1 use crate::config::Config; 2 use crate::cranelift_arbitrary::CraneliftArbitrary; 3 use crate::target_isa_extras::TargetIsaExtras; 4 use anyhow::Result; 5 use arbitrary::{Arbitrary, Unstructured}; 6 use cranelift::codegen::data_value::DataValue; 7 use cranelift::codegen::ir::immediates::Offset32; 8 use cranelift::codegen::ir::instructions::{InstructionFormat, ResolvedConstraint}; 9 use cranelift::codegen::ir::stackslot::StackSize; 10 11 use cranelift::codegen::ir::{ 12 AliasRegion, AtomicRmwOp, Block, BlockArg, ConstantData, Endianness, ExternalName, FuncRef, 13 Function, LibCall, Opcode, SigRef, Signature, StackSlot, UserExternalName, UserFuncName, Value, 14 types::*, 15 }; 16 use cranelift::codegen::isa::CallConv; 17 use cranelift::frontend::{FunctionBuilder, FunctionBuilderContext, Switch, Variable}; 18 use cranelift::prelude::isa::OwnedTargetIsa; 19 use cranelift::prelude::{ 20 ExtFuncData, FloatCC, InstBuilder, IntCC, JumpTableData, MemFlags, StackSlotData, StackSlotKind, 21 }; 22 use std::collections::HashMap; 23 use std::ops::RangeInclusive; 24 use std::str::FromStr; 25 use std::sync::LazyLock; 26 use target_lexicon::{Architecture, Triple}; 27 28 type BlockSignature = Vec<Type>; 29 30 fn insert_opcode( 31 fgen: &mut FunctionGenerator, 32 builder: &mut FunctionBuilder, 33 opcode: Opcode, 34 args: &[Type], 35 rets: &[Type], 36 ) -> Result<()> { 37 let mut vals = Vec::with_capacity(args.len()); 38 for &arg in args.into_iter() { 39 let var = fgen.get_variable_of_type(arg)?; 40 let val = builder.use_var(var); 41 vals.push(val); 42 } 43 44 // Some opcodes require us to look at their input arguments to determine the 45 // controlling type. This is not the general case, but we can neatly check this 46 // using `requires_typevar_operand`. 47 let ctrl_type = if opcode.constraints().requires_typevar_operand() { 48 args.first() 49 } else { 50 rets.first() 51 } 52 .copied() 53 .unwrap_or(INVALID); 54 55 // Choose the appropriate instruction format for this opcode 56 let (inst, dfg) = match opcode.format() { 57 InstructionFormat::NullAry => builder.ins().NullAry(opcode, ctrl_type), 58 InstructionFormat::Unary => builder.ins().Unary(opcode, ctrl_type, vals[0]), 59 InstructionFormat::Binary => builder.ins().Binary(opcode, ctrl_type, vals[0], vals[1]), 60 InstructionFormat::Ternary => builder 61 .ins() 62 .Ternary(opcode, ctrl_type, vals[0], vals[1], vals[2]), 63 _ => unimplemented!(), 64 }; 65 let results = dfg.inst_results(inst).to_vec(); 66 67 for (val, &ty) in results.into_iter().zip(rets) { 68 let var = fgen.get_variable_of_type(ty)?; 69 builder.def_var(var, val); 70 } 71 Ok(()) 72 } 73 74 fn insert_call_to_function( 75 fgen: &mut FunctionGenerator, 76 builder: &mut FunctionBuilder, 77 call_opcode: Opcode, 78 sig: &Signature, 79 sig_ref: SigRef, 80 func_ref: FuncRef, 81 ) -> Result<()> { 82 let actuals = fgen.generate_values_for_signature( 83 builder, 84 sig.params.iter().map(|abi_param| abi_param.value_type), 85 )?; 86 87 let addr_ty = fgen.isa.pointer_type(); 88 let call = match call_opcode { 89 Opcode::Call => builder.ins().call(func_ref, &actuals), 90 Opcode::ReturnCall => builder.ins().return_call(func_ref, &actuals), 91 Opcode::CallIndirect => { 92 let addr = builder.ins().func_addr(addr_ty, func_ref); 93 builder.ins().call_indirect(sig_ref, addr, &actuals) 94 } 95 Opcode::ReturnCallIndirect => { 96 let addr = builder.ins().func_addr(addr_ty, func_ref); 97 builder.ins().return_call_indirect(sig_ref, addr, &actuals) 98 } 99 _ => unreachable!(), 100 }; 101 102 // Assign the return values to random variables 103 let ret_values = builder.inst_results(call).to_vec(); 104 let ret_types = sig.returns.iter().map(|p| p.value_type); 105 for (ty, val) in ret_types.zip(ret_values) { 106 let var = fgen.get_variable_of_type(ty)?; 107 builder.def_var(var, val); 108 } 109 110 Ok(()) 111 } 112 113 fn insert_call( 114 fgen: &mut FunctionGenerator, 115 builder: &mut FunctionBuilder, 116 opcode: Opcode, 117 _args: &[Type], 118 _rets: &[Type], 119 ) -> Result<()> { 120 assert!(matches!(opcode, Opcode::Call | Opcode::CallIndirect)); 121 let (sig, sig_ref, func_ref) = fgen.u.choose(&fgen.resources.func_refs)?.clone(); 122 insert_call_to_function(fgen, builder, opcode, &sig, sig_ref, func_ref) 123 } 124 125 fn insert_stack_load( 126 fgen: &mut FunctionGenerator, 127 builder: &mut FunctionBuilder, 128 _opcode: Opcode, 129 _args: &[Type], 130 rets: &[Type], 131 ) -> Result<()> { 132 let typevar = rets[0]; 133 let type_size = typevar.bytes(); 134 let (slot, slot_size, _align, category) = fgen.stack_slot_with_size(type_size)?; 135 136 // `stack_load` doesn't support setting MemFlags, and it does not set any 137 // alias analysis bits, so we can only emit it for `Other` slots. 138 if category != AACategory::Other { 139 return Err(arbitrary::Error::IncorrectFormat.into()); 140 } 141 142 let offset = fgen.u.int_in_range(0..=(slot_size - type_size))? as i32; 143 144 let val = builder.ins().stack_load(typevar, slot, offset); 145 let var = fgen.get_variable_of_type(typevar)?; 146 builder.def_var(var, val); 147 148 Ok(()) 149 } 150 151 fn insert_stack_store( 152 fgen: &mut FunctionGenerator, 153 builder: &mut FunctionBuilder, 154 _opcode: Opcode, 155 args: &[Type], 156 _rets: &[Type], 157 ) -> Result<()> { 158 let typevar = args[0]; 159 let type_size = typevar.bytes(); 160 161 let (slot, slot_size, _align, category) = fgen.stack_slot_with_size(type_size)?; 162 163 // `stack_store` doesn't support setting MemFlags, and it does not set any 164 // alias analysis bits, so we can only emit it for `Other` slots. 165 if category != AACategory::Other { 166 return Err(arbitrary::Error::IncorrectFormat.into()); 167 } 168 169 let offset = fgen.u.int_in_range(0..=(slot_size - type_size))? as i32; 170 171 let arg0 = fgen.get_variable_of_type(typevar)?; 172 let arg0 = builder.use_var(arg0); 173 174 builder.ins().stack_store(arg0, slot, offset); 175 Ok(()) 176 } 177 178 fn insert_cmp( 179 fgen: &mut FunctionGenerator, 180 builder: &mut FunctionBuilder, 181 opcode: Opcode, 182 args: &[Type], 183 rets: &[Type], 184 ) -> Result<()> { 185 let lhs = fgen.get_variable_of_type(args[0])?; 186 let lhs = builder.use_var(lhs); 187 188 let rhs = fgen.get_variable_of_type(args[1])?; 189 let rhs = builder.use_var(rhs); 190 191 let res = if opcode == Opcode::Fcmp { 192 let cc = *fgen.u.choose(FloatCC::all())?; 193 194 // We filter out condition codes that aren't supported by the target at 195 // this point after randomly choosing one, instead of randomly choosing a 196 // supported one, to avoid invalidating the corpus when these get implemented. 197 let unimplemented_cc = match (fgen.isa.triple().architecture, cc) { 198 // Some FloatCC's are not implemented on AArch64, see: 199 // https://github.com/bytecodealliance/wasmtime/issues/4850 200 (Architecture::Aarch64(_), FloatCC::OrderedNotEqual) => true, 201 (Architecture::Aarch64(_), FloatCC::UnorderedOrEqual) => true, 202 203 // These are not implemented on x86_64, for vectors. 204 (Architecture::X86_64, FloatCC::UnorderedOrEqual | FloatCC::OrderedNotEqual) => { 205 args[0].is_vector() 206 } 207 _ => false, 208 }; 209 if unimplemented_cc { 210 return Err(arbitrary::Error::IncorrectFormat.into()); 211 } 212 213 builder.ins().fcmp(cc, lhs, rhs) 214 } else { 215 let cc = *fgen.u.choose(IntCC::all())?; 216 builder.ins().icmp(cc, lhs, rhs) 217 }; 218 219 let var = fgen.get_variable_of_type(rets[0])?; 220 builder.def_var(var, res); 221 222 Ok(()) 223 } 224 225 fn insert_const( 226 fgen: &mut FunctionGenerator, 227 builder: &mut FunctionBuilder, 228 _opcode: Opcode, 229 _args: &[Type], 230 rets: &[Type], 231 ) -> Result<()> { 232 let typevar = rets[0]; 233 let var = fgen.get_variable_of_type(typevar)?; 234 let val = fgen.generate_const(builder, typevar)?; 235 builder.def_var(var, val); 236 Ok(()) 237 } 238 239 fn insert_bitcast( 240 fgen: &mut FunctionGenerator, 241 builder: &mut FunctionBuilder, 242 args: &[Type], 243 rets: &[Type], 244 ) -> Result<()> { 245 let from_var = fgen.get_variable_of_type(args[0])?; 246 let from_val = builder.use_var(from_var); 247 248 let to_var = fgen.get_variable_of_type(rets[0])?; 249 250 // TODO: We can generate little/big endian flags here. 251 let mut memflags = MemFlags::new(); 252 253 // When bitcasting between vectors of different lane counts, we need to 254 // specify the endianness. 