1//=- AArch64InstrAtomics.td - AArch64 Atomic codegen support -*- tablegen -*-=// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// AArch64 Atomic operand code-gen constructs. 11// 12//===----------------------------------------------------------------------===// 13 14//===---------------------------------- 15// Atomic fences 16//===---------------------------------- 17def : Pat<(atomic_fence (i64 4), (imm)), (DMB (i32 0x9))>; 18def : Pat<(atomic_fence (imm), (imm)), (DMB (i32 0xb))>; 19 20//===---------------------------------- 21// Atomic loads 22//===---------------------------------- 23 24// When they're actually atomic, only one addressing mode (GPR64sp) is 25// supported, but when they're relaxed and anything can be used, all the 26// standard modes would be valid and may give efficiency gains. 27 28// A atomic load operation that actually needs acquire semantics. 29class acquiring_load<PatFrag base> 30 : PatFrag<(ops node:$ptr), (base node:$ptr), [{ 31 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering(); 32 assert(Ordering != AcquireRelease && "unexpected load ordering"); 33 return Ordering == Acquire || Ordering == SequentiallyConsistent; 34}]>; 35 36// An atomic load operation that does not need either acquire or release 37// semantics. 38class relaxed_load<PatFrag base> 39 : PatFrag<(ops node:$ptr), (base node:$ptr), [{ 40 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering(); 41 return Ordering == Monotonic || Ordering == Unordered; 42}]>; 43 44// 8-bit loads 45def : Pat<(acquiring_load<atomic_load_8> GPR64sp:$ptr), (LDARB GPR64sp:$ptr)>; 46def : Pat<(relaxed_load<atomic_load_8> (ro_Windexed8 GPR64sp:$Rn, GPR32:$Rm, 47 ro_Wextend8:$offset)), 48 (LDRBBroW GPR64sp:$Rn, GPR32:$Rm, ro_Wextend8:$offset)>; 49def : Pat<(relaxed_load<atomic_load_8> (ro_Xindexed8 GPR64sp:$Rn, GPR64:$Rm, 50 ro_Xextend8:$offset)), 51 (LDRBBroX GPR64sp:$Rn, GPR64:$Rm, ro_Xextend8:$offset)>; 52def : Pat<(relaxed_load<atomic_load_8> (am_indexed8 GPR64sp:$Rn, 53 uimm12s1:$offset)), 54 (LDRBBui GPR64sp:$Rn, uimm12s1:$offset)>; 55def : Pat<(relaxed_load<atomic_load_8> 56 (am_unscaled8 GPR64sp:$Rn, simm9:$offset)), 57 (LDURBBi GPR64sp:$Rn, simm9:$offset)>; 58 59// 16-bit loads 60def : Pat<(acquiring_load<atomic_load_16> GPR64sp:$ptr), (LDARH GPR64sp:$ptr)>; 61def : Pat<(relaxed_load<atomic_load_16> (ro_Windexed16 GPR64sp:$Rn, GPR32:$Rm, 62 ro_Wextend16:$extend)), 63 (LDRHHroW GPR64sp:$Rn, GPR32:$Rm, ro_Wextend16:$extend)>; 64def : Pat<(relaxed_load<atomic_load_16> (ro_Xindexed16 GPR64sp:$Rn, GPR64:$Rm, 65 ro_Xextend16:$extend)), 66 (LDRHHroX GPR64sp:$Rn, GPR64:$Rm, ro_Xextend16:$extend)>; 67def : Pat<(relaxed_load<atomic_load_16> (am_indexed16 GPR64sp:$Rn, 68 uimm12s2:$offset)), 69 (LDRHHui GPR64sp:$Rn, uimm12s2:$offset)>; 70def : Pat<(relaxed_load<atomic_load_16> 71 (am_unscaled16 GPR64sp:$Rn, simm9:$offset)), 72 (LDURHHi GPR64sp:$Rn, simm9:$offset)>; 73 74// 32-bit loads 75def : Pat<(acquiring_load<atomic_load_32> GPR64sp:$ptr), (LDARW GPR64sp:$ptr)>; 76def : Pat<(relaxed_load<atomic_load_32> (ro_Windexed32 GPR64sp:$Rn, GPR32:$Rm, 77 