X86ISelLowering.cpp revision 826f36ff80e318e8dc72f62ac72579137e30dc23
1//===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by Chris Lattner and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file defines the interfaces that X86 uses to lower LLVM code into a 11// selection DAG. 12// 13//===----------------------------------------------------------------------===// 14 15#include "X86.h" 16#include "X86InstrBuilder.h" 17#include "X86ISelLowering.h" 18#include "X86MachineFunctionInfo.h" 19#include "X86TargetMachine.h" 20#include "llvm/CallingConv.h" 21#include "llvm/Constants.h" 22#include "llvm/DerivedTypes.h" 23#include "llvm/Function.h" 24#include "llvm/Intrinsics.h" 25#include "llvm/ADT/VectorExtras.h" 26#include "llvm/Analysis/ScalarEvolutionExpressions.h" 27#include "llvm/CodeGen/CallingConvLower.h" 28#include "llvm/CodeGen/MachineFrameInfo.h" 29#include "llvm/CodeGen/MachineFunction.h" 30#include "llvm/CodeGen/MachineInstrBuilder.h" 31#include "llvm/CodeGen/SelectionDAG.h" 32#include "llvm/CodeGen/SSARegMap.h" 33#include "llvm/Support/MathExtras.h" 34#include "llvm/Target/TargetOptions.h" 35#include "llvm/ADT/StringExtras.h" 36using namespace llvm; 37 38X86TargetLowering::X86TargetLowering(TargetMachine &TM) 39 : TargetLowering(TM) { 40 Subtarget = &TM.getSubtarget<X86Subtarget>(); 41 X86ScalarSSE = Subtarget->hasSSE2(); 42 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP; 43 44 // Set up the TargetLowering object. 45 46 // X86 is weird, it always uses i8 for shift amounts and setcc results. 47 setShiftAmountType(MVT::i8); 48 setSetCCResultType(MVT::i8); 49 setSetCCResultContents(ZeroOrOneSetCCResult); 50 setSchedulingPreference(SchedulingForRegPressure); 51 setShiftAmountFlavor(Mask); // shl X, 32 == shl X, 0 52 setStackPointerRegisterToSaveRestore(X86StackPtr); 53 54 if (Subtarget->isTargetDarwin()) { 55 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp. 56 setUseUnderscoreSetJmp(false); 57 setUseUnderscoreLongJmp(false); 58 } else if (Subtarget->isTargetMingw()) { 59 // MS runtime is weird: it exports _setjmp, but longjmp! 60 setUseUnderscoreSetJmp(true); 61 setUseUnderscoreLongJmp(false); 62 } else { 63 setUseUnderscoreSetJmp(true); 64 setUseUnderscoreLongJmp(true); 65 } 66 67 // Set up the register classes. 68 addRegisterClass(MVT::i8, X86::GR8RegisterClass); 69 addRegisterClass(MVT::i16, X86::GR16RegisterClass); 70 addRegisterClass(MVT::i32, X86::GR32RegisterClass); 71 if (Subtarget->is64Bit()) 72 addRegisterClass(MVT::i64, X86::GR64RegisterClass); 73 74 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand); 75 76 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this 77 // operation. 78 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote); 79 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote); 80 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote); 81 82 if (Subtarget->is64Bit()) { 83 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand); 84 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote); 85 } else { 86 if (X86ScalarSSE) 87 // If SSE i64 SINT_TO_FP is not available, expand i32 UINT_TO_FP. 88 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Expand); 89 else 90 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote); 91 } 92 93 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have 94 // this operation. 95 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote); 96 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote); 97 // SSE has no i16 to fp conversion, only i32 98 if (X86ScalarSSE) 99 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote); 100 else { 101 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom); 102 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom); 103 } 104 105 if (!Subtarget->is64Bit()) { 106 // Custom lower SINT_TO_FP and FP_TO_SINT from/to i64 in 32-bit mode. 107 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom); 108 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom); 109 } 110 111 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have 112 // this operation. 113 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote); 114 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote); 115 116 if (X86ScalarSSE) { 117 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote); 118 } else { 119 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom); 120 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom); 121 } 122 123 // Handle FP_TO_UINT by promoting the destination to a larger signed 124 // conversion. 125 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote); 126 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote); 127 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote); 128 129 if (Subtarget->is64Bit()) { 130 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand); 131 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote); 132 } else { 133 if (X86ScalarSSE && !Subtarget->hasSSE3()) 134 // Expand FP_TO_UINT into a select. 135 // FIXME: We would like to use a Custom expander here eventually to do 136 // the optimal thing for SSE vs. the default expansion in the legalizer. 137 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand); 138 else 139 // With SSE3 we can use fisttpll to convert to a signed i64. 140 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote); 141 } 142 143 // TODO: when we have SSE, these could be more efficient, by using movd/movq. 144 if (!X86ScalarSSE) { 145 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand); 146 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand); 147 } 148 149 setOperationAction(ISD::BR_JT , MVT::Other, Expand); 150 setOperationAction(ISD::BRCOND , MVT::Other, Custom); 151 setOperationAction(ISD::BR_CC , MVT::Other, Expand); 152 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand); 153 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand); 154 if (Subtarget->is64Bit()) 155 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand); 156 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Expand); 157 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Expand); 158 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand); 159 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand); 160 setOperationAction(ISD::FREM , MVT::f64 , Expand); 161 162 setOperationAction(ISD::CTPOP , MVT::i8 , Expand); 163 setOperationAction(ISD::CTTZ , MVT::i8 , Expand); 164 setOperationAction(ISD::CTLZ , MVT::i8 , Expand); 165 setOperationAction(ISD::CTPOP , MVT::i16 , Expand); 166 setOperationAction(ISD::CTTZ , MVT::i16 , Expand); 167 setOperationAction(ISD::CTLZ , MVT::i16 , Expand); 168 setOperationAction(ISD::CTPOP , MVT::i32 , Expand); 169 setOperationAction(ISD::CTTZ , MVT::i32 , Expand); 170 setOperationAction(ISD::CTLZ , MVT::i32 , Expand); 171 if (Subtarget->is64Bit()) { 172 setOperationAction(ISD::CTPOP , MVT::i64 , Expand); 173 setOperationAction(ISD::CTTZ , MVT::i64 , Expand); 174 setOperationAction(ISD::CTLZ , MVT::i64 , Expand); 175 } 176 177 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom); 178 setOperationAction(ISD::BSWAP , MVT::i16 , Expand); 179 180 // These should be promoted to a larger select which is supported. 181 setOperationAction(ISD::SELECT , MVT::i1 , Promote); 182 setOperationAction(ISD::SELECT , MVT::i8 , Promote); 183 // X86 wants to expand cmov itself. 184 setOperationAction(ISD::SELECT , MVT::i16 , Custom); 185 setOperationAction(ISD::SELECT , MVT::i32 , Custom); 186 setOperationAction(ISD::SELECT , MVT::f32 , Custom); 187 setOperationAction(ISD::SELECT , MVT::f64 , Custom); 188 setOperationAction(ISD::SETCC , MVT::i8 , Custom); 189 setOperationAction(ISD::SETCC , MVT::i16 , Custom); 190 setOperationAction(ISD::SETCC , MVT::i32 , Custom); 191 setOperationAction(ISD::SETCC , MVT::f32 , Custom); 192 setOperationAction(ISD::SETCC , MVT::f64 , Custom); 193 if (Subtarget->is64Bit()) { 194 setOperationAction(ISD::SELECT , MVT::i64 , Custom); 195 setOperationAction(ISD::SETCC , MVT::i64 , Custom); 196 } 197 // X86 ret instruction may pop stack. 198 setOperationAction(ISD::RET , MVT::Other, Custom); 199 // Darwin ABI issue. 200 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom); 201 setOperationAction(ISD::JumpTable , MVT::i32 , Custom); 202 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom); 203 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom); 204 if (Subtarget->is64Bit()) { 205 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom); 206 setOperationAction(ISD::JumpTable , MVT::i64 , Custom); 207 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom); 208 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom); 209 } 210 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86) 211 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom); 212 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom); 213 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom); 214 // X86 wants to expand memset / memcpy itself. 215 setOperationAction(ISD::MEMSET , MVT::Other, Custom); 216 setOperationAction(ISD::MEMCPY , MVT::Other, Custom); 217 218 // We don't have line number support yet. 219 setOperationAction(ISD::LOCATION, MVT::Other, Expand); 220 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand); 221 // FIXME - use subtarget debug flags 222 if (!Subtarget->isTargetDarwin() && 223 !Subtarget->isTargetELF() && 224 !Subtarget->isTargetCygMing()) 225 setOperationAction(ISD::LABEL, MVT::Other, Expand); 226 227 // VASTART needs to be custom lowered to use the VarArgsFrameIndex 228 setOperationAction(ISD::VASTART , MVT::Other, Custom); 229 setOperationAction(ISD::VAARG , MVT::Other, Expand); 230 setOperationAction(ISD::VAEND , MVT::Other, Expand); 231 if (Subtarget->is64Bit()) 232 setOperationAction(ISD::VACOPY , MVT::Other, Custom); 233 else 234 setOperationAction(ISD::VACOPY , MVT::Other, Expand); 235 236 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 237 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 238 if (Subtarget->is64Bit()) 239 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand); 240 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand); 241 242 if (X86ScalarSSE) { 243 // Set up the FP register classes. 244 addRegisterClass(MVT::f32, X86::FR32RegisterClass); 245 addRegisterClass(MVT::f64, X86::FR64RegisterClass); 246 247 // Use ANDPD to simulate FABS. 248 setOperationAction(ISD::FABS , MVT::f64, Custom); 249 setOperationAction(ISD::FABS , MVT::f32, Custom); 250 251 // Use XORP to simulate FNEG. 252 setOperationAction(ISD::FNEG , MVT::f64, Custom); 253 setOperationAction(ISD::FNEG , MVT::f32, Custom); 254 255 // Use ANDPD and ORPD to simulate FCOPYSIGN. 256 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 257 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 258 259 // We don't support sin/cos/fmod 260 setOperationAction(ISD::FSIN , MVT::f64, Expand); 261 setOperationAction(ISD::FCOS , MVT::f64, Expand); 262 setOperationAction(ISD::FREM , MVT::f64, Expand); 263 setOperationAction(ISD::FSIN , MVT::f32, Expand); 264 setOperationAction(ISD::FCOS , MVT::f32, Expand); 265 setOperationAction(ISD::FREM , MVT::f32, Expand); 266 267 // Expand FP immediates into loads from the stack, except for the special 268 // cases we handle. 269 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 270 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 271 addLegalFPImmediate(+0.0); // xorps / xorpd 272 } else { 273 // Set up the FP register classes. 274 addRegisterClass(MVT::f64, X86::RFPRegisterClass); 275 276 setOperationAction(ISD::UNDEF, MVT::f64, Expand); 277 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); 278 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand); 279 280 if (!UnsafeFPMath) { 281 setOperationAction(ISD::FSIN , MVT::f64 , Expand); 282 setOperationAction(ISD::FCOS , MVT::f64 , Expand); 283 } 284 285 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 286 addLegalFPImmediate(+0.0); // FLD0 287 addLegalFPImmediate(+1.0); // FLD1 288 addLegalFPImmediate(-0.0); // FLD0/FCHS 289 addLegalFPImmediate(-1.0); // FLD1/FCHS 290 } 291 292 // First set operation action for all vector types to expand. Then we 293 // will selectively turn on ones that can be effectively codegen'd. 294 for (unsigned VT = (unsigned)MVT::Vector + 1; 295 VT != (unsigned)MVT::LAST_VALUETYPE; VT++) { 296 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand); 297 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand); 298 setOperationAction(ISD::FADD, (MVT::ValueType)VT, Expand); 299 setOperationAction(ISD::FSUB, (MVT::ValueType)VT, Expand); 300 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand); 301 setOperationAction(ISD::FMUL, (MVT::ValueType)VT, Expand); 302 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand); 303 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand); 304 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand); 305 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand); 306 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand); 307 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand); 308 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Expand); 309 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand); 310 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand); 311 } 312 313 if (Subtarget->hasMMX()) { 314 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass); 315 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass); 316 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass); 317 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass); 318 319 // FIXME: add MMX packed arithmetics 320 321 setOperationAction(ISD::ADD, MVT::v8i8, Legal); 322 setOperationAction(ISD::ADD, MVT::v4i16, Legal); 323 setOperationAction(ISD::ADD, MVT::v2i32, Legal); 324 325 setOperationAction(ISD::SUB, MVT::v8i8, Legal); 326 setOperationAction(ISD::SUB, MVT::v4i16, Legal); 327 setOperationAction(ISD::SUB, MVT::v2i32, Legal); 328 329 setOperationAction(ISD::MULHS, MVT::v4i16, Legal); 330 setOperationAction(ISD::MUL, MVT::v4i16, Legal); 331 332 setOperationAction(ISD::AND, MVT::v8i8, Promote); 333 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64); 334 setOperationAction(ISD::AND, MVT::v4i16, Promote); 335 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64); 336 setOperationAction(ISD::AND, MVT::v2i32, Promote); 337 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64); 338 setOperationAction(ISD::AND, MVT::v1i64, Legal); 339 340 setOperationAction(ISD::OR, MVT::v8i8, Promote); 341 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64); 342 setOperationAction(ISD::OR, MVT::v4i16, Promote); 343 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64); 344 setOperationAction(ISD::OR, MVT::v2i32, Promote); 345 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64); 346 setOperationAction(ISD::OR, MVT::v1i64, Legal); 347 348 setOperationAction(ISD::XOR, MVT::v8i8, Promote); 349 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64); 350 setOperationAction(ISD::XOR, MVT::v4i16, Promote); 351 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64); 352 setOperationAction(ISD::XOR, MVT::v2i32, Promote); 353 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64); 354 setOperationAction(ISD::XOR, MVT::v1i64, Legal); 355 356 setOperationAction(ISD::LOAD, MVT::v8i8, Promote); 357 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64); 358 setOperationAction(ISD::LOAD, MVT::v4i16, Promote); 359 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64); 360 setOperationAction(ISD::LOAD, MVT::v2i32, Promote); 361 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64); 362 setOperationAction(ISD::LOAD, MVT::v1i64, Legal); 363 364 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom); 365 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom); 366 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom); 367 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom); 368 369 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom); 370 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom); 371 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom); 372 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom); 373 374 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom); 375 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom); 376 } 377 378 if (Subtarget->hasSSE1()) { 379 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass); 380 381 setOperationAction(ISD::FADD, MVT::v4f32, Legal); 382 setOperationAction(ISD::FSUB, MVT::v4f32, Legal); 383 setOperationAction(ISD::FMUL, MVT::v4f32, Legal); 384 setOperationAction(ISD::FDIV, MVT::v4f32, Legal); 385 setOperationAction(ISD::LOAD, MVT::v4f32, Legal); 386 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom); 387 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom); 388 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom); 389 setOperationAction(ISD::SELECT, MVT::v4f32, Custom); 390 } 391 392 if (Subtarget->hasSSE2()) { 393 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass); 394 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass); 395 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass); 396 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass); 397 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass); 398 399 setOperationAction(ISD::ADD, MVT::v16i8, Legal); 400 setOperationAction(ISD::ADD, MVT::v8i16, Legal); 401 setOperationAction(ISD::ADD, MVT::v4i32, Legal); 402 setOperationAction(ISD::ADD, MVT::v2i64, Legal); 403 setOperationAction(ISD::SUB, MVT::v16i8, Legal); 404 setOperationAction(ISD::SUB, MVT::v8i16, Legal); 405 setOperationAction(ISD::SUB, MVT::v4i32, Legal); 406 setOperationAction(ISD::SUB, MVT::v2i64, Legal); 407 setOperationAction(ISD::MUL, MVT::v8i16, Legal); 408 setOperationAction(ISD::FADD, MVT::v2f64, Legal); 409 setOperationAction(ISD::FSUB, MVT::v2f64, Legal); 410 setOperationAction(ISD::FMUL, MVT::v2f64, Legal); 411 setOperationAction(ISD::FDIV, MVT::v2f64, Legal); 412 413 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom); 414 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom); 415 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom); 416 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom); 417 // Implement v4f32 insert_vector_elt in terms of SSE2 v8i16 ones. 418 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom); 419 420 // Custom lower build_vector, vector_shuffle, and extract_vector_elt. 421 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) { 422 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom); 423 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom); 424 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom); 425 } 426 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom); 427 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom); 428 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom); 429 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom); 430 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom); 431 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom); 432 433 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64. 434 for (unsigned VT = (unsigned)MVT::v16i8; VT != (unsigned)MVT::v2i64; VT++) { 435 setOperationAction(ISD::AND, (MVT::ValueType)VT, Promote); 436 AddPromotedToType (ISD::AND, (MVT::ValueType)VT, MVT::v2i64); 437 setOperationAction(ISD::OR, (MVT::ValueType)VT, Promote); 438 AddPromotedToType (ISD::OR, (MVT::ValueType)VT, MVT::v2i64); 439 setOperationAction(ISD::XOR, (MVT::ValueType)VT, Promote); 440 AddPromotedToType (ISD::XOR, (MVT::ValueType)VT, MVT::v2i64); 441 setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Promote); 442 AddPromotedToType (ISD::LOAD, (MVT::ValueType)VT, MVT::v2i64); 443 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote); 444 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v2i64); 445 } 446 447 // Custom lower v2i64 and v2f64 selects. 448 setOperationAction(ISD::LOAD, MVT::v2f64, Legal); 449 setOperationAction(ISD::LOAD, MVT::v2i64, Legal); 450 setOperationAction(ISD::SELECT, MVT::v2f64, Custom); 451 setOperationAction(ISD::SELECT, MVT::v2i64, Custom); 452 } 453 454 // We want to custom lower some of our intrinsics. 455 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 456 457 // We have target-specific dag combine patterns for the following nodes: 458 setTargetDAGCombine(ISD::VECTOR_SHUFFLE); 459 setTargetDAGCombine(ISD::SELECT); 460 461 computeRegisterProperties(); 462 463 // FIXME: These should be based on subtarget info. Plus, the values should 464 // be smaller when we are in optimizing for size mode. 465 maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores 466 maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores 467 maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores 468 allowUnalignedMemoryAccesses = true; // x86 supports it! 469} 470 471 472//===----------------------------------------------------------------------===// 473// Return Value Calling Convention Implementation 474//===----------------------------------------------------------------------===// 475 476#include "X86GenCallingConv.inc" 477 478/// LowerRET - Lower an ISD::RET node. 479SDOperand X86TargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG) { 480 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args"); 481 482 SmallVector<CCValAssign, 16> RVLocs; 483 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv(); 484 CCState CCInfo(CC, getTargetMachine(), RVLocs); 485 CCInfo.AnalyzeReturn(Op.Val, RetCC_X86); 486 487 488 // If this is the first return lowered for this function, add the regs to the 489 // liveout set for the function. 490 if (DAG.getMachineFunction().liveout_empty()) { 491 for (unsigned i = 0; i != RVLocs.size(); ++i) 492 if (RVLocs[i].isRegLoc()) 493 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg()); 494 } 495 496 SDOperand Chain = Op.getOperand(0); 497 SDOperand Flag; 498 499 // Copy the result values into the output registers. 500 if (RVLocs.size() != 1 || !RVLocs[0].isRegLoc() || 501 RVLocs[0].getLocReg() != X86::ST0) { 502 for (unsigned i = 0; i != RVLocs.size(); ++i) { 503 CCValAssign &VA = RVLocs[i]; 504 assert(VA.isRegLoc() && "Can only return in registers!"); 505 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), 506 Flag); 507 Flag = Chain.getValue(1); 508 } 509 } else { 510 // We need to handle a destination of ST0 specially, because it isn't really 511 // a register. 512 SDOperand Value = Op.getOperand(1); 513 514 // If this is an FP return with ScalarSSE, we need to move the value from 515 // an XMM register onto the fp-stack. 516 if (X86ScalarSSE) { 517 SDOperand MemLoc; 518 519 // If this is a load into a scalarsse value, don't store the loaded value 520 // back to the stack, only to reload it: just replace the scalar-sse load. 521 if (ISD::isNON_EXTLoad(Value.Val) && 522 (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) { 523 Chain = Value.getOperand(0); 524 MemLoc = Value.getOperand(1); 525 } else { 526 // Spill the value to memory and reload it into top of stack. 