255 if args[0].lane_count() != rets[0].lane_count() { 256 memflags.set_endianness(Endianness::Little); 257 } 258 259 let res = builder.ins().bitcast(rets[0], memflags, from_val); 260 builder.def_var(to_var, res); 261 Ok(()) 262 } 263 264 fn insert_load_store( 265 fgen: &mut FunctionGenerator, 266 builder: &mut FunctionBuilder, 267 opcode: Opcode, 268 args: &[Type], 269 rets: &[Type], 270 ) -> Result<()> { 271 if opcode == Opcode::Bitcast { 272 return insert_bitcast(fgen, builder, args, rets); 273 } 274 275 let ctrl_type = *rets.first().or(args.first()).unwrap(); 276 let type_size = ctrl_type.bytes(); 277 278 let is_atomic = [Opcode::AtomicLoad, Opcode::AtomicStore].contains(&opcode); 279 let (address, flags, offset) = 280 fgen.generate_address_and_memflags(builder, type_size, is_atomic)?; 281 282 // The variable being loaded or stored into 283 let var = fgen.get_variable_of_type(ctrl_type)?; 284 285 match opcode.format() { 286 InstructionFormat::LoadNoOffset => { 287 let (inst, dfg) = builder 288 .ins() 289 .LoadNoOffset(opcode, ctrl_type, flags, address); 290 291 let new_val = dfg.first_result(inst); 292 builder.def_var(var, new_val); 293 } 294 InstructionFormat::StoreNoOffset => { 295 let val = builder.use_var(var); 296 297 builder 298 .ins() 299 .StoreNoOffset(opcode, ctrl_type, flags, val, address); 300 } 301 InstructionFormat::Store => { 302 let val = builder.use_var(var); 303 304 builder 305 .ins() 306 .Store(opcode, ctrl_type, flags, offset, val, address); 307 } 308 InstructionFormat::Load => { 309 let (inst, dfg) = builder 310 .ins() 311 .Load(opcode, ctrl_type, flags, offset, address); 312 313 let new_val = dfg.first_result(inst); 314 builder.def_var(var, new_val); 315 } 316 _ => unimplemented!(), 317 } 318 319 Ok(()) 320 } 321 322 fn insert_atomic_rmw( 323 fgen: &mut FunctionGenerator, 324 builder: &mut FunctionBuilder, 325 _: Opcode, 326 _: &[Type], 327 rets: &[Type], 328 ) -> Result<()> { 329 let ctrl_type = *rets.first().unwrap(); 330 let type_size = ctrl_type.bytes(); 331 332 let rmw_op = *fgen.u.choose(AtomicRmwOp::all())?; 333 334 let (address, flags, offset) = fgen.generate_address_and_memflags(builder, type_size, true)?; 335 336 // AtomicRMW does not directly support offsets, so add the offset to the address separately. 337 let address = builder.ins().iadd_imm(address, i64::from(offset)); 338 339 // Load and store target variables 340 let source_var = fgen.get_variable_of_type(ctrl_type)?; 341 let target_var = fgen.get_variable_of_type(ctrl_type)?; 342 343 let source_val = builder.use_var(source_var); 344 let new_val = builder 345 .ins() 346 .atomic_rmw(ctrl_type, flags, rmw_op, address, source_val); 347 348 builder.def_var(target_var, new_val); 349 Ok(()) 350 } 351 352 fn insert_atomic_cas( 353 fgen: &mut FunctionGenerator, 354 builder: &mut FunctionBuilder, 355 _: Opcode, 356 _: &[Type], 357 rets: &[Type], 358 ) -> Result<()> { 359 let ctrl_type = *rets.first().unwrap(); 360 let type_size = ctrl_type.bytes(); 361 362 let (address, flags, offset) = fgen.generate_address_and_memflags(builder, type_size, true)?; 363 364 // AtomicCas does not directly support offsets, so add the offset to the address separately. 365 let address = builder.ins().iadd_imm(address, i64::from(offset)); 366 367 // Source and Target variables 368 let expected_var = fgen.get_variable_of_type(ctrl_type)?; 369 let store_var = fgen.get_variable_of_type(ctrl_type)?; 370 let loaded_var = fgen.get_variable_of_type(ctrl_type)?; 371 372 let expected_val = builder.use_var(expected_var); 373 let store_val = builder.use_var(store_var); 374 let new_val = builder 375 .ins() 376 .atomic_cas(flags, address, expected_val, store_val); 377 378 builder.def_var(loaded_var, new_val); 379 Ok(()) 380 } 381 382 fn insert_shuffle( 383 fgen: &mut FunctionGenerator, 384 builder: &mut FunctionBuilder, 385 opcode: Opcode, 386 _: &[Type], 387 rets: &[Type], 388 ) -> Result<()> { 389 let ctrl_type = *rets.first().unwrap(); 390 391 let lhs = builder.use_var(fgen.get_variable_of_type(ctrl_type)?); 392 let rhs = builder.use_var(fgen.get_variable_of_type(ctrl_type)?); 393 394 let mask = { 395 let mut lanes = [0u8; 16]; 396 for lane in lanes.iter_mut() { 397 *lane = fgen.u.int_in_range(0..=31)?; 398 } 399 let lanes = ConstantData::from(lanes.as_ref()); 400 builder.func.dfg.immediates.push(lanes) 401 }; 402 403 // This function is called for any `InstructionFormat::Shuffle`. Which today is just 404 // `shuffle`, but lets assert that, just to be sure we don't accidentally insert 405 // something else. 406 assert_eq!(opcode, Opcode::Shuffle); 407 let res = builder.ins().shuffle(lhs, rhs, mask); 408 409 let target_var = fgen.get_variable_of_type(ctrl_type)?; 410 builder.def_var(target_var, res); 411 412 Ok(()) 413 } 414 415 fn insert_ins_ext_lane( 416 fgen: &mut FunctionGenerator, 417 builder: &mut FunctionBuilder, 418 opcode: Opcode, 419 args: &[Type], 420 rets: &[Type], 421 ) -> Result<()> { 422 let vector_type = *args.first().unwrap(); 423 let ret_type = *rets.first().unwrap(); 424 425 let lhs = builder.use_var(fgen.get_variable_of_type(vector_type)?); 426 let max_lane = (vector_type.lane_count() as u8) - 1; 427 let lane = fgen.u.int_in_range(0..=max_lane)?; 428 429 let res = match opcode { 430 Opcode::Insertlane => { 431 let rhs = builder.use_var(fgen.get_variable_of_type(args[1])?); 432 builder.ins().insertlane(lhs, rhs, lane) 433 } 434 Opcode::Extractlane => builder.ins().extractlane(lhs, lane), 435 _ => todo!(), 436 }; 437 438 let target_var = fgen.get_variable_of_type(ret_type)?; 439 builder.def_var(target_var, res); 440 441 Ok(()) 442 } 443 444 type OpcodeInserter = fn( 445 fgen: &mut FunctionGenerator, 446 builder: &mut FunctionBuilder, 447 Opcode, 448 &[Type], 449 &[Type], 450 ) -> Result<()>; 451 452 macro_rules! exceptions { 453 ($op:expr, $args:expr, $rets:expr, $(($($cases:pat),*)),* $(,)?) => { 454 match ($op, $args, $rets) { 455 $( ($($cases,)* ..) => return false, )* 456 _ => true, 457 } 458 } 459 } 460 461 /// Returns true if we believe this `OpcodeSignature` should compile correctly 462 /// for the given target triple. We currently have a range of known issues 463 /// with specific lowerings on specific backends, and we don't want to get 464 /// fuzz bug reports for those. Over time our goal is to eliminate all of these 465 /// exceptions. 466 fn valid_for_target(triple: &Triple, op: Opcode, args: &[Type], rets: &[Type]) -> bool { 467 // Rule out invalid combinations that we don't yet have a good way of rejecting with the 468 // instruction DSL type constraints. 469 match op { 470 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat => { 471 assert_eq!(args.len(), 1); 472 assert_eq!(rets.len(), 1); 473 474 let arg = args[0]; 475 let ret = rets[0]; 476 477 // Vector arguments must produce vector results, and scalar arguments must produce 478 // scalar results. 479 if arg.is_vector() != ret.is_vector() { 480 return false; 481 } 482 483 if arg.is_vector() && ret.is_vector() { 484 // Vector conversions must have the same number of lanes, and the lanes must be the 485 // same bit-width. 