1 /* 2 * Copyright (C) 2012 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 #include "codegen_x86.h" 18 #include "dex/quick/mir_to_lir-inl.h" 19 #include "x86_lir.h" 20 21 namespace art { 22 23 void X86Mir2Lir::GenArithOpFloat(Instruction::Code opcode, 24 RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2) { 25 X86OpCode op = kX86Nop; 26 RegLocation rl_result; 27 28 /* 29 * Don't attempt to optimize register usage since these opcodes call out to 30 * the handlers. 31 */ 32 switch (opcode) { 33 case Instruction::ADD_FLOAT_2ADDR: 34 case Instruction::ADD_FLOAT: 35 op = kX86AddssRR; 36 break; 37 case Instruction::SUB_FLOAT_2ADDR: 38 case Instruction::SUB_FLOAT: 39 op = kX86SubssRR; 40 break; 41 case Instruction::DIV_FLOAT_2ADDR: 42 case Instruction::DIV_FLOAT: 43 op = kX86DivssRR; 44 break; 45 case Instruction::MUL_FLOAT_2ADDR: 46 case Instruction::MUL_FLOAT: 47 op = kX86MulssRR; 48 break; 49 case Instruction::REM_FLOAT_2ADDR: 50 case Instruction::REM_FLOAT: 51 FlushAllRegs(); // Send everything to home location 52 CallRuntimeHelperRegLocationRegLocation(QUICK_ENTRYPOINT_OFFSET(pFmodf), rl_src1, rl_src2, 53 false); 54 rl_result = GetReturn(true); 55 StoreValue(rl_dest, rl_result); 56 return; 57 case Instruction::NEG_FLOAT: 58 GenNegFloat(rl_dest, rl_src1); 59 return; 60 default: 61 LOG(FATAL) << "Unexpected opcode: " << opcode; 62 } 63 rl_src1 = LoadValue(rl_src1, kFPReg); 64 rl_src2 = LoadValue(rl_src2, kFPReg); 65 rl_result = EvalLoc(rl_dest, kFPReg, true); 66 int r_dest = rl_result.low_reg; 67 int r_src1 = rl_src1.low_reg; 68 int r_src2 = rl_src2.low_reg; 69 if (r_dest == r_src2) { 70 r_src2 = AllocTempFloat(); 71 OpRegCopy(r_src2, r_dest); 72 } 73 OpRegCopy(r_dest, r_src1); 74 NewLIR2(op, r_dest, r_src2); 75 StoreValue(rl_dest, rl_result); 76 } 77 78 void X86Mir2Lir::GenArithOpDouble(Instruction::Code opcode, 79 RegLocation rl_dest, RegLocation rl_src1, RegLocation rl_src2) { 80 X86OpCode op = kX86Nop; 81 RegLocation rl_result; 82 83 switch (opcode) { 84 case Instruction::ADD_DOUBLE_2ADDR: 85 case Instruction::ADD_DOUBLE: 86 op = kX86AddsdRR; 87 break; 88 case Instruction::SUB_DOUBLE_2ADDR: 89 case Instruction::SUB_DOUBLE: 90 op = kX86SubsdRR; 91 break; 92 case Instruction::DIV_DOUBLE_2ADDR: 93 case Instruction::DIV_DOUBLE: 94 op = kX86DivsdRR; 95 break; 96 case Instruction::MUL_DOUBLE_2ADDR: 97 case Instruction::MUL_DOUBLE: 98 op = kX86MulsdRR; 99 break; 100 case Instruction::REM_DOUBLE_2ADDR: 101 case Instruction::REM_DOUBLE: 102 FlushAllRegs(); // Send everything to home location 103 CallRuntimeHelperRegLocationRegLocation(QUICK_ENTRYPOINT_OFFSET(pFmod), rl_src1, rl_src2, 104 false); 105 rl_result = GetReturnWide(true); 106 StoreValueWide(rl_dest, rl_result); 107 return; 108 case Instruction::NEG_DOUBLE: 109 GenNegDouble(rl_dest, rl_src1); 110 return; 111 default: 112 LOG(FATAL) << "Unexpected opcode: " << opcode; 113 } 114 rl_src1 = LoadValueWide(rl_src1, kFPReg); 115 DCHECK(rl_src1.wide); 116 rl_src2 = LoadValueWide(rl_src2, kFPReg); 117 DCHECK(rl_src2.wide); 