ro_Wextend32:$extend)), 78 (LDRWroW GPR64sp:$Rn, GPR32:$Rm, ro_Wextend32:$extend)>; 79def : Pat<(relaxed_load<atomic_load_32> (ro_Xindexed32 GPR64sp:$Rn, GPR64:$Rm, 80 ro_Xextend32:$extend)), 81 (LDRWroX GPR64sp:$Rn, GPR64:$Rm, ro_Xextend32:$extend)>; 82def : Pat<(relaxed_load<atomic_load_32> (am_indexed32 GPR64sp:$Rn, 83 uimm12s4:$offset)), 84 (LDRWui GPR64sp:$Rn, uimm12s4:$offset)>; 85def : Pat<(relaxed_load<atomic_load_32> 86 (am_unscaled32 GPR64sp:$Rn, simm9:$offset)), 87 (LDURWi GPR64sp:$Rn, simm9:$offset)>; 88 89// 64-bit loads 90def : Pat<(acquiring_load<atomic_load_64> GPR64sp:$ptr), (LDARX GPR64sp:$ptr)>; 91def : Pat<(relaxed_load<atomic_load_64> (ro_Windexed64 GPR64sp:$Rn, GPR32:$Rm, 92 ro_Wextend64:$extend)), 93 (LDRXroW GPR64sp:$Rn, GPR32:$Rm, ro_Wextend64:$extend)>; 94def : Pat<(relaxed_load<atomic_load_64> (ro_Xindexed64 GPR64sp:$Rn, GPR64:$Rm, 95 ro_Xextend64:$extend)), 96 (LDRXroX GPR64sp:$Rn, GPR64:$Rm, ro_Xextend64:$extend)>; 97def : Pat<(relaxed_load<atomic_load_64> (am_indexed64 GPR64sp:$Rn, 98 uimm12s8:$offset)), 99 (LDRXui GPR64sp:$Rn, uimm12s8:$offset)>; 100def : Pat<(relaxed_load<atomic_load_64> 101 (am_unscaled64 GPR64sp:$Rn, simm9:$offset)), 102 (LDURXi GPR64sp:$Rn, simm9:$offset)>; 103 104//===---------------------------------- 105// Atomic stores 106//===---------------------------------- 107 108// When they're actually atomic, only one addressing mode (GPR64sp) is 109// supported, but when they're relaxed and anything can be used, all the 110// standard modes would be valid and may give efficiency gains. 111 112// A store operation that actually needs release semantics. 113class releasing_store<PatFrag base> 114 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{ 115 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering(); 116 assert(Ordering != AcquireRelease && "unexpected store ordering"); 117 return Ordering == Release || Ordering == SequentiallyConsistent; 118}]>; 119 120// An atomic store operation that doesn't actually need to be atomic on AArch64. 121class relaxed_store<PatFrag base> 122 : PatFrag<(ops node:$ptr, node:$val), (base node:$ptr, node:$val), [{ 123 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering(); 124 return Ordering == Monotonic || Ordering == Unordered; 125}]>; 126 127// 8-bit stores 128def : Pat<(releasing_store<atomic_store_8> GPR64sp:$ptr, GPR32:$val), 129 (STLRB GPR32:$val, GPR64sp:$ptr)>; 130def : Pat<(relaxed_store<atomic_store_8> 131 (ro_Windexed8 GPR64sp:$Rn, GPR32:$Rm, ro_Wextend8:$extend), 132 GPR32:$val), 133 (STRBBroW GPR32:$val, GPR64sp:$Rn, GPR32:$Rm, ro_Wextend8:$extend)>; 134def : Pat<(relaxed_store<atomic_store_8> 135 (ro_Xindexed8 GPR64sp:$Rn, GPR64:$Rm, ro_Xextend8:$extend), 136 GPR32:$val), 137 (STRBBroX GPR32:$val, GPR64sp:$Rn, GPR64:$Rm, ro_Xextend8:$extend)>; 138def : Pat<(relaxed_store<atomic_store_8> 139 (am_indexed8 GPR64sp:$Rn, uimm12s1:$offset), GPR32:$val), 140 (STRBBui GPR32:$val, GPR64sp:$Rn, uimm12s1:$offset)>; 141def : Pat<(relaxed_store<atomic_store_8> 142 (am_unscaled8 GPR64sp:$Rn, simm9:$offset), GPR32:$val), 143 (STURBBi