486 if arg.lane_count() != ret.lane_count() { 487 return false; 488 } 489 490 if arg.lane_of().bits() != ret.lane_of().bits() { 491 return false; 492 } 493 } 494 } 495 496 Opcode::Bitcast => { 497 assert_eq!(args.len(), 1); 498 assert_eq!(rets.len(), 1); 499 500 let arg = args[0]; 501 let ret = rets[0]; 502 503 // The opcode generator still allows bitcasts between different sized types, but these 504 // are rejected in the verifier. 505 if arg.bits() != ret.bits() { 506 return false; 507 } 508 } 509 510 // This requires precise runtime integration so it's not supported at 511 // all in fuzzgen just yet. 512 Opcode::StackSwitch => return false, 513 514 _ => {} 515 } 516 517 match triple.architecture { 518 Architecture::X86_64 => { 519 exceptions!( 520 op, 521 args, 522 rets, 523 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 524 (Opcode::Imul, &[I8X16, I8X16]), 525 // https://github.com/bytecodealliance/wasmtime/issues/4756 526 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 527 // https://github.com/bytecodealliance/wasmtime/issues/5474 528 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 529 // https://github.com/bytecodealliance/wasmtime/issues/3370 530 ( 531 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 532 &[I128, I128] 533 ), 534 // https://github.com/bytecodealliance/wasmtime/issues/5107 535 (Opcode::Cls, &[I8], &[I8]), 536 (Opcode::Cls, &[I16], &[I16]), 537 (Opcode::Cls, &[I32], &[I32]), 538 (Opcode::Cls, &[I64], &[I64]), 539 (Opcode::Cls, &[I128], &[I128]), 540 // TODO 541 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 542 // https://github.com/bytecodealliance/wasmtime/issues/4897 543 // https://github.com/bytecodealliance/wasmtime/issues/4899 544 ( 545 Opcode::FcvtToUint 546 | Opcode::FcvtToUintSat 547 | Opcode::FcvtToSint 548 | Opcode::FcvtToSintSat, 549 &[F32 | F64], 550 &[I8 | I16 | I128] 551 ), 552 (Opcode::FcvtToUint | Opcode::FcvtToSint, &[F32X4], &[I32X4]), 553 ( 554 Opcode::FcvtToUint 555 | Opcode::FcvtToUintSat 556 | Opcode::FcvtToSint 557 | Opcode::FcvtToSintSat, 558 &[F64X2], 559 &[I64X2] 560 ), 561 // https://github.com/bytecodealliance/wasmtime/issues/4900 562 (Opcode::FcvtFromUint, &[I128], &[F32 | F64]), 563 // This has a lowering, but only when preceded by `uwiden_low`. 564 (Opcode::FcvtFromUint, &[I64X2], &[F64X2]), 565 // https://github.com/bytecodealliance/wasmtime/issues/4900 566 (Opcode::FcvtFromSint, &[I128], &[F32 | F64]), 567 (Opcode::FcvtFromSint, &[I64X2], &[F64X2]), 568 ( 569 Opcode::Umulhi | Opcode::Smulhi, 570 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 571 ), 572 ( 573 Opcode::UaddSat | Opcode::SaddSat | Opcode::UsubSat | Opcode::SsubSat, 574 &([I32X4, I32X4] | [I64X2, I64X2]) 575 ), 576 (Opcode::Fcopysign, &([F32X4, F32X4] | [F64X2, F64X2])), 577 (Opcode::Popcnt, &([I8X16] | [I16X8] | [I32X4] | [I64X2])), 578 ( 579 Opcode::Umax | Opcode::Smax | Opcode::Umin | Opcode::Smin, 580 &[I64X2, I64X2] 581 ), 582 // https://github.com/bytecodealliance/wasmtime/issues/6104 583 (Opcode::Bitcast, &[I128], &[_]), 584 (Opcode::Bitcast, &[_], &[I128]), 585 (Opcode::Uunarrow), 586 (Opcode::Snarrow | Opcode::Unarrow, &[I64X2, I64X2]), 587 (Opcode::SqmulRoundSat, &[I32X4, I32X4]), 588 // This Icmp is not implemented: #5529 589 (Opcode::Icmp, &[I64X2, I64X2]), 590 // IaddPairwise is implemented, but only for some types, and with some preceding ops. 591 (Opcode::IaddPairwise), 592 // Nothing wrong with this select. But we have an isle rule that can optimize it 593 // into a `min`/`max` instructions, which we don't have implemented yet. 594 (Opcode::Select, &[_, I128, I128]), 595 // These stack accesses can cause segfaults if they are merged into an SSE instruction. 596 // See: #5922 597 ( 598 Opcode::StackStore, 599 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 600 ), 601 ( 602 Opcode::StackLoad, 603 &[], 604 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 605 ), 606 // TODO 607 ( 608 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 609 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 610 ), 611 ( 612 Opcode::Rotr | Opcode::Rotl, 613 &[I8X16 | I16X8 | I32X4 | I64X2, _] 614 ), 615 ) 616 } 617 618 Architecture::Aarch64(_) => { 619 exceptions!( 620 op, 621 args, 622 rets, 623 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 624 // https://github.com/bytecodealliance/wasmtime/issues/4864 625 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 626 // https://github.com/bytecodealliance/wasmtime/issues/5472 627 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 628 // https://github.com/bytecodealliance/wasmtime/issues/4313 629 ( 630 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 631 &[I128, I128] 632 ), 633 // https://github.com/bytecodealliance/wasmtime/issues/4870 634 (Opcode::Bnot, &[F32 | F64]), 635 ( 636 Opcode::Band 637 | Opcode::Bor 638 | Opcode::Bxor 639 | Opcode::BandNot 640 | Opcode::BorNot 641 | Opcode::BxorNot, 642 &([F32, F32] | [F64, F64]) 643 ), 644 // https://github.com/bytecodealliance/wasmtime/issues/5198 645 (Opcode::Bitselect, &[I128, I128, I128]), 646 // https://github.com/bytecodealliance/wasmtime/issues/4934 647 ( 648 Opcode::FcvtToUint 649 | Opcode::FcvtToUintSat 650 | Opcode::FcvtToSint 651 | Opcode::FcvtToSintSat, 652 &[F32 | F64], 653 &[I128] 654 ), 655 // https://github.com/bytecodealliance/wasmtime/issues/4933 656 ( 657 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 658 &[I128], 659 &[F32 | F64] 660 ), 661 ( 662 Opcode::Umulhi | Opcode::Smulhi, 663 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 664 ), 665 (Opcode::Popcnt, &[I16X8 | I32X4 | I64X2]), 666 // Nothing wrong with this select. But we have an isle rule that can optimize it 667 // into a `min`/`max` instructions, which we don't have implemented yet. 668 (Opcode::Select, &[I8, I128, I128]), 669 // https://github.com/bytecodealliance/wasmtime/issues/6104 670 (Opcode::Bitcast, &[I128], &[_]), 671 (Opcode::Bitcast, &[_], &[I128]), 672 // TODO 673 ( 674 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 675 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 676 ), 677 ( 678 Opcode::Rotr | Opcode::Rotl, 679 &[I8X16 | I16X8 | I32X4 | I64X2, _] 680 ), 681 // TODO 682 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 683 (Opcode::VhighBits, &[F32X4 | F64X2]), 684 ) 685 } 686 687 Architecture::S390x => { 688 exceptions!( 689 op, 690 args, 691 rets, 692 (Opcode::UaddOverflow | Opcode::SaddOverflow), 693 (Opcode::UsubOverflow | Opcode::SsubOverflow), 694 (Opcode::UmulOverflow | Opcode::SmulOverflow), 695 ( 696 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 697 &[I128, I128] 698 ), 699 (Opcode::Bnot, &[F32 | F64]), 700 ( 701 Opcode::Band 702 | Opcode::Bor 703 | Opcode::Bxor 704 | Opcode::BandNot 705 | Opcode::BorNot 706 | Opcode::BxorNot, 707 &([F32, F32] | [F64, F64]) 708 ), 709 ( 710 Opcode::FcvtToUint 711 | Opcode::FcvtToUintSat 712 | Opcode::FcvtToSint 713 | Opcode::FcvtToSintSat, 714 &[F32 | F64], 715 &[I128] 716 ), 717 ( 718 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 719 &[I128], 720 &[F32 | F64] 721 ), 722 (Opcode::SsubSat | Opcode::SaddSat, &[I64X2, I64X2]), 723 // https://github.com/bytecodealliance/wasmtime/issues/6104 724 (Opcode::Bitcast, &[I128], &[_]), 725 (Opcode::Bitcast, &[_], &[I128]), 726 // TODO 727 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 728 ) 729 } 730 731 Architecture::Riscv64(_) => { 732 exceptions!