118 rl_result = EvalLoc(rl_dest, kFPReg, true); 119 DCHECK(rl_dest.wide); 120 DCHECK(rl_result.wide); 121 int r_dest = S2d(rl_result.low_reg, rl_result.high_reg); 122 int r_src1 = S2d(rl_src1.low_reg, rl_src1.high_reg); 123 int r_src2 = S2d(rl_src2.low_reg, rl_src2.high_reg); 124 if (r_dest == r_src2) { 125 r_src2 = AllocTempDouble() | X86_FP_DOUBLE; 126 OpRegCopy(r_src2, r_dest); 127 } 128 OpRegCopy(r_dest, r_src1); 129 NewLIR2(op, r_dest, r_src2); 130 StoreValueWide(rl_dest, rl_result); 131 } 132 133 void X86Mir2Lir::GenConversion(Instruction::Code opcode, RegLocation rl_dest, 134 RegLocation rl_src) { 135 RegisterClass rcSrc = kFPReg; 136 X86OpCode op = kX86Nop; 137 int src_reg; 138 RegLocation rl_result; 139 switch (opcode) { 140 case Instruction::INT_TO_FLOAT: 141 rcSrc = kCoreReg; 142 op = kX86Cvtsi2ssRR; 143 break; 144 case Instruction::DOUBLE_TO_FLOAT: 145 rcSrc = kFPReg; 146 op = kX86Cvtsd2ssRR; 147 break; 148 case Instruction::FLOAT_TO_DOUBLE: 149 rcSrc = kFPReg; 150 op = kX86Cvtss2sdRR; 151 break; 152 case Instruction::INT_TO_DOUBLE: 153 rcSrc = kCoreReg; 154 op = kX86Cvtsi2sdRR; 155 break; 156 case Instruction::FLOAT_TO_INT: { 157 rl_src = LoadValue(rl_src, kFPReg); 158 src_reg = rl_src.low_reg; 159 // In case result vreg is also src vreg, break association to avoid useless copy by EvalLoc() 160 ClobberSReg(rl_dest.s_reg_low); 161 rl_result = EvalLoc(rl_dest, kCoreReg, true); 162 int temp_reg = AllocTempFloat(); 163 164 LoadConstant(rl_result.low_reg, 0x7fffffff); 165 NewLIR2(kX86Cvtsi2ssRR, temp_reg, rl_result.low_reg); 166 NewLIR2(kX86ComissRR, src_reg, temp_reg); 167 LIR* branch_pos_overflow = NewLIR2(kX86Jcc8, 0, kX86CondA); 168 LIR* branch_na_n = NewLIR2(kX86Jcc8, 0, kX86CondP); 169 NewLIR2(kX86Cvttss2siRR, rl_result.low_reg, src_reg); 170 LIR* branch_normal = NewLIR1(kX86Jmp8, 0); 171 branch_na_n->target = NewLIR0(kPseudoTargetLabel); 172 NewLIR2(kX86Xor32RR, rl_result.low_reg, rl_result.low_reg); 173 branch_pos_overflow->target = NewLIR0(kPseudoTargetLabel); 174 branch_normal->target = NewLIR0(kPseudoTargetLabel); 175 StoreValue(rl_dest, rl_result); 176 return; 177 } 178 case Instruction::DOUBLE_TO_INT: { 179 rl_src = LoadValueWide(rl_src, kFPReg); 180 src_reg = rl_src.low_reg; 181 // In case result vreg is also src vreg, break association to avoid useless copy by EvalLoc() 182 ClobberSReg(rl_dest.s_reg_low); 183 rl_result = EvalLoc(rl_dest, kCoreReg, true); 184 int temp_reg = AllocTempDouble() | X86_FP_DOUBLE; 185 186 LoadConstant(rl_result.low_reg, 0x7fffffff); 187 NewLIR2(kX86Cvtsi2sdRR, temp_reg, rl_result.low_reg); 188 NewLIR2(kX86ComisdRR, src_reg, temp_reg); 189 LIR* branch_pos_overflow = NewLIR2(kX86Jcc8, 0, kX86CondA); 190 LIR* branch_na_n = NewLIR2(kX86Jcc8, 0, kX86CondP); 191 NewLIR2(kX86Cvttsd2siRR, rl_result.low_reg, src_reg); 192 LIR* branch_normal = NewLIR1(kX86Jmp8, 0); 193 branch_na_n->target = NewLIR0(kPseudoTargetLabel); 194 NewLIR2(kX86Xor32RR, rl_result.low_reg, rl_result.low_reg); 195 branch_pos_overflow->target = NewLIR0(kPseudoTargetLabel); 196 branch_normal->target = NewLIR0(kPseudoTargetLabel); 197 StoreValue(rl_dest, rl_result); 198 return; 199 } 200 case Instruction::LONG_TO_DOUBLE: 201 GenConversionCall(QUICK_ENTRYPOINT_OFFSET(pL2d), rl_dest, rl_src); 202 return; 203 case Instruction::LONG_TO_FLOAT: 204 // TODO: inline by using memory as a 64-bit source. Be careful about promoted registers. 