\d+>> Add [<<Add1>>,<<Add2>>] loop:<<Loop>> outer_loop:none 61 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add3>>,<<I1>>] loop:<<Loop>> outer_loop:none 62 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 63 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 64 // 65 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 66 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 67 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 68 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Int16 rounded:false loop:<<Loop>> outer_loop:none 69 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 70 private static void halving_add_signed_alt(short[] b1, short[] b2, short[] bo) { 71 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 72 for (int i = 0; i < min_length; i++) { 73 // Cancelling constant computations do not confuse recognition. 74 bo[i] = (short) (((b1[i] + 10) + (b2[i] - 10)) >> 1); 75 } 76 } 77 78 /// CHECK-START: void Main.halving_add_unsigned(short[], short[], short[]) instruction_simplifier (before) 79 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 80 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 81 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 82 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 83 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 84 /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 85 /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 86 /// CHECK-DAG: <<Add:i\d+>> Add [<<And1>>,<<And2>>] loop:<<Loop>> outer_loop:none 87 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 88 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 89 /// CHECK-DAG: ArraySet [{{l\d+}},{{i\d+}},<<Cnv>>] loop:<<Loop>> outer_loop:none 90 // 91 /// CHECK-START: void Main.halving_add_unsigned(short[], short[], short[]) loop_optimization (before) 92 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 93 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 94 /// CHECK-DAG: <<Get1:c\d+>> ArrayGet loop:<<Loop>> outer_loop:none 95 /// CHECK-DAG: <<Get2:c\d+>> ArrayGet loop:<<Loop>> outer_loop:none 96 /// CHECK-DAG: <<Add:i\d+>> Add [<<Get1>>,<<Get2>>] loop:<<Loop>> outer_loop:none 97 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 98 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 99 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 100 // 101 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 102 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 103 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 104 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Uint16 rounded:false loop:<<Loop>> outer_loop:none 105 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 106 private static void halving_add_unsigned(short[] b1, short[] b2, short[] bo) { 107 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 108 for (int i = 0; i < min_length; i++) { 109 bo[i] = (short) (((b1[i] & 0xffff) + (b2[i] & 0xffff)) >> 1); 110 } 111 } 112 113 /// CHECK-START: void Main.rounding_halving_add_signed(short[], short[], short[]) loop_optimization (before) 114 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 115 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 116 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 117 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 118 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<Get2>>] loop:<<Loop>> outer_loop:none 119 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<I1>>] loop:<<Loop>> outer_loop:none 120 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 121 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 122 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 123 // 124 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.rounding_halving_add_signed(short[], short[], short[]) loop_optimization (after) 125 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 126 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 127 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Int16 rounded:true loop:<<Loop>> outer_loop:none 128 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 129 private static void rounding_halving_add_signed(short[] b1, short[] b2, short[] bo) { 130 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 131 for (int i = 0; i < min_length; i++) { 132 bo[i] = (short) ((b1[i] + b2[i] + 1) >> 1); 133 } 134 } 135 136 /// CHECK-START: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (before) 137 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 138 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 139 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 