527 unsigned Size = MVT::getSizeInBits(RVLocs[0].getValVT())/8; 528 MachineFunction &MF = DAG.getMachineFunction(); 529 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size); 530 MemLoc = DAG.getFrameIndex(SSFI, getPointerTy()); 531 Chain = DAG.getStore(Op.getOperand(0), Value, MemLoc, NULL, 0); 532 } 533 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other); 534 SDOperand Ops[] = {Chain, MemLoc, DAG.getValueType(RVLocs[0].getValVT())}; 535 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3); 536 Chain = Value.getValue(1); 537 } 538 539 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag); 540 SDOperand Ops[] = { Chain, Value }; 541 Chain = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops, 2); 542 Flag = Chain.getValue(1); 543 } 544 545 SDOperand BytesToPop = DAG.getConstant(getBytesToPopOnReturn(), MVT::i16); 546 if (Flag.Val) 547 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop, Flag); 548 else 549 return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Chain, BytesToPop); 550} 551 552 553/// LowerCallResult - Lower the result values of an ISD::CALL into the 554/// appropriate copies out of appropriate physical registers. This assumes that 555/// Chain/InFlag are the input chain/flag to use, and that TheCall is the call 556/// being lowered. The returns a SDNode with the same number of values as the 557/// ISD::CALL. 558SDNode *X86TargetLowering:: 559LowerCallResult(SDOperand Chain, SDOperand InFlag, SDNode *TheCall, 560 unsigned CallingConv, SelectionDAG &DAG) { 561 562 // Assign locations to each value returned by this call. 563 SmallVector<CCValAssign, 16> RVLocs; 564 CCState CCInfo(CallingConv, getTargetMachine(), RVLocs); 565 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86); 566 567 568 SmallVector<SDOperand, 8> ResultVals; 569 570 // Copy all of the result registers out of their specified physreg. 571 if (RVLocs.size() != 1 || RVLocs[0].getLocReg() != X86::ST0) { 572 for (unsigned i = 0; i != RVLocs.size(); ++i) { 573 Chain = DAG.getCopyFromReg(Chain, RVLocs[i].getLocReg(), 574 RVLocs[i].getValVT(), InFlag).getValue(1); 575 InFlag = Chain.getValue(2); 576 ResultVals.push_back(Chain.getValue(0)); 577 } 578 } else { 579 // Copies from the FP stack are special, as ST0 isn't a valid register 580 // before the fp stackifier runs. 581 582 // Copy ST0 into an RFP register with FP_GET_RESULT. 583 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag); 584 SDOperand GROps[] = { Chain, InFlag }; 585 SDOperand RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, GROps, 2); 586 Chain = RetVal.getValue(1); 587 InFlag = RetVal.getValue(2); 588 589 // If we are using ScalarSSE, store ST(0) to the stack and reload it into 590 // an XMM register. 591 if (X86ScalarSSE) { 592 // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This 593 // shouldn't be necessary except that RFP cannot be live across 594 // multiple blocks. When stackifier is fixed, they can be uncoupled. 595 MachineFunction &MF = DAG.getMachineFunction(); 596 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8); 597 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); 598 SDOperand Ops[] = { 599 Chain, RetVal, StackSlot, DAG.getValueType(RVLocs[0].getValVT()), InFlag 600 }; 601 Chain = DAG.getNode(X86ISD::FST, MVT::Other, Ops, 5); 602 RetVal = DAG.getLoad(RVLocs[0].getValVT(), Chain, StackSlot, NULL, 0); 603 Chain = RetVal.getValue(1); 604 } 605 606 if (RVLocs[0].getValVT() == MVT::f32 && !X86ScalarSSE) 607 // FIXME: we would really like to remember that this FP_ROUND 608 // operation is okay to eliminate if we allow excess FP precision. 609 RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal); 610 ResultVals.push_back(RetVal); 611 } 612 613 // Merge everything together with a MERGE_VALUES node. 614 ResultVals.push_back(Chain); 615 return DAG.getNode(ISD::MERGE_VALUES, TheCall->getVTList(), 616 &ResultVals[0], ResultVals.size()).Val; 617} 618 619 620//===----------------------------------------------------------------------===// 621// C & StdCall Calling Convention implementation 622//===----------------------------------------------------------------------===// 623// StdCall calling convention seems to be standard for many Windows' API 624// routines and around. It differs from C calling convention just a little: 625// callee should clean up the stack, not caller. Symbols should be also 626// decorated in some fancy way :) It doesn't support any vector arguments. 627 628/// AddLiveIn - This helper function adds the specified physical register to the 629/// MachineFunction as a live in value. It also creates a corresponding virtual 630/// register for it. 631static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg, 632 const TargetRegisterClass *RC) { 633 assert(RC->contains(PReg) && "Not the correct regclass!"); 634 unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC); 635 MF.addLiveIn(PReg, VReg); 636 return VReg; 637} 638 639SDOperand X86TargetLowering::LowerCCCArguments(SDOperand Op, SelectionDAG &DAG, 640 bool isStdCall) { 641 unsigned NumArgs = Op.Val->getNumValues() - 1; 642 MachineFunction &MF = DAG.getMachineFunction(); 643 MachineFrameInfo *MFI = MF.getFrameInfo(); 644 SDOperand Root = Op.getOperand(0); 645 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0; 646 647 // Assign locations to all of the incoming arguments. 648 SmallVector<CCValAssign, 16> ArgLocs; 649 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(), 650 ArgLocs); 651 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_C); 652 653 SmallVector<SDOperand, 8> ArgValues; 654 unsigned LastVal = ~0U; 655 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 656 CCValAssign &VA = ArgLocs[i]; 657 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later 658 // places. 659 assert(VA.getValNo() != LastVal && 660 "Don't support value assigned to multiple locs yet"); 661 LastVal = VA.getValNo(); 662 663 if (VA.isRegLoc()) { 664 MVT::ValueType RegVT = VA.getLocVT(); 665 TargetRegisterClass *RC; 666 if (RegVT == MVT::i32) 667 RC = X86::GR32RegisterClass; 668 else { 669 assert(MVT::isVector(RegVT)); 670 RC = X86::VR128RegisterClass; 671 } 672 673 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC); 674 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT); 675 676 // If this is an 8 or 16-bit value, it is really passed promoted to 32 677 // bits. Insert an assert[sz]ext to capture this, then truncate to the 678 // right size. 679 if (VA.getLocInfo() == CCValAssign::SExt) 680 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue, 681 DAG.getValueType(VA.getValVT())); 682 else if (VA.getLocInfo() == CCValAssign::ZExt) 683 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue, 684 DAG.getValueType(VA.getValVT())); 685 686 if (VA.getLocInfo() != CCValAssign::Full) 687 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue); 688 689 ArgValues.push_back(ArgValue); 690 } else { 691 assert(VA.isMemLoc()); 692 693 // Create the nodes corresponding to a load from this parameter slot. 694 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8, 695 VA.getLocMemOffset()); 696 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy()); 697 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0)); 698 } 699 } 700 701 unsigned StackSize = CCInfo.getNextStackOffset(); 702 703 ArgValues.push_back(Root); 704 705 // If the function takes variable number of arguments, make a frame index for 706 // the start of the first vararg value... for expansion of llvm.va_start. 707 if (isVarArg) 708 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize); 709 710 if (isStdCall && !isVarArg) { 711 BytesToPopOnReturn = StackSize; // Callee pops everything.. 712 BytesCallerReserves = 0; 713 } else { 714 BytesToPopOnReturn = 0; // Callee pops nothing. 715 716 // If this is an sret function, the return should pop the hidden pointer. 717 if (NumArgs && 718 (cast<ConstantSDNode>(Op.getOperand(3))->getValue() & 719 ISD::ParamFlags::StructReturn)) 720 BytesToPopOnReturn = 4; 721 722 BytesCallerReserves = StackSize; 723 } 724 725 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only. 726 ReturnAddrIndex = 0; // No return address slot generated yet. 727 728 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn); 729 730 // Return the new list of results. 731 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), 732 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo); 733} 734 735SDOperand X86TargetLowering::LowerCCCCallTo(SDOperand Op, SelectionDAG &DAG, 736 unsigned CC) { 737 SDOperand Chain = Op.getOperand(0); 738 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0; 739 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0; 740 SDOperand Callee = Op.getOperand(4); 741 unsigned NumOps = (Op.getNumOperands() - 5) / 2; 742 743 // Analyze operands of the call, assigning locations to each operand. 744 SmallVector<CCValAssign, 16> ArgLocs; 745 CCState CCInfo(CC, getTargetMachine(), ArgLocs); 746 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_C); 747 748 // Get a count of how many bytes are to be pushed on the stack. 749 unsigned NumBytes = CCInfo.getNextStackOffset(); 750 751 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy())); 752 753 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass; 754 SmallVector<SDOperand, 8> MemOpChains; 755 756 SDOperand StackPtr; 757 758 // Walk the register/memloc assignments, inserting copies/loads. 759 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 760 CCValAssign &VA = ArgLocs[i]; 761 SDOperand Arg = Op.getOperand(5+2*VA.getValNo()); 762 763 // Promote the value if needed. 764 switch (VA.getLocInfo()) { 765 default: assert(0 && "Unknown loc info!"); 766 case CCValAssign::Full: break; 767 case CCValAssign::SExt: 768 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg); 769 break; 770 case CCValAssign::ZExt: 771 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg); 772 break; 773 case CCValAssign::AExt: 774 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg); 775 break; 776 } 777 778 if (VA.isRegLoc()) { 779 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 780 } else { 781 assert(VA.isMemLoc()); 782 if (StackPtr.Val == 0) 783 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy()); 784 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy()); 785 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff); 786 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0)); 787 } 788 } 789 790 // If the first argument is an sret pointer, remember it. 791 bool isSRet = NumOps && 792 (cast<ConstantSDNode>(Op.getOperand(6))->getValue() & 793 ISD::ParamFlags::StructReturn); 794 795 if (!MemOpChains.empty()) 796 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, 797 &MemOpChains[0], MemOpChains.size()); 798 799 // Build a sequence of copy-to-reg nodes chained together with token chain 800 // and flag operands which copy the outgoing args into registers. 801 SDOperand InFlag; 802 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 803 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second, 804 InFlag); 805 InFlag = Chain.getValue(1); 806 } 807 808 // ELF / PIC requires GOT in the EBX register before function calls via PLT 809 // GOT pointer. 810 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 811 Subtarget->isPICStyleGOT()) { 812 Chain = DAG.getCopyToReg(Chain, X86::EBX, 813 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), 814 InFlag); 815 InFlag = Chain.getValue(1); 816 } 817 818 // If the callee is a GlobalAddress node (quite common, every direct call is) 819 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 820 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 821 // We should use extra load for direct calls to dllimported functions in 822 // non-JIT mode. 823 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(), 824 getTargetMachine(), true)) 825 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy()); 826 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) 827 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy()); 828 829 // Returns a chain & a flag for retval copy to use. 830 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 831 SmallVector<SDOperand, 8> Ops; 832 Ops.push_back(Chain); 833 Ops.push_back(Callee); 834 835 // Add argument registers to the end of the list so that they are known live 836 // into the call. 837 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 838 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 839 RegsToPass[i].second.getValueType())); 840 841 // Add an implicit use GOT pointer in EBX. 842 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 843 Subtarget->isPICStyleGOT()) 844 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy())); 845 846 if (InFlag.Val) 847 Ops.push_back(InFlag); 848 849 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL, 850 NodeTys, &Ops[0], Ops.size()); 851 InFlag = Chain.getValue(1); 852 853 // Create the CALLSEQ_END node. 854 unsigned NumBytesForCalleeToPush = 0; 855 856 if (CC == CallingConv::X86_StdCall) { 857 if (isVarArg) 858 NumBytesForCalleeToPush = isSRet ? 4 : 0; 859 else 860 NumBytesForCalleeToPush = NumBytes; 861 } else { 862 // If this is is a call to a struct-return function, the callee 863 // pops the hidden struct pointer, so we have to push it back. 864 // This is common for Darwin/X86, Linux & Mingw32 targets. 865 NumBytesForCalleeToPush = isSRet ? 4 : 0; 866 } 867 868 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 869 Ops.clear(); 870 Ops.push_back(Chain); 871 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy())); 872 Ops.push_back(DAG.getConstant(NumBytesForCalleeToPush, getPointerTy())); 873 Ops.push_back(InFlag); 874 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size()); 875 InFlag = Chain.getValue(1); 876 877 // Handle result values, copying them out of physregs into vregs that we 878 // return. 879 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo); 880} 881 882 883//===----------------------------------------------------------------------===// 884// FastCall Calling Convention implementation 885//===----------------------------------------------------------------------===// 886// 887// The X86 'fastcall' calling convention passes up to two integer arguments in 888// registers (an appropriate portion of ECX/EDX), passes arguments in C order, 889// and requires that the callee pop its arguments off the stack (allowing proper 890// tail calls), and has the same return value conventions as C calling convs. 891// 892// This calling convention always arranges for the callee pop value to be 8n+4 893// bytes, which is needed for tail recursion elimination and stack alignment 894// reasons. 895SDOperand 896X86TargetLowering::LowerFastCCArguments(SDOperand Op, SelectionDAG &DAG) { 897 MachineFunction &MF = DAG.getMachineFunction(); 898 MachineFrameInfo *MFI = MF.getFrameInfo(); 899 SDOperand Root = Op.getOperand(0); 900 901 // Assign locations to all of the incoming arguments. 902 SmallVector<CCValAssign, 16> ArgLocs; 903 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(), 904 ArgLocs); 905 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_32_FastCall); 906 907 SmallVector<SDOperand, 8> ArgValues; 908 unsigned LastVal = ~0U; 909 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 910 CCValAssign &VA = ArgLocs[i]; 911 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later 912 // places. 913 assert(VA.getValNo() != LastVal && 914 "Don't support value assigned to multiple locs yet"); 915 LastVal = VA.getValNo(); 916 917 if (VA.isRegLoc()) { 918 MVT::ValueType RegVT = VA.getLocVT(); 919 TargetRegisterClass *RC; 920 if (RegVT == MVT::i32) 921 RC = X86::GR32RegisterClass; 922 else { 923 assert(MVT::isVector(RegVT)); 924 RC = X86::VR128RegisterClass; 925 } 926 927 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC); 928 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT); 929 930 // If this is an 8 or 16-bit value, it is really passed promoted to 32 931 // bits. Insert an assert[sz]ext to capture this, then truncate to the 932 // right size. 933 if (VA.getLocInfo() == CCValAssign::SExt) 934 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue, 935 DAG.getValueType(VA.getValVT())); 936 else if (VA.getLocInfo() == CCValAssign::ZExt) 937 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue, 938 DAG.getValueType(VA.getValVT())); 939 940 if (VA.getLocInfo() != CCValAssign::Full) 941 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue); 942 943 ArgValues.push_back(ArgValue); 944 } else { 945 assert(VA.isMemLoc()); 946 947 // Create the nodes corresponding to a load from this parameter slot. 948 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8, 949 VA.getLocMemOffset()); 950 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy()); 951 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0)); 952 } 953 } 954 955 ArgValues.push_back(Root); 956 957 unsigned StackSize = CCInfo.getNextStackOffset(); 958 959 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) { 960 // Make sure the instruction takes 8n+4 bytes to make sure the start of the 961 // arguments and the arguments after the retaddr has been pushed are aligned. 962 if ((StackSize & 7) == 0) 963 StackSize += 4; 964 } 965 966 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs. 967 RegSaveFrameIndex = 0xAAAAAAA; // X86-64 only. 968 ReturnAddrIndex = 0; // No return address slot generated yet. 969 BytesToPopOnReturn = StackSize; // Callee pops all stack arguments. 970 BytesCallerReserves = 0; 971 972 MF.getInfo<X86FunctionInfo>()->setBytesToPopOnReturn(BytesToPopOnReturn); 973 974 // Return the new list of results. 975 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), 976 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo); 977} 978 979SDOperand X86TargetLowering::LowerFastCCCallTo(SDOperand Op, SelectionDAG &DAG, 980 unsigned CC) { 981 SDOperand Chain = Op.getOperand(0); 982 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0; 983 SDOperand Callee = Op.getOperand(4); 984 985 // Analyze operands of the call, assigning locations to each operand. 986 SmallVector<CCValAssign, 16> ArgLocs; 987 CCState CCInfo(CC, getTargetMachine(), ArgLocs); 988 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_32_FastCall); 989 990 // Get a count of how many bytes are to be pushed on the stack. 991 unsigned NumBytes = CCInfo.getNextStackOffset(); 992 993 if (!Subtarget->isTargetCygMing() && !Subtarget->isTargetWindows()) { 994 // Make sure the instruction takes 8n+4 bytes to make sure the start of the 995 // arguments and the arguments after the retaddr has been pushed are aligned. 996 if ((NumBytes & 7) == 0) 997 NumBytes += 4; 998 } 999 1000 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy())); 1001 1002 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass; 1003 SmallVector<SDOperand, 8> MemOpChains; 1004 1005 SDOperand StackPtr; 1006 1007 // Walk the register/memloc assignments, inserting copies/loads. 1008 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1009 CCValAssign &VA = ArgLocs[i]; 1010 SDOperand Arg = Op.getOperand(5+2*VA.getValNo()); 1011 1012 // Promote the value if needed. 1013 switch (VA.getLocInfo()) { 1014 default: assert(0 && "Unknown loc info!"); 1015 case CCValAssign::Full: break; 1016 case CCValAssign::SExt: 1017 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg); 1018 break; 1019 case CCValAssign::ZExt: 1020 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg); 1021 break; 1022 case CCValAssign::AExt: 1023 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg); 1024 break; 1025 } 1026 1027 if (VA.isRegLoc()) { 1028 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 1029 } else { 1030 assert(VA.isMemLoc()); 1031 if (StackPtr.Val == 0) 1032 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy()); 1033 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy()); 1034 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff); 1035 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0)); 1036 } 1037 } 1038 1039 if (!MemOpChains.empty()) 1040 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, 1041 &MemOpChains[0], MemOpChains.size()); 1042 1043 // Build a sequence of copy-to-reg nodes chained together with token chain 1044 // and flag operands which copy the outgoing args into registers. 1045 SDOperand InFlag; 1046 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1047 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second, 1048 InFlag); 1049 InFlag = Chain.getValue(1); 1050 } 1051 1052 // If the callee is a GlobalAddress node (quite common, every direct call is) 1053 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 1054 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1055 // We should use extra load for direct calls to dllimported functions in 1056 // non-JIT mode. 1057 if (!Subtarget->GVRequiresExtraLoad(G->getGlobal(), 1058 getTargetMachine(), true)) 1059 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy()); 1060 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) 1061 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy()); 1062 1063 // ELF / PIC requires GOT in the EBX register before function calls via PLT 1064 // GOT pointer. 1065 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 1066 Subtarget->isPICStyleGOT()) { 1067 Chain = DAG.getCopyToReg(Chain, X86::EBX, 1068 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), 1069 InFlag); 1070 InFlag = Chain.getValue(1); 1071 } 1072 1073 // Returns a chain & a flag for retval copy to use. 1074 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 1075 SmallVector<SDOperand, 8> Ops; 1076 Ops.push_back(Chain); 1077 Ops.push_back(Callee); 1078 1079 // Add argument registers to the end of the list so that they are known live 1080 // into the call. 1081 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1082 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1083 RegsToPass[i].second.getValueType())); 1084 1085 // Add an implicit use GOT pointer in EBX. 1086 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 1087 Subtarget->isPICStyleGOT()) 1088 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy())); 1089 1090 if (InFlag.Val) 1091 Ops.push_back(InFlag); 1092 1093 // FIXME: Do not generate X86ISD::TAILCALL for now. 1094 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL, 1095 NodeTys, &Ops[0], Ops.size()); 1096 InFlag = Chain.getValue(1); 1097 1098 // Returns a flag for retval copy to use. 1099 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 1100 Ops.clear(); 1101 Ops.push_back(Chain); 1102 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy())); 1103 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy())); 1104 Ops.push_back(InFlag); 1105 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size()); 1106 InFlag = Chain.getValue(1); 1107 1108 // Handle result values, copying them out of physregs into vregs that we 1109 // return. 