( 733 op, 734 args, 735 rets, 736 // TODO 737 (Opcode::UaddOverflow | Opcode::SaddOverflow), 738 (Opcode::UsubOverflow | Opcode::SsubOverflow), 739 (Opcode::UmulOverflow | Opcode::SmulOverflow), 740 // TODO 741 ( 742 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 743 &[I128, I128] 744 ), 745 // TODO 746 (Opcode::Iabs, &[I128]), 747 // TODO 748 (Opcode::Bitselect, &[I128, I128, I128]), 749 // https://github.com/bytecodealliance/wasmtime/issues/5528 750 ( 751 Opcode::FcvtToUint | Opcode::FcvtToSint, 752 [F32 | F64], 753 &[I128] 754 ), 755 ( 756 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat, 757 &[F32 | F64], 758 &[I128] 759 ), 760 // https://github.com/bytecodealliance/wasmtime/issues/5528 761 ( 762 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 763 &[I128], 764 &[F32 | F64] 765 ), 766 // TODO 767 ( 768 Opcode::SelectSpectreGuard, 769 &[_, _, _], 770 &[F32 | F64 | I8X16 | I16X8 | I32X4 | I64X2 | F64X2 | F32X4] 771 ), 772 // TODO 773 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 774 ( 775 Opcode::Rotr | Opcode::Rotl, 776 &[I8X16 | I16X8 | I32X4 | I64X2, _] 777 ), 778 ) 779 } 780 781 _ => true, 782 } 783 } 784 785 type OpcodeSignature = (Opcode, Vec<Type>, Vec<Type>); 786 787 static OPCODE_SIGNATURES: LazyLock<Vec<OpcodeSignature>> = LazyLock::new(|| { 788 let types = &[ 789 I8, I16, I32, I64, I128, // Scalar Integers 790 F32, F64, // Scalar Floats 791 I8X16, I16X8, I32X4, I64X2, // SIMD Integers 792 F32X4, F64X2, // SIMD Floats 793 ]; 794 795 // When this env variable is passed, we only generate instructions for the opcodes listed in 796 // the comma-separated list. This is useful for debugging, as it allows us to focus on a few 797 // specific opcodes. 798 let allowed_opcodes = std::env::var("FUZZGEN_ALLOWED_OPS").ok().map(|s| { 799 s.split(',') 800 .map(|s| s.trim()) 801 .filter(|s| !s.is_empty()) 802 .map(|s| Opcode::from_str(s).expect("Unrecoginzed opcode")) 803 .collect::<Vec<_>>() 804 }); 805 806 Opcode::all() 807 .iter() 808 .filter(|op| { 809 match op { 810 // Control flow opcodes should not be generated through `generate_instructions`. 811 Opcode::BrTable 812 | Opcode::Brif 813 | Opcode::Jump 814 | Opcode::Return 815 | Opcode::ReturnCall 816 | Opcode::ReturnCallIndirect 817 | Opcode::TryCall 818 | Opcode::TryCallIndirect => false, 819 820 // Constants are generated outside of `generate_instructions` 821 Opcode::Iconst => false, 822 823 // TODO: extract_vector raises exceptions during return type generation because it 824 // uses dynamic vectors. 825 Opcode::ExtractVector => false, 826 827 _ => true, 828 } 829 }) 830 .flat_map(|op| { 831 let constraints = op.constraints(); 832 833 let ctrl_types = if let Some(ctrls) = constraints.ctrl_typeset() { 834 Vec::from_iter(types.iter().copied().filter(|ty| ctrls.contains(*ty))) 835 } else { 836 vec![INVALID] 837 }; 838 839 ctrl_types.into_iter().flat_map(move |ctrl_type| { 840 let rets = Vec::from_iter( 841 (0..constraints.num_fixed_results()) 842 .map(|i| constraints.result_type(i, ctrl_type)), 843 ); 844 845 // Cols is a vector whose length will match `num_fixed_value_arguments`, and whose 846 // elements will be vectors of types that are valid for that fixed argument 847 // position. 848 let mut cols = vec![]; 849 850 for i in 0..constraints.num_fixed_value_arguments() { 851 match constraints.value_argument_constraint(i, ctrl_type) { 852 ResolvedConstraint::Bound(ty) => cols.push(Vec::from([ty])), 853 ResolvedConstraint::Free(tys) => cols.push(Vec::from_iter( 854 types.iter().copied().filter(|ty| tys.contains(*ty)), 855 )), 856 } 857 } 858 859 // Generate the cartesian product of cols to produce a vector of argument lists, 860 // argss. The argss vector is seeded with the empty argument list, so there's an 861 // initial value to be extended in the loop below. 862 let mut argss = vec![vec![]]; 863 let mut cols = cols.as_slice(); 864 while let Some((col, rest)) = cols.split_last() { 865 cols = rest; 866 867 let mut next = vec![]; 868 for current in argss.iter() { 869 // Extend the front of each argument candidate with every type in `col`. 870 for ty in col { 871 let mut args = vec![*ty]; 872 args.extend_from_slice(¤t); 873 next.push(args); 874 } 875 } 876 877 let _ = std::mem::replace(&mut argss, next); 878 } 879 880 argss.into_iter().map(move |args| (*op, args, rets.clone())) 881 }) 882 }) 883 .filter(|(op, args, rets)| { 884 // These op/signature combinations need to be vetted 885 exceptions!( 886 op, 887 args.as_slice(), 888 rets.as_slice(), 889 (Opcode::Debugtrap), 890 (Opcode::Trap), 891 (Opcode::Trapz), 892 (Opcode::Trapnz), 893 (Opcode::CallIndirect, &[I32]), 894 (Opcode::FuncAddr), 895 (Opcode::X86Pshufb), 896 (Opcode::AvgRound), 897 (Opcode::Uload8x8), 898 (Opcode::Sload8x8), 899 (Opcode::Uload16x4), 900 (Opcode::Sload16x4), 901 (Opcode::Uload32x2), 902 (Opcode::Sload32x2), 903 (Opcode::StackAddr), 904 (Opcode::DynamicStackLoad), 905 (Opcode::DynamicStackStore), 906 (Opcode::DynamicStackAddr), 907 (Opcode::GlobalValue), 908 (Opcode::SymbolValue), 909 (Opcode::TlsValue), 910 (Opcode::GetPinnedReg), 911 (Opcode::SetPinnedReg), 912 (Opcode::GetFramePointer), 913 (Opcode::GetStackPointer), 914 (Opcode::GetReturnAddress), 915 (Opcode::X86Blendv), 916 (Opcode::IcmpImm), 917 (Opcode::X86Pmulhrsw), 918 (Opcode::IaddImm), 919 (Opcode::ImulImm), 920 (Opcode::UdivImm), 921 (Opcode::SdivImm), 922 (Opcode::UremImm), 923 (Opcode::SremImm), 924 (Opcode::IrsubImm), 925 (Opcode::UaddOverflowCin), 926 (Opcode::SaddOverflowCin), 927 (Opcode::UaddOverflowTrap), 928 (Opcode::UsubOverflowBin), 929 (Opcode::SsubOverflowBin), 930 (Opcode::BandImm), 931 (Opcode::BorImm), 932 (Opcode::BxorImm), 933 (Opcode::RotlImm), 934 (Opcode::RotrImm), 935 (Opcode::IshlImm), 936 (Opcode::UshrImm), 937 (Opcode::SshrImm), 938 (Opcode::ScalarToVector), 939 (Opcode::X86Pmaddubsw), 940 (Opcode::X86Cvtt2dq), 941 (Opcode::Umulhi, &[I128, I128], &[I128]), 942 (Opcode::Smulhi, &[I128, I128], &[I128]), 943 // https://github.com/bytecodealliance/wasmtime/issues/6073 944 (Opcode::Iconcat, &[I32, I32], &[I64]), 945 (Opcode::Iconcat, &[I16, I16], &[I32]), 946 (Opcode::Iconcat, &[I8, I8], &[I16]), 947 // https://github.com/bytecodealliance/wasmtime/issues/6073 948 (Opcode::Isplit, &[I64], &[I32, I32]), 949 (Opcode::Isplit, &[I32], &[I16, I16]), 950 (Opcode::Isplit, &[I16], &[I8, I8]), 951 (Opcode::Fmin, &[F32X4, F32X4], &[F32X4]), 952 (Opcode::Fmin, &[F64X2, F64X2], &[F64X2]), 953 (Opcode::Fmax, &[F32X4, F32X4], &[F32X4]), 954 (Opcode::Fmax, &[F64X2, F64X2], &[F64X2]), 955 (Opcode::FcvtToUintSat, &[F32X4], &[I8]), 956 (Opcode::FcvtToUintSat, &[F64X2], &[I8]), 957 (Opcode::FcvtToUintSat, &[F32X4], &[I16]), 958 (Opcode::FcvtToUintSat, &[F64X2], &[I16]), 959 (Opcode::FcvtToUintSat, &[F32X4], &[I32]), 960 (Opcode::FcvtToUintSat, &[F64X2], &[I32]), 961 (Opcode::FcvtToUintSat, &[F32X4], &[I64]), 962 (Opcode::FcvtToUintSat, &[F64X2], &[I64]), 963 (Opcode::FcvtToUintSat, &[F32X4], &[I128]), 964 (Opcode::FcvtToUintSat, &[F64X2], &[I128]), 965 (Opcode::FcvtToUintSat, &[F32], &[I8X16]), 966 (Opcode::FcvtToUintSat, &[F64], &[I8X16]), 967 (Opcode::FcvtToUintSat, &[F32X4], &[I8X16]), 968 (Opcode::FcvtToUintSat, &[F64X2], &[I8X16]), 969 (Opcode::FcvtToUintSat, &[F32], &[I16X8]), 970 (Opcode::FcvtToUintSat, &[F64], &[I16X8]), 971 (Opcode::FcvtToUintSat, &[F32X4], &[I16X8]), 972 (Opcode::FcvtToUintSat, &[F64X2], &[I16X8]), 973 (Opcode::FcvtToUintSat, &[F32], &[I32X4]), 974 (Opcode::FcvtToUintSat, &[F64], &[I32X4]), 975 (Opcode::FcvtToUintSat, &[F64X2], &[I32X4]), 976 (Opcode::FcvtToUintSat, &[F32], &[I64X2]), 977 (Opcode::FcvtToUintSat, &[F64], &[I64X2]), 978 (Opcode::FcvtToUintSat, &[F32X4], &[I64X2]), 979 (Opcode::FcvtToSintSat, &[F32X4], &[I8]), 980 (Opcode::FcvtToSintSat, &[F64X2], &[I8]), 981 (Opcode::FcvtToSintSat, &[F32X4], &[I16]), 982 (Opcode::FcvtToSintSat, &[F64X2], &[I16]), 983 (Opcode::FcvtToSintSat, &[F32X4], &[I32]), 984 (Opcode::FcvtToSintSat, &[F64X2], &[I32]), 985 (Opcode::FcvtToSintSat, &[F32X4], &[I64]), 986 (Opcode::FcvtToSintSat, &[F64X2], &[I64]), 987 (Opcode::FcvtToSintSat, &[F32X4], &[I128]), 