140 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 141 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I1>>] loop:<<Loop>> outer_loop:none 142 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<Get2>>] loop:<<Loop>> outer_loop:none 143 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 144 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 145 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 146 // 147 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.rounding_halving_add_signed_alt(short[], short[], short[]) loop_optimization (after) 148 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 149 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 150 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Int16 rounded:true loop:<<Loop>> outer_loop:none 151 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 152 private static void rounding_halving_add_signed_alt(short[] b1, short[] b2, short[] bo) { 153 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 154 for (int i = 0; i < min_length; i++) { 155 // Slightly different order in idiom does not confuse recognition. 156 bo[i] = (short) (((1 + b1[i]) + b2[i]) >> 1); 157 } 158 } 159 160 /// CHECK-START: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (before) 161 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 162 /// CHECK-DAG: <<I10:i\d+>> IntConstant 10 loop:none 163 /// CHECK-DAG: <<M9:i\d+>> IntConstant -9 loop:none 164 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 165 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 166 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 167 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<I10>>] loop:<<Loop>> outer_loop:none 168 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Get2>>,<<M9>>] loop:<<Loop>> outer_loop:none 169 /// CHECK-DAG: <<Add3:i\d+>> Add [<<Add1>>,<<Add2>>] loop:<<Loop>> outer_loop:none 170 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add3>>,<<I1>>] loop:<<Loop>> outer_loop:none 171 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 172 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 173 // 174 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.rounding_halving_add_signed_alt2(short[], short[], short[]) loop_optimization (after) 175 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 176 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 177 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Int16 rounded:true loop:<<Loop>> outer_loop:none 178 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 179 private static void rounding_halving_add_signed_alt2(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 // Computations that cancel to adding 1 also do not confuse recognition. 183 bo[i] = (short) (((b1[i] + 10) + (b2[i] - 9)) >> 1); 184 } 185 } 186 187 /// CHECK-START: void Main.rounding_halving_add_unsigned(short[], short[], short[]) instruction_simplifier (before) 188 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 189 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 190 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 191 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 192 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 193 /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 194 /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 195 /// CHECK-DAG: <<Add1:i\d+>> Add [<<And1>>,<<And2>>] loop:<<Loop>> outer_loop:none 196 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<I1>>] loop:<<Loop>> outer_loop:none 197 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 198 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 199 /// CHECK-DAG: ArraySet [{{l\d+}},{{i\d+}},<<Cnv>>] loop:<<Loop>> outer_loop:none 200 // 201 /// CHECK-START: void Main.rounding_halving_add_unsigned(short[], short[], short[]) loop_optimization (before) 202 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 203 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 204 /// CHECK-DAG: <<Get1:c\d+>> ArrayGet loop:<<Loop>> outer_loop:none 205 /// CHECK-DAG: <<Get2:c\d+>> ArrayGet loop:<<Loop>> outer_loop:none 206 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<Get2>>] loop:<<Loop>> outer_loop:none 207 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<I1>>] loop:<<Loop>> outer_loop:none 208 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 209 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 210 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 211 // 212 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.rounding_halving_add_unsigned(short[], short[], short[]) loop_optimization (after) 213 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 214 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 215 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Uint16 rounded:true loop:<<Loop>> outer_loop:none 216 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 217 private static void rounding_halving_add_unsigned(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 bo[i] = (short) (((b1[i] & 0xffff) + (b2[i] & 0xffff) + 1) >> 1); 221 } 222 } 223 224 /// CHECK-START: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) instruction_simplifier (before) 225 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 226 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 227 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 228 /// CHECK-DAG: <<Get1:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 229 /// CHECK-DAG: <<Get2:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 230 /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 231 /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 232 /// CHECK-DAG: <<Add1:i\d+>> Add [<<And2>>,<<I1>>] loop:<<Loop>> outer_loop:none 233 /// CHECK-DAG: <<Add2:i\d+>> Add [<<And1>>,<<Add1>>] loop:<<Loop>> outer_loop:none 234 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 235 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 236 /// CHECK-DAG: ArraySet [{{l\d+}},{{i\d+}},<<Cnv>>] loop:<<Loop>> outer_loop:none 237 // 238 /// CHECK-START: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (before) 239 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 240 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 241 /// CHECK-DAG: <<Get1:c\d+>> ArrayGet loop:<<Loop>> outer_loop:none 242 /// CHECK-DAG: <<Get2:c\d+>> ArrayGet loop:<<Loop>> outer_loop:none 243 /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get2>>,<<I1>>] loop:<<Loop>> outer_loop:none 244 /// CHECK-DAG: <<Add2:i\d+>> Add [<<Get1>>,<<Add1>>] loop:<<Loop>> outer_loop:none 245 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add2>>,<<I1>>] loop:<<Loop>> outer_loop:none 246 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 247 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 248 // 249 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.rounding_halving_add_unsigned_alt(short[], short[], short[]) loop_optimization (after) 250 /// CHECK-DAG: <<Get1:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 251 /// CHECK-DAG: <<Get2:d\d+>> VecLoad loop:<<Loop>> outer_loop:none 252 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] packed_type:Uint16 rounded:true loop:<<Loop>> outer_loop:none 253 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 254 private static void rounding_halving_add_unsigned_alt(short[] b1, short[] b2, short[] bo) { 255 int min_length = Math.min(bo.length, Math.min(b1.length, b2.length)); 256 for (int i = 0; i < min_length; i++) { 257 // Slightly different order in idiom does not confuse recognition. 258 bo[i] = (short) ((b1[i] & 0xffff) + ((b2[i] & 0xffff) + 1) >> 1); 259 } 260 } 261 262 /// CHECK-START: void Main.halving_add_signed_constant(short[], short[]) loop_optimization (before) 263 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 264 /// CHECK-DAG: <<SMAX:i\d+>> IntConstant 32767 loop:none 265 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 266 /// CHECK-DAG: <<Get:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 267 /// CHECK-DAG: <<Add:i\d+>> Add [<<Get>>,<<SMAX>>] loop:<<Loop>> outer_loop:none 268 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 269 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 270 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 271 // 272 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.halving_add_signed_constant(short[], short[]) loop_optimization (after) 273 /// CHECK-DAG: <<SMAX:i\d+>> IntConstant 32767 loop:none 274 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<SMAX>>] loop:none 275 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 276 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] packed_type:Int16 rounded:false loop:<<Loop>> outer_loop:none 277 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 278 private static void halving_add_signed_constant(short[] b1, short[] bo) { 279 int min_length = Math.min(bo.length, b1.length); 280 for (int i = 0; i < min_length; i++) { 281 bo[i] = (short) ((b1[i] + 0x7fff) >> 1); 282 } 283 } 284 285 /// CHECK-START: void Main.halving_add_unsigned_constant(short[], short[]) instruction_simplifier (before) 286 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 287 