1110 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo); 1111} 1112 1113 1114//===----------------------------------------------------------------------===// 1115// X86-64 C Calling Convention implementation 1116//===----------------------------------------------------------------------===// 1117 1118SDOperand 1119X86TargetLowering::LowerX86_64CCCArguments(SDOperand Op, SelectionDAG &DAG) { 1120 MachineFunction &MF = DAG.getMachineFunction(); 1121 MachineFrameInfo *MFI = MF.getFrameInfo(); 1122 SDOperand Root = Op.getOperand(0); 1123 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0; 1124 1125 static const unsigned GPR64ArgRegs[] = { 1126 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9 1127 }; 1128 static const unsigned XMMArgRegs[] = { 1129 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, 1130 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 1131 }; 1132 1133 1134 // Assign locations to all of the incoming arguments. 1135 SmallVector<CCValAssign, 16> ArgLocs; 1136 CCState CCInfo(MF.getFunction()->getCallingConv(), getTargetMachine(), 1137 ArgLocs); 1138 CCInfo.AnalyzeFormalArguments(Op.Val, CC_X86_64_C); 1139 1140 SmallVector<SDOperand, 8> ArgValues; 1141 unsigned LastVal = ~0U; 1142 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1143 CCValAssign &VA = ArgLocs[i]; 1144 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later 1145 // places. 1146 assert(VA.getValNo() != LastVal && 1147 "Don't support value assigned to multiple locs yet"); 1148 LastVal = VA.getValNo(); 1149 1150 if (VA.isRegLoc()) { 1151 MVT::ValueType RegVT = VA.getLocVT(); 1152 TargetRegisterClass *RC; 1153 if (RegVT == MVT::i32) 1154 RC = X86::GR32RegisterClass; 1155 else if (RegVT == MVT::i64) 1156 RC = X86::GR64RegisterClass; 1157 else if (RegVT == MVT::f32) 1158 RC = X86::FR32RegisterClass; 1159 else if (RegVT == MVT::f64) 1160 RC = X86::FR64RegisterClass; 1161 else { 1162 assert(MVT::isVector(RegVT)); 1163 RC = X86::VR128RegisterClass; 1164 } 1165 1166 unsigned Reg = AddLiveIn(DAG.getMachineFunction(), VA.getLocReg(), RC); 1167 SDOperand ArgValue = DAG.getCopyFromReg(Root, Reg, RegVT); 1168 1169 // If this is an 8 or 16-bit value, it is really passed promoted to 32 1170 // bits. Insert an assert[sz]ext to capture this, then truncate to the 1171 // right size. 1172 if (VA.getLocInfo() == CCValAssign::SExt) 1173 ArgValue = DAG.getNode(ISD::AssertSext, RegVT, ArgValue, 1174 DAG.getValueType(VA.getValVT())); 1175 else if (VA.getLocInfo() == CCValAssign::ZExt) 1176 ArgValue = DAG.getNode(ISD::AssertZext, RegVT, ArgValue, 1177 DAG.getValueType(VA.getValVT())); 1178 1179 if (VA.getLocInfo() != CCValAssign::Full) 1180 ArgValue = DAG.getNode(ISD::TRUNCATE, VA.getValVT(), ArgValue); 1181 1182 ArgValues.push_back(ArgValue); 1183 } else { 1184 assert(VA.isMemLoc()); 1185 1186 // Create the nodes corresponding to a load from this parameter slot. 1187 int FI = MFI->CreateFixedObject(MVT::getSizeInBits(VA.getValVT())/8, 1188 VA.getLocMemOffset()); 1189 SDOperand FIN = DAG.getFrameIndex(FI, getPointerTy()); 1190 ArgValues.push_back(DAG.getLoad(VA.getValVT(), Root, FIN, NULL, 0)); 1191 } 1192 } 1193 1194 unsigned StackSize = CCInfo.getNextStackOffset(); 1195 1196 // If the function takes variable number of arguments, make a frame index for 1197 // the start of the first vararg value... for expansion of llvm.va_start. 1198 if (isVarArg) { 1199 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs, 6); 1200 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8); 1201 1202 // For X86-64, if there are vararg parameters that are passed via 1203 // registers, then we must store them to their spots on the stack so they 1204 // may be loaded by deferencing the result of va_next. 1205 VarArgsGPOffset = NumIntRegs * 8; 1206 VarArgsFPOffset = 6 * 8 + NumXMMRegs * 16; 1207 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize); 1208 RegSaveFrameIndex = MFI->CreateStackObject(6 * 8 + 8 * 16, 16); 1209 1210 // Store the integer parameter registers. 1211 SmallVector<SDOperand, 8> MemOps; 1212 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy()); 1213 SDOperand FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN, 1214 DAG.getConstant(VarArgsGPOffset, getPointerTy())); 1215 for (; NumIntRegs != 6; ++NumIntRegs) { 1216 unsigned VReg = AddLiveIn(MF, GPR64ArgRegs[NumIntRegs], 1217 X86::GR64RegisterClass); 1218 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i64); 1219 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0); 1220 MemOps.push_back(Store); 1221 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, 1222 DAG.getConstant(8, getPointerTy())); 1223 } 1224 1225 // Now store the XMM (fp + vector) parameter registers. 1226 FIN = DAG.getNode(ISD::ADD, getPointerTy(), RSFIN, 1227 DAG.getConstant(VarArgsFPOffset, getPointerTy())); 1228 for (; NumXMMRegs != 8; ++NumXMMRegs) { 1229 unsigned VReg = AddLiveIn(MF, XMMArgRegs[NumXMMRegs], 1230 X86::VR128RegisterClass); 1231 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::v4f32); 1232 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0); 1233 MemOps.push_back(Store); 1234 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, 1235 DAG.getConstant(16, getPointerTy())); 1236 } 1237 if (!MemOps.empty()) 1238 Root = DAG.getNode(ISD::TokenFactor, MVT::Other, 1239 &MemOps[0], MemOps.size()); 1240 } 1241 1242 ArgValues.push_back(Root); 1243 1244 ReturnAddrIndex = 0; // No return address slot generated yet. 1245 BytesToPopOnReturn = 0; // Callee pops nothing. 1246 BytesCallerReserves = StackSize; 1247 1248 // Return the new list of results. 1249 return DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(), 1250 &ArgValues[0], ArgValues.size()).getValue(Op.ResNo); 1251} 1252 1253SDOperand 1254X86TargetLowering::LowerX86_64CCCCallTo(SDOperand Op, SelectionDAG &DAG, 1255 unsigned CC) { 1256 SDOperand Chain = Op.getOperand(0); 1257 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0; 1258 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0; 1259 SDOperand Callee = Op.getOperand(4); 1260 1261 // Analyze operands of the call, assigning locations to each operand. 1262 SmallVector<CCValAssign, 16> ArgLocs; 1263 CCState CCInfo(CC, getTargetMachine(), ArgLocs); 1264 CCInfo.AnalyzeCallOperands(Op.Val, CC_X86_64_C); 1265 1266 // Get a count of how many bytes are to be pushed on the stack. 1267 unsigned NumBytes = CCInfo.getNextStackOffset(); 1268 Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy())); 1269 1270 SmallVector<std::pair<unsigned, SDOperand>, 8> RegsToPass; 1271 SmallVector<SDOperand, 8> MemOpChains; 1272 1273 SDOperand StackPtr; 1274 1275 // Walk the register/memloc assignments, inserting copies/loads. 1276 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1277 CCValAssign &VA = ArgLocs[i]; 1278 SDOperand Arg = Op.getOperand(5+2*VA.getValNo()); 1279 1280 // Promote the value if needed. 1281 switch (VA.getLocInfo()) { 1282 default: assert(0 && "Unknown loc info!"); 1283 case CCValAssign::Full: break; 1284 case CCValAssign::SExt: 1285 Arg = DAG.getNode(ISD::SIGN_EXTEND, VA.getLocVT(), Arg); 1286 break; 1287 case CCValAssign::ZExt: 1288 Arg = DAG.getNode(ISD::ZERO_EXTEND, VA.getLocVT(), Arg); 1289 break; 1290 case CCValAssign::AExt: 1291 Arg = DAG.getNode(ISD::ANY_EXTEND, VA.getLocVT(), Arg); 1292 break; 1293 } 1294 1295 if (VA.isRegLoc()) { 1296 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 1297 } else { 1298 assert(VA.isMemLoc()); 1299 if (StackPtr.Val == 0) 1300 StackPtr = DAG.getRegister(getStackPtrReg(), getPointerTy()); 1301 SDOperand PtrOff = DAG.getConstant(VA.getLocMemOffset(), getPointerTy()); 1302 PtrOff = DAG.getNode(ISD::ADD, getPointerTy(), StackPtr, PtrOff); 1303 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0)); 1304 } 1305 } 1306 1307 if (!MemOpChains.empty()) 1308 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, 1309 &MemOpChains[0], MemOpChains.size()); 1310 1311 // Build a sequence of copy-to-reg nodes chained together with token chain 1312 // and flag operands which copy the outgoing args into registers. 1313 SDOperand InFlag; 1314 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 1315 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second, 1316 InFlag); 1317 InFlag = Chain.getValue(1); 1318 } 1319 1320 if (isVarArg) { 1321 // From AMD64 ABI document: 1322 // For calls that may call functions that use varargs or stdargs 1323 // (prototype-less calls or calls to functions containing ellipsis (...) in 1324 // the declaration) %al is used as hidden argument to specify the number 1325 // of SSE registers used. The contents of %al do not need to match exactly 1326 // the number of registers, but must be an ubound on the number of SSE 1327 // registers used and is in the range 0 - 8 inclusive. 1328 1329 // Count the number of XMM registers allocated. 1330 static const unsigned XMMArgRegs[] = { 1331 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, 1332 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 1333 }; 1334 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8); 1335 1336 Chain = DAG.getCopyToReg(Chain, X86::AL, 1337 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag); 1338 InFlag = Chain.getValue(1); 1339 } 1340 1341 // If the callee is a GlobalAddress node (quite common, every direct call is) 1342 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. 1343 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 1344 // We should use extra load for direct calls to dllimported functions in 1345 // non-JIT mode. 1346 if (getTargetMachine().getCodeModel() != CodeModel::Large 1347 && !Subtarget->GVRequiresExtraLoad(G->getGlobal(), 1348 getTargetMachine(), true)) 1349 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy()); 1350 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) 1351 if (getTargetMachine().getCodeModel() != CodeModel::Large) 1352 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy()); 1353 1354 // Returns a chain & a flag for retval copy to use. 1355 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 1356 SmallVector<SDOperand, 8> Ops; 1357 Ops.push_back(Chain); 1358 Ops.push_back(Callee); 1359 1360 // Add argument registers to the end of the list so that they are known live 1361 // into the call. 1362 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1363 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1364 RegsToPass[i].second.getValueType())); 1365 1366 if (InFlag.Val) 1367 Ops.push_back(InFlag); 1368 1369 // FIXME: Do not generate X86ISD::TAILCALL for now. 1370 Chain = DAG.getNode(isTailCall ? X86ISD::TAILCALL : X86ISD::CALL, 1371 NodeTys, &Ops[0], Ops.size()); 1372 InFlag = Chain.getValue(1); 1373 1374 // Returns a flag for retval copy to use. 1375 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag); 1376 Ops.clear(); 1377 Ops.push_back(Chain); 1378 Ops.push_back(DAG.getConstant(NumBytes, getPointerTy())); 1379 Ops.push_back(DAG.getConstant(0, getPointerTy())); 1380 Ops.push_back(InFlag); 1381 Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, &Ops[0], Ops.size()); 1382 InFlag = Chain.getValue(1); 1383 1384 // Handle result values, copying them out of physregs into vregs that we 1385 // return. 1386 return SDOperand(LowerCallResult(Chain, InFlag, Op.Val, CC, DAG), Op.ResNo); 1387} 1388 1389 1390//===----------------------------------------------------------------------===// 1391// Other Lowering Hooks 1392//===----------------------------------------------------------------------===// 1393 1394 1395SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) { 1396 if (ReturnAddrIndex == 0) { 1397 // Set up a frame object for the return address. 1398 MachineFunction &MF = DAG.getMachineFunction(); 1399 if (Subtarget->is64Bit()) 1400 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(8, -8); 1401 else 1402 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4); 1403 } 1404 1405 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy()); 1406} 1407 1408 1409 1410/// translateX86CC - do a one to one translation of a ISD::CondCode to the X86 1411/// specific condition code. It returns a false if it cannot do a direct 1412/// translation. X86CC is the translated CondCode. LHS/RHS are modified as 1413/// needed. 1414static bool translateX86CC(ISD::CondCode SetCCOpcode, bool isFP, 1415 unsigned &X86CC, SDOperand &LHS, SDOperand &RHS, 1416 SelectionDAG &DAG) { 1417 X86CC = X86::COND_INVALID; 1418 if (!isFP) { 1419 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) { 1420 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) { 1421 // X > -1 -> X == 0, jump !sign. 1422 RHS = DAG.getConstant(0, RHS.getValueType()); 1423 X86CC = X86::COND_NS; 1424 return true; 1425 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) { 1426 // X < 0 -> X == 0, jump on sign. 1427 X86CC = X86::COND_S; 1428 return true; 1429 } 1430 } 1431 1432 switch (SetCCOpcode) { 1433 default: break; 1434 case ISD::SETEQ: X86CC = X86::COND_E; break; 1435 case ISD::SETGT: X86CC = X86::COND_G; break; 1436 case ISD::SETGE: X86CC = X86::COND_GE; break; 1437 case ISD::SETLT: X86CC = X86::COND_L; break; 1438 case ISD::SETLE: X86CC = X86::COND_LE; break; 1439 case ISD::SETNE: X86CC = X86::COND_NE; break; 1440 case ISD::SETULT: X86CC = X86::COND_B; break; 1441 case ISD::SETUGT: X86CC = X86::COND_A; break; 1442 case ISD::SETULE: X86CC = X86::COND_BE; break; 1443 case ISD::SETUGE: X86CC = X86::COND_AE; break; 1444 } 1445 } else { 1446 // On a floating point condition, the flags are set as follows: 1447 // ZF PF CF op 1448 // 0 | 0 | 0 | X > Y 1449 // 0 | 0 | 1 | X < Y 1450 // 1 | 0 | 0 | X == Y 1451 // 1 | 1 | 1 | unordered 1452 bool Flip = false; 1453 switch (SetCCOpcode) { 1454 default: break; 1455 case ISD::SETUEQ: 1456 case ISD::SETEQ: X86CC = X86::COND_E; break; 1457 case ISD::SETOLT: Flip = true; // Fallthrough 1458 case ISD::SETOGT: 1459 case ISD::SETGT: X86CC = X86::COND_A; break; 1460 case ISD::SETOLE: Flip = true; // Fallthrough 1461 case ISD::SETOGE: 1462 case ISD::SETGE: X86CC = X86::COND_AE; break; 1463 case ISD::SETUGT: Flip = true; // Fallthrough 1464 case ISD::SETULT: 1465 case ISD::SETLT: X86CC = X86::COND_B; break; 1466 case ISD::SETUGE: Flip = true; // Fallthrough 1467 case ISD::SETULE: 1468 case ISD::SETLE: X86CC = X86::COND_BE; break; 1469 case ISD::SETONE: 1470 case ISD::SETNE: X86CC = X86::COND_NE; break; 1471 case ISD::SETUO: X86CC = X86::COND_P; break; 1472 case ISD::SETO: X86CC = X86::COND_NP; break; 1473 } 1474 if (Flip) 1475 std::swap(LHS, RHS); 1476 } 1477 1478 return X86CC != X86::COND_INVALID; 1479} 1480 1481/// hasFPCMov - is there a floating point cmov for the specific X86 condition 1482/// code. Current x86 isa includes the following FP cmov instructions: 1483/// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu. 1484static bool hasFPCMov(unsigned X86CC) { 1485 switch (X86CC) { 1486 default: 1487 return false; 1488 case X86::COND_B: 1489 case X86::COND_BE: 1490 case X86::COND_E: 1491 case X86::COND_P: 1492 case X86::COND_A: 1493 case X86::COND_AE: 1494 case X86::COND_NE: 1495 case X86::COND_NP: 1496 return true; 1497 } 1498} 1499 1500/// isUndefOrInRange - Op is either an undef node or a ConstantSDNode. Return 1501/// true if Op is undef or if its value falls within the specified range (L, H]. 1502static bool isUndefOrInRange(SDOperand Op, unsigned Low, unsigned Hi) { 1503 if (Op.getOpcode() == ISD::UNDEF) 1504 return true; 1505 1506 unsigned Val = cast<ConstantSDNode>(Op)->getValue(); 1507 return (Val >= Low && Val < Hi); 1508} 1509 1510/// isUndefOrEqual - Op is either an undef node or a ConstantSDNode. Return 1511/// true if Op is undef or if its value equal to the specified value. 1512static bool isUndefOrEqual(SDOperand Op, unsigned Val) { 1513 if (Op.getOpcode() == ISD::UNDEF) 1514 return true; 1515 return cast<ConstantSDNode>(Op)->getValue() == Val; 1516} 1517 1518/// isPSHUFDMask - Return true if the specified VECTOR_SHUFFLE operand 1519/// specifies a shuffle of elements that is suitable for input to PSHUFD. 1520bool X86::isPSHUFDMask(SDNode *N) { 1521 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1522 1523 if (N->getNumOperands() != 4) 1524 return false; 1525 1526 // Check if the value doesn't reference the second vector. 1527 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 1528 SDOperand Arg = N->getOperand(i); 1529 if (Arg.getOpcode() == ISD::UNDEF) continue; 1530 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1531 if (cast<ConstantSDNode>(Arg)->getValue() >= 4) 1532 return false; 1533 } 1534 1535 return true; 1536} 1537 1538/// isPSHUFHWMask - Return true if the specified VECTOR_SHUFFLE operand 1539/// specifies a shuffle of elements that is suitable for input to PSHUFHW. 1540bool X86::isPSHUFHWMask(SDNode *N) { 1541 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1542 1543 if (N->getNumOperands() != 8) 1544 return false; 1545 1546 // Lower quadword copied in order. 1547 for (unsigned i = 0; i != 4; ++i) { 1548 SDOperand Arg = N->getOperand(i); 1549 if (Arg.getOpcode() == ISD::UNDEF) continue; 1550 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1551 if (cast<ConstantSDNode>(Arg)->getValue() != i) 1552 return false; 1553 } 1554 1555 // Upper quadword shuffled. 1556 for (unsigned i = 4; i != 8; ++i) { 1557 SDOperand Arg = N->getOperand(i); 1558 if (Arg.getOpcode() == ISD::UNDEF) continue; 1559 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1560 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 1561 if (Val < 4 || Val > 7) 1562 return false; 1563 } 1564 1565 return true; 1566} 1567 1568/// isPSHUFLWMask - Return true if the specified VECTOR_SHUFFLE operand 1569/// specifies a shuffle of elements that is suitable for input to PSHUFLW. 1570bool X86::isPSHUFLWMask(SDNode *N) { 1571 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1572 1573 if (N->getNumOperands() != 8) 1574 return false; 1575 1576 // Upper quadword copied in order. 1577 for (unsigned i = 4; i != 8; ++i) 1578 if (!isUndefOrEqual(N->getOperand(i), i)) 1579 return false; 1580 1581 // Lower quadword shuffled. 1582 for (unsigned i = 0; i != 4; ++i) 1583 if (!isUndefOrInRange(N->getOperand(i), 0, 4)) 1584 return false; 1585 1586 return true; 1587} 1588 1589/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand 1590/// specifies a shuffle of elements that is suitable for input to SHUFP*. 1591static bool isSHUFPMask(const SDOperand *Elems, unsigned NumElems) { 1592 if (NumElems != 2 && NumElems != 4) return false; 1593 1594 unsigned Half = NumElems / 2; 1595 for (unsigned i = 0; i < Half; ++i) 1596 if (!isUndefOrInRange(Elems[i], 0, NumElems)) 1597 return false; 1598 for (unsigned i = Half; i < NumElems; ++i) 1599 if (!isUndefOrInRange(Elems[i], NumElems, NumElems*2)) 1600 return false; 1601 1602 return true; 1603} 1604 1605bool X86::isSHUFPMask(SDNode *N) { 1606 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1607 return ::isSHUFPMask(N->op_begin(), N->getNumOperands()); 1608} 1609 1610/// isCommutedSHUFP - Returns true if the shuffle mask is except 1611/// the reverse of what x86 shuffles want. x86 shuffles requires the lower 1612/// half elements to come from vector 1 (which would equal the dest.) and 1613/// the upper half to come from vector 2. 1614static bool isCommutedSHUFP(const SDOperand *Ops, unsigned NumOps) { 1615 if (NumOps != 2 && NumOps != 4) return false; 1616 1617 unsigned Half = NumOps / 2; 1618 for (unsigned i = 0; i < Half; ++i) 1619 if (!isUndefOrInRange(Ops[i], NumOps, NumOps*2)) 1620 return false; 1621 for (unsigned i = Half; i < NumOps; ++i) 1622 if (!isUndefOrInRange(Ops[i], 0, NumOps)) 1623 return false; 1624 return true; 1625} 1626 1627static bool isCommutedSHUFP(SDNode *N) { 1628 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1629 return isCommutedSHUFP(N->op_begin(), N->getNumOperands()); 1630} 1631 1632/// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand 1633/// specifies a shuffle of elements that is suitable for input to MOVHLPS. 1634bool X86::isMOVHLPSMask(SDNode *N) { 1635 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1636 1637 if (N->getNumOperands() != 4) 1638 return false; 1639 1640 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3 1641 return isUndefOrEqual(N->getOperand(0), 6) && 1642 isUndefOrEqual(N->getOperand(1), 7) && 1643 isUndefOrEqual(N->getOperand(2), 2) && 1644 isUndefOrEqual(N->getOperand(3), 3); 1645} 1646 1647/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form 1648/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef, 1649/// <2, 3, 2, 3> 1650bool X86::isMOVHLPS_v_undef_Mask(SDNode *N) { 1651 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1652 1653 if (N->getNumOperands() != 4) 1654 return false; 1655 1656 // Expect bit0 == 2, bit1 == 3, bit2 == 2, bit3 == 3 1657 return isUndefOrEqual(N->getOperand(0), 2) && 1658 isUndefOrEqual(N->getOperand(1), 3) && 1659 isUndefOrEqual(N->getOperand(2), 2) && 1660 isUndefOrEqual(N->getOperand(3), 3); 1661} 1662 1663/// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand 1664/// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}. 1665bool X86::isMOVLPMask(SDNode *N) { 1666 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1667 1668 unsigned NumElems = N->getNumOperands(); 1669 if (NumElems != 2 && NumElems != 4) 1670 return false; 1671 1672 for (unsigned i = 0; i < NumElems/2; ++i) 1673 if (!isUndefOrEqual(N->getOperand(i), i + NumElems)) 1674 return false; 1675 1676 for (unsigned i = NumElems/2; i < NumElems; ++i) 1677 if (!isUndefOrEqual(N->getOperand(i), i)) 1678 return false; 1679 1680 return true; 1681} 1682 1683/// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand 1684/// specifies a shuffle of elements that is suitable for input to MOVHP{S|D} 1685/// and MOVLHPS. 1686bool X86::isMOVHPMask(SDNode *N) { 1687 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1688 1689 unsigned NumElems = N->getNumOperands(); 1690 if (NumElems != 2 && NumElems != 4) 1691 return false; 1692 1693 for (unsigned i = 0; i < NumElems/2; ++i) 1694 if (!isUndefOrEqual(N->getOperand(i), i)) 1695 return false; 1696 1697 for (unsigned i = 0; i < NumElems/2; ++i) { 1698 SDOperand Arg = N->getOperand(i + NumElems/2); 1699 if (!isUndefOrEqual(Arg, i + NumElems)) 1700 return false; 1701 } 1702 1703 return true; 1704} 1705 1706/// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand 1707/// specifies a shuffle of elements that is suitable for input to UNPCKL. 1708bool static isUNPCKLMask(const SDOperand *Elts, unsigned NumElts, 1709 bool V2IsSplat = false) { 1710 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16) 1711 return false; 1712 1713 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) { 1714 SDOperand BitI = Elts[i]; 1715 SDOperand BitI1 = Elts[i+1]; 1716 if (!isUndefOrEqual(BitI, j)) 1717 return false; 1718 if (V2IsSplat) { 1719 if (isUndefOrEqual(BitI1, NumElts)) 1720 return false; 1721 } else { 1722 if (!isUndefOrEqual(BitI1, j + NumElts)) 1723 return false; 1724 } 1725 } 1726 1727 return true; 1728} 1729 1730bool X86::isUNPCKLMask(SDNode *N, bool V2IsSplat) { 1731 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1732 return ::isUNPCKLMask(N->op_begin(), N->getNumOperands(), V2IsSplat); 1733} 1734 1735/// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand 1736/// specifies a shuffle of elements that is suitable for input to UNPCKH. 