988 (Opcode::FcvtToSintSat, &[F64X2], &[I128]), 989 (Opcode::FcvtToSintSat, &[F32], &[I8X16]), 990 (Opcode::FcvtToSintSat, &[F64], &[I8X16]), 991 (Opcode::FcvtToSintSat, &[F32X4], &[I8X16]), 992 (Opcode::FcvtToSintSat, &[F64X2], &[I8X16]), 993 (Opcode::FcvtToSintSat, &[F32], &[I16X8]), 994 (Opcode::FcvtToSintSat, &[F64], &[I16X8]), 995 (Opcode::FcvtToSintSat, &[F32X4], &[I16X8]), 996 (Opcode::FcvtToSintSat, &[F64X2], &[I16X8]), 997 (Opcode::FcvtToSintSat, &[F32], &[I32X4]), 998 (Opcode::FcvtToSintSat, &[F64], &[I32X4]), 999 (Opcode::FcvtToSintSat, &[F64X2], &[I32X4]), 1000 (Opcode::FcvtToSintSat, &[F32], &[I64X2]), 1001 (Opcode::FcvtToSintSat, &[F64], &[I64X2]), 1002 (Opcode::FcvtToSintSat, &[F32X4], &[I64X2]), 1003 (Opcode::FcvtFromUint, &[I8X16], &[F32]), 1004 (Opcode::FcvtFromUint, &[I16X8], &[F32]), 1005 (Opcode::FcvtFromUint, &[I32X4], &[F32]), 1006 (Opcode::FcvtFromUint, &[I64X2], &[F32]), 1007 (Opcode::FcvtFromUint, &[I8X16], &[F64]), 1008 (Opcode::FcvtFromUint, &[I16X8], &[F64]), 1009 (Opcode::FcvtFromUint, &[I32X4], &[F64]), 1010 (Opcode::FcvtFromUint, &[I64X2], &[F64]), 1011 (Opcode::FcvtFromUint, &[I8], &[F32X4]), 1012 (Opcode::FcvtFromUint, &[I16], &[F32X4]), 1013 (Opcode::FcvtFromUint, &[I32], &[F32X4]), 1014 (Opcode::FcvtFromUint, &[I64], &[F32X4]), 1015 (Opcode::FcvtFromUint, &[I128], &[F32X4]), 1016 (Opcode::FcvtFromUint, &[I8X16], &[F32X4]), 1017 (Opcode::FcvtFromUint, &[I16X8], &[F32X4]), 1018 (Opcode::FcvtFromUint, &[I64X2], &[F32X4]), 1019 (Opcode::FcvtFromUint, &[I8], &[F64X2]), 1020 (Opcode::FcvtFromUint, &[I16], &[F64X2]), 1021 (Opcode::FcvtFromUint, &[I32], &[F64X2]), 1022 (Opcode::FcvtFromUint, &[I64], &[F64X2]), 1023 (Opcode::FcvtFromUint, &[I128], &[F64X2]), 1024 (Opcode::FcvtFromUint, &[I8X16], &[F64X2]), 1025 (Opcode::FcvtFromUint, &[I16X8], &[F64X2]), 1026 (Opcode::FcvtFromUint, &[I32X4], &[F64X2]), 1027 (Opcode::FcvtFromSint, &[I8X16], &[F32]), 1028 (Opcode::FcvtFromSint, &[I16X8], &[F32]), 1029 (Opcode::FcvtFromSint, &[I32X4], &[F32]), 1030 (Opcode::FcvtFromSint, &[I64X2], &[F32]), 1031 (Opcode::FcvtFromSint, &[I8X16], &[F64]), 1032 (Opcode::FcvtFromSint, &[I16X8], &[F64]), 1033 (Opcode::FcvtFromSint, &[I32X4], &[F64]), 1034 (Opcode::FcvtFromSint, &[I64X2], &[F64]), 1035 (Opcode::FcvtFromSint, &[I8], &[F32X4]), 1036 (Opcode::FcvtFromSint, &[I16], &[F32X4]), 1037 (Opcode::FcvtFromSint, &[I32], &[F32X4]), 1038 (Opcode::FcvtFromSint, &[I64], &[F32X4]), 1039 (Opcode::FcvtFromSint, &[I128], &[F32X4]), 1040 (Opcode::FcvtFromSint, &[I8X16], &[F32X4]), 1041 (Opcode::FcvtFromSint, &[I16X8], &[F32X4]), 1042 (Opcode::FcvtFromSint, &[I64X2], &[F32X4]), 1043 (Opcode::FcvtFromSint, &[I8], &[F64X2]), 1044 (Opcode::FcvtFromSint, &[I16], &[F64X2]), 1045 (Opcode::FcvtFromSint, &[I32], &[F64X2]), 1046 (Opcode::FcvtFromSint, &[I64], &[F64X2]), 1047 (Opcode::FcvtFromSint, &[I128], &[F64X2]), 1048 (Opcode::FcvtFromSint, &[I8X16], &[F64X2]), 1049 (Opcode::FcvtFromSint, &[I16X8], &[F64X2]), 1050 (Opcode::FcvtFromSint, &[I32X4], &[F64X2]), 1051 // Only supported on x64 with a feature at this time, so 128-bit 1052 // atomics are not suitable to fuzz yet. 1053 (Opcode::AtomicRmw, _, &[I128]), 1054 (Opcode::AtomicCas, _, &[I128]), 1055 (Opcode::AtomicLoad, _, &[I128]), 1056 (Opcode::AtomicStore, &[I128, _], _), 1057 (Opcode::SequencePoint), 1058 ) 1059 }) 1060 .filter(|(op, ..)| { 1061 allowed_opcodes 1062 .as_ref() 1063 .map_or(true, |opcodes| opcodes.contains(op)) 1064 }) 1065 .collect() 1066 }); 1067 1068 fn inserter_for_format(fmt: InstructionFormat) -> OpcodeInserter { 1069 match fmt { 1070 InstructionFormat::AtomicCas => insert_atomic_cas, 1071 InstructionFormat::AtomicRmw => insert_atomic_rmw, 1072 InstructionFormat::Binary => insert_opcode, 1073 InstructionFormat::BinaryImm64 => todo!(), 1074 InstructionFormat::BinaryImm8 => insert_ins_ext_lane, 1075 InstructionFormat::Call => insert_call, 1076 InstructionFormat::CallIndirect => insert_call, 1077 InstructionFormat::CondTrap => todo!(), 1078 InstructionFormat::DynamicStackLoad => todo!(), 1079 InstructionFormat::DynamicStackStore => todo!(), 1080 InstructionFormat::FloatCompare => insert_cmp, 1081 InstructionFormat::FuncAddr => todo!(), 1082 InstructionFormat::IntAddTrap => todo!(), 1083 InstructionFormat::IntCompare => insert_cmp, 1084 InstructionFormat::IntCompareImm => todo!(), 1085 InstructionFormat::Load => insert_load_store, 1086 InstructionFormat::LoadNoOffset => insert_load_store, 1087 InstructionFormat::NullAry => insert_opcode, 1088 InstructionFormat::Shuffle => insert_shuffle, 1089 InstructionFormat::StackLoad => insert_stack_load, 1090 InstructionFormat::StackStore => insert_stack_store, 1091 InstructionFormat::Store => insert_load_store, 1092 InstructionFormat::StoreNoOffset => insert_load_store, 1093 InstructionFormat::Ternary => insert_opcode, 1094 InstructionFormat::TernaryImm8 => insert_ins_ext_lane, 1095 InstructionFormat::Trap => todo!(), 1096 InstructionFormat::Unary => insert_opcode, 1097 InstructionFormat::UnaryConst => insert_const, 1098 InstructionFormat::UnaryGlobalValue => todo!(), 1099 InstructionFormat::UnaryIeee16 => insert_const, 1100 InstructionFormat::UnaryIeee32 => insert_const, 1101 InstructionFormat::UnaryIeee64 => insert_const, 1102 InstructionFormat::UnaryImm => insert_const, 1103 InstructionFormat::ExceptionHandlerAddress => insert_const, 1104 1105 InstructionFormat::BranchTable 1106 | InstructionFormat::Brif 1107 | InstructionFormat::Jump 1108 | InstructionFormat::MultiAry 1109 | InstructionFormat::TryCall 1110 | InstructionFormat::TryCallIndirect => { 1111 panic!("Control-flow instructions should be handled by 'insert_terminator': {fmt:?}") 1112 } 1113 } 1114 } 1115 1116 pub struct FunctionGenerator<'r, 'data> 1117 where 1118 'data: 'r, 1119 { 1120 u: &'r mut Unstructured<'data>, 1121 config: &'r Config, 1122 resources: Resources, 1123 isa: OwnedTargetIsa, 1124 name: UserFuncName, 1125 signature: Signature, 1126 } 1127 1128 #[derive(Debug, Clone)] 1129 enum BlockTerminator { 1130 Return, 1131 Jump(Block), 1132 Br(Block, Block), 1133 BrTable(Block, Vec<Block>), 1134 Switch(Type, Block, HashMap<u128, Block>), 1135 TailCall(FuncRef), 1136 TailCallIndirect(FuncRef), 1137 } 1138 1139 #[derive(Debug, Clone)] 1140 enum BlockTerminatorKind { 1141 Return, 1142 Jump, 1143 Br, 1144 BrTable, 1145 Switch, 1146 TailCall, 1147 TailCallIndirect, 1148 } 1149 1150 /// Alias Analysis Category 1151 /// 1152 /// Our alias analysis pass supports 4 categories of accesses to distinguish 1153 /// different regions. The "Other" region is the general case, and is the default 1154 /// Although they have highly suggestive names there is no difference between any 1155 /// of the categories. 1156 /// 1157 /// We assign each stack slot a category when we first generate them, and then 1158 /// ensure that all accesses to that stack slot are correctly tagged. We already 1159 /// ensure that memory accesses never cross stack slots, so there is no risk 1160 /// of a memory access being tagged with the wrong category. 1161 #[derive(Debug, PartialEq, Clone, Copy)] 1162 enum AACategory { 1163 Other, 1164 Heap, 1165 Table, 1166 VmCtx, 1167 } 1168 1169 impl AACategory { 1170 pub fn all() -> &'static [Self] { 1171 &[ 1172 AACategory::Other, 1173 AACategory::Heap, 1174 AACategory::Table, 1175 AACategory::VmCtx, 1176 ] 1177 } 1178 1179 pub fn update_memflags(&self, flags: &mut MemFlags) { 1180 flags.set_alias_region(match self { 1181 AACategory::Other => None, 1182 AACategory::Heap => Some(AliasRegion::Heap), 1183 AACategory::Table => Some(AliasRegion::Table), 1184 AACategory::VmCtx => Some(AliasRegion::Vmctx), 1185 }) 1186 } 1187 } 1188 1189 pub type StackAlignment = StackSize; 1190 1191 #[derive(Default)] 1192 struct Resources { 1193 vars: HashMap<Type, Vec<Variable>>, 1194 blocks: Vec<(Block, BlockSignature)>, 1195 blocks_without_params: Vec<Block>, 1196 block_terminators: Vec<BlockTerminator>, 1197 func_refs: Vec<(Signature, SigRef, FuncRef)>, 1198 /// This field is required to be sorted by stack slot size at all times. 