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 288 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 289 /// CHECK-DAG: <<Get:s\d+>> ArrayGet loop:<<Loop>> outer_loop:none 290 /// CHECK-DAG: <<And:i\d+>> And [<<Get>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 291 /// CHECK-DAG: <<Add:i\d+>> Add [<<And>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 292 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 293 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 294 /// CHECK-DAG: ArraySet [{{l\d+}},{{i\d+}},<<Cnv>>] loop:<<Loop>> outer_loop:none 295 // 296 /// CHECK-START: void Main.halving_add_unsigned_constant(short[], short[]) loop_optimization (before) 297 /// CHECK-DAG: <<I1:i\d+>> IntConstant 1 loop:none 298 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 299 /// CHECK-DAG: <<Phi:i\d+>> Phi loop:<<Loop:B\d+>> outer_loop:none 300 /// CHECK-DAG: <<Get:c\d+>> ArrayGet loop:<<Loop>> outer_loop:none 301 /// CHECK-DAG: <<Add:i\d+>> Add [<<Get>>,<<UMAX>>] loop:<<Loop>> outer_loop:none 302 /// CHECK-DAG: <<Shr:i\d+>> Shr [<<Add>>,<<I1>>] loop:<<Loop>> outer_loop:none 303 /// CHECK-DAG: <<Cnv:s\d+>> TypeConversion [<<Shr>>] loop:<<Loop>> outer_loop:none 304 /// CHECK-DAG: ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>> outer_loop:none 305 // 306 /// CHECK-START-{ARM,ARM64,MIPS64}: void Main.halving_add_unsigned_constant(short[], short[]) loop_optimization (after) 307 /// CHECK-DAG: <<UMAX:i\d+>> IntConstant 65535 loop:none 308 /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<UMAX>>] loop:none 309 /// CHECK-DAG: <<Get:d\d+>> VecLoad loop:<<Loop:B\d+>> outer_loop:none 310 /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] packed_type:Uint16 rounded:false loop:<<Loop>> outer_loop:none 311 /// CHECK-DAG: VecStore [{{l\d+}},{{i\d+}},<<HAdd>>] loop:<<Loop>> outer_loop:none 312 private static void halving_add_unsigned_constant(short[] b1, short[] bo) { 313 int min_length = Math.min(bo.length, b1.length); 314 for (int i = 0; i < min_length; i++) { 315 bo[i] = (short) (((b1[i] & 0xffff) + 0xffff) >> 1); 316 } 317 } 318 319 public static void main(String[] args) { 320 // Some interesting values. 321 short[] interesting = { 322 (short) 0x0000, 323 (short) 0x0001, 324 (short) 0x0002, 325 (short) 0x1234, 326 (short) 0x8000, 327 (short) 0x8001, 328 (short) 0x7fff, 329 (short) 0xffff 330 }; 331 // Initialize cross-values to test all cases, and also 332 // set up some extra values to exercise the cleanup loop. 333 for (int i = 0; i < M; i++) { 334 sB1[i] = (short) i; 335 sB2[i] = interesting[i & 7]; 336 } 337 338 // Test halving add idioms. 339 halving_add_signed(sB1, sB2, sBo); 340 for (int i = 0; i < M; i++) { 341 short e = (short) ((sB1[i] + sB2[i]) >> 1); 342 expectEquals(e, sBo[i]); 343 } 344 halving_add_signed_alt(sB1, sB2, sBo); 345 for (int i = 0; i < M; i++) { 346 short e = (short) ((sB1[i] + sB2[i]) >> 1); 347 expectEquals(e, sBo[i]); 348 } 349 halving_add_unsigned(sB1, sB2, sBo); 350 for (int i = 0; i < M; i++) { 351 short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff)) >> 1); 352 expectEquals(e, sBo[i]); 353 } 354 rounding_halving_add_signed(sB1, sB2, sBo); 355 for (int i = 0; i < M; i++) { 356 short e = (short) ((sB1[i] + sB2[i] + 1) >> 1); 357 expectEquals(e, sBo[i]); 358 } 359 rounding_halving_add_signed_alt(sB1, sB2, sBo); 360 for (int i = 0; i < M; i++) { 361 short e = (short) ((sB1[i] + sB2[i] + 1) >> 1); 362 expectEquals(e, sBo[i]); 363 } 364 rounding_halving_add_signed_alt2(sB1, sB2, sBo); 365 for (int i = 0; i < M; i++) { 366 short e = (short) ((sB1[i] + sB2[i] + 1) >> 1); 367 expectEquals(e, sBo[i]); 368 } 369 rounding_halving_add_unsigned(sB1, sB2, sBo); 370 for (int i = 0; i < M; i++) { 371 short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff) + 1) >> 1); 372 expectEquals(e, sBo[i]); 373 } 374 rounding_halving_add_unsigned_alt(sB1, sB2, sBo); 375 for (int i = 0; i < M; i++) { 376 short e = (short) (((sB1[i] & 0xffff) + (sB2[i] & 0xffff) + 1) >> 1); 377 expectEquals(e, sBo[i]); 378 } 379 halving_add_signed_constant(sB1, sBo); 380 for (int i = 0; i < M; i++) { 381 short e = (short) ((sB1[i] + 0x7fff) >> 1); 382 expectEquals(e, sBo[i]); 383 } 384 halving_add_unsigned_constant(sB1, sBo); 385 for (int i = 0; i < M; i++) { 386 short e = (short) (((sB1[i] & 0xffff) + 0xffff) >> 1); 387 expectEquals(e, sBo[i]); 388 } 389 390 System.out.println("passed"); 391 } 392 393 private static void expectEquals(int expected, int result) { 394 if (expected != result) { 395 throw new Error("Expected: " + expected + ", found: " + result); 396 } 397 } 398 } 399