1737bool static isUNPCKHMask(const SDOperand *Elts, unsigned NumElts, 1738 bool V2IsSplat = false) { 1739 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16) 1740 return false; 1741 1742 for (unsigned i = 0, j = 0; i != NumElts; i += 2, ++j) { 1743 SDOperand BitI = Elts[i]; 1744 SDOperand BitI1 = Elts[i+1]; 1745 if (!isUndefOrEqual(BitI, j + NumElts/2)) 1746 return false; 1747 if (V2IsSplat) { 1748 if (isUndefOrEqual(BitI1, NumElts)) 1749 return false; 1750 } else { 1751 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts)) 1752 return false; 1753 } 1754 } 1755 1756 return true; 1757} 1758 1759bool X86::isUNPCKHMask(SDNode *N, bool V2IsSplat) { 1760 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1761 return ::isUNPCKHMask(N->op_begin(), N->getNumOperands(), V2IsSplat); 1762} 1763 1764/// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form 1765/// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef, 1766/// <0, 0, 1, 1> 1767bool X86::isUNPCKL_v_undef_Mask(SDNode *N) { 1768 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1769 1770 unsigned NumElems = N->getNumOperands(); 1771 if (NumElems != 4 && NumElems != 8 && NumElems != 16) 1772 return false; 1773 1774 for (unsigned i = 0, j = 0; i != NumElems; i += 2, ++j) { 1775 SDOperand BitI = N->getOperand(i); 1776 SDOperand BitI1 = N->getOperand(i+1); 1777 1778 if (!isUndefOrEqual(BitI, j)) 1779 return false; 1780 if (!isUndefOrEqual(BitI1, j)) 1781 return false; 1782 } 1783 1784 return true; 1785} 1786 1787/// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand 1788/// specifies a shuffle of elements that is suitable for input to MOVSS, 1789/// MOVSD, and MOVD, i.e. setting the lowest element. 1790static bool isMOVLMask(const SDOperand *Elts, unsigned NumElts) { 1791 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16) 1792 return false; 1793 1794 if (!isUndefOrEqual(Elts[0], NumElts)) 1795 return false; 1796 1797 for (unsigned i = 1; i < NumElts; ++i) { 1798 if (!isUndefOrEqual(Elts[i], i)) 1799 return false; 1800 } 1801 1802 return true; 1803} 1804 1805bool X86::isMOVLMask(SDNode *N) { 1806 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1807 return ::isMOVLMask(N->op_begin(), N->getNumOperands()); 1808} 1809 1810/// isCommutedMOVL - Returns true if the shuffle mask is except the reverse 1811/// of what x86 movss want. X86 movs requires the lowest element to be lowest 1812/// element of vector 2 and the other elements to come from vector 1 in order. 1813static bool isCommutedMOVL(const SDOperand *Ops, unsigned NumOps, 1814 bool V2IsSplat = false, 1815 bool V2IsUndef = false) { 1816 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16) 1817 return false; 1818 1819 if (!isUndefOrEqual(Ops[0], 0)) 1820 return false; 1821 1822 for (unsigned i = 1; i < NumOps; ++i) { 1823 SDOperand Arg = Ops[i]; 1824 if (!(isUndefOrEqual(Arg, i+NumOps) || 1825 (V2IsUndef && isUndefOrInRange(Arg, NumOps, NumOps*2)) || 1826 (V2IsSplat && isUndefOrEqual(Arg, NumOps)))) 1827 return false; 1828 } 1829 1830 return true; 1831} 1832 1833static bool isCommutedMOVL(SDNode *N, bool V2IsSplat = false, 1834 bool V2IsUndef = false) { 1835 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1836 return isCommutedMOVL(N->op_begin(), N->getNumOperands(), 1837 V2IsSplat, V2IsUndef); 1838} 1839 1840/// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand 1841/// specifies a shuffle of elements that is suitable for input to MOVSHDUP. 1842bool X86::isMOVSHDUPMask(SDNode *N) { 1843 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1844 1845 if (N->getNumOperands() != 4) 1846 return false; 1847 1848 // Expect 1, 1, 3, 3 1849 for (unsigned i = 0; i < 2; ++i) { 1850 SDOperand Arg = N->getOperand(i); 1851 if (Arg.getOpcode() == ISD::UNDEF) continue; 1852 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1853 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 1854 if (Val != 1) return false; 1855 } 1856 1857 bool HasHi = false; 1858 for (unsigned i = 2; i < 4; ++i) { 1859 SDOperand Arg = N->getOperand(i); 1860 if (Arg.getOpcode() == ISD::UNDEF) continue; 1861 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1862 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 1863 if (Val != 3) return false; 1864 HasHi = true; 1865 } 1866 1867 // Don't use movshdup if it can be done with a shufps. 1868 return HasHi; 1869} 1870 1871/// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand 1872/// specifies a shuffle of elements that is suitable for input to MOVSLDUP. 1873bool X86::isMOVSLDUPMask(SDNode *N) { 1874 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1875 1876 if (N->getNumOperands() != 4) 1877 return false; 1878 1879 // Expect 0, 0, 2, 2 1880 for (unsigned i = 0; i < 2; ++i) { 1881 SDOperand Arg = N->getOperand(i); 1882 if (Arg.getOpcode() == ISD::UNDEF) continue; 1883 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1884 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 1885 if (Val != 0) return false; 1886 } 1887 1888 bool HasHi = false; 1889 for (unsigned i = 2; i < 4; ++i) { 1890 SDOperand Arg = N->getOperand(i); 1891 if (Arg.getOpcode() == ISD::UNDEF) continue; 1892 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1893 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 1894 if (Val != 2) return false; 1895 HasHi = true; 1896 } 1897 1898 // Don't use movshdup if it can be done with a shufps. 1899 return HasHi; 1900} 1901 1902/// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies 1903/// a splat of a single element. 1904static bool isSplatMask(SDNode *N) { 1905 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1906 1907 // This is a splat operation if each element of the permute is the same, and 1908 // if the value doesn't reference the second vector. 1909 unsigned NumElems = N->getNumOperands(); 1910 SDOperand ElementBase; 1911 unsigned i = 0; 1912 for (; i != NumElems; ++i) { 1913 SDOperand Elt = N->getOperand(i); 1914 if (isa<ConstantSDNode>(Elt)) { 1915 ElementBase = Elt; 1916 break; 1917 } 1918 } 1919 1920 if (!ElementBase.Val) 1921 return false; 1922 1923 for (; i != NumElems; ++i) { 1924 SDOperand Arg = N->getOperand(i); 1925 if (Arg.getOpcode() == ISD::UNDEF) continue; 1926 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 1927 if (Arg != ElementBase) return false; 1928 } 1929 1930 // Make sure it is a splat of the first vector operand. 1931 return cast<ConstantSDNode>(ElementBase)->getValue() < NumElems; 1932} 1933 1934/// isSplatMask - Return true if the specified VECTOR_SHUFFLE operand specifies 1935/// a splat of a single element and it's a 2 or 4 element mask. 1936bool X86::isSplatMask(SDNode *N) { 1937 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1938 1939 // We can only splat 64-bit, and 32-bit quantities with a single instruction. 1940 if (N->getNumOperands() != 4 && N->getNumOperands() != 2) 1941 return false; 1942 return ::isSplatMask(N); 1943} 1944 1945/// isSplatLoMask - Return true if the specified VECTOR_SHUFFLE operand 1946/// specifies a splat of zero element. 1947bool X86::isSplatLoMask(SDNode *N) { 1948 assert(N->getOpcode() == ISD::BUILD_VECTOR); 1949 1950 for (unsigned i = 0, e = N->getNumOperands(); i < e; ++i) 1951 if (!isUndefOrEqual(N->getOperand(i), 0)) 1952 return false; 1953 return true; 1954} 1955 1956/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle 1957/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP* 1958/// instructions. 1959unsigned X86::getShuffleSHUFImmediate(SDNode *N) { 1960 unsigned NumOperands = N->getNumOperands(); 1961 unsigned Shift = (NumOperands == 4) ? 2 : 1; 1962 unsigned Mask = 0; 1963 for (unsigned i = 0; i < NumOperands; ++i) { 1964 unsigned Val = 0; 1965 SDOperand Arg = N->getOperand(NumOperands-i-1); 1966 if (Arg.getOpcode() != ISD::UNDEF) 1967 Val = cast<ConstantSDNode>(Arg)->getValue(); 1968 if (Val >= NumOperands) Val -= NumOperands; 1969 Mask |= Val; 1970 if (i != NumOperands - 1) 1971 Mask <<= Shift; 1972 } 1973 1974 return Mask; 1975} 1976 1977/// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle 1978/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW 1979/// instructions. 1980unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) { 1981 unsigned Mask = 0; 1982 // 8 nodes, but we only care about the last 4. 1983 for (unsigned i = 7; i >= 4; --i) { 1984 unsigned Val = 0; 1985 SDOperand Arg = N->getOperand(i); 1986 if (Arg.getOpcode() != ISD::UNDEF) 1987 Val = cast<ConstantSDNode>(Arg)->getValue(); 1988 Mask |= (Val - 4); 1989 if (i != 4) 1990 Mask <<= 2; 1991 } 1992 1993 return Mask; 1994} 1995 1996/// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle 1997/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW 1998/// instructions. 1999unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) { 2000 unsigned Mask = 0; 2001 // 8 nodes, but we only care about the first 4. 2002 for (int i = 3; i >= 0; --i) { 2003 unsigned Val = 0; 2004 SDOperand Arg = N->getOperand(i); 2005 if (Arg.getOpcode() != ISD::UNDEF) 2006 Val = cast<ConstantSDNode>(Arg)->getValue(); 2007 Mask |= Val; 2008 if (i != 0) 2009 Mask <<= 2; 2010 } 2011 2012 return Mask; 2013} 2014 2015/// isPSHUFHW_PSHUFLWMask - true if the specified VECTOR_SHUFFLE operand 2016/// specifies a 8 element shuffle that can be broken into a pair of 2017/// PSHUFHW and PSHUFLW. 2018static bool isPSHUFHW_PSHUFLWMask(SDNode *N) { 2019 assert(N->getOpcode() == ISD::BUILD_VECTOR); 2020 2021 if (N->getNumOperands() != 8) 2022 return false; 2023 2024 // Lower quadword shuffled. 2025 for (unsigned i = 0; i != 4; ++i) { 2026 SDOperand Arg = N->getOperand(i); 2027 if (Arg.getOpcode() == ISD::UNDEF) continue; 2028 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 2029 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 2030 if (Val > 4) 2031 return false; 2032 } 2033 2034 // Upper quadword shuffled. 2035 for (unsigned i = 4; i != 8; ++i) { 2036 SDOperand Arg = N->getOperand(i); 2037 if (Arg.getOpcode() == ISD::UNDEF) continue; 2038 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 2039 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 2040 if (Val < 4 || Val > 7) 2041 return false; 2042 } 2043 2044 return true; 2045} 2046 2047/// CommuteVectorShuffle - Swap vector_shuffle operandsas well as 2048/// values in ther permute mask. 2049static SDOperand CommuteVectorShuffle(SDOperand Op, SDOperand &V1, 2050 SDOperand &V2, SDOperand &Mask, 2051 SelectionDAG &DAG) { 2052 MVT::ValueType VT = Op.getValueType(); 2053 MVT::ValueType MaskVT = Mask.getValueType(); 2054 MVT::ValueType EltVT = MVT::getVectorBaseType(MaskVT); 2055 unsigned NumElems = Mask.getNumOperands(); 2056 SmallVector<SDOperand, 8> MaskVec; 2057 2058 for (unsigned i = 0; i != NumElems; ++i) { 2059 SDOperand Arg = Mask.getOperand(i); 2060 if (Arg.getOpcode() == ISD::UNDEF) { 2061 MaskVec.push_back(DAG.getNode(ISD::UNDEF, EltVT)); 2062 continue; 2063 } 2064 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!"); 2065 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 2066 if (Val < NumElems) 2067 MaskVec.push_back(DAG.getConstant(Val + NumElems, EltVT)); 2068 else 2069 MaskVec.push_back(DAG.getConstant(Val - NumElems, EltVT)); 2070 } 2071 2072 std::swap(V1, V2); 2073 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size()); 2074 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask); 2075} 2076 2077/// ShouldXformToMOVHLPS - Return true if the node should be transformed to 2078/// match movhlps. The lower half elements should come from upper half of 2079/// V1 (and in order), and the upper half elements should come from the upper 2080/// half of V2 (and in order). 2081static bool ShouldXformToMOVHLPS(SDNode *Mask) { 2082 unsigned NumElems = Mask->getNumOperands(); 2083 if (NumElems != 4) 2084 return false; 2085 for (unsigned i = 0, e = 2; i != e; ++i) 2086 if (!isUndefOrEqual(Mask->getOperand(i), i+2)) 2087 return false; 2088 for (unsigned i = 2; i != 4; ++i) 2089 if (!isUndefOrEqual(Mask->getOperand(i), i+4)) 2090 return false; 2091 return true; 2092} 2093 2094/// isScalarLoadToVector - Returns true if the node is a scalar load that 2095/// is promoted to a vector. 2096static inline bool isScalarLoadToVector(SDNode *N) { 2097 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) { 2098 N = N->getOperand(0).Val; 2099 return ISD::isNON_EXTLoad(N); 2100 } 2101 return false; 2102} 2103 2104/// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to 2105/// match movlp{s|d}. The lower half elements should come from lower half of 2106/// V1 (and in order), and the upper half elements should come from the upper 2107/// half of V2 (and in order). And since V1 will become the source of the 2108/// MOVLP, it must be either a vector load or a scalar load to vector. 2109static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2, SDNode *Mask) { 2110 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1)) 2111 return false; 2112 // Is V2 is a vector load, don't do this transformation. We will try to use 2113 // load folding shufps op. 2114 if (ISD::isNON_EXTLoad(V2)) 2115 return false; 2116 2117 unsigned NumElems = Mask->getNumOperands(); 2118 if (NumElems != 2 && NumElems != 4) 2119 return false; 2120 for (unsigned i = 0, e = NumElems/2; i != e; ++i) 2121 if (!isUndefOrEqual(Mask->getOperand(i), i)) 2122 return false; 2123 for (unsigned i = NumElems/2; i != NumElems; ++i) 2124 if (!isUndefOrEqual(Mask->getOperand(i), i+NumElems)) 2125 return false; 2126 return true; 2127} 2128 2129/// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are 2130/// all the same. 2131static bool isSplatVector(SDNode *N) { 2132 if (N->getOpcode() != ISD::BUILD_VECTOR) 2133 return false; 2134 2135 SDOperand SplatValue = N->getOperand(0); 2136 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i) 2137 if (N->getOperand(i) != SplatValue) 2138 return false; 2139 return true; 2140} 2141 2142/// isUndefShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved 2143/// to an undef. 2144static bool isUndefShuffle(SDNode *N) { 2145 if (N->getOpcode() != ISD::BUILD_VECTOR) 2146 return false; 2147 2148 SDOperand V1 = N->getOperand(0); 2149 SDOperand V2 = N->getOperand(1); 2150 SDOperand Mask = N->getOperand(2); 2151 unsigned NumElems = Mask.getNumOperands(); 2152 for (unsigned i = 0; i != NumElems; ++i) { 2153 SDOperand Arg = Mask.getOperand(i); 2154 if (Arg.getOpcode() != ISD::UNDEF) { 2155 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 2156 if (Val < NumElems && V1.getOpcode() != ISD::UNDEF) 2157 return false; 2158 else if (Val >= NumElems && V2.getOpcode() != ISD::UNDEF) 2159 return false; 2160 } 2161 } 2162 return true; 2163} 2164 2165/// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements 2166/// that point to V2 points to its first element. 2167static SDOperand NormalizeMask(SDOperand Mask, SelectionDAG &DAG) { 2168 assert(Mask.getOpcode() == ISD::BUILD_VECTOR); 2169 2170 bool Changed = false; 2171 SmallVector<SDOperand, 8> MaskVec; 2172 unsigned NumElems = Mask.getNumOperands(); 2173 for (unsigned i = 0; i != NumElems; ++i) { 2174 SDOperand Arg = Mask.getOperand(i); 2175 if (Arg.getOpcode() != ISD::UNDEF) { 2176 unsigned Val = cast<ConstantSDNode>(Arg)->getValue(); 2177 if (Val > NumElems) { 2178 Arg = DAG.getConstant(NumElems, Arg.getValueType()); 2179 Changed = true; 2180 } 2181 } 2182 MaskVec.push_back(Arg); 2183 } 2184 2185 if (Changed) 2186 Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(), 2187 &MaskVec[0], MaskVec.size()); 2188 return Mask; 2189} 2190 2191/// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd 2192/// operation of specified width. 2193static SDOperand getMOVLMask(unsigned NumElems, SelectionDAG &DAG) { 2194 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems); 2195 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT); 2196 2197 SmallVector<SDOperand, 8> MaskVec; 2198 MaskVec.push_back(DAG.getConstant(NumElems, BaseVT)); 2199 for (unsigned i = 1; i != NumElems; ++i) 2200 MaskVec.push_back(DAG.getConstant(i, BaseVT)); 2201 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size()); 2202} 2203 2204/// getUnpacklMask - Returns a vector_shuffle mask for an unpackl operation 2205/// of specified width. 2206static SDOperand getUnpacklMask(unsigned NumElems, SelectionDAG &DAG) { 2207 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems); 2208 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT); 2209 SmallVector<SDOperand, 8> MaskVec; 2210 for (unsigned i = 0, e = NumElems/2; i != e; ++i) { 2211 MaskVec.push_back(DAG.getConstant(i, BaseVT)); 2212 MaskVec.push_back(DAG.getConstant(i + NumElems, BaseVT)); 2213 } 2214 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size()); 2215} 2216 2217/// getUnpackhMask - Returns a vector_shuffle mask for an unpackh operation 2218/// of specified width. 2219static SDOperand getUnpackhMask(unsigned NumElems, SelectionDAG &DAG) { 2220 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems); 2221 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT); 2222 unsigned Half = NumElems/2; 2223 SmallVector<SDOperand, 8> MaskVec; 2224 for (unsigned i = 0; i != Half; ++i) { 2225 MaskVec.push_back(DAG.getConstant(i + Half, BaseVT)); 2226 MaskVec.push_back(DAG.getConstant(i + NumElems + Half, BaseVT)); 2227 } 2228 return DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0], MaskVec.size()); 2229} 2230 2231/// getZeroVector - Returns a vector of specified type with all zero elements. 2232/// 2233static SDOperand getZeroVector(MVT::ValueType VT, SelectionDAG &DAG) { 2234 assert(MVT::isVector(VT) && "Expected a vector type"); 2235 unsigned NumElems = getVectorNumElements(VT); 2236 MVT::ValueType EVT = MVT::getVectorBaseType(VT); 2237 bool isFP = MVT::isFloatingPoint(EVT); 2238 SDOperand Zero = isFP ? DAG.getConstantFP(0.0, EVT) : DAG.getConstant(0, EVT); 2239 SmallVector<SDOperand, 8> ZeroVec(NumElems, Zero); 2240 return DAG.getNode(ISD::BUILD_VECTOR, VT, &ZeroVec[0], ZeroVec.size()); 2241} 2242 2243/// PromoteSplat - Promote a splat of v8i16 or v16i8 to v4i32. 2244/// 2245static SDOperand PromoteSplat(SDOperand Op, SelectionDAG &DAG) { 2246 SDOperand V1 = Op.getOperand(0); 2247 SDOperand Mask = Op.getOperand(2); 2248 MVT::ValueType VT = Op.getValueType(); 2249 unsigned NumElems = Mask.getNumOperands(); 2250 Mask = getUnpacklMask(NumElems, DAG); 2251 while (NumElems != 4) { 2252 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, Mask); 2253 NumElems >>= 1; 2254 } 2255 V1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, V1); 2256 2257 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4); 2258 Mask = getZeroVector(MaskVT, DAG); 2259 SDOperand Shuffle = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v4i32, V1, 2260 DAG.getNode(ISD::UNDEF, MVT::v4i32), Mask); 2261 return DAG.getNode(ISD::BIT_CONVERT, VT, Shuffle); 2262} 2263 2264/// isZeroNode - Returns true if Elt is a constant zero or a floating point 2265/// constant +0.0. 2266static inline bool isZeroNode(SDOperand Elt) { 2267 return ((isa<ConstantSDNode>(Elt) && 2268 cast<ConstantSDNode>(Elt)->getValue() == 0) || 2269 (isa<ConstantFPSDNode>(Elt) && 2270 cast<ConstantFPSDNode>(Elt)->isExactlyValue(0.0))); 2271} 2272 2273/// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified 2274/// vector and zero or undef vector. 2275static SDOperand getShuffleVectorZeroOrUndef(SDOperand V2, MVT::ValueType VT, 2276 unsigned NumElems, unsigned Idx, 2277 bool isZero, SelectionDAG &DAG) { 2278 SDOperand V1 = isZero ? getZeroVector(VT, DAG) : DAG.getNode(ISD::UNDEF, VT); 2279 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems); 2280 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT); 2281 SDOperand Zero = DAG.getConstant(0, EVT); 2282 SmallVector<SDOperand, 8> MaskVec(NumElems, Zero); 2283 MaskVec[Idx] = DAG.getConstant(NumElems, EVT); 2284 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2285 &MaskVec[0], MaskVec.size()); 2286 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask); 2287} 2288 2289/// LowerBuildVectorv8i8 - Custom lower build_vector of v8i8. 2290/// 2291static SDOperand LowerBuildVectorv8i8(SDOperand Op, unsigned NonZeros, 2292 unsigned NumNonZero, unsigned NumZero, 2293 SelectionDAG &DAG, TargetLowering &TLI) { 2294 if (NumNonZero > 8) 2295 return SDOperand(); 2296 2297 SDOperand V(0, 0); 2298 bool First = true; 2299 for (unsigned i = 0; i < 8; ++i) { 2300 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0; 2301 if (ThisIsNonZero && First) { 2302 if (NumZero) 2303 V = getZeroVector(MVT::v4i16, DAG); 2304 else 2305 V = DAG.getNode(ISD::UNDEF, MVT::v4i16); 2306 First = false; 2307 } 2308 2309 if ((i & 1) != 0) { 2310 SDOperand ThisElt(0, 0), LastElt(0, 0); 2311 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0; 2312 if (LastIsNonZero) { 2313 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1)); 2314 } 2315 if (ThisIsNonZero) { 2316 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i)); 2317 ThisElt = DAG.getNode(ISD::SHL, MVT::i16, 2318 ThisElt, DAG.getConstant(8, MVT::i8)); 2319 if (LastIsNonZero) 2320 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt); 2321 } else 2322 ThisElt = LastElt; 2323 2324 if (ThisElt.Val) 2325 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v4i16, V, ThisElt, 2326 DAG.getConstant(i/2, TLI.getPointerTy())); 2327 } 2328 } 2329 2330 return DAG.getNode(ISD::BIT_CONVERT, MVT::v8i8, V); 2331} 2332 2333/// LowerBuildVectorv4i16 - Custom lower build_vector of v4i16. 2334/// 2335static SDOperand LowerBuildVectorv4i16(SDOperand Op, unsigned NonZeros, 2336 unsigned NumNonZero, unsigned NumZero, 2337 SelectionDAG &DAG, TargetLowering &TLI) { 2338 if (NumNonZero > 4) 2339 return SDOperand(); 2340 2341 SDOperand V(0, 0); 2342 bool First = true; 2343 for (unsigned i = 0; i < 4; ++i) { 2344 bool isNonZero = (NonZeros & (1 << i)) != 0; 2345 if (isNonZero) { 2346 if (First) { 2347 if (NumZero) 2348 V = getZeroVector(MVT::v4i16, DAG); 2349 else 2350 V = DAG.getNode(ISD::UNDEF, MVT::v4i16); 2351 First = false; 2352 } 2353 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v4i16, V, Op.getOperand(i), 2354 DAG.getConstant(i, TLI.getPointerTy())); 2355 } 2356 } 2357 2358 return V; 2359} 2360 2361/// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8. 