1199 /// We use this invariant when searching for stack slots with a given size. 1200 /// See [FunctionGenerator::stack_slot_with_size] 1201 stack_slots: Vec<(StackSlot, StackSize, StackAlignment, AACategory)>, 1202 usercalls: Vec<(UserExternalName, Signature)>, 1203 libcalls: Vec<LibCall>, 1204 } 1205 1206 impl Resources { 1207 /// Partitions blocks at `block`. Only blocks that can be targeted by branches are considered. 1208 /// 1209 /// The first slice includes all blocks up to and including `block`. 1210 /// The second slice includes all remaining blocks. 1211 fn partition_target_blocks( 1212 &self, 1213 block: Block, 1214 ) -> (&[(Block, BlockSignature)], &[(Block, BlockSignature)]) { 1215 // Blocks are stored in-order and have no gaps, this means that we can simply index them by 1216 // their number. We also need to exclude the entry block since it isn't a valid target. 1217 let target_blocks = &self.blocks[1..]; 1218 target_blocks.split_at(block.as_u32() as usize) 1219 } 1220 1221 /// Returns blocks forward of `block`. Only blocks that can be targeted by branches are considered. 1222 fn forward_blocks(&self, block: Block) -> &[(Block, BlockSignature)] { 1223 let (_, forward_blocks) = self.partition_target_blocks(block); 1224 forward_blocks 1225 } 1226 1227 /// Generates a slice of `blocks_without_params` ahead of `block` 1228 fn forward_blocks_without_params(&self, block: Block) -> &[Block] { 1229 let partition_point = self.blocks_without_params.partition_point(|b| *b <= block); 1230 &self.blocks_without_params[partition_point..] 1231 } 1232 1233 /// Generates an iterator of all valid tail call targets. This includes all functions with both 1234 /// the `tail` calling convention and the same return values as the caller. 1235 fn tail_call_targets<'a>( 1236 &'a self, 1237 caller_sig: &'a Signature, 1238 ) -> impl Iterator<Item = &'a (Signature, SigRef, FuncRef)> { 1239 self.func_refs.iter().filter(|(sig, _, _)| { 1240 sig.call_conv == CallConv::Tail && sig.returns == caller_sig.returns 1241 }) 1242 } 1243 } 1244 1245 impl<'r, 'data> FunctionGenerator<'r, 'data> 1246 where 1247 'data: 'r, 1248 { 1249 pub fn new( 1250 u: &'r mut Unstructured<'data>, 1251 config: &'r Config, 1252 isa: OwnedTargetIsa, 1253 name: UserFuncName, 1254 signature: Signature, 1255 usercalls: Vec<(UserExternalName, Signature)>, 1256 libcalls: Vec<LibCall>, 1257 ) -> Self { 1258 Self { 1259 u, 1260 config, 1261 resources: Resources { 1262 usercalls, 1263 libcalls, 1264 ..Resources::default() 1265 }, 1266 isa, 1267 name, 1268 signature, 1269 } 1270 } 1271 1272 /// Generates a random value for config `param` 1273 fn param(&mut self, param: &RangeInclusive<usize>) -> Result<usize> { 1274 Ok(self.u.int_in_range(param.clone())?) 1275 } 1276 1277 fn system_callconv(&mut self) -> CallConv { 1278 // TODO: This currently only runs on linux, so this is the only choice 1279 // We should improve this once we generate flags and targets 1280 CallConv::SystemV 1281 } 1282 1283 /// Finds a stack slot with size of at least n bytes 1284 fn stack_slot_with_size( 1285 &mut self, 1286 n: u32, 1287 ) -> Result<(StackSlot, StackSize, StackAlignment, AACategory)> { 1288 let first = self 1289 .resources 1290 .stack_slots 1291 .partition_point(|&(_slot, size, _align, _category)| size < n); 1292 Ok(*self.u.choose(&self.resources.stack_slots[first..])?) 1293 } 1294 1295 /// Generates an address that should allow for a store or a load. 1296 /// 1297 /// Addresses aren't generated like other values. They are never stored in variables so that 1298 /// we don't run the risk of returning them from a function, which would make the fuzzer 1299 /// complain since they are different from the interpreter to the backend. 1300 /// 1301 /// `min_size`: Controls the amount of space that the address should have. 1302 /// 1303 /// `aligned`: When passed as true, the resulting address is guaranteed to be aligned 1304 /// on an 8 byte boundary. 1305 /// 1306 /// Returns a valid address and the maximum possible offset that still respects `min_size`. 1307 fn generate_load_store_address( 1308 &mut self, 1309 builder: &mut FunctionBuilder, 1310 min_size: u32, 1311 aligned: bool, 1312 ) -> Result<(Value, u32, AACategory)> { 1313 // TODO: Currently our only source of addresses is stack_addr, but we 1314 // should add global_value, symbol_value eventually 1315 let (addr, available_size, category) = { 1316 let (ss, slot_size, _align, category) = self.stack_slot_with_size(min_size)?; 1317 1318 // stack_slot_with_size guarantees that slot_size >= min_size 1319 let max_offset = slot_size - min_size; 1320 let offset = if aligned { 1321 self.u.int_in_range(0..=max_offset / min_size)? * min_size 1322 } else { 1323 self.u.int_in_range(0..=max_offset)? 1324 }; 1325 1326 let base_addr = builder.ins().stack_addr(I64, ss, offset as i32); 1327 let available_size = slot_size.saturating_sub(offset); 1328 (base_addr, available_size, category) 1329 }; 1330 1331 // TODO: Insert a bunch of amode opcodes here to modify the address! 1332 1333 // Now that we have an address and a size, we just choose a random offset to return to the 1334 // caller. Preserving min_size bytes. 1335 let max_offset = available_size.saturating_sub(min_size); 1336 Ok((addr, max_offset, category)) 1337 } 1338 1339 // Generates an address and memflags for a load or store. 1340 fn generate_address_and_memflags( 1341 &mut self, 1342 builder: &mut FunctionBuilder, 1343 min_size: u32, 1344 is_atomic: bool, 1345 ) -> Result<(Value, MemFlags, Offset32)> { 1346 // Should we generate an aligned address 1347 // Some backends have issues with unaligned atomics. 1348 // AArch64: https://github.com/bytecodealliance/wasmtime/issues/5483 1349 // RISCV: https://github.com/bytecodealliance/wasmtime/issues/5882 1350 let requires_aligned_atomics = matches!( 1351 self.isa.triple().architecture, 1352 Architecture::Aarch64(_) | Architecture::Riscv64(_) 1353 ); 1354 let aligned = if is_atomic && requires_aligned_atomics { 1355 true 1356 } else if min_size > 8 { 1357 // TODO: We currently can't guarantee that a stack_slot will be aligned on a 16 byte 1358 // boundary. We don't have a way to specify alignment when creating stack slots, and 1359 // cranelift only guarantees 8 byte alignment between stack slots. 1360 // See: https://github.com/bytecodealliance/wasmtime/issues/5922#issuecomment-1457926624 1361 false 1362 } else { 1363 bool::arbitrary(self.u)? 1364 }; 1365 1366 let mut flags = MemFlags::new(); 1367 // Even if we picked an aligned address, we can always generate unaligned memflags 1368 if aligned && bool::arbitrary(self.u)? { 1369 flags.set_aligned(); 1370 } 1371 // If the address is aligned, then we know it won't trap 1372 if aligned && bool::arbitrary(self.u)? { 1373 flags.set_notrap(); 1374 } 1375 1376 let (address, max_offset, category) = 1377 self.generate_load_store_address(builder, min_size, aligned)?; 1378 1379 // Set the Alias Analysis bits on the memflags 1380 category.update_memflags(&mut flags); 1381 1382 // Pick an offset to pass into the load/store. 1383 let offset = if aligned { 1384 0 1385 } else { 1386 self.u.int_in_range(0..