205 GenConversionCall(QUICK_ENTRYPOINT_OFFSET(pL2f), rl_dest, rl_src); 206 return; 207 case Instruction::FLOAT_TO_LONG: 208 GenConversionCall(QUICK_ENTRYPOINT_OFFSET(pF2l), rl_dest, rl_src); 209 return; 210 case Instruction::DOUBLE_TO_LONG: 211 GenConversionCall(QUICK_ENTRYPOINT_OFFSET(pD2l), rl_dest, rl_src); 212 return; 213 default: 214 LOG(INFO) << "Unexpected opcode: " << opcode; 215 } 216 if (rl_src.wide) { 217 rl_src = LoadValueWide(rl_src, rcSrc); 218 src_reg = S2d(rl_src.low_reg, rl_src.high_reg); 219 } else { 220 rl_src = LoadValue(rl_src, rcSrc); 221 src_reg = rl_src.low_reg; 222 } 223 if (rl_dest.wide) { 224 rl_result = EvalLoc(rl_dest, kFPReg, true); 225 NewLIR2(op, S2d(rl_result.low_reg, rl_result.high_reg), src_reg); 226 StoreValueWide(rl_dest, rl_result); 227 } else { 228 rl_result = EvalLoc(rl_dest, kFPReg, true); 229 NewLIR2(op, rl_result.low_reg, src_reg); 230 StoreValue(rl_dest, rl_result); 231 } 232 } 233 234 void X86Mir2Lir::GenCmpFP(Instruction::Code code, RegLocation rl_dest, 235 RegLocation rl_src1, RegLocation rl_src2) { 236 bool single = (code == Instruction::CMPL_FLOAT) || (code == Instruction::CMPG_FLOAT); 237 bool unordered_gt = (code == Instruction::CMPG_DOUBLE) || (code == Instruction::CMPG_FLOAT); 238 int src_reg1; 239 int src_reg2; 240 if (single) { 241 rl_src1 = LoadValue(rl_src1, kFPReg); 242 src_reg1 = rl_src1.low_reg; 243 rl_src2 = LoadValue(rl_src2, kFPReg); 244 src_reg2 = rl_src2.low_reg; 245 } else { 246 rl_src1 = LoadValueWide(rl_src1, kFPReg); 247 src_reg1 = S2d(rl_src1.low_reg, rl_src1.high_reg); 248 rl_src2 = LoadValueWide(rl_src2, kFPReg); 249 src_reg2 = S2d(rl_src2.low_reg, rl_src2.high_reg); 250 } 251 // In case result vreg is also src vreg, break association to avoid useless copy by EvalLoc() 252 ClobberSReg(rl_dest.s_reg_low); 253 RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true); 254 LoadConstantNoClobber(rl_result.low_reg, unordered_gt ? 1 : 0); 255 if (single) { 256 NewLIR2(kX86UcomissRR, src_reg1, src_reg2); 257 } else { 258 NewLIR2(kX86UcomisdRR, src_reg1, src_reg2); 259 } 260 LIR* branch = NULL; 261 if (unordered_gt) { 262 branch = NewLIR2(kX86Jcc8, 0, kX86CondPE); 263 } 264 // If the result reg can't be byte accessed, use a jump and move instead of a set. 265 if (rl_result.low_reg >= 4) { 266 LIR* branch2 = NULL; 267 if (unordered_gt) { 268 branch2 = NewLIR2(kX86Jcc8, 0, kX86CondA); 269 NewLIR2(kX86Mov32RI, rl_result.low_reg, 0x0); 270 } else { 271 branch2 = NewLIR2(kX86Jcc8, 0, kX86CondBe); 272 NewLIR2(kX86Mov32RI, rl_result.low_reg, 0x1); 273 } 274 branch2->target = NewLIR0(kPseudoTargetLabel); 275 } else { 276 