GPR32:$val, GPR64sp:$Rn, simm9:$offset)>; 144 145// 16-bit stores 146def : Pat<(releasing_store<atomic_store_16> GPR64sp:$ptr, GPR32:$val), 147 (STLRH GPR32:$val, GPR64sp:$ptr)>; 148def : Pat<(relaxed_store<atomic_store_16> (ro_Windexed16 GPR64sp:$Rn, GPR32:$Rm, 149 ro_Wextend16:$extend), 150 GPR32:$val), 151 (STRHHroW GPR32:$val, GPR64sp:$Rn, GPR32:$Rm, ro_Wextend16:$extend)>; 152def : Pat<(relaxed_store<atomic_store_16> (ro_Xindexed16 GPR64sp:$Rn, GPR64:$Rm, 153 ro_Xextend16:$extend), 154 GPR32:$val), 155 (STRHHroX GPR32:$val, GPR64sp:$Rn, GPR64:$Rm, ro_Xextend16:$extend)>; 156def : Pat<(relaxed_store<atomic_store_16> 157 (am_indexed16 GPR64sp:$Rn, uimm12s2:$offset), GPR32:$val), 158 (STRHHui GPR32:$val, GPR64sp:$Rn, uimm12s2:$offset)>; 159def : Pat<(relaxed_store<atomic_store_16> 160 (am_unscaled16 GPR64sp:$Rn, simm9:$offset), GPR32:$val), 161 (STURHHi GPR32:$val, GPR64sp:$Rn, simm9:$offset)>; 162 163// 32-bit stores 164def : Pat<(releasing_store<atomic_store_32> GPR64sp:$ptr, GPR32:$val), 165 (STLRW GPR32:$val, GPR64sp:$ptr)>; 166def : Pat<(relaxed_store<atomic_store_32> (ro_Windexed32 GPR64sp:$Rn, GPR32:$Rm, 167 ro_Wextend32:$extend), 168 GPR32:$val), 169 (STRWroW GPR32:$val, GPR64sp:$Rn, GPR32:$Rm, ro_Wextend32:$extend)>; 170def : Pat<(relaxed_store<atomic_store_32> (ro_Xindexed32 GPR64sp:$Rn, GPR64:$Rm, 171 ro_Xextend32:$extend), 172 GPR32:$val), 173 (STRWroX GPR32:$val, GPR64sp:$Rn, GPR64:$Rm, ro_Xextend32:$extend)>; 174def : Pat<(relaxed_store<atomic_store_32> 175 (am_indexed32 GPR64sp:$Rn, uimm12s4:$offset), GPR32:$val), 176 (STRWui GPR32:$val, GPR64sp:$Rn, uimm12s4:$offset)>; 177def : Pat<(relaxed_store<atomic_store_32> 178 (am_unscaled32 GPR64sp:$Rn, simm9:$offset), GPR32:$val), 179 (STURWi GPR32:$val, GPR64sp:$Rn, simm9:$offset)>; 180 181// 64-bit stores 182def : Pat<(releasing_store<atomic_store_64> GPR64sp:$ptr, GPR64:$val), 183 (STLRX GPR64:$val, GPR64sp:$ptr)>; 184def : Pat<(relaxed_store<atomic_store_64> (ro_Windexed64 GPR64sp:$Rn, GPR32:$Rm, 185 ro_Wextend16:$extend), 186 GPR64:$val), 187 (STRXroW GPR64:$val, GPR64sp:$Rn, GPR32:$Rm, ro_Wextend64:$extend)>; 188def : Pat<(relaxed_store<atomic_store_64> (ro_Xindexed64 GPR64sp:$Rn, GPR64:$Rm, 189 ro_Xextend16:$extend), 190 GPR64:$val), 191 (STRXroX GPR64:$val, GPR64sp:$Rn, GPR64:$Rm, ro_Xextend64:$extend)>; 192def : Pat<(relaxed_store<atomic_store_64> 193 (am_indexed64 GPR64sp:$Rn, uimm12s8:$offset), GPR64:$val), 194 (STRXui GPR64:$val, GPR64sp:$Rn, uimm12s8:$offset)>; 195def : Pat<(relaxed_store<atomic_store_64> 196 (am_unscaled64 GPR64sp:$Rn, simm9:$offset), GPR64:$val), 197 (STURXi GPR64:$val, GPR64sp:$Rn, simm9:$offset)>; 198 199//===---------------------------------- 200// Low-level exclusive operations 201//===---------------------------------- 202 203// Load-exclusives. 