=max_offset)? as i32 1387 } 1388 .into(); 1389 1390 Ok((address, flags, offset)) 1391 } 1392 1393 /// Get a variable of type `ty` from the current function 1394 fn get_variable_of_type(&mut self, ty: Type) -> Result<Variable> { 1395 let opts = self.resources.vars.get(&ty).map_or(&[][..], Vec::as_slice); 1396 let var = self.u.choose(opts)?; 1397 Ok(*var) 1398 } 1399 1400 /// Generates an instruction(`iconst`/`fconst`/etc...) to introduce a constant value 1401 fn generate_const(&mut self, builder: &mut FunctionBuilder, ty: Type) -> Result<Value> { 1402 Ok(match self.u.datavalue(ty)? { 1403 DataValue::I8(i) => builder.ins().iconst(ty, i as u8 as i64), 1404 DataValue::I16(i) => builder.ins().iconst(ty, i as u16 as i64), 1405 DataValue::I32(i) => builder.ins().iconst(ty, i as u32 as i64), 1406 DataValue::I64(i) => builder.ins().iconst(ty, i), 1407 DataValue::I128(i) => { 1408 let hi = builder.ins().iconst(I64, (i >> 64) as i64); 1409 let lo = builder.ins().iconst(I64, i as i64); 1410 builder.ins().iconcat(lo, hi) 1411 } 1412 DataValue::F16(f) => builder.ins().f16const(f), 1413 DataValue::F32(f) => builder.ins().f32const(f), 1414 DataValue::F64(f) => builder.ins().f64const(f), 1415 DataValue::F128(f) => { 1416 let handle = builder.func.dfg.constants.insert(f.into()); 1417 builder.ins().f128const(handle) 1418 } 1419 DataValue::V128(bytes) => { 1420 let data = bytes.to_vec().into(); 1421 let handle = builder.func.dfg.constants.insert(data); 1422 builder.ins().vconst(ty, handle) 1423 } 1424 _ => unimplemented!(), 1425 }) 1426 } 1427 1428 /// Chooses a random block which can be targeted by a jump / branch. 1429 /// This means any block that is not the first block. 1430 fn generate_target_block(&mut self, source_block: Block) -> Result<Block> { 1431 // We try to mostly generate forward branches to avoid generating an excessive amount of 1432 // infinite loops. But they are still important, so give them a small chance of existing. 1433 let (backwards_blocks, forward_blocks) = 1434 self.resources.partition_target_blocks(source_block); 1435 let ratio = self.config.backwards_branch_ratio; 1436 let block_targets = if !backwards_blocks.is_empty() && self.u.ratio(ratio.0, ratio.1)? { 1437 backwards_blocks 1438 } else { 1439 forward_blocks 1440 }; 1441 assert!(!block_targets.is_empty()); 1442 1443 let (block, _) = self.u.choose(block_targets)?.clone(); 1444 Ok(block) 1445 } 1446 1447 fn generate_values_for_block( 1448 &mut self, 1449 builder: &mut FunctionBuilder, 1450 block: Block, 1451 ) -> Result<Vec<BlockArg>> { 1452 let (_, sig) = self.resources.blocks[block.as_u32() as usize].clone(); 1453 Ok(self 1454 .generate_values_for_signature(builder, sig.iter().copied())? 1455 .into_iter() 1456 .map(|val| BlockArg::Value(val)) 1457 .collect::<Vec<_>>()) 1458 } 1459 1460 fn generate_values_for_signature<I: Iterator<Item = Type>>( 1461 &mut self, 1462 builder: &mut FunctionBuilder, 1463 signature: I, 1464 ) -> Result<Vec<Value>> { 1465 signature 1466 .map(|ty| { 1467 let var = self.get_variable_of_type(ty)?; 1468 let val = builder.use_var(var); 1469 Ok(val) 1470 }) 1471 .collect() 1472 } 1473 1474 /// The terminator that we need to insert has already been picked ahead of time 1475 /// we just need to build the instructions for it 1476 fn insert_terminator( 1477 &mut self, 1478 builder: &mut FunctionBuilder, 1479 source_block: Block, 1480 ) -> Result<()> { 1481 let terminator = self.resources.block_terminators[source_block.as_u32() as usize].clone(); 1482 1483 match terminator { 1484 BlockTerminator::Return => { 1485 let types: Vec<Type> = { 1486 let rets = &builder.func.signature.returns; 1487 rets.iter().map(|p| p.value_type).collect() 1488 }; 1489 let vals = self.generate_values_for_signature(builder, types.into_iter())?; 1490 1491 builder.ins().return_(&vals[..]); 1492 } 1493 BlockTerminator::Jump(target) => { 1494 let args = self.generate_values_for_block(builder, target)?; 1495 builder.ins().jump(target, &args[..]); 1496 } 1497 BlockTerminator::Br(left, right) => { 1498 let left_args = self.generate_values_for_block(builder, left)?; 1499 let right_args = self.generate_values_for_block(builder, right)?; 1500 1501 let condbr_types = [I8, I16, I32, I64, I128]; 1502 let _type = *self.u.choose(&condbr_types[..])?; 1503 let val = builder.use_var(self.get_variable_of_type(_type)?); 1504 builder 1505 .ins() 1506 .brif(val, left, &left_args[..], right, &right_args[..]); 1507 } 1508 BlockTerminator::BrTable(default, targets) => { 1509 // Create jump tables on demand 1510 let mut jt = Vec::with_capacity(targets.len()); 1511 for block in targets { 1512 let args = self.generate_values_for_block(builder, block)?; 1513 jt.push(builder.func.dfg.block_call(block, &args)) 1514 } 1515 1516 let args = self.generate_values_for_block(builder, default)?; 1517 let jt_data = JumpTableData::new(builder.func.dfg.block_call(default, &args), &jt); 1518 let jt = builder.create_jump_table(jt_data); 1519 1520 // br_table only supports I32 1521 let val = builder.use_var(self.get_variable_of_type(I32)?); 1522 1523 builder.ins().br_table(val, jt); 1524 } 1525 BlockTerminator::Switch(_type, default, entries) => { 1526 let mut switch = Switch::new(); 1527 for (&entry, &block) in entries.iter() { 1528 switch.set_entry(entry, block); 1529 } 1530 1531 let switch_val = builder.use_var(self.get_variable_of_type(_type)?); 1532 1533 switch.emit(builder, switch_val, default); 1534 } 1535 BlockTerminator::TailCall(target) | BlockTerminator::TailCallIndirect(target) => { 1536 let (sig, sig_ref, func_ref) = self 1537 .resources 1538 .func_refs 1539 .iter() 1540 .find(|(_, _, f)| *f == target) 1541 .expect("Failed to find previously selected function") 1542 .clone(); 1543 1544 let opcode = match terminator { 1545 BlockTerminator::TailCall(_) => Opcode::ReturnCall, 1546 BlockTerminator::TailCallIndirect(_) => Opcode::ReturnCallIndirect, 1547 _ => unreachable!(), 1548 }; 1549 1550 insert_call_to_function(self, builder, opcode, &sig, sig_ref, func_ref)?; 1551 } 1552 } 1553 1554 Ok(()) 1555 } 1556 1557 /// Fills the current block with random instructions 1558 fn generate_instructions(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1559 for _ in 0..self.param(&self.config.instructions_per_block)? { 1560 let (op, args, rets) = self.u.choose(&OPCODE_SIGNATURES)?; 1561 1562 // We filter out instructions that aren't supported by the target at this point instead 1563 // of building a single vector of valid instructions at the beginning of function 1564 // generation, to avoid invalidating the corpus when instructions are enabled/disabled. 1565 if !valid_for_target(&self.isa.triple(), *op, &args, &rets) { 1566 return Err(arbitrary::Error::IncorrectFormat.into()); 1567 } 1568 1569 let inserter = inserter_for_format(op.format()); 1570 inserter(self, builder, *op, &args, &rets)?; 1571 } 1572 1573 Ok(()) 1574 } 1575 1576 fn generate_funcrefs(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1577 let usercalls: Vec<_> = self 1578 .resources 1579 .usercalls 1580 .iter() 1581 .map(|(name, signature)| { 1582 let user_func_ref = builder.func.declare_imported_user_function(name.clone()); 1583 let name = ExternalName::User(user_func_ref); 1584 (name, signature.clone()) 1585 }) 1586 .collect(); 1587 1588 let lib_callconv = self.system_callconv(); 1589 let libcalls: Vec<_> = self 1590 .resources 1591 .libcalls 1592 .iter() 1593 .map(|libcall| { 1594 let pointer_type = Type::int_with_byte_size( 1595 self.isa.triple().pointer_width().unwrap().bytes().into(), 1596 ) 1597 .unwrap(); 1598 let signature = libcall.signature(lib_callconv, pointer_type); 1599 let name = ExternalName::LibCall(*libcall); 1600 (name, signature) 1601 }) 1602 .collect(); 1603 1604 for (name, signature) in usercalls.into_iter().chain(libcalls) { 1605 let sig_ref = builder.import_signature(signature.clone()); 1606 let func_ref = builder.import_function(ExtFuncData { 1607 name, 1608 signature: sig_ref, 1609 1610 // Libcalls can't be colocated because they can be very far away 1611 // from allocated memory at runtime, and additionally at this 1612 // time cranelift-jit puts all functions in their own mmap so 1613 // they also cannot be colocated. 