NewLIR2(kX86Set8R, rl_result.low_reg, kX86CondA /* above - unsigned > */); 277 } 278 NewLIR2(kX86Sbb32RI, rl_result.low_reg, 0); 279 if (unordered_gt) { 280 branch->target = NewLIR0(kPseudoTargetLabel); 281 } 282 StoreValue(rl_dest, rl_result); 283 } 284 285 void X86Mir2Lir::GenFusedFPCmpBranch(BasicBlock* bb, MIR* mir, bool gt_bias, 286 bool is_double) { 287 LIR* taken = &block_label_list_[bb->taken->id]; 288 LIR* not_taken = &block_label_list_[bb->fall_through->id]; 289 LIR* branch = NULL; 290 RegLocation rl_src1; 291 RegLocation rl_src2; 292 if (is_double) { 293 rl_src1 = mir_graph_->GetSrcWide(mir, 0); 294 rl_src2 = mir_graph_->GetSrcWide(mir, 2); 295 rl_src1 = LoadValueWide(rl_src1, kFPReg); 296 rl_src2 = LoadValueWide(rl_src2, kFPReg); 297 NewLIR2(kX86UcomisdRR, S2d(rl_src1.low_reg, rl_src1.high_reg), 298 S2d(rl_src2.low_reg, rl_src2.high_reg)); 299 } else { 300 rl_src1 = mir_graph_->GetSrc(mir, 0); 301 rl_src2 = mir_graph_->GetSrc(mir, 1); 302 rl_src1 = LoadValue(rl_src1, kFPReg); 303 rl_src2 = LoadValue(rl_src2, kFPReg); 304 NewLIR2(kX86UcomissRR, rl_src1.low_reg, rl_src2.low_reg); 305 } 306 ConditionCode ccode = static_cast<ConditionCode>(mir->dalvikInsn.arg[0]); 307 switch (ccode) { 308 case kCondEq: 309 if (!gt_bias) { 310 branch = NewLIR2(kX86Jcc8, 0, kX86CondPE); 311 branch->target = not_taken; 312 } 313 break; 314 case kCondNe: 315 if (!gt_bias) { 316 branch = NewLIR2(kX86Jcc8, 0, kX86CondPE); 317 branch->target = taken; 318 } 319 break; 320 case kCondLt: 321 if (gt_bias) { 322 branch = NewLIR2(kX86Jcc8, 0, kX86CondPE); 323 branch->target = not_taken; 324 } 325 ccode = kCondCs; 326 break; 327 case kCondLe: 328 if (gt_bias) { 329 branch = NewLIR2(kX86Jcc8, 0, kX86CondPE); 330 branch->target = not_taken; 331 } 332 ccode = kCondLs; 333 break; 334 case kCondGt: 335 if (gt_bias) { 336 branch = NewLIR2(kX86Jcc8, 0, kX86CondPE); 337 branch->target = taken; 338 } 339 ccode = kCondHi; 340 break; 341 case kCondGe: 342 if (gt_bias) { 343 branch = NewLIR2(kX86Jcc8, 0, kX86CondPE); 344 branch->target = taken; 345 } 346 ccode = kCondCc; 347 break; 348 default: 349 LOG(FATAL) << "Unexpected ccode: " << ccode; 350 } 351 OpCondBranch(ccode, taken); 352 } 353 354 void X86Mir2Lir::GenNegFloat(RegLocation rl_dest, RegLocation rl_src) { 355 RegLocation rl_result; 356 rl_src = LoadValue(rl_src, kCoreReg); 357 rl_result = EvalLoc(rl_dest, kCoreReg, true); 358 OpRegRegImm(kOpAdd, rl_result.low_reg, rl_src.low_reg, 0x80000000); 359 StoreValue(rl_dest, rl_result); 360 } 361 362 void X86Mir2Lir::GenNegDouble(RegLocation rl_dest, RegLocation rl_src) { 363 RegLocation rl_result; 364 rl_src = LoadValueWide(rl_src, kCoreReg); 365 rl_result = EvalLoc(rl_dest, kCoreReg, true); 366 OpRegRegImm(kOpAdd, rl_result.high_reg, rl_src.high_reg, 0x80000000); 367 OpRegCopy(rl_result.low_reg, rl_src.low_reg); 368 StoreValueWide(rl_dest, rl_result); 369 } 370 371 bool X86Mir2Lir::GenInlinedSqrt(CallInfo* info) { 372 DCHECK_NE(cu_->instruction_set, kThumb2); 373 return false; 374 } 375 376 377 378 } // namespace art 379