2362/// 2363static SDOperand LowerBuildVectorv16i8(SDOperand Op, unsigned NonZeros, 2364 unsigned NumNonZero, unsigned NumZero, 2365 SelectionDAG &DAG, TargetLowering &TLI) { 2366 if (NumNonZero > 8) 2367 return SDOperand(); 2368 2369 SDOperand V(0, 0); 2370 bool First = true; 2371 for (unsigned i = 0; i < 16; ++i) { 2372 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0; 2373 if (ThisIsNonZero && First) { 2374 if (NumZero) 2375 V = getZeroVector(MVT::v8i16, DAG); 2376 else 2377 V = DAG.getNode(ISD::UNDEF, MVT::v8i16); 2378 First = false; 2379 } 2380 2381 if ((i & 1) != 0) { 2382 SDOperand ThisElt(0, 0), LastElt(0, 0); 2383 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0; 2384 if (LastIsNonZero) { 2385 LastElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i-1)); 2386 } 2387 if (ThisIsNonZero) { 2388 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, Op.getOperand(i)); 2389 ThisElt = DAG.getNode(ISD::SHL, MVT::i16, 2390 ThisElt, DAG.getConstant(8, MVT::i8)); 2391 if (LastIsNonZero) 2392 ThisElt = DAG.getNode(ISD::OR, MVT::i16, ThisElt, LastElt); 2393 } else 2394 ThisElt = LastElt; 2395 2396 if (ThisElt.Val) 2397 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, ThisElt, 2398 DAG.getConstant(i/2, TLI.getPointerTy())); 2399 } 2400 } 2401 2402 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, V); 2403} 2404 2405/// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16. 2406/// 2407static SDOperand LowerBuildVectorv8i16(SDOperand Op, unsigned NonZeros, 2408 unsigned NumNonZero, unsigned NumZero, 2409 SelectionDAG &DAG, TargetLowering &TLI) { 2410 if (NumNonZero > 4) 2411 return SDOperand(); 2412 2413 SDOperand V(0, 0); 2414 bool First = true; 2415 for (unsigned i = 0; i < 8; ++i) { 2416 bool isNonZero = (NonZeros & (1 << i)) != 0; 2417 if (isNonZero) { 2418 if (First) { 2419 if (NumZero) 2420 V = getZeroVector(MVT::v8i16, DAG); 2421 else 2422 V = DAG.getNode(ISD::UNDEF, MVT::v8i16); 2423 First = false; 2424 } 2425 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, MVT::v8i16, V, Op.getOperand(i), 2426 DAG.getConstant(i, TLI.getPointerTy())); 2427 } 2428 } 2429 2430 return V; 2431} 2432 2433SDOperand 2434X86TargetLowering::LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) { 2435 // All zero's are handled with pxor. 2436 if (ISD::isBuildVectorAllZeros(Op.Val)) 2437 return Op; 2438 2439 // All one's are handled with pcmpeqd. 2440 if (ISD::isBuildVectorAllOnes(Op.Val)) 2441 return Op; 2442 2443 MVT::ValueType VT = Op.getValueType(); 2444 MVT::ValueType EVT = MVT::getVectorBaseType(VT); 2445 unsigned EVTBits = MVT::getSizeInBits(EVT); 2446 2447 unsigned NumElems = Op.getNumOperands(); 2448 unsigned NumZero = 0; 2449 unsigned NumNonZero = 0; 2450 unsigned NonZeros = 0; 2451 std::set<SDOperand> Values; 2452 for (unsigned i = 0; i < NumElems; ++i) { 2453 SDOperand Elt = Op.getOperand(i); 2454 if (Elt.getOpcode() != ISD::UNDEF) { 2455 Values.insert(Elt); 2456 if (isZeroNode(Elt)) 2457 NumZero++; 2458 else { 2459 NonZeros |= (1 << i); 2460 NumNonZero++; 2461 } 2462 } 2463 } 2464 2465 if (NumNonZero == 0) 2466 // Must be a mix of zero and undef. Return a zero vector. 2467 return getZeroVector(VT, DAG); 2468 2469 // Splat is obviously ok. Let legalizer expand it to a shuffle. 2470 if (Values.size() == 1) 2471 return SDOperand(); 2472 2473 // Special case for single non-zero element. 2474 if (NumNonZero == 1) { 2475 unsigned Idx = CountTrailingZeros_32(NonZeros); 2476 SDOperand Item = Op.getOperand(Idx); 2477 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Item); 2478 if (Idx == 0) 2479 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector. 2480 return getShuffleVectorZeroOrUndef(Item, VT, NumElems, Idx, 2481 NumZero > 0, DAG); 2482 2483 if (EVTBits == 32) { 2484 // Turn it into a shuffle of zero and zero-extended scalar to vector. 2485 Item = getShuffleVectorZeroOrUndef(Item, VT, NumElems, 0, NumZero > 0, 2486 DAG); 2487 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems); 2488 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT); 2489 SmallVector<SDOperand, 8> MaskVec; 2490 for (unsigned i = 0; i < NumElems; i++) 2491 MaskVec.push_back(DAG.getConstant((i == Idx) ? 0 : 1, MaskEVT)); 2492 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2493 &MaskVec[0], MaskVec.size()); 2494 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, Item, 2495 DAG.getNode(ISD::UNDEF, VT), Mask); 2496 } 2497 } 2498 2499 // Let legalizer expand 2-wide build_vector's. 2500 if (EVTBits == 64) 2501 return SDOperand(); 2502 2503 // If element VT is < 32 bits, convert it to inserts into a zero vector. 2504 if (EVTBits == 8 && NumElems == 16) { 2505 SDOperand V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG, 2506 *this); 2507 if (V.Val) return V; 2508 } 2509 2510 if (EVTBits == 16 && NumElems == 8) { 2511 SDOperand V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG, 2512 *this); 2513 if (V.Val) return V; 2514 } 2515 2516 // If element VT is == 32 bits, turn it into a number of shuffles. 2517 SmallVector<SDOperand, 8> V; 2518 V.resize(NumElems); 2519 if (NumElems == 4 && NumZero > 0) { 2520 for (unsigned i = 0; i < 4; ++i) { 2521 bool isZero = !(NonZeros & (1 << i)); 2522 if (isZero) 2523 V[i] = getZeroVector(VT, DAG); 2524 else 2525 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i)); 2526 } 2527 2528 for (unsigned i = 0; i < 2; ++i) { 2529 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) { 2530 default: break; 2531 case 0: 2532 V[i] = V[i*2]; // Must be a zero vector. 2533 break; 2534 case 1: 2535 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2+1], V[i*2], 2536 getMOVLMask(NumElems, DAG)); 2537 break; 2538 case 2: 2539 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1], 2540 getMOVLMask(NumElems, DAG)); 2541 break; 2542 case 3: 2543 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i*2], V[i*2+1], 2544 getUnpacklMask(NumElems, DAG)); 2545 break; 2546 } 2547 } 2548 2549 // Take advantage of the fact GR32 to VR128 scalar_to_vector (i.e. movd) 2550 // clears the upper bits. 2551 // FIXME: we can do the same for v4f32 case when we know both parts of 2552 // the lower half come from scalar_to_vector (loadf32). We should do 2553 // that in post legalizer dag combiner with target specific hooks. 2554 if (MVT::isInteger(EVT) && (NonZeros & (0x3 << 2)) == 0) 2555 return V[0]; 2556 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems); 2557 MVT::ValueType EVT = MVT::getVectorBaseType(MaskVT); 2558 SmallVector<SDOperand, 8> MaskVec; 2559 bool Reverse = (NonZeros & 0x3) == 2; 2560 for (unsigned i = 0; i < 2; ++i) 2561 if (Reverse) 2562 MaskVec.push_back(DAG.getConstant(1-i, EVT)); 2563 else 2564 MaskVec.push_back(DAG.getConstant(i, EVT)); 2565 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2; 2566 for (unsigned i = 0; i < 2; ++i) 2567 if (Reverse) 2568 MaskVec.push_back(DAG.getConstant(1-i+NumElems, EVT)); 2569 else 2570 MaskVec.push_back(DAG.getConstant(i+NumElems, EVT)); 2571 SDOperand ShufMask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2572 &MaskVec[0], MaskVec.size()); 2573 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[0], V[1], ShufMask); 2574 } 2575 2576 if (Values.size() > 2) { 2577 // Expand into a number of unpckl*. 2578 // e.g. for v4f32 2579 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0> 2580 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1> 2581 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0> 2582 SDOperand UnpckMask = getUnpacklMask(NumElems, DAG); 2583 for (unsigned i = 0; i < NumElems; ++i) 2584 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, Op.getOperand(i)); 2585 NumElems >>= 1; 2586 while (NumElems != 0) { 2587 for (unsigned i = 0; i < NumElems; ++i) 2588 V[i] = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V[i], V[i + NumElems], 2589 UnpckMask); 2590 NumElems >>= 1; 2591 } 2592 return V[0]; 2593 } 2594 2595 return SDOperand(); 2596} 2597 2598SDOperand 2599X86TargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) { 2600 SDOperand V1 = Op.getOperand(0); 2601 SDOperand V2 = Op.getOperand(1); 2602 SDOperand PermMask = Op.getOperand(2); 2603 MVT::ValueType VT = Op.getValueType(); 2604 unsigned NumElems = PermMask.getNumOperands(); 2605 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF; 2606 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF; 2607 bool V1IsSplat = false; 2608 bool V2IsSplat = false; 2609 2610 if (isUndefShuffle(Op.Val)) 2611 return DAG.getNode(ISD::UNDEF, VT); 2612 2613 if (isSplatMask(PermMask.Val)) { 2614 if (NumElems <= 4) return Op; 2615 // Promote it to a v4i32 splat. 2616 return PromoteSplat(Op, DAG); 2617 } 2618 2619 if (X86::isMOVLMask(PermMask.Val)) 2620 return (V1IsUndef) ? V2 : Op; 2621 2622 if (X86::isMOVSHDUPMask(PermMask.Val) || 2623 X86::isMOVSLDUPMask(PermMask.Val) || 2624 X86::isMOVHLPSMask(PermMask.Val) || 2625 X86::isMOVHPMask(PermMask.Val) || 2626 X86::isMOVLPMask(PermMask.Val)) 2627 return Op; 2628 2629 if (ShouldXformToMOVHLPS(PermMask.Val) || 2630 ShouldXformToMOVLP(V1.Val, V2.Val, PermMask.Val)) 2631 return CommuteVectorShuffle(Op, V1, V2, PermMask, DAG); 2632 2633 bool Commuted = false; 2634 V1IsSplat = isSplatVector(V1.Val); 2635 V2IsSplat = isSplatVector(V2.Val); 2636 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) { 2637 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG); 2638 std::swap(V1IsSplat, V2IsSplat); 2639 std::swap(V1IsUndef, V2IsUndef); 2640 Commuted = true; 2641 } 2642 2643 if (isCommutedMOVL(PermMask.Val, V2IsSplat, V2IsUndef)) { 2644 if (V2IsUndef) return V1; 2645 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG); 2646 if (V2IsSplat) { 2647 // V2 is a splat, so the mask may be malformed. That is, it may point 2648 // to any V2 element. The instruction selectior won't like this. Get 2649 // a corrected mask and commute to form a proper MOVS{S|D}. 2650 SDOperand NewMask = getMOVLMask(NumElems, DAG); 2651 if (NewMask.Val != PermMask.Val) 2652 Op = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask); 2653 } 2654 return Op; 2655 } 2656 2657 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) || 2658 X86::isUNPCKLMask(PermMask.Val) || 2659 X86::isUNPCKHMask(PermMask.Val)) 2660 return Op; 2661 2662 if (V2IsSplat) { 2663 // Normalize mask so all entries that point to V2 points to its first 2664 // element then try to match unpck{h|l} again. If match, return a 2665 // new vector_shuffle with the corrected mask. 2666 SDOperand NewMask = NormalizeMask(PermMask, DAG); 2667 if (NewMask.Val != PermMask.Val) { 2668 if (X86::isUNPCKLMask(PermMask.Val, true)) { 2669 SDOperand NewMask = getUnpacklMask(NumElems, DAG); 2670 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask); 2671 } else if (X86::isUNPCKHMask(PermMask.Val, true)) { 2672 SDOperand NewMask = getUnpackhMask(NumElems, DAG); 2673 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, NewMask); 2674 } 2675 } 2676 } 2677 2678 // Normalize the node to match x86 shuffle ops if needed 2679 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(PermMask.Val)) 2680 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG); 2681 2682 if (Commuted) { 2683 // Commute is back and try unpck* again. 2684 Op = CommuteVectorShuffle(Op, V1, V2, PermMask, DAG); 2685 if (X86::isUNPCKL_v_undef_Mask(PermMask.Val) || 2686 X86::isUNPCKLMask(PermMask.Val) || 2687 X86::isUNPCKHMask(PermMask.Val)) 2688 return Op; 2689 } 2690 2691 // If VT is integer, try PSHUF* first, then SHUFP*. 2692 if (MVT::isInteger(VT)) { 2693 if (X86::isPSHUFDMask(PermMask.Val) || 2694 X86::isPSHUFHWMask(PermMask.Val) || 2695 X86::isPSHUFLWMask(PermMask.Val)) { 2696 if (V2.getOpcode() != ISD::UNDEF) 2697 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, 2698 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask); 2699 return Op; 2700 } 2701 2702 if (X86::isSHUFPMask(PermMask.Val)) 2703 return Op; 2704 2705 // Handle v8i16 shuffle high / low shuffle node pair. 2706 if (VT == MVT::v8i16 && isPSHUFHW_PSHUFLWMask(PermMask.Val)) { 2707 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(NumElems); 2708 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT); 2709 SmallVector<SDOperand, 8> MaskVec; 2710 for (unsigned i = 0; i != 4; ++i) 2711 MaskVec.push_back(PermMask.getOperand(i)); 2712 for (unsigned i = 4; i != 8; ++i) 2713 MaskVec.push_back(DAG.getConstant(i, BaseVT)); 2714 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2715 &MaskVec[0], MaskVec.size()); 2716 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask); 2717 MaskVec.clear(); 2718 for (unsigned i = 0; i != 4; ++i) 2719 MaskVec.push_back(DAG.getConstant(i, BaseVT)); 2720 for (unsigned i = 4; i != 8; ++i) 2721 MaskVec.push_back(PermMask.getOperand(i)); 2722 Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, &MaskVec[0],MaskVec.size()); 2723 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask); 2724 } 2725 } else { 2726 // Floating point cases in the other order. 2727 if (X86::isSHUFPMask(PermMask.Val)) 2728 return Op; 2729 if (X86::isPSHUFDMask(PermMask.Val) || 2730 X86::isPSHUFHWMask(PermMask.Val) || 2731 X86::isPSHUFLWMask(PermMask.Val)) { 2732 if (V2.getOpcode() != ISD::UNDEF) 2733 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, 2734 DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask); 2735 return Op; 2736 } 2737 } 2738 2739 if (NumElems == 4) { 2740 MVT::ValueType MaskVT = PermMask.getValueType(); 2741 MVT::ValueType MaskEVT = MVT::getVectorBaseType(MaskVT); 2742 SmallVector<std::pair<int, int>, 8> Locs; 2743 Locs.reserve(NumElems); 2744 SmallVector<SDOperand, 8> Mask1(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT)); 2745 SmallVector<SDOperand, 8> Mask2(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT)); 2746 unsigned NumHi = 0; 2747 unsigned NumLo = 0; 2748 // If no more than two elements come from either vector. This can be 2749 // implemented with two shuffles. First shuffle gather the elements. 2750 // The second shuffle, which takes the first shuffle as both of its 2751 // vector operands, put the elements into the right order. 2752 for (unsigned i = 0; i != NumElems; ++i) { 2753 SDOperand Elt = PermMask.getOperand(i); 2754 if (Elt.getOpcode() == ISD::UNDEF) { 2755 Locs[i] = std::make_pair(-1, -1); 2756 } else { 2757 unsigned Val = cast<ConstantSDNode>(Elt)->getValue(); 2758 if (Val < NumElems) { 2759 Locs[i] = std::make_pair(0, NumLo); 2760 Mask1[NumLo] = Elt; 2761 NumLo++; 2762 } else { 2763 Locs[i] = std::make_pair(1, NumHi); 2764 if (2+NumHi < NumElems) 2765 Mask1[2+NumHi] = Elt; 2766 NumHi++; 2767 } 2768 } 2769 } 2770 if (NumLo <= 2 && NumHi <= 2) { 2771 V1 = DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, 2772 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2773 &Mask1[0], Mask1.size())); 2774 for (unsigned i = 0; i != NumElems; ++i) { 2775 if (Locs[i].first == -1) 2776 continue; 2777 else { 2778 unsigned Idx = (i < NumElems/2) ? 0 : NumElems; 2779 Idx += Locs[i].first * (NumElems/2) + Locs[i].second; 2780 Mask2[i] = DAG.getConstant(Idx, MaskEVT); 2781 } 2782 } 2783 2784 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V1, 2785 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2786 &Mask2[0], Mask2.size())); 2787 } 2788 2789 // Break it into (shuffle shuffle_hi, shuffle_lo). 2790 Locs.clear(); 2791 SmallVector<SDOperand,8> LoMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT)); 2792 SmallVector<SDOperand,8> HiMask(NumElems, DAG.getNode(ISD::UNDEF, MaskEVT)); 2793 SmallVector<SDOperand,8> *MaskPtr = &LoMask; 2794 unsigned MaskIdx = 0; 2795 unsigned LoIdx = 0; 2796 unsigned HiIdx = NumElems/2; 2797 for (unsigned i = 0; i != NumElems; ++i) { 2798 if (i == NumElems/2) { 2799 MaskPtr = &HiMask; 2800 MaskIdx = 1; 2801 LoIdx = 0; 2802 HiIdx = NumElems/2; 2803 } 2804 SDOperand Elt = PermMask.getOperand(i); 2805 if (Elt.getOpcode() == ISD::UNDEF) { 2806 Locs[i] = std::make_pair(-1, -1); 2807 } else if (cast<ConstantSDNode>(Elt)->getValue() < NumElems) { 2808 Locs[i] = std::make_pair(MaskIdx, LoIdx); 2809 (*MaskPtr)[LoIdx] = Elt; 2810 LoIdx++; 2811 } else { 2812 Locs[i] = std::make_pair(MaskIdx, HiIdx); 2813 (*MaskPtr)[HiIdx] = Elt; 2814 HiIdx++; 2815 } 2816 } 2817 2818 SDOperand LoShuffle = 2819 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, 2820 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2821 &LoMask[0], LoMask.size())); 2822 SDOperand HiShuffle = 2823 DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, 2824 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2825 &HiMask[0], HiMask.size())); 2826 SmallVector<SDOperand, 8> MaskOps; 2827 for (unsigned i = 0; i != NumElems; ++i) { 2828 if (Locs[i].first == -1) { 2829 MaskOps.push_back(DAG.getNode(ISD::UNDEF, MaskEVT)); 2830 } else { 2831 unsigned Idx = Locs[i].first * NumElems + Locs[i].second; 2832 MaskOps.push_back(DAG.getConstant(Idx, MaskEVT)); 2833 } 2834 } 2835 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, LoShuffle, HiShuffle, 2836 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2837 &MaskOps[0], MaskOps.size())); 2838 } 2839 2840 return SDOperand(); 2841} 2842 2843SDOperand 2844X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) { 2845 if (!isa<ConstantSDNode>(Op.getOperand(1))) 2846 return SDOperand(); 2847 2848 MVT::ValueType VT = Op.getValueType(); 2849 // TODO: handle v16i8. 2850 if (MVT::getSizeInBits(VT) == 16) { 2851 // Transform it so it match pextrw which produces a 32-bit result. 2852 MVT::ValueType EVT = (MVT::ValueType)(VT+1); 2853 SDOperand Extract = DAG.getNode(X86ISD::PEXTRW, EVT, 2854 Op.getOperand(0), Op.getOperand(1)); 2855 SDOperand Assert = DAG.getNode(ISD::AssertZext, EVT, Extract, 2856 DAG.getValueType(VT)); 2857 return DAG.getNode(ISD::TRUNCATE, VT, Assert); 2858 } else if (MVT::getSizeInBits(VT) == 32) { 2859 SDOperand Vec = Op.getOperand(0); 2860 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue(); 2861 if (Idx == 0) 2862 return Op; 2863 // SHUFPS the element to the lowest double word, then movss. 2864 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4); 2865 SmallVector<SDOperand, 8> IdxVec; 2866 IdxVec.push_back(DAG.getConstant(Idx, MVT::getVectorBaseType(MaskVT))); 2867 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT))); 2868 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT))); 2869 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT))); 2870 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2871 &IdxVec[0], IdxVec.size()); 2872 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(), 2873 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask); 2874 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec, 2875 DAG.getConstant(0, getPointerTy())); 2876 } else if (MVT::getSizeInBits(VT) == 64) { 2877 SDOperand Vec = Op.getOperand(0); 2878 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue(); 2879 if (Idx == 0) 2880 return Op; 2881 2882 // UNPCKHPD the element to the lowest double word, then movsd. 2883 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored 2884 // to a f64mem, the whole operation is folded into a single MOVHPDmr. 2885 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4); 2886 SmallVector<SDOperand, 8> IdxVec; 2887 IdxVec.push_back(DAG.getConstant(1, MVT::getVectorBaseType(MaskVT))); 2888 IdxVec.push_back(DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(MaskVT))); 2889 SDOperand Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2890 &IdxVec[0], IdxVec.size()); 2891 Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(), 2892 Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask); 2893 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec, 2894 DAG.getConstant(0, getPointerTy())); 2895 } 2896 2897 return SDOperand(); 2898} 2899 2900SDOperand 2901X86TargetLowering::LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) { 2902 // Transform it so it match pinsrw which expects a 16-bit value in a GR32 2903 // as its second argument. 2904 MVT::ValueType VT = Op.getValueType(); 2905 MVT::ValueType BaseVT = MVT::getVectorBaseType(VT); 2906 SDOperand N0 = Op.getOperand(0); 2907 SDOperand N1 = Op.getOperand(1); 2908 SDOperand N2 = Op.getOperand(2); 2909 if (MVT::getSizeInBits(BaseVT) == 16) { 2910 if (N1.getValueType() != MVT::i32) 2911 N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1); 2912 if (N2.getValueType() != MVT::i32) 2913 N2 = DAG.getConstant(cast<ConstantSDNode>(N2)->getValue(), MVT::i32); 2914 return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2); 2915 } else if (MVT::getSizeInBits(BaseVT) == 32) { 2916 unsigned Idx = cast<ConstantSDNode>(N2)->getValue(); 2917 if (Idx == 0) { 2918 // Use a movss. 2919 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, N1); 2920 MVT::ValueType MaskVT = MVT::getIntVectorWithNumElements(4); 2921 MVT::ValueType BaseVT = MVT::getVectorBaseType(MaskVT); 2922 SmallVector<SDOperand, 8> MaskVec; 2923 MaskVec.push_back(DAG.getConstant(4, BaseVT)); 2924 for (unsigned i = 1; i <= 3; ++i) 2925 MaskVec.push_back(DAG.getConstant(i, BaseVT)); 2926 return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, N0, N1, 2927 DAG.getNode(ISD::BUILD_VECTOR, MaskVT, 2928 &MaskVec[0], MaskVec.size())); 2929 } else { 2930 // Use two pinsrw instructions to insert a 32 bit value. 2931 Idx <<= 1; 2932 if (MVT::isFloatingPoint(N1.getValueType())) { 2933 if (ISD::isNON_EXTLoad(N1.Val)) { 2934 // Just load directly from f32mem to GR32. 2935 LoadSDNode *LD = cast<LoadSDNode>(N1); 2936 N1 = DAG.getLoad(MVT::i32, LD->getChain(), LD->getBasePtr(), 2937 LD->getSrcValue(), LD->getSrcValueOffset()); 2938 } else { 2939 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v4f32, N1); 2940 N1 = DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, N1); 2941 N1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, N1, 2942 DAG.getConstant(0, getPointerTy())); 2943 } 2944 } 2945 N0 = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, N0); 2946 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1, 2947 DAG.getConstant(Idx, getPointerTy())); 2948 N1 = DAG.getNode(ISD::SRL, MVT::i32, N1, DAG.getConstant(16, MVT::i8)); 2949 N0 = DAG.getNode(X86ISD::PINSRW, MVT::v8i16, N0, N1, 2950 DAG.getConstant(Idx+1, getPointerTy())); 2951 return DAG.getNode(ISD::BIT_CONVERT, VT, N0); 2952 } 2953 } 2954 2955 return SDOperand(); 2956} 2957 2958SDOperand 2959X86TargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) { 2960 SDOperand AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0)); 2961 return DAG.getNode(X86ISD::S2VEC, Op.getValueType(), AnyExt); 2962} 2963 2964// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as 2965// their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is 2966// one of the above mentioned nodes. It has to be wrapped because otherwise 2967// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only 2968// be used to form addressing mode. These wrapped nodes will be selected 2969// into MOV32ri. 2970SDOperand 2971X86TargetLowering::LowerConstantPool(SDOperand Op, SelectionDAG &DAG) { 2972 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 2973 SDOperand Result = DAG.getTargetConstantPool(CP->getConstVal(), 2974 getPointerTy(), 2975 CP->getAlignment()); 2976 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result); 2977 // With PIC, the address is actually $g + Offset. 2978 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 2979 !Subtarget->isPICStyleRIPRel()) { 2980 Result = DAG.getNode(ISD::ADD, getPointerTy(), 2981 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), 2982 Result); 2983 } 2984 2985 return Result; 2986} 2987 2988SDOperand 2989X86TargetLowering::LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) { 2990 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 2991 SDOperand Result = DAG.getTargetGlobalAddress(GV, getPointerTy()); 2992 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result); 2993 // With PIC, the address is actually $g + Offset. 