204 205def ldxr_1 : PatFrag<(ops node:$ptr), (int_aarch64_ldxr node:$ptr), [{ 206 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i8; 207}]>; 208 209def ldxr_2 : PatFrag<(ops node:$ptr), (int_aarch64_ldxr node:$ptr), [{ 210 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i16; 211}]>; 212 213def ldxr_4 : PatFrag<(ops node:$ptr), (int_aarch64_ldxr node:$ptr), [{ 214 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i32; 215}]>; 216 217def ldxr_8 : PatFrag<(ops node:$ptr), (int_aarch64_ldxr node:$ptr), [{ 218 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i64; 219}]>; 220 221def : Pat<(ldxr_1 GPR64sp:$addr), 222 (SUBREG_TO_REG (i64 0), (LDXRB GPR64sp:$addr), sub_32)>; 223def : Pat<(ldxr_2 GPR64sp:$addr), 224 (SUBREG_TO_REG (i64 0), (LDXRH GPR64sp:$addr), sub_32)>; 225def : Pat<(ldxr_4 GPR64sp:$addr), 226 (SUBREG_TO_REG (i64 0), (LDXRW GPR64sp:$addr), sub_32)>; 227def : Pat<(ldxr_8 GPR64sp:$addr), (LDXRX GPR64sp:$addr)>; 228 229def : Pat<(and (ldxr_1 GPR64sp:$addr), 0xff), 230 (SUBREG_TO_REG (i64 0), (LDXRB GPR64sp:$addr), sub_32)>; 231def : Pat<(and (ldxr_2 GPR64sp:$addr), 0xffff), 232 (SUBREG_TO_REG (i64 0), (LDXRH GPR64sp:$addr), sub_32)>; 233def : Pat<(and (ldxr_4 GPR64sp:$addr), 0xffffffff), 234 (SUBREG_TO_REG (i64 0), (LDXRW GPR64sp:$addr), sub_32)>; 235 236// Load-exclusives. 237 238def ldaxr_1 : PatFrag<(ops node:$ptr), (int_aarch64_ldaxr node:$ptr), [{ 239 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i8; 240}]>; 241 242def ldaxr_2 : PatFrag<(ops node:$ptr), (int_aarch64_ldaxr node:$ptr), [{ 243 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i16; 244}]>; 245 246def ldaxr_4 : PatFrag<(ops node:$ptr), (int_aarch64_ldaxr node:$ptr), [{ 247 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i32; 248}]>; 249 250def ldaxr_8 : PatFrag<(ops node:$ptr), (int_aarch64_ldaxr node:$ptr), [{ 251 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i64; 252}]>; 253 254def : Pat<(ldaxr_1 GPR64sp:$addr), 255 (SUBREG_TO_REG (i64 0), (LDAXRB GPR64sp:$addr), sub_32)>; 256def : Pat<(ldaxr_2 GPR64sp:$addr), 257 (SUBREG_TO_REG (i64 0), (LDAXRH GPR64sp:$addr), sub_32)>; 258def : Pat<(ldaxr_4 GPR64sp:$addr), 259 (SUBREG_TO_REG (i64 0), (LDAXRW GPR64sp:$addr), sub_32)>; 260def : Pat<(ldaxr_8 GPR64sp:$addr), (LDAXRX GPR64sp:$addr)>; 261 262def : Pat<(and (ldaxr_1 GPR64sp:$addr), 0xff), 263 (SUBREG_TO_REG (i64 0), (LDAXRB GPR64sp:$addr), sub_32)>; 264def : Pat<(and (ldaxr_2 GPR64sp:$addr), 0xffff), 265 (SUBREG_TO_REG (i64 0), (LDAXRH GPR64sp:$addr), sub_32)>; 266def : Pat<(and (ldaxr_4 GPR64sp:$addr), 0xffffffff), 267 (SUBREG_TO_REG (i64 0), (LDAXRW GPR64sp:$addr), sub_32)>; 268 269// Store-exclusives. 270 271def stxr_1 : PatFrag<(ops node:$val, node:$ptr), 272 (int_aarch64_stxr node:$val, node:$ptr), [{ 273 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i8; 274}]>; 275 276def stxr_2 : PatFrag<(ops node:$val, node:$ptr), 277 (int_aarch64_stxr node:$val, node:$ptr), [{ 278 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i16; 279}]>; 280 281def stxr_4 : PatFrag<(ops node:$val, node:$ptr), 282 (int_aarch64_stxr node:$val, node:$ptr), [{ 283 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i32; 284}]>; 285 286def stxr_8 : PatFrag<(ops node:$val, node:$ptr), 287 (int_aarch64_stxr node:$val, node:$ptr), [{ 288 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i64; 289}]>; 290 291 292def : Pat<(stxr_1 GPR64:$val, GPR64sp:$addr), 293 (STXRB (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 294def : Pat<(stxr_2 GPR64:$val, GPR64sp:$addr), 295 (STXRH (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 296def : Pat<(stxr_4 GPR64:$val, GPR64sp:$addr), 297 (STXRW (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 298def : Pat<(stxr_8 GPR64:$val, GPR64sp:$addr), 299 (STXRX GPR64:$val, GPR64sp:$addr)>; 300 301def : Pat<(stxr_1 (zext (and GPR32:$val, 0xff)), GPR64sp:$addr), 302 (STXRB GPR32:$val, GPR64sp:$addr)>; 303def : Pat<(stxr_2 (zext (and GPR32:$val, 0xffff)), GPR64sp:$addr), 304 (STXRH GPR32:$val, GPR64sp:$addr)>; 305def : Pat<(stxr_4 (zext GPR32:$val), GPR64sp:$addr), 306 (STXRW GPR32:$val, GPR64sp:$addr)>; 307 308def : Pat<(stxr_1 (and GPR64:$val, 0xff), GPR64sp:$addr), 309 (STXRB (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 310def : Pat<(stxr_2 (and GPR64:$val, 0xffff), GPR64sp:$addr), 311 (STXRH (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 312def : Pat<(stxr_4 (and GPR64:$val, 0xffffffff), GPR64sp:$addr), 313 (STXRW (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 314 315// Store-release-exclusives. 316 317def stlxr_1 : PatFrag<(ops node:$val, node:$ptr), 318 (int_aarch64_stlxr node:$val, node:$ptr), [{ 319 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i8; 320}]>; 321 322def stlxr_2 : PatFrag<(ops node:$val, node:$ptr), 323 (int_aarch64_stlxr node:$val, node:$ptr), [{ 324 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i16; 325}]>; 326 327def stlxr_4 : PatFrag<(ops node:$val, node:$ptr), 328 (int_aarch64_stlxr node:$val, node:$ptr), [{ 329 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i32; 330}]>; 331 332def stlxr_8 : PatFrag<(ops node:$val, node:$ptr), 333 (int_aarch64_stlxr node:$val, node:$ptr), [{ 334 return cast<MemIntrinsicSDNode>(N)->getMemoryVT() == MVT::i64; 335}]>; 336 337 338def : Pat<(stlxr_1 GPR64:$val, GPR64sp:$addr), 339 (STLXRB (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 340def : Pat<(stlxr_2 GPR64:$val, GPR64sp:$addr), 341 (STLXRH (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 342def : Pat<(stlxr_4 GPR64:$val, GPR64sp:$addr), 343 (STLXRW (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 344def : Pat<(stlxr_8 GPR64:$val, GPR64sp:$addr), 345 (STLXRX GPR64:$val, GPR64sp:$addr)>; 346 347def : Pat<(stlxr_1 (zext (and GPR32:$val, 0xff)), GPR64sp:$addr), 348 (STLXRB GPR32:$val, GPR64sp:$addr)>; 349def : Pat<(stlxr_2 (zext (and GPR32:$val, 0xffff)), GPR64sp:$addr), 350 (STLXRH GPR32:$val, GPR64sp:$addr)>; 351def : Pat<(stlxr_4 (zext GPR32:$val), GPR64sp:$addr), 352 (STLXRW GPR32:$val, GPR64sp:$addr)>; 353 354def : Pat<(stlxr_1 (and GPR64:$val, 0xff), GPR64sp:$addr), 355 (STLXRB (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 356def : Pat<(stlxr_2 (and GPR64:$val, 0xffff), GPR64sp:$addr), 357 (STLXRH (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 358def : Pat<(stlxr_4 (and GPR64:$val, 0xffffffff), GPR64sp:$addr), 359 (STLXRW (EXTRACT_SUBREG GPR64:$val, sub_32), GPR64sp:$addr)>; 360 361 362// And clear exclusive. 363 364def : Pat<(int_aarch64_clrex), (CLREX 0xf)>; 365