2994 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 2995 !Subtarget->isPICStyleRIPRel()) { 2996 Result = DAG.getNode(ISD::ADD, getPointerTy(), 2997 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), 2998 Result); 2999 } 3000 3001 // For Darwin & Mingw32, external and weak symbols are indirect, so we want to 3002 // load the value at address GV, not the value of GV itself. This means that 3003 // the GlobalAddress must be in the base or index register of the address, not 3004 // the GV offset field. Platform check is inside GVRequiresExtraLoad() call 3005 // The same applies for external symbols during PIC codegen 3006 if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false)) 3007 Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result, NULL, 0); 3008 3009 return Result; 3010} 3011 3012SDOperand 3013X86TargetLowering::LowerExternalSymbol(SDOperand Op, SelectionDAG &DAG) { 3014 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol(); 3015 SDOperand Result = DAG.getTargetExternalSymbol(Sym, getPointerTy()); 3016 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result); 3017 // With PIC, the address is actually $g + Offset. 3018 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 3019 !Subtarget->isPICStyleRIPRel()) { 3020 Result = DAG.getNode(ISD::ADD, getPointerTy(), 3021 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), 3022 Result); 3023 } 3024 3025 return Result; 3026} 3027 3028SDOperand X86TargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) { 3029 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); 3030 SDOperand Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy()); 3031 Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result); 3032 // With PIC, the address is actually $g + Offset. 3033 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ && 3034 !Subtarget->isPICStyleRIPRel()) { 3035 Result = DAG.getNode(ISD::ADD, getPointerTy(), 3036 DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), 3037 Result); 3038 } 3039 3040 return Result; 3041} 3042 3043SDOperand X86TargetLowering::LowerShift(SDOperand Op, SelectionDAG &DAG) { 3044 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 && 3045 "Not an i64 shift!"); 3046 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS; 3047 SDOperand ShOpLo = Op.getOperand(0); 3048 SDOperand ShOpHi = Op.getOperand(1); 3049 SDOperand ShAmt = Op.getOperand(2); 3050 SDOperand Tmp1 = isSRA ? 3051 DAG.getNode(ISD::SRA, MVT::i32, ShOpHi, DAG.getConstant(31, MVT::i8)) : 3052 DAG.getConstant(0, MVT::i32); 3053 3054 SDOperand Tmp2, Tmp3; 3055 if (Op.getOpcode() == ISD::SHL_PARTS) { 3056 Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt); 3057 Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt); 3058 } else { 3059 Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt); 3060 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt); 3061 } 3062 3063 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag); 3064 SDOperand AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt, 3065 DAG.getConstant(32, MVT::i8)); 3066 SDOperand COps[]={DAG.getEntryNode(), AndNode, DAG.getConstant(0, MVT::i8)}; 3067 SDOperand InFlag = DAG.getNode(X86ISD::CMP, VTs, 2, COps, 3).getValue(1); 3068 3069 SDOperand Hi, Lo; 3070 SDOperand CC = DAG.getConstant(X86::COND_NE, MVT::i8); 3071 3072 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::Flag); 3073 SmallVector<SDOperand, 4> Ops; 3074 if (Op.getOpcode() == ISD::SHL_PARTS) { 3075 Ops.push_back(Tmp2); 3076 Ops.push_back(Tmp3); 3077 Ops.push_back(CC); 3078 Ops.push_back(InFlag); 3079 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size()); 3080 InFlag = Hi.getValue(1); 3081 3082 Ops.clear(); 3083 Ops.push_back(Tmp3); 3084 Ops.push_back(Tmp1); 3085 Ops.push_back(CC); 3086 Ops.push_back(InFlag); 3087 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size()); 3088 } else { 3089 Ops.push_back(Tmp2); 3090 Ops.push_back(Tmp3); 3091 Ops.push_back(CC); 3092 Ops.push_back(InFlag); 3093 Lo = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size()); 3094 InFlag = Lo.getValue(1); 3095 3096 Ops.clear(); 3097 Ops.push_back(Tmp3); 3098 Ops.push_back(Tmp1); 3099 Ops.push_back(CC); 3100 Ops.push_back(InFlag); 3101 Hi = DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size()); 3102 } 3103 3104 VTs = DAG.getNodeValueTypes(MVT::i32, MVT::i32); 3105 Ops.clear(); 3106 Ops.push_back(Lo); 3107 Ops.push_back(Hi); 3108 return DAG.getNode(ISD::MERGE_VALUES, VTs, 2, &Ops[0], Ops.size()); 3109} 3110 3111SDOperand X86TargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) { 3112 assert(Op.getOperand(0).getValueType() <= MVT::i64 && 3113 Op.getOperand(0).getValueType() >= MVT::i16 && 3114 "Unknown SINT_TO_FP to lower!"); 3115 3116 SDOperand Result; 3117 MVT::ValueType SrcVT = Op.getOperand(0).getValueType(); 3118 unsigned Size = MVT::getSizeInBits(SrcVT)/8; 3119 MachineFunction &MF = DAG.getMachineFunction(); 3120 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size); 3121 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); 3122 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0), 3123 StackSlot, NULL, 0); 3124 3125 // Build the FILD 3126 SDVTList Tys; 3127 if (X86ScalarSSE) 3128 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag); 3129 else 3130 Tys = DAG.getVTList(MVT::f64, MVT::Other); 3131 SmallVector<SDOperand, 8> Ops; 3132 Ops.push_back(Chain); 3133 Ops.push_back(StackSlot); 3134 Ops.push_back(DAG.getValueType(SrcVT)); 3135 Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD, 3136 Tys, &Ops[0], Ops.size()); 3137 3138 if (X86ScalarSSE) { 3139 Chain = Result.getValue(1); 3140 SDOperand InFlag = Result.getValue(2); 3141 3142 // FIXME: Currently the FST is flagged to the FILD_FLAG. This 3143 // shouldn't be necessary except that RFP cannot be live across 3144 // multiple blocks. When stackifier is fixed, they can be uncoupled. 3145 MachineFunction &MF = DAG.getMachineFunction(); 3146 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8); 3147 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); 3148 Tys = DAG.getVTList(MVT::Other); 3149 SmallVector<SDOperand, 8> Ops; 3150 Ops.push_back(Chain); 3151 Ops.push_back(Result); 3152 Ops.push_back(StackSlot); 3153 Ops.push_back(DAG.getValueType(Op.getValueType())); 3154 Ops.push_back(InFlag); 3155 Chain = DAG.getNode(X86ISD::FST, Tys, &Ops[0], Ops.size()); 3156 Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot, NULL, 0); 3157 } 3158 3159 return Result; 3160} 3161 3162SDOperand X86TargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) { 3163 assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 && 3164 "Unknown FP_TO_SINT to lower!"); 3165 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary 3166 // stack slot. 3167 MachineFunction &MF = DAG.getMachineFunction(); 3168 unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8; 3169 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize); 3170 SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); 3171 3172 unsigned Opc; 3173 switch (Op.getValueType()) { 3174 default: assert(0 && "Invalid FP_TO_SINT to lower!"); 3175 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break; 3176 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break; 3177 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break; 3178 } 3179 3180 SDOperand Chain = DAG.getEntryNode(); 3181 SDOperand Value = Op.getOperand(0); 3182 if (X86ScalarSSE) { 3183 assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!"); 3184 Chain = DAG.getStore(Chain, Value, StackSlot, NULL, 0); 3185 SDVTList Tys = DAG.getVTList(MVT::f64, MVT::Other); 3186 SDOperand Ops[] = { 3187 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType()) 3188 }; 3189 Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3); 3190 Chain = Value.getValue(1); 3191 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize); 3192 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy()); 3193 } 3194 3195 // Build the FP_TO_INT*_IN_MEM 3196 SDOperand Ops[] = { Chain, Value, StackSlot }; 3197 SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops, 3); 3198 3199 // Load the result. 3200 return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0); 3201} 3202 3203SDOperand X86TargetLowering::LowerFABS(SDOperand Op, SelectionDAG &DAG) { 3204 MVT::ValueType VT = Op.getValueType(); 3205 const Type *OpNTy = MVT::getTypeForValueType(VT); 3206 std::vector<Constant*> CV; 3207 if (VT == MVT::f64) { 3208 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63)))); 3209 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3210 } else { 3211 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31)))); 3212 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3213 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3214 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3215 } 3216 Constant *CS = ConstantStruct::get(CV); 3217 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4); 3218 SDVTList Tys = DAG.getVTList(VT, MVT::Other); 3219 SmallVector<SDOperand, 3> Ops; 3220 Ops.push_back(DAG.getEntryNode()); 3221 Ops.push_back(CPIdx); 3222 Ops.push_back(DAG.getSrcValue(NULL)); 3223 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size()); 3224 return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask); 3225} 3226 3227SDOperand X86TargetLowering::LowerFNEG(SDOperand Op, SelectionDAG &DAG) { 3228 MVT::ValueType VT = Op.getValueType(); 3229 const Type *OpNTy = MVT::getTypeForValueType(VT); 3230 std::vector<Constant*> CV; 3231 if (VT == MVT::f64) { 3232 CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63))); 3233 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3234 } else { 3235 CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31))); 3236 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3237 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3238 CV.push_back(ConstantFP::get(OpNTy, 0.0)); 3239 } 3240 Constant *CS = ConstantStruct::get(CV); 3241 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4); 3242 SDVTList Tys = DAG.getVTList(VT, MVT::Other); 3243 SmallVector<SDOperand, 3> Ops; 3244 Ops.push_back(DAG.getEntryNode()); 3245 Ops.push_back(CPIdx); 3246 Ops.push_back(DAG.getSrcValue(NULL)); 3247 SDOperand Mask = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size()); 3248 return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask); 3249} 3250 3251SDOperand X86TargetLowering::LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) { 3252 SDOperand Op0 = Op.getOperand(0); 3253 SDOperand Op1 = Op.getOperand(1); 3254 MVT::ValueType VT = Op.getValueType(); 3255 MVT::ValueType SrcVT = Op1.getValueType(); 3256 const Type *SrcTy = MVT::getTypeForValueType(SrcVT); 3257 3258 // If second operand is smaller, extend it first. 3259 if (MVT::getSizeInBits(SrcVT) < MVT::getSizeInBits(VT)) { 3260 Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1); 3261 SrcVT = VT; 3262 } 3263 3264 // First get the sign bit of second operand. 3265 std::vector<Constant*> CV; 3266 if (SrcVT == MVT::f64) { 3267 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(1ULL << 63))); 3268 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3269 } else { 3270 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(1U << 31))); 3271 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3272 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3273 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3274 } 3275 Constant *CS = ConstantStruct::get(CV); 3276 SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4); 3277 SDVTList Tys = DAG.getVTList(SrcVT, MVT::Other); 3278 SmallVector<SDOperand, 3> Ops; 3279 Ops.push_back(DAG.getEntryNode()); 3280 Ops.push_back(CPIdx); 3281 Ops.push_back(DAG.getSrcValue(NULL)); 3282 SDOperand Mask1 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size()); 3283 SDOperand SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1); 3284 3285 // Shift sign bit right or left if the two operands have different types. 3286 if (MVT::getSizeInBits(SrcVT) > MVT::getSizeInBits(VT)) { 3287 // Op0 is MVT::f32, Op1 is MVT::f64. 3288 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit); 3289 SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit, 3290 DAG.getConstant(32, MVT::i32)); 3291 SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit); 3292 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit, 3293 DAG.getConstant(0, getPointerTy())); 3294 } 3295 3296 // Clear first operand sign bit. 3297 CV.clear(); 3298 if (VT == MVT::f64) { 3299 CV.push_back(ConstantFP::get(SrcTy, BitsToDouble(~(1ULL << 63)))); 3300 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3301 } else { 3302 CV.push_back(ConstantFP::get(SrcTy, BitsToFloat(~(1U << 31)))); 3303 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3304 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3305 CV.push_back(ConstantFP::get(SrcTy, 0.0)); 3306 } 3307 CS = ConstantStruct::get(CV); 3308 CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4); 3309 Tys = DAG.getVTList(VT, MVT::Other); 3310 Ops.clear(); 3311 Ops.push_back(DAG.getEntryNode()); 3312 Ops.push_back(CPIdx); 3313 Ops.push_back(DAG.getSrcValue(NULL)); 3314 SDOperand Mask2 = DAG.getNode(X86ISD::LOAD_PACK, Tys, &Ops[0], Ops.size()); 3315 SDOperand Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2); 3316 3317 // Or the value with the sign bit. 3318 return DAG.getNode(X86ISD::FOR, VT, Val, SignBit); 3319} 3320 3321SDOperand X86TargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG, 3322 SDOperand Chain) { 3323 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer"); 3324 SDOperand Cond; 3325 SDOperand Op0 = Op.getOperand(0); 3326 SDOperand Op1 = Op.getOperand(1); 3327 SDOperand CC = Op.getOperand(2); 3328 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); 3329 const MVT::ValueType *VTs1 = DAG.getNodeValueTypes(MVT::Other, MVT::Flag); 3330 const MVT::ValueType *VTs2 = DAG.getNodeValueTypes(MVT::i8, MVT::Flag); 3331 bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType()); 3332 unsigned X86CC; 3333 3334 if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC, 3335 Op0, Op1, DAG)) { 3336 SDOperand Ops1[] = { Chain, Op0, Op1 }; 3337 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, Ops1, 3).getValue(1); 3338 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond }; 3339 return DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2); 3340 } 3341 3342 assert(isFP && "Illegal integer SetCC!"); 3343 3344 SDOperand COps[] = { Chain, Op0, Op1 }; 3345 Cond = DAG.getNode(X86ISD::CMP, VTs1, 2, COps, 3).getValue(1); 3346 3347 switch (SetCCOpcode) { 3348 default: assert(false && "Illegal floating point SetCC!"); 3349 case ISD::SETOEQ: { // !PF & ZF 3350 SDOperand Ops1[] = { DAG.getConstant(X86::COND_NP, MVT::i8), Cond }; 3351 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2); 3352 SDOperand Ops2[] = { DAG.getConstant(X86::COND_E, MVT::i8), 3353 Tmp1.getValue(1) }; 3354 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2); 3355 return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2); 3356 } 3357 case ISD::SETUNE: { // PF | !ZF 3358 SDOperand Ops1[] = { DAG.getConstant(X86::COND_P, MVT::i8), Cond }; 3359 SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops1, 2); 3360 SDOperand Ops2[] = { DAG.getConstant(X86::COND_NE, MVT::i8), 3361 Tmp1.getValue(1) }; 3362 SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, VTs2, 2, Ops2, 2); 3363 return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2); 3364 } 3365 } 3366} 3367 3368SDOperand X86TargetLowering::LowerSELECT(SDOperand Op, SelectionDAG &DAG) { 3369 bool addTest = true; 3370 SDOperand Chain = DAG.getEntryNode(); 3371 SDOperand Cond = Op.getOperand(0); 3372 SDOperand CC; 3373 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag); 3374 3375 if (Cond.getOpcode() == ISD::SETCC) 3376 Cond = LowerSETCC(Cond, DAG, Chain); 3377 3378 if (Cond.getOpcode() == X86ISD::SETCC) { 3379 CC = Cond.getOperand(0); 3380 3381 // If condition flag is set by a X86ISD::CMP, then make a copy of it 3382 // (since flag operand cannot be shared). Use it as the condition setting 3383 // operand in place of the X86ISD::SETCC. 3384 // If the X86ISD::SETCC has more than one use, then perhaps it's better 3385 // to use a test instead of duplicating the X86ISD::CMP (for register 3386 // pressure reason)? 3387 SDOperand Cmp = Cond.getOperand(1); 3388 unsigned Opc = Cmp.getOpcode(); 3389 bool IllegalFPCMov = !X86ScalarSSE && 3390 MVT::isFloatingPoint(Op.getValueType()) && 3391 !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended()); 3392 if ((Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) && 3393 !IllegalFPCMov) { 3394 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) }; 3395 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3); 3396 addTest = false; 3397 } 3398 } 3399 3400 if (addTest) { 3401 CC = DAG.getConstant(X86::COND_NE, MVT::i8); 3402 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) }; 3403 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3); 3404 } 3405 3406 VTs = DAG.getNodeValueTypes(Op.getValueType(), MVT::Flag); 3407 SmallVector<SDOperand, 4> Ops; 3408 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if 3409 // condition is true. 3410 Ops.push_back(Op.getOperand(2)); 3411 Ops.push_back(Op.getOperand(1)); 3412 Ops.push_back(CC); 3413 Ops.push_back(Cond.getValue(1)); 3414 return DAG.getNode(X86ISD::CMOV, VTs, 2, &Ops[0], Ops.size()); 3415} 3416 3417SDOperand X86TargetLowering::LowerBRCOND(SDOperand Op, SelectionDAG &DAG) { 3418 bool addTest = true; 3419 SDOperand Chain = Op.getOperand(0); 3420 SDOperand Cond = Op.getOperand(1); 3421 SDOperand Dest = Op.getOperand(2); 3422 SDOperand CC; 3423 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag); 3424 3425 if (Cond.getOpcode() == ISD::SETCC) 3426 Cond = LowerSETCC(Cond, DAG, Chain); 3427 3428 if (Cond.getOpcode() == X86ISD::SETCC) { 3429 CC = Cond.getOperand(0); 3430 3431 // If condition flag is set by a X86ISD::CMP, then make a copy of it 3432 // (since flag operand cannot be shared). Use it as the condition setting 3433 // operand in place of the X86ISD::SETCC. 3434 // If the X86ISD::SETCC has more than one use, then perhaps it's better 3435 // to use a test instead of duplicating the X86ISD::CMP (for register 3436 // pressure reason)? 3437 SDOperand Cmp = Cond.getOperand(1); 3438 unsigned Opc = Cmp.getOpcode(); 3439 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI) { 3440 SDOperand Ops[] = { Chain, Cmp.getOperand(1), Cmp.getOperand(2) }; 3441 Cond = DAG.getNode(Opc, VTs, 2, Ops, 3); 3442 addTest = false; 3443 } 3444 } 3445 3446 if (addTest) { 3447 CC = DAG.getConstant(X86::COND_NE, MVT::i8); 3448 SDOperand Ops[] = { Chain, Cond, DAG.getConstant(0, MVT::i8) }; 3449 Cond = DAG.getNode(X86ISD::CMP, VTs, 2, Ops, 3); 3450 } 3451 return DAG.getNode(X86ISD::BRCOND, Op.getValueType(), 3452 Cond, Op.getOperand(2), CC, Cond.getValue(1)); 3453} 3454 3455SDOperand X86TargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) { 3456 unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue(); 3457 3458 if (Subtarget->is64Bit()) 3459 return LowerX86_64CCCCallTo(Op, DAG, CallingConv); 3460 else 3461 switch (CallingConv) { 3462 default: 3463 assert(0 && "Unsupported calling convention"); 3464 case CallingConv::Fast: 3465 // TODO: Implement fastcc 3466 // Falls through 3467 case CallingConv::C: 3468 case CallingConv::X86_StdCall: 3469 return LowerCCCCallTo(Op, DAG, CallingConv); 3470 case CallingConv::X86_FastCall: 3471 return LowerFastCCCallTo(Op, DAG, CallingConv); 3472 } 3473} 3474 3475SDOperand 3476X86TargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) { 3477 MachineFunction &MF = DAG.getMachineFunction(); 3478 const Function* Fn = MF.getFunction(); 3479 if (Fn->hasExternalLinkage() && 3480 Subtarget->isTargetCygMing() && 3481 Fn->getName() == "main") 3482 MF.getInfo<X86FunctionInfo>()->setForceFramePointer(true); 3483 3484 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getValue(); 3485 if (Subtarget->is64Bit()) 3486 return LowerX86_64CCCArguments(Op, DAG); 3487 else 3488 switch(CC) { 3489 default: 3490 assert(0 && "Unsupported calling convention"); 3491 case CallingConv::Fast: 3492 // TODO: implement fastcc. 3493 3494 // Falls through 3495 case CallingConv::C: 3496 return LowerCCCArguments(Op, DAG); 3497 case CallingConv::X86_StdCall: 3498 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(StdCall); 3499 return LowerCCCArguments(Op, DAG, true); 3500 case CallingConv::X86_FastCall: 3501 MF.getInfo<X86FunctionInfo>()->setDecorationStyle(FastCall); 3502 return LowerFastCCArguments(Op, DAG); 3503 } 3504} 3505 3506SDOperand X86TargetLowering::LowerMEMSET(SDOperand Op, SelectionDAG &DAG) { 3507 SDOperand InFlag(0, 0); 3508 SDOperand Chain = Op.getOperand(0); 3509 unsigned Align = 3510 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue(); 3511 if (Align == 0) Align = 1; 3512 3513 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3)); 3514 // If not DWORD aligned, call memset if size is less than the threshold. 3515 // It knows how to align to the right boundary first. 3516 if ((Align & 3) != 0 || 3517 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) { 3518 MVT::ValueType IntPtr = getPointerTy(); 3519 const Type *IntPtrTy = getTargetData()->getIntPtrType(); 3520 TargetLowering::ArgListTy Args; 3521 TargetLowering::ArgListEntry Entry; 3522 Entry.Node = Op.getOperand(1); 3523 Entry.Ty = IntPtrTy; 3524 Args.push_back(Entry); 3525 // Extend the unsigned i8 argument to be an int value for the call. 3526 Entry.Node = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2)); 3527 Entry.Ty = IntPtrTy; 3528 Args.push_back(Entry); 3529 Entry.Node = Op.getOperand(3); 3530 Args.push_back(Entry); 3531 std::pair<SDOperand,SDOperand> CallResult = 3532 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false, 3533 DAG.getExternalSymbol("memset", IntPtr), Args, DAG); 3534 return CallResult.second; 3535 } 3536 3537 MVT::ValueType AVT; 3538 SDOperand Count; 3539 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2)); 3540 unsigned BytesLeft = 0; 3541 bool TwoRepStos = false; 3542 if (ValC) { 3543 unsigned ValReg; 3544 uint64_t Val = ValC->getValue() & 255; 3545 3546 // If the value is a constant, then we can potentially use larger sets. 3547 switch (Align & 3) { 3548 case 2: // WORD aligned 3549 AVT = MVT::i16; 3550 ValReg = X86::AX; 3551 Val = (Val << 8) | Val; 3552 break; 3553 case 0: // DWORD aligned 3554 AVT = MVT::i32; 3555 ValReg = X86::EAX; 3556 Val = (Val << 8) | Val; 3557 Val = (Val << 16) | Val; 3558 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) { // QWORD aligned 3559 AVT = MVT::i64; 3560 ValReg = X86::RAX; 3561 Val = (Val << 32) | Val; 3562 } 3563 break; 3564 default: // Byte aligned 3565 AVT = MVT::i8; 3566 ValReg = X86::AL; 3567 Count = Op.getOperand(3); 3568 break; 3569 } 3570 3571 if (AVT > MVT::i8) { 3572 if (I) { 3573 unsigned UBytes = MVT::getSizeInBits(AVT) / 8; 3574 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy()); 3575 BytesLeft = I->getValue() % UBytes; 3576 } else { 3577 assert(AVT >= MVT::i32 && 3578 "Do not use rep;stos if not at least DWORD aligned"); 3579 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(), 3580 Op.getOperand(3), DAG.getConstant(2, MVT::i8)); 3581 TwoRepStos = true; 3582 } 3583 } 3584 3585 Chain = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT), 3586 InFlag); 3587 InFlag = Chain.getValue(1); 3588 } else { 3589 AVT = MVT::i8; 3590 Count = Op.getOperand(3); 3591 Chain = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag); 3592 InFlag = Chain.getValue(1); 3593 } 3594 3595 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX, 3596 Count, InFlag); 3597 InFlag = Chain.getValue(1); 3598 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI, 3599 Op.getOperand(1), InFlag); 3600 InFlag = Chain.getValue(1); 3601 3602 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag); 3603 SmallVector<SDOperand, 8> Ops; 3604 Ops.push_back(Chain); 3605 Ops.push_back(DAG.getValueType(AVT)); 3606 Ops.push_back(InFlag); 3607 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size()); 3608 3609 if (TwoRepStos) { 3610 InFlag = Chain.getValue(1); 3611 Count = Op.getOperand(3); 3612 MVT::ValueType CVT = Count.getValueType(); 3613 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count, 3614 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT)); 3615 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX, 3616 Left, InFlag); 3617 InFlag = Chain.getValue(1); 3618 Tys = DAG.getVTList(MVT::Other, MVT::Flag); 3619 Ops.clear(); 3620 Ops.push_back(Chain); 3621 Ops.push_back(DAG.getValueType(MVT::i8)); 3622 Ops.push_back(InFlag); 3623 Chain = DAG.getNode(X86ISD::REP_STOS, Tys, &Ops[0], Ops.size()); 3624 } else if (BytesLeft) { 3625 // Issue stores for the last 1 - 7 bytes. 3626 SDOperand Value; 3627 unsigned Val = ValC->getValue() & 255; 3628 unsigned Offset = I->getValue() - BytesLeft; 3629 SDOperand DstAddr = Op.getOperand(1); 3630 MVT::ValueType AddrVT = DstAddr.getValueType(); 3631 if (BytesLeft >= 4) { 3632 Val = (Val << 8) | Val; 3633 Val = (Val << 16) | Val; 3634 Value = DAG.getConstant(Val, MVT::i32); 3635 Chain = DAG.getStore(Chain, Value, 3636 DAG.getNode(ISD::ADD, AddrVT, DstAddr, 3637 DAG.getConstant(Offset, AddrVT)), 3638 NULL, 0); 3639 BytesLeft -= 4; 3640 Offset += 4; 3641 } 3642 if (BytesLeft >= 2) { 3643 Value = DAG.getConstant((Val << 8) | Val, MVT::i16); 3644 Chain = DAG.getStore(Chain, Value, 3645 DAG.getNode(ISD::ADD, AddrVT, DstAddr, 3646 DAG.getConstant(Offset, AddrVT)), 3647 NULL, 0); 3648 BytesLeft -= 2; 3649 Offset += 2; 3650 } 3651 if (BytesLeft == 1) { 3652 Value = DAG.getConstant(Val, MVT::i8); 3653 Chain = DAG.getStore(Chain, Value, 3654 DAG.getNode(ISD::ADD, AddrVT, DstAddr, 3655 DAG.getConstant(Offset, AddrVT)), 3656 NULL, 0); 3657 } 3658 } 3659 3660 return Chain; 3661} 3662 3663SDOperand X86TargetLowering::LowerMEMCPY(SDOperand Op, SelectionDAG &DAG) { 3664 SDOperand Chain = Op.getOperand(0); 3665 unsigned Align = 3666 (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue(); 3667 if (Align == 0) Align = 1; 3668 3669 ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3)); 3670 // If not DWORD aligned, call memcpy if size is less than the threshold. 3671 // It knows how to align to the right boundary first. 3672 if ((Align & 3) != 0 || 3673 (I && I->getValue() < Subtarget->getMinRepStrSizeThreshold())) { 3674 MVT::ValueType IntPtr = getPointerTy(); 3675 TargetLowering::ArgListTy Args; 3676 TargetLowering::ArgListEntry Entry; 3677 Entry.Ty = getTargetData()->getIntPtrType(); 3678 Entry.Node = Op.getOperand(1); Args.push_back(Entry); 3679 Entry.Node = Op.getOperand(2); Args.push_back(Entry); 3680 Entry.Node = Op.getOperand(3); Args.push_back(Entry); 3681 std::pair<SDOperand,SDOperand> CallResult = 3682 LowerCallTo(Chain, Type::VoidTy, false, false, CallingConv::C, false, 3683 DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG); 3684 return CallResult.second; 3685 } 3686 3687 MVT::ValueType AVT; 3688 SDOperand Count; 3689 unsigned BytesLeft = 0; 3690 bool TwoRepMovs = false; 3691 switch (Align & 3) { 3692 case 2: // WORD aligned 3693 AVT = MVT::i16; 3694 break; 3695 case 0: // DWORD aligned 3696 AVT = MVT::i32; 3697 if (Subtarget->is64Bit() && ((Align & 0xF) == 0)) // QWORD aligned 3698 AVT = MVT::i64; 3699 break; 3700 default: // Byte aligned 3701 AVT = MVT::i8; 3702 Count = Op.getOperand(3); 3703 break; 3704 } 3705 3706 if (AVT > MVT::i8) { 3707 if (I) { 3708 unsigned UBytes = MVT::getSizeInBits(AVT) / 8; 3709 Count = DAG.getConstant(I->getValue() / UBytes, getPointerTy()); 3710 BytesLeft = I->getValue() % UBytes; 3711 } else { 3712 assert(AVT >= MVT::i32 && 3713 "Do not use rep;movs if not at least DWORD aligned"); 3714 Count = DAG.getNode(ISD::SRL, Op.getOperand(3).getValueType(), 3715 Op.getOperand(3), DAG.getConstant(2, MVT::i8)); 3716 TwoRepMovs = true; 3717 } 3718 } 3719 3720 SDOperand InFlag(0, 0); 3721 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RCX : X86::ECX, 3722 Count, InFlag); 3723 InFlag = Chain.getValue(1); 3724 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RDI : X86::EDI, 3725 Op.getOperand(1), InFlag); 3726 InFlag = Chain.getValue(1); 3727 Chain = DAG.getCopyToReg(Chain, Subtarget->is64Bit() ? X86::RSI : X86::ESI, 3728 Op.getOperand(2), InFlag); 3729 InFlag = Chain.getValue(1); 3730 3731 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag); 3732 SmallVector<SDOperand, 8> Ops; 3733 Ops.push_back(Chain); 3734 Ops.push_back(DAG.getValueType(AVT)); 3735 Ops.push_back(InFlag); 3736 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size()); 3737 3738 if (TwoRepMovs) { 3739 InFlag = Chain.getValue(1); 3740 Count = Op.getOperand(3); 3741 MVT::ValueType CVT = Count.getValueType(); 3742 SDOperand Left = DAG.getNode(ISD::AND, CVT, Count, 3743 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT)); 3744 Chain = DAG.getCopyToReg(Chain, (CVT == MVT::i64) ? X86::RCX : X86::ECX, 3745 Left, InFlag); 3746 InFlag = Chain.getValue(1); 3747 Tys = DAG.getVTList(MVT::Other, MVT::Flag); 3748 Ops.clear(); 3749 Ops.push_back(Chain); 3750 Ops.push_back(DAG.getValueType(MVT::i8)); 3751 Ops.push_back(InFlag); 3752 Chain = DAG.getNode(X86ISD::REP_MOVS, Tys, &Ops[0], Ops.size()); 3753 } else if (BytesLeft) { 3754 // Issue loads and stores for the last 1 - 7 bytes. 3755 unsigned Offset = I->getValue() - BytesLeft; 3756 SDOperand DstAddr = Op.getOperand(1); 3757 MVT::ValueType DstVT = DstAddr.getValueType(); 3758 SDOperand SrcAddr = Op.getOperand(2); 3759 MVT::ValueType SrcVT = SrcAddr.getValueType(); 3760 SDOperand Value; 3761 if (BytesLeft >= 4) { 3762 Value = DAG.getLoad(MVT::i32, Chain, 3763 DAG.getNode(ISD::ADD, SrcVT, SrcAddr, 3764 DAG.getConstant(Offset, SrcVT)), 3765 NULL, 0); 3766 Chain = Value.getValue(1); 3767 Chain = DAG.getStore(Chain, Value, 3768 DAG.getNode(ISD::ADD, DstVT, DstAddr, 3769 DAG.getConstant(Offset, DstVT)), 3770 NULL, 0); 3771 BytesLeft -= 4; 3772 Offset += 4; 3773 } 3774 if (BytesLeft >= 2) { 3775 Value = DAG.getLoad(MVT::i16, Chain, 3776 DAG.getNode(ISD::ADD, SrcVT, SrcAddr, 3777 DAG.getConstant(Offset, SrcVT)), 3778 NULL, 0); 3779 Chain = Value.getValue(1); 3780 Chain = DAG.getStore(Chain, Value, 3781 DAG.getNode(ISD::ADD, DstVT, DstAddr, 3782 DAG.getConstant(Offset, DstVT)), 3783 NULL, 0); 3784 BytesLeft -= 2; 3785 Offset += 2; 3786 } 3787 3788 if (BytesLeft == 1) { 3789 Value = DAG.getLoad(MVT::i8, Chain, 3790 DAG.getNode(ISD::ADD, SrcVT, SrcAddr, 3791 DAG.getConstant(Offset, SrcVT)), 3792 NULL, 0); 3793 Chain = Value.getValue(1); 3794 Chain = DAG.getStore(Chain, Value, 3795 DAG.getNode(ISD::ADD, DstVT, DstAddr, 3796 DAG.getConstant(Offset, DstVT)), 3797 NULL, 0); 3798 } 3799 } 3800 3801 return Chain; 3802} 3803 3804SDOperand 3805X86TargetLowering::LowerREADCYCLCECOUNTER(SDOperand Op, SelectionDAG &DAG) { 3806 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag); 3807 SDOperand TheOp = Op.getOperand(0); 3808 SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, &TheOp, 1); 3809 if (Subtarget->is64Bit()) { 3810 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::RAX, MVT::i64, rd.getValue(1)); 3811 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::RDX, 3812 MVT::i64, Copy1.getValue(2)); 3813 SDOperand Tmp = DAG.getNode(ISD::SHL, MVT::i64, Copy2, 3814 DAG.getConstant(32, MVT::i8)); 3815 SDOperand Ops[] = { 3816 DAG.getNode(ISD::OR, MVT::i64, Copy1, Tmp), Copy2.getValue(1) 3817 }; 3818 3819 Tys = DAG.getVTList(MVT::i64, MVT::Other); 3820 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 2); 3821 } 3822 3823 SDOperand Copy1 = DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1)); 3824 SDOperand Copy2 = DAG.getCopyFromReg(Copy1.getValue(1), X86::EDX, 3825 MVT::i32, Copy1.getValue(2)); 3826 SDOperand Ops[] = { Copy1, Copy2, Copy2.getValue(1) }; 3827 Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); 3828 return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops, 3); 3829} 3830 3831SDOperand X86TargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG) { 3832 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2)); 3833 3834 if (!Subtarget->is64Bit()) { 3835 // vastart just stores the address of the VarArgsFrameIndex slot into the 3836 // memory location argument. 3837 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy()); 3838 return DAG.getStore(Op.getOperand(0), FR,Op.getOperand(1), SV->getValue(), 3839 SV->getOffset()); 3840 } 3841 3842 // __va_list_tag: 3843 // gp_offset (0 - 6 * 8) 3844 // fp_offset (48 - 48 + 8 * 16) 3845 // overflow_arg_area (point to parameters coming in memory). 3846 // reg_save_area 3847 SmallVector<SDOperand, 8> MemOps; 3848 SDOperand FIN = Op.getOperand(1); 3849 // Store gp_offset 3850 SDOperand Store = DAG.getStore(Op.getOperand(0), 3851 DAG.getConstant(VarArgsGPOffset, MVT::i32), 3852 FIN, SV->getValue(), SV->getOffset()); 3853 MemOps.push_back(Store); 3854 3855 // Store fp_offset 3856 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, 3857 DAG.getConstant(4, getPointerTy())); 3858 Store = DAG.getStore(Op.getOperand(0), 3859 DAG.getConstant(VarArgsFPOffset, MVT::i32), 3860 FIN, SV->getValue(), SV->getOffset()); 3861 MemOps.push_back(Store); 3862 3863 // Store ptr to overflow_arg_area 3864 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, 3865 DAG.getConstant(4, getPointerTy())); 3866 SDOperand OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy()); 3867 Store = DAG.getStore(Op.getOperand(0), OVFIN, FIN, SV->getValue(), 3868 SV->getOffset()); 3869 MemOps.push_back(Store); 3870 3871 // Store ptr to reg_save_area. 3872 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN, 3873 DAG.getConstant(8, getPointerTy())); 3874 SDOperand RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy()); 3875 Store = DAG.getStore(Op.getOperand(0), RSFIN, FIN, SV->getValue(), 3876 SV->getOffset()); 3877 MemOps.push_back(Store); 3878 return DAG.getNode(ISD::TokenFactor, MVT::Other, &MemOps[0], MemOps.size()); 3879} 3880 3881SDOperand X86TargetLowering::LowerVACOPY(SDOperand Op, SelectionDAG &DAG) { 3882 // X86-64 va_list is a struct { i32, i32, i8*, i8* }. 3883 SDOperand Chain = Op.getOperand(0); 3884 SDOperand DstPtr = Op.getOperand(1); 3885 SDOperand SrcPtr = Op.getOperand(2); 3886 SrcValueSDNode *DstSV = cast<SrcValueSDNode>(Op.getOperand(3)); 3887 SrcValueSDNode *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4)); 3888 3889 SrcPtr = DAG.getLoad(getPointerTy(), Chain, SrcPtr, 3890 SrcSV->getValue(), SrcSV->getOffset()); 3891 Chain = SrcPtr.getValue(1); 3892 for (unsigned i = 0; i < 3; ++i) { 3893 SDOperand Val = DAG.getLoad(MVT::i64, Chain, SrcPtr, 3894 SrcSV->getValue(), SrcSV->getOffset()); 3895 Chain = Val.getValue(1); 3896 Chain = DAG.getStore(Chain, Val, DstPtr, 3897 DstSV->getValue(), DstSV->getOffset()); 3898 if (i == 2) 3899 break; 3900 SrcPtr = DAG.getNode(ISD::ADD, getPointerTy(), SrcPtr, 3901 DAG.getConstant(8, getPointerTy())); 3902 DstPtr = DAG.getNode(ISD::ADD, getPointerTy(), DstPtr, 3903 DAG.getConstant(8, getPointerTy())); 3904 } 3905 return Chain; 3906} 3907 3908SDOperand 3909X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) { 3910 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue(); 3911 switch (IntNo) { 3912 default: return SDOperand(); // Don't custom lower most intrinsics. 3913 // Comparison intrinsics. 3914 case Intrinsic::x86_sse_comieq_ss: 3915 case Intrinsic::x86_sse_comilt_ss: 3916 case Intrinsic::x86_sse_comile_ss: 3917 case Intrinsic::x86_sse_comigt_ss: 3918 case Intrinsic::x86_sse_comige_ss: 3919 case Intrinsic::x86_sse_comineq_ss: 3920 case Intrinsic::x86_sse_ucomieq_ss: 3921 case Intrinsic::x86_sse_ucomilt_ss: 3922 case Intrinsic::x86_sse_ucomile_ss: 3923 case Intrinsic::x86_sse_ucomigt_ss: 3924 case Intrinsic::x86_sse_ucomige_ss: 3925 case Intrinsic::x86_sse_ucomineq_ss: 3926 case Intrinsic::x86_sse2_comieq_sd: 3927 case Intrinsic::x86_sse2_comilt_sd: 3928 case Intrinsic::x86_sse2_comile_sd: 3929 case Intrinsic::x86_sse2_comigt_sd: 3930 case Intrinsic::x86_sse2_comige_sd: 3931 case Intrinsic::x86_sse2_comineq_sd: 3932 case Intrinsic::x86_sse2_ucomieq_sd: 3933 case Intrinsic::x86_sse2_ucomilt_sd: 3934 case Intrinsic::x86_sse2_ucomile_sd: 3935 case Intrinsic::x86_sse2_ucomigt_sd: 3936 case Intrinsic::x86_sse2_ucomige_sd: 3937 case Intrinsic::x86_sse2_ucomineq_sd: { 3938 unsigned Opc = 0; 3939 ISD::CondCode CC = ISD::SETCC_INVALID; 3940 switch (IntNo) { 3941 default: break; 3942 case Intrinsic::x86_sse_comieq_ss: 3943 case Intrinsic::x86_sse2_comieq_sd: 3944 Opc = X86ISD::COMI; 3945 CC = ISD::SETEQ; 3946 break; 3947 case Intrinsic::x86_sse_comilt_ss: 3948 case Intrinsic::x86_sse2_comilt_sd: 3949 Opc = X86ISD::COMI; 3950 CC = ISD::SETLT; 3951 break; 3952 case Intrinsic::x86_sse_comile_ss: 3953 case Intrinsic::x86_sse2_comile_sd: 3954 Opc = X86ISD::COMI; 3955 CC = ISD::SETLE; 3956 break; 3957 case Intrinsic::x86_sse_comigt_ss: 3958 case Intrinsic::x86_sse2_comigt_sd: 3959 Opc = X86ISD::COMI; 3960 CC = ISD::SETGT; 3961 break; 3962 case Intrinsic::x86_sse_comige_ss: 3963 case Intrinsic::x86_sse2_comige_sd: 3964 Opc = X86ISD::COMI; 3965 CC = ISD::SETGE; 3966 break; 3967 case Intrinsic::x86_sse_comineq_ss: 3968 case Intrinsic::x86_sse2_comineq_sd: 3969 Opc = X86ISD::COMI; 3970 CC = ISD::SETNE; 3971 break; 3972 case Intrinsic::x86_sse_ucomieq_ss: 3973 case Intrinsic::x86_sse2_ucomieq_sd: 3974 Opc = X86ISD::UCOMI; 3975 CC = ISD::SETEQ; 3976 break; 3977 case Intrinsic::x86_sse_ucomilt_ss: 3978 case Intrinsic::x86_sse2_ucomilt_sd: 3979 Opc = X86ISD::UCOMI; 3980 CC = ISD::SETLT; 3981 break; 3982 case Intrinsic::x86_sse_ucomile_ss: 3983 case Intrinsic::x86_sse2_ucomile_sd: 3984 Opc = X86ISD::UCOMI; 3985 CC = ISD::SETLE; 3986 break; 3987 case Intrinsic::x86_sse_ucomigt_ss: 3988 case Intrinsic::x86_sse2_ucomigt_sd: 3989 Opc = X86ISD::UCOMI; 3990 CC = ISD::SETGT; 3991 break; 3992 case Intrinsic::x86_sse_ucomige_ss: 3993 case Intrinsic::x86_sse2_ucomige_sd: 3994 Opc = X86ISD::UCOMI; 3995 CC = ISD::SETGE; 3996 break; 3997 case Intrinsic::x86_sse_ucomineq_ss: 3998 case Intrinsic::x86_sse2_ucomineq_sd: 3999 Opc = X86ISD::UCOMI; 4000 CC = ISD::SETNE; 4001 break; 4002 } 4003 4004 unsigned X86CC; 4005 SDOperand LHS = Op.getOperand(1); 4006 SDOperand RHS = Op.getOperand(2); 4007 translateX86CC(CC, true, X86CC, LHS, RHS, DAG); 4008 4009 const MVT::ValueType *VTs = DAG.getNodeValueTypes(MVT::Other, MVT::Flag); 4010 SDOperand Ops1[] = { DAG.getEntryNode(), LHS, RHS }; 4011 SDOperand Cond = DAG.getNode(Opc, VTs, 2, Ops1, 3); 4012 VTs = DAG.getNodeValueTypes(MVT::i8, MVT::Flag); 4013 SDOperand Ops2[] = { DAG.getConstant(X86CC, MVT::i8), Cond }; 4014 SDOperand SetCC = DAG.getNode(X86ISD::SETCC, VTs, 2, Ops2, 2); 4015 return DAG.getNode(ISD::ANY_EXTEND, MVT::i32, SetCC); 4016 } 4017 } 4018} 4019 4020SDOperand X86TargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) { 4021 // Depths > 0 not supported yet! 4022 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0) 4023 return SDOperand(); 4024 4025 // Just load the return address 4026 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG); 4027 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0); 4028} 4029 4030SDOperand X86TargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) { 4031 // Depths > 0 not supported yet! 4032 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0) 4033 return SDOperand(); 4034 4035 SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG); 4036 return DAG.getNode(ISD::SUB, getPointerTy(), RetAddrFI, 4037 DAG.getConstant(4, getPointerTy())); 4038} 4039 4040/// LowerOperation - Provide custom lowering hooks for some operations. 4041/// 4042SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) { 4043 switch (Op.getOpcode()) { 4044 default: assert(0 && "Should not custom lower this!"); 4045 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); 4046 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); 4047 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); 4048 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); 4049 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG); 4050 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 4051 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); 4052 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG); 4053 case ISD::SHL_PARTS: 4054 case ISD::SRA_PARTS: 4055 case ISD::SRL_PARTS: return LowerShift(Op, DAG); 4056 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG); 4057 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG); 4058 case ISD::FABS: return LowerFABS(Op, DAG); 4059 case ISD::FNEG: return LowerFNEG(Op, DAG); 4060 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); 4061 case ISD::SETCC: return LowerSETCC(Op, DAG, DAG.getEntryNode()); 4062 case ISD::SELECT: return LowerSELECT(Op, DAG); 4063 case ISD::BRCOND: return LowerBRCOND(Op, DAG); 4064 case ISD::JumpTable: return LowerJumpTable(Op, DAG); 4065 case ISD::CALL: return LowerCALL(Op, DAG); 4066 case ISD::RET: return LowerRET(Op, DAG); 4067 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG); 4068 case ISD::MEMSET: return LowerMEMSET(Op, DAG); 4069 case ISD::MEMCPY: return LowerMEMCPY(Op, DAG); 4070 case ISD::READCYCLECOUNTER: return LowerREADCYCLCECOUNTER(Op, DAG); 4071 case ISD::VASTART: return LowerVASTART(Op, DAG); 4072 case ISD::VACOPY: return LowerVACOPY(Op, DAG); 4073 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); 4074 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); 4075 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 4076 } 4077 return SDOperand(); 4078} 4079 4080const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const { 4081 switch (Opcode) { 4082 default: return NULL; 4083 case X86ISD::SHLD: return "X86ISD::SHLD"; 4084 case X86ISD::SHRD: return "X86ISD::SHRD"; 4085 case X86ISD::FAND: return "X86ISD::FAND"; 4086 case X86ISD::FOR: return "X86ISD::FOR"; 4087 case X86ISD::FXOR: return "X86ISD::FXOR"; 4088 case X86ISD::FSRL: return "X86ISD::FSRL"; 4089 case X86ISD::FILD: return "X86ISD::FILD"; 4090 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG"; 4091 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM"; 4092 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM"; 4093 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM"; 4094 case X86ISD::FLD: return "X86ISD::FLD"; 4095 case X86ISD::FST: return "X86ISD::FST"; 4096 case X86ISD::FP_GET_RESULT: return "X86ISD::FP_GET_RESULT"; 4097 case X86ISD::FP_SET_RESULT: return "X86ISD::FP_SET_RESULT"; 4098 case X86ISD::CALL: return "X86ISD::CALL"; 4099 case X86ISD::TAILCALL: return "X86ISD::TAILCALL"; 4100 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG"; 4101 case X86ISD::CMP: return "X86ISD::CMP"; 4102 case X86ISD::COMI: return "X86ISD::COMI"; 4103 case X86ISD::UCOMI: return "X86ISD::UCOMI"; 4104 case X86ISD::SETCC: return "X86ISD::SETCC"; 4105 case X86ISD::CMOV: return "X86ISD::CMOV"; 4106 case X86ISD::BRCOND: return "X86ISD::BRCOND"; 4107 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG"; 4108 case X86ISD::REP_STOS: return "X86ISD::REP_STOS"; 4109 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS"; 4110 case X86ISD::LOAD_PACK: return "X86ISD::LOAD_PACK"; 4111 case X86ISD::LOAD_UA: return "X86ISD::LOAD_UA"; 4112 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg"; 4113 case X86ISD::Wrapper: return "X86ISD::Wrapper"; 4114 case X86ISD::S2VEC: return "X86ISD::S2VEC"; 4115 case X86ISD::PEXTRW: return "X86ISD::PEXTRW"; 4116 case X86ISD::PINSRW: return "X86ISD::PINSRW"; 4117 case X86ISD::FMAX: return "X86ISD::FMAX"; 4118 case X86ISD::FMIN: return "X86ISD::FMIN"; 4119 } 4120} 4121 4122/// isLegalAddressImmediate - Return true if the integer value can be used 4123/// as the offset of the target addressing mode for load / store of the 4124/// given type. 4125bool X86TargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{ 4126 // X86 allows a sign-extended 32-bit immediate field. 4127 return (V > -(1LL << 32) && V < (1LL << 32)-1); 4128} 4129 4130/// isLegalAddressImmediate - Return true if the GlobalValue can be used as 4131/// the offset of the target addressing mode. 4132bool X86TargetLowering::isLegalAddressImmediate(GlobalValue *GV) const { 4133 // In 64-bit mode, GV is 64-bit so it won't fit in the 32-bit displacement 4134 // field unless we are in small code model. 4135 if (Subtarget->is64Bit() && 4136 getTargetMachine().getCodeModel() != CodeModel::Small) 4137 return false; 4138 4139 return (!Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false)); 4140} 4141 4142/// isLegalAddressScale - Return true if the integer value can be used as the 4143/// scale of the target addressing mode for load / store of the given type. 4144bool X86TargetLowering::isLegalAddressScale(int64_t S, const Type *Ty) const { 4145 switch (S) { 4146 default: 4147 return false; 4148 case 2: case 4: case 8: 4149 return true; 4150 // FIXME: These require both scale + index last and thus more expensive. 4151 // How to tell LSR to try for 2, 4, 8 first? 4152 case 3: case 5: case 9: 4153 return true; 4154 } 4155} 4156 4157/// isLegalAddressScaleAndImm - Return true if S works for IsLegalAddressScale 4158/// and V works for isLegalAddressImmediate _and_ both can be applied 4159/// simultaneously to the same instruction. 4160bool X86TargetLowering::isLegalAddressScaleAndImm(int64_t S, int64_t V, 4161 const Type* Ty) const { 4162 return isLegalAddressScale(S, Ty) && isLegalAddressImmediate(V, Ty); 4163} 4164 4165/// isLegalAddressScaleAndImm - Return true if S works for IsLegalAddressScale 4166/// and GV works for isLegalAddressImmediate _and_ both can be applied 4167/// simultaneously to the same instruction. 4168bool X86TargetLowering::isLegalAddressScaleAndImm(int64_t S, GlobalValue *GV, 4169 const Type* Ty) const { 4170 return isLegalAddressScale(S, Ty) && isLegalAddressImmediate(GV); 4171} 4172 4173/// isShuffleMaskLegal - Targets can use this to indicate that they only 4174/// support *some* VECTOR_SHUFFLE operations, those with specific masks. 4175/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 4176/// are assumed to be legal. 4177bool 4178X86TargetLowering::isShuffleMaskLegal(SDOperand Mask, MVT::ValueType VT) const { 4179 // Only do shuffles on 128-bit vector types for now. 4180 if (MVT::getSizeInBits(VT) == 64) return false; 4181 return (Mask.Val->getNumOperands() <= 4 || 4182 isSplatMask(Mask.Val) || 4183 isPSHUFHW_PSHUFLWMask(Mask.Val) || 4184 X86::isUNPCKLMask(Mask.Val) || 4185 X86::isUNPCKL_v_undef_Mask(Mask.Val) || 4186 X86::isUNPCKHMask(Mask.Val)); 4187} 4188 4189bool X86TargetLowering::isVectorClearMaskLegal(std::vector<SDOperand> &BVOps, 4190 MVT::ValueType EVT, 4191 SelectionDAG &DAG) const { 4192 unsigned NumElts = BVOps.size(); 4193 // Only do shuffles on 128-bit vector types for now. 4194 if (MVT::getSizeInBits(EVT) * NumElts == 64) return false; 4195 if (NumElts == 2) return true; 4196 if (NumElts == 4) { 4197 return (isMOVLMask(&BVOps[0], 4) || 4198 isCommutedMOVL(&BVOps[0], 4, true) || 4199 isSHUFPMask(&BVOps[0], 4) || 4200 isCommutedSHUFP(&BVOps[0], 4)); 4201 } 4202 return false; 4203} 4204 4205//===----------------------------------------------------------------------===// 4206// X86 Scheduler Hooks 4207//===----------------------------------------------------------------------===// 4208 4209MachineBasicBlock * 4210X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI, 4211 MachineBasicBlock *BB) { 4212 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 4213 switch (MI->getOpcode()) { 4214 default: assert(false && "Unexpected instr type to insert"); 4215 case X86::CMOV_FR32: 4216 case X86::CMOV_FR64: 4217 case X86::CMOV_V4F32: 4218 case X86::CMOV_V2F64: 4219 case X86::CMOV_V2I64: { 4220 // To "insert" a SELECT_CC instruction, we actually have to insert the 4221 // diamond control-flow pattern. The incoming instruction knows the 4222 // destination vreg to set, the condition code register to branch on, the 4223 // true/false values to select between, and a branch opcode to use. 4224 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 4225 ilist<MachineBasicBlock>::iterator It = BB; 4226 ++It; 4227 4228 // thisMBB: 4229 // ... 4230 // TrueVal = ... 4231 // cmpTY ccX, r1, r2 4232 // bCC copy1MBB 4233 // fallthrough --> copy0MBB 4234 MachineBasicBlock *thisMBB = BB; 4235 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB); 4236 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB); 4237 unsigned Opc = 4238 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm()); 4239 BuildMI(BB, TII->get(Opc)).addMBB(sinkMBB); 4240 MachineFunction *F = BB->getParent(); 4241 F->getBasicBlockList().insert(It, copy0MBB); 4242 F->getBasicBlockList().insert(It, sinkMBB); 4243 // Update machine-CFG edges by first adding all successors of the current 4244 // block to the new block which will contain the Phi node for the select. 4245 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(), 4246 e = BB->succ_end(); i != e; ++i) 4247 sinkMBB->addSuccessor(*i); 4248 // Next, remove all successors of the current block, and add the true 4249 // and fallthrough blocks as its successors. 4250 while(!BB->succ_empty()) 4251 BB->removeSuccessor(BB->succ_begin()); 4252 BB->addSuccessor(copy0MBB); 4253 BB->addSuccessor(sinkMBB); 4254 4255 // copy0MBB: 4256 // %FalseValue = ... 4257 // # fallthrough to sinkMBB 4258 BB = copy0MBB; 4259 4260 // Update machine-CFG edges 4261 BB->addSuccessor(sinkMBB); 4262 4263 // sinkMBB: 4264 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 4265 // ... 4266 BB = sinkMBB; 4267 BuildMI(BB, TII->get(X86::PHI), MI->getOperand(0).getReg()) 4268 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB) 4269 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); 4270 4271 delete MI; // The pseudo instruction is gone now. 4272 return BB; 4273 } 4274 4275 case X86::FP_TO_INT16_IN_MEM: 4276 case X86::FP_TO_INT32_IN_MEM: 4277 case X86::FP_TO_INT64_IN_MEM: { 4278 // Change the floating point control register to use "round towards zero" 4279 // mode when truncating to an integer value. 4280 MachineFunction *F = BB->getParent(); 4281 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2); 4282 addFrameReference(BuildMI(BB, TII->get(X86::FNSTCW16m)), CWFrameIdx); 4283 4284 // Load the old value of the high byte of the control word... 4285 unsigned OldCW = 4286 F->getSSARegMap()->createVirtualRegister(X86::GR16RegisterClass); 4287 addFrameReference(BuildMI(BB, TII->get(X86::MOV16rm), OldCW), CWFrameIdx); 4288 4289 // Set the high part to be round to zero... 4290 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mi)), CWFrameIdx) 4291 .addImm(0xC7F); 4292 4293 // Reload the modified control word now... 4294 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx); 4295 4296 // Restore the memory image of control word to original value 4297 addFrameReference(BuildMI(BB, TII->get(X86::MOV16mr)), CWFrameIdx) 4298 .addReg(OldCW); 4299 4300 // Get the X86 opcode to use. 4301 unsigned Opc; 4302 switch (MI->getOpcode()) { 4303 default: assert(0 && "illegal opcode!"); 4304 case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break; 4305 case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break; 4306 case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break; 4307 } 4308 4309 X86AddressMode AM; 4310 MachineOperand &Op = MI->getOperand(0); 4311 if (Op.isRegister()) { 4312 AM.BaseType = X86AddressMode::RegBase; 4313 AM.Base.Reg = Op.getReg(); 4314 } else { 4315 AM.BaseType = X86AddressMode::FrameIndexBase; 4316 AM.Base.FrameIndex = Op.getFrameIndex(); 4317 } 4318 Op = MI->getOperand(1); 4319 if (Op.isImmediate()) 4320 AM.Scale = Op.getImm(); 4321 Op = MI->getOperand(2); 4322 if (Op.isImmediate()) 4323 AM.IndexReg = Op.getImm(); 4324 Op = MI->getOperand(3); 4325 if (Op.isGlobalAddress()) { 4326 AM.GV = Op.getGlobal(); 4327 } else { 4328 AM.Disp = Op.getImm(); 4329 } 4330 addFullAddress(BuildMI(BB, TII->get(Opc)), AM) 4331 .addReg(MI->getOperand(4).getReg()); 4332 4333 // Reload the original control word now. 4334 addFrameReference(BuildMI(BB, TII->get(X86::FLDCW16m)), CWFrameIdx); 4335 4336 delete MI; // The pseudo instruction is gone now. 4337 return BB; 4338 } 4339 } 4340} 4341 4342//===----------------------------------------------------------------------===// 4343// X86 Optimization Hooks 4344//===----------------------------------------------------------------------===// 4345 4346void X86TargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op, 4347 uint64_t Mask, 4348 uint64_t &KnownZero, 4349 uint64_t &KnownOne, 4350 unsigned Depth) const { 4351 unsigned Opc = Op.getOpcode(); 4352 assert((Opc >= ISD::BUILTIN_OP_END || 4353 Opc == ISD::INTRINSIC_WO_CHAIN || 4354 Opc == ISD::INTRINSIC_W_CHAIN || 4355 Opc == ISD::INTRINSIC_VOID) && 4356 "Should use MaskedValueIsZero if you don't know whether Op" 4357 " is a target node!"); 4358 4359 KnownZero = KnownOne = 0; // Don't know anything. 4360 switch (Opc) { 4361 default: break; 4362 case X86ISD::SETCC: 4363 KnownZero |= (MVT::getIntVTBitMask(Op.getValueType()) ^ 1ULL); 4364 break; 4365 } 4366} 4367 4368/// getShuffleScalarElt - Returns the scalar element that will make up the ith 4369/// element of the result of the vector shuffle. 4370static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) { 4371 MVT::ValueType VT = N->getValueType(0); 4372 SDOperand PermMask = N->getOperand(2); 4373 unsigned NumElems = PermMask.getNumOperands(); 4374 SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1); 4375 i %= NumElems; 4376 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) { 4377 return (i == 0) 4378 ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT)); 4379 } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) { 4380 SDOperand Idx = PermMask.getOperand(i); 4381 if (Idx.getOpcode() == ISD::UNDEF) 4382 return DAG.getNode(ISD::UNDEF, MVT::getVectorBaseType(VT)); 4383 return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG); 4384 } 4385 return SDOperand(); 4386} 4387 4388/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 4389/// node is a GlobalAddress + an offset. 4390static bool isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) { 4391 unsigned Opc = N->getOpcode(); 4392 if (Opc == X86ISD::Wrapper) { 4393 if (dyn_cast<GlobalAddressSDNode>(N->getOperand(0))) { 4394 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal(); 4395 return true; 4396 } 4397 } else if (Opc == ISD::ADD) { 4398 SDOperand N1 = N->getOperand(0); 4399 SDOperand N2 = N->getOperand(1); 4400 if (isGAPlusOffset(N1.Val, GA, Offset)) { 4401 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 4402 if (V) { 4403 Offset += V->getSignExtended(); 4404 return true; 4405 } 4406 } else if (isGAPlusOffset(N2.Val, GA, Offset)) { 4407 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 4408 if (V) { 4409 Offset += V->getSignExtended(); 4410 return true; 4411 } 4412 } 4413 } 4414 return false; 4415} 4416 4417/// isConsecutiveLoad - Returns true if N is loading from an address of Base 4418/// + Dist * Size. 4419static bool isConsecutiveLoad(SDNode *N, SDNode *Base, int Dist, int Size, 4420 MachineFrameInfo *MFI) { 4421 if (N->getOperand(0).Val != Base->getOperand(0).Val) 4422 return false; 4423 4424 SDOperand Loc = N->getOperand(1); 4425 SDOperand BaseLoc = Base->getOperand(1); 4426 if (Loc.getOpcode() == ISD::FrameIndex) { 4427 if (BaseLoc.getOpcode() != ISD::FrameIndex) 4428 return false; 4429 int FI = dyn_cast<FrameIndexSDNode>(Loc)->getIndex(); 4430 int BFI = dyn_cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 4431 int FS = MFI->getObjectSize(FI); 4432 int BFS = MFI->getObjectSize(BFI); 4433 if (FS != BFS || FS != Size) return false; 4434 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Size); 4435 } else { 4436 GlobalValue *GV1 = NULL; 4437 GlobalValue *GV2 = NULL; 4438 int64_t Offset1 = 0; 4439 int64_t Offset2 = 0; 4440 bool isGA1 = isGAPlusOffset(Loc.Val, GV1, Offset1); 4441 bool isGA2 = isGAPlusOffset(BaseLoc.Val, GV2, Offset2); 4442 if (isGA1 && isGA2 && GV1 == GV2) 4443 return Offset1 == (Offset2 + Dist*Size); 4444 } 4445 4446 return false; 4447} 4448 4449static bool isBaseAlignment16(SDNode *Base, MachineFrameInfo *MFI, 4450 const X86Subtarget *Subtarget) { 4451 GlobalValue *GV; 4452 int64_t Offset; 4453 if (isGAPlusOffset(Base, GV, Offset)) 4454 return (GV->getAlignment() >= 16 && (Offset % 16) == 0); 4455 else { 4456 assert(Base->getOpcode() == ISD::FrameIndex && "Unexpected base node!"); 4457 int BFI = dyn_cast<FrameIndexSDNode>(Base)->getIndex(); 4458 if (BFI < 0) 4459 // Fixed objects do not specify alignment, however the offsets are known. 4460 return ((Subtarget->getStackAlignment() % 16) == 0 && 4461 (MFI->getObjectOffset(BFI) % 16) == 0); 4462 else 4463 return MFI->getObjectAlignment(BFI) >= 16; 4464 } 4465 return false; 4466} 4467 4468 4469/// PerformShuffleCombine - Combine a vector_shuffle that is equal to 4470/// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load 4471/// if the load addresses are consecutive, non-overlapping, and in the right 4472/// order. 4473static SDOperand PerformShuffleCombine(SDNode *N, SelectionDAG &DAG, 4474 const X86Subtarget *Subtarget) { 4475 MachineFunction &MF = DAG.getMachineFunction(); 4476 MachineFrameInfo *MFI = MF.getFrameInfo(); 4477 MVT::ValueType VT = N->getValueType(0); 4478 MVT::ValueType EVT = MVT::getVectorBaseType(VT); 4479 SDOperand PermMask = N->getOperand(2); 4480 int NumElems = (int)PermMask.getNumOperands(); 4481 SDNode *Base = NULL; 4482 for (int i = 0; i < NumElems; ++i) { 4483 SDOperand Idx = PermMask.getOperand(i); 4484 if (Idx.getOpcode() == ISD::UNDEF) { 4485 if (!Base) return SDOperand(); 4486 } else { 4487 SDOperand Arg = 4488 getShuffleScalarElt(N, cast<ConstantSDNode>(Idx)->getValue(), DAG); 4489 if (!Arg.Val || !ISD::isNON_EXTLoad(Arg.Val)) 4490 return SDOperand(); 4491 if (!Base) 4492 Base = Arg.Val; 4493 else if (!isConsecutiveLoad(Arg.Val, Base, 4494 i, MVT::getSizeInBits(EVT)/8,MFI)) 4495 return SDOperand(); 4496 } 4497 } 4498 4499 bool isAlign16 = isBaseAlignment16(Base->getOperand(1).Val, MFI, Subtarget); 4500 if (isAlign16) { 4501 LoadSDNode *LD = cast<LoadSDNode>(Base); 4502 return DAG.getLoad(VT, LD->getChain(), LD->getBasePtr(), LD->getSrcValue(), 4503 LD->getSrcValueOffset()); 4504 } else { 4505 // Just use movups, it's shorter. 4506 SDVTList Tys = DAG.getVTList(MVT::v4f32, MVT::Other); 4507 SmallVector<SDOperand, 3> Ops; 4508 Ops.push_back(Base->getOperand(0)); 4509 Ops.push_back(Base->getOperand(1)); 4510 Ops.push_back(Base->getOperand(2)); 4511 return DAG.getNode(ISD::BIT_CONVERT, VT, 4512 DAG.getNode(X86ISD::LOAD_UA, Tys, &Ops[0], Ops.size())); 4513 } 4514} 4515 4516/// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes. 4517static SDOperand PerformSELECTCombine(SDNode *N, SelectionDAG &DAG, 4518 const X86Subtarget *Subtarget) { 4519 SDOperand Cond = N->getOperand(0); 4520 4521 // If we have SSE[12] support, try to form min/max nodes. 4522 if (Subtarget->hasSSE2() && 4523 (N->getValueType(0) == MVT::f32 || N->getValueType(0) == MVT::f64)) { 4524 if (Cond.getOpcode() == ISD::SETCC) { 4525 // Get the LHS/RHS of the select. 4526 SDOperand LHS = N->getOperand(1); 4527 SDOperand RHS = N->getOperand(2); 4528 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); 4529 4530 unsigned Opcode = 0; 4531 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) { 4532 switch (CC) { 4533 default: break; 4534 case ISD::SETOLE: // (X <= Y) ? X : Y -> min 4535 case ISD::SETULE: 4536 case ISD::SETLE: 4537 if (!UnsafeFPMath) break; 4538 // FALL THROUGH. 4539 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min 4540 case ISD::SETLT: 4541 Opcode = X86ISD::FMIN; 4542 break; 4543 4544 case ISD::SETOGT: // (X > Y) ? X : Y -> max 4545 case ISD::SETUGT: 4546 case ISD::SETGT: 4547 if (!UnsafeFPMath) break; 4548 // FALL THROUGH. 4549 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max 4550 case ISD::SETGE: 4551 Opcode = X86ISD::FMAX; 4552 break; 4553 } 4554 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) { 4555 switch (CC) { 4556 default: break; 4557 case ISD::SETOGT: // (X > Y) ? Y : X -> min 4558 case ISD::SETUGT: 4559 case ISD::SETGT: 4560 if (!UnsafeFPMath) break; 4561 // FALL THROUGH. 4562 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min 4563 case ISD::SETGE: 4564 Opcode = X86ISD::FMIN; 4565 break; 4566 4567 case ISD::SETOLE: // (X <= Y) ? Y : X -> max 4568 case ISD::SETULE: 4569 case ISD::SETLE: 4570 if (!UnsafeFPMath) break; 4571 // FALL THROUGH. 4572 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max 4573 case ISD::SETLT: 4574 Opcode = X86ISD::FMAX; 4575 break; 4576 } 4577 } 4578 4579 if (Opcode) 4580 return DAG.getNode(Opcode, N->getValueType(0), LHS, RHS); 4581 } 4582 4583 } 4584 4585 return SDOperand(); 4586} 4587 4588 4589SDOperand X86TargetLowering::PerformDAGCombine(SDNode *N, 4590 DAGCombinerInfo &DCI) const { 4591 SelectionDAG &DAG = DCI.DAG; 4592 switch (N->getOpcode()) { 4593 default: break; 4594 case ISD::VECTOR_SHUFFLE: 4595 return PerformShuffleCombine(N, DAG, Subtarget); 4596 case ISD::SELECT: 4597 return PerformSELECTCombine(N, DAG, Subtarget); 4598 } 4599 4600 return SDOperand(); 4601} 4602 4603//===----------------------------------------------------------------------===// 4604// X86 Inline Assembly Support 4605//===----------------------------------------------------------------------===// 4606 4607/// getConstraintType - Given a constraint letter, return the type of 4608/// constraint it is for this target. 4609X86TargetLowering::ConstraintType 4610X86TargetLowering::getConstraintType(const std::string &Constraint) const { 4611 if (Constraint.size() == 1) { 4612 switch (Constraint[0]) { 4613 case 'A': 4614 case 'r': 4615 case 'R': 4616 case 'l': 4617 case 'q': 4618 case 'Q': 4619 case 'x': 4620 case 'Y': 4621 return C_RegisterClass; 4622 default: 4623 break; 4624 } 4625 } 4626 return TargetLowering::getConstraintType(Constraint); 4627} 4628 4629/// isOperandValidForConstraint - Return the specified operand (possibly 4630/// modified) if the specified SDOperand is valid for the specified target 4631/// constraint letter, otherwise return null. 4632SDOperand X86TargetLowering:: 4633isOperandValidForConstraint(SDOperand Op, char Constraint, SelectionDAG &DAG) { 4634 switch (Constraint) { 4635 default: break; 4636 case 'I': 4637 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4638 if (C->getValue() <= 31) 4639 return Op; 4640 } 4641 return SDOperand(0,0); 4642 case 'N': 4643 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) { 4644 if (C->getValue() <= 255) 4645 return Op; 4646 } 4647 return SDOperand(0,0); 4648 case 'i': 4649 // Literal immediates are always ok. 4650 if (isa<ConstantSDNode>(Op)) return Op; 4651 4652 // If we are in non-pic codegen mode, we allow the address of a global to 4653 // be used with 'i'. 4654 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) { 4655 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) 4656 return SDOperand(0, 0); 4657 4658 if (GA->getOpcode() != ISD::TargetGlobalAddress) 4659 Op = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0), 4660 GA->getOffset()); 4661 return Op; 4662 } 4663 4664 // Otherwise, not valid for this mode. 4665 return SDOperand(0, 0); 4666 } 4667 return TargetLowering::isOperandValidForConstraint(Op, Constraint, DAG); 4668} 4669 4670 4671std::vector<unsigned> X86TargetLowering:: 4672getRegClassForInlineAsmConstraint(const std::string &Constraint, 4673 MVT::ValueType VT) const { 4674 if (Constraint.size() == 1) { 4675 // FIXME: not handling fp-stack yet! 4676 // FIXME: not handling MMX registers yet ('y' constraint). 4677 switch (Constraint[0]) { // GCC X86 Constraint Letters 4678 default: break; // Unknown constraint letter 4679 case 'A': // EAX/EDX 4680 if (VT == MVT::i32 || VT == MVT::i64) 4681 return make_vector<unsigned>(X86::EAX, X86::EDX, 0); 4682 break; 4683 case 'r': // GENERAL_REGS 4684 case 'R': // LEGACY_REGS 4685 if (VT == MVT::i64 && Subtarget->is64Bit()) 4686 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 4687 X86::RSI, X86::RDI, X86::RBP, X86::RSP, 4688 X86::R8, X86::R9, X86::R10, X86::R11, 4689 X86::R12, X86::R13, X86::R14, X86::R15, 0); 4690 if (VT == MVT::i32) 4691 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 4692 X86::ESI, X86::EDI, X86::EBP, X86::ESP, 0); 4693 else if (VT == MVT::i16) 4694 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 4695 X86::SI, X86::DI, X86::BP, X86::SP, 0); 4696 else if (VT == MVT::i8) 4697 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0); 4698 break; 4699 case 'l': // INDEX_REGS 4700 if (VT == MVT::i32) 4701 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 4702 X86::ESI, X86::EDI, X86::EBP, 0); 4703 else if (VT == MVT::i16) 4704 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 4705 X86::SI, X86::DI, X86::BP, 0); 4706 else if (VT == MVT::i8) 4707 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0); 4708 break; 4709 case 'q': // Q_REGS (GENERAL_REGS in 64-bit mode) 4710 case 'Q': // Q_REGS 4711 if (VT == MVT::i32) 4712 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0); 4713 else if (VT == MVT::i16) 4714 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0); 4715 else if (VT == MVT::i8) 4716 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::DL, 0); 4717 break; 4718 case 'x': // SSE_REGS if SSE1 allowed 4719 if (Subtarget->hasSSE1()) 4720 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, 4721 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7, 4722 0); 4723 return std::vector<unsigned>(); 4724 case 'Y': // SSE_REGS if SSE2 allowed 4725 if (Subtarget->hasSSE2()) 4726 return make_vector<unsigned>(X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, 4727 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7, 4728 0); 4729 return std::vector<unsigned>(); 4730 } 4731 } 4732 4733 return std::vector<unsigned>(); 4734} 4735 4736std::pair<unsigned, const TargetRegisterClass*> 4737X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, 4738 MVT::ValueType VT) const { 4739 // Use the default implementation in TargetLowering to convert the register 4740 // constraint into a member of a register class. 4741 std::pair<unsigned, const TargetRegisterClass*> Res; 4742 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 4743 4744 // Not found as a standard register? 4745 if (Res.second == 0) { 4746 // GCC calls "st(0)" just plain "st". 4747 if (StringsEqualNoCase("{st}", Constraint)) { 4748 Res.first = X86::ST0; 4749 Res.second = X86::RSTRegisterClass; 4750 } 4751 4752 return Res; 4753 } 4754 4755 // Otherwise, check to see if this is a register class of the wrong value 4756 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to 4757 // turn into {ax},{dx}. 4758 if (Res.second->hasType(VT)) 4759 return Res; // Correct type already, nothing to do. 4760 4761 // All of the single-register GCC register classes map their values onto 4762 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we 4763 // really want an 8-bit or 32-bit register, map to the appropriate register 4764 // class and return the appropriate register. 4765 if (Res.second != X86::GR16RegisterClass) 4766 return Res; 4767 4768 if (VT == MVT::i8) { 4769 unsigned DestReg = 0; 4770 switch (Res.first) { 4771 default: break; 4772 case X86::AX: DestReg = X86::AL; break; 4773 case X86::DX: DestReg = X86::DL; break; 4774 case X86::CX: DestReg = X86::CL; break; 4775 case X86::BX: DestReg = X86::BL; break; 4776 } 4777 if (DestReg) { 4778 Res.first = DestReg; 4779 Res.second = Res.second = X86::GR8RegisterClass; 4780 } 4781 } else if (VT == MVT::i32) { 4782 unsigned DestReg = 0; 4783 switch (Res.first) { 4784 default: break; 4785 case X86::AX: DestReg = X86::EAX; break; 4786 case X86::DX: DestReg = X86::EDX; break; 4787 case X86::CX: DestReg = X86::ECX; break; 4788 case X86::BX: DestReg = X86::EBX; break; 4789 case X86::SI: DestReg = X86::ESI; break; 4790 case X86::DI: DestReg = X86::EDI; break; 4791 case X86::BP: DestReg = X86::EBP; break; 4792 case X86::SP: DestReg = X86::ESP; break; 4793 } 4794 if (DestReg) { 4795 Res.first = DestReg; 4796 Res.second = Res.second = X86::GR32RegisterClass; 4797 } 4798 } else if (VT == MVT::i64) { 4799 unsigned DestReg = 0; 4800 switch (Res.first) { 4801 default: break; 4802 case X86::AX: DestReg = X86::RAX; break; 4803 case X86::DX: DestReg = X86::RDX; break; 4804 case X86::CX: DestReg = X86::RCX; break; 4805 case X86::BX: DestReg = X86::RBX; break; 4806 case X86::SI: DestReg = X86::RSI; break; 4807 case X86::DI: DestReg = X86::RDI; break; 4808 case X86::BP: DestReg = X86::RBP; break; 4809 case X86::SP: DestReg = X86::RSP; break; 4810 } 4811 if (DestReg) { 4812 Res.first = DestReg; 4813 Res.second = Res.second = X86::GR64RegisterClass; 4814 } 4815 } 4816 4817 return Res; 4818} 4819