1614 colocated: false, 1615 patchable: false, 1616 }); 1617 1618 self.resources 1619 .func_refs 1620 .push((signature, sig_ref, func_ref)); 1621 } 1622 1623 Ok(()) 1624 } 1625 1626 fn generate_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1627 for _ in 0..self.param(&self.config.static_stack_slots_per_function)? { 1628 let bytes = self.param(&self.config.static_stack_slot_size)? as u32; 1629 let alignment = self.param(&self.config.stack_slot_alignment_log2)? as u8; 1630 let alignment_bytes = 1 << alignment; 1631 1632 let ss_data = StackSlotData::new(StackSlotKind::ExplicitSlot, bytes, alignment); 1633 let slot = builder.create_sized_stack_slot(ss_data); 1634 1635 // Generate one Alias Analysis Category for each slot 1636 let category = *self.u.choose(AACategory::all())?; 1637 1638 self.resources 1639 .stack_slots 1640 .push((slot, bytes, alignment_bytes, category)); 1641 } 1642 1643 self.resources 1644 .stack_slots 1645 .sort_unstable_by_key(|&(_slot, bytes, _align, _category)| bytes); 1646 1647 Ok(()) 1648 } 1649 1650 /// Zero initializes the stack slot by inserting `stack_store`'s. 1651 fn initialize_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1652 let i8_zero = builder.ins().iconst(I8, 0); 1653 let i16_zero = builder.ins().iconst(I16, 0); 1654 let i32_zero = builder.ins().iconst(I32, 0); 1655 let i64_zero = builder.ins().iconst(I64, 0); 1656 let i128_zero = builder.ins().uextend(I128, i64_zero); 1657 1658 for &(slot, init_size, _align, category) in self.resources.stack_slots.iter() { 1659 let mut size = init_size; 1660 1661 // Insert the largest available store for the remaining size. 1662 while size != 0 { 1663 let offset = (init_size - size) as i32; 1664 let (val, filled) = match size { 1665 sz if sz / 16 > 0 => (i128_zero, 16), 1666 sz if sz / 8 > 0 => (i64_zero, 8), 1667 sz if sz / 4 > 0 => (i32_zero, 4), 1668 sz if sz / 2 > 0 => (i16_zero, 2), 1669 _ => (i8_zero, 1), 1670 }; 1671 let addr = builder.ins().stack_addr(I64, slot, offset); 1672 1673 // Each stack slot has an associated category, that means we have to set the 1674 // correct memflags for it. So we can't use `stack_store` directly. 1675 let mut flags = MemFlags::new(); 1676 flags.set_notrap(); 1677 category.update_memflags(&mut flags); 1678 1679 builder.ins().store(flags, val, addr, 0); 1680 1681 size -= filled; 1682 } 1683 } 1684 Ok(()) 1685 } 1686 1687 /// Creates a random amount of blocks in this function 1688 fn generate_blocks(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1689 let extra_block_count = self.param(&self.config.blocks_per_function)?; 1690 1691 // We must always have at least one block, so we generate the "extra" blocks and add 1 for 1692 // the entry block. 1693 let block_count = 1 + extra_block_count; 1694 1695 // Blocks need to be sorted in ascending order 1696 self.resources.blocks = (0..block_count) 1697 .map(|i| { 1698 let is_entry = i == 0; 1699 let block = builder.create_block(); 1700 1701 // Optionally mark blocks that are not the entry block as cold 1702 if !is_entry { 1703 if bool::arbitrary(self.u)? { 1704 builder.set_cold_block(block); 1705 } 1706 } 1707 1708 // The first block has to have the function signature, but for the rest of them we generate 1709 // a random signature; 1710 if is_entry { 1711 builder.append_block_params_for_function_params(block); 1712 Ok(( 1713 block, 1714 self.signature.params.iter().map(|a| a.value_type).collect(), 1715 )) 1716 } else { 1717 let sig = self.generate_block_signature()?; 1718 sig.iter().for_each(|ty| { 1719 builder.append_block_param(block, *ty); 1720 }); 1721 Ok((block, sig)) 1722 } 1723 }) 1724 .collect::<Result<Vec<_>>>()?; 1725 1726 // Valid blocks for jump tables have to have no parameters in the signature, and must also 1727 // not be the first block. 1728 self.resources.blocks_without_params = self.resources.blocks[1..] 1729 .iter() 1730 .filter(|(_, sig)| sig.len() == 0) 1731 .map(|(b, _)| *b) 1732 .collect(); 1733 1734 // Compute the block CFG 1735 // 1736 // cranelift-frontend requires us to never generate unreachable blocks 1737 // To ensure this property we start by constructing a main "spine" of blocks. So block1 can 1738 // always jump to block2, and block2 can always jump to block3, etc... 1739 // 1740 // That is not a very interesting CFG, so we introduce variations on that, but always 1741 // ensuring that the property of pointing to the next block is maintained whatever the 1742 // branching mechanism we use. 1743 let blocks = self.resources.blocks.clone(); 1744 self.resources.block_terminators = blocks 1745 .iter() 1746 .map(|&(block, _)| { 1747 let next_block = Block::with_number(block.as_u32() + 1).unwrap(); 1748 let forward_blocks = self.resources.forward_blocks(block); 1749 let paramless_targets = self.resources.forward_blocks_without_params(block); 1750 let has_paramless_targets = !paramless_targets.is_empty(); 1751 let next_block_is_paramless = paramless_targets.contains(&next_block); 1752 1753 let mut valid_terminators = vec![]; 1754 1755 if forward_blocks.is_empty() { 1756 // Return is only valid on the last block. 1757 valid_terminators.push(BlockTerminatorKind::Return); 1758 } else { 1759 // If we have more than one block we can allow terminators that target blocks. 1760 // TODO: We could add some kind of BrReturn here, to explore edges where we 1761 // exit in the middle of the function 1762 valid_terminators.extend_from_slice(&[ 1763 BlockTerminatorKind::Jump, 1764 BlockTerminatorKind::Br, 1765 BlockTerminatorKind::BrTable, 1766 ]); 1767 } 1768 1769 // As the Switch interface only allows targeting blocks without params we need 1770 // to ensure that the next block has no params, since that one is guaranteed to be 1771 // picked in either case. 1772 if has_paramless_targets && next_block_is_paramless { 1773 valid_terminators.push(BlockTerminatorKind::Switch); 1774 } 1775 1776 // Tail Calls are a block terminator, so we should insert them as any other block 1777 // terminator. We should ensure that we can select at least one target before considering 1778 // them as candidate instructions. 1779 let has_tail_callees = self 1780 .resources 1781 .tail_call_targets(&self.signature) 1782 .next() 1783 .is_some(); 1784 let is_tail_caller = self.signature.call_conv == CallConv::Tail; 1785 1786 let supports_tail_calls = match self.isa.triple().architecture { 1787 Architecture::Aarch64(_) | Architecture::Riscv64(_) => true, 1788 // TODO: x64 currently requires frame pointers for tail calls. 1789 Architecture::X86_64 => self.isa.flags().preserve_frame_pointers(), 1790 // TODO: Other platforms do not support tail calls yet. 1791 _ => false, 1792 }; 1793 1794 if is_tail_caller && has_tail_callees && supports_tail_calls { 1795 valid_terminators.extend([ 1796 BlockTerminatorKind::TailCall, 1797 BlockTerminatorKind::TailCallIndirect, 1798 ]); 1799 } 1800 1801 let terminator = self.u.choose(&valid_terminators)?; 1802 1803 // Choose block targets for the terminators that we picked above 1804 Ok(match terminator { 1805 BlockTerminatorKind::Return => BlockTerminator::Return, 1806 BlockTerminatorKind::Jump => BlockTerminator::Jump(next_block), 1807 BlockTerminatorKind::Br => { 1808 BlockTerminator::Br(next_block, self.generate_target_block(block)?) 1809 } 1810 // TODO: Allow generating backwards branches here 1811 BlockTerminatorKind::BrTable => { 1812 // Make the default the next block, and then we don't have to worry 1813 // that we can reach it via the targets 1814 let default = next_block; 1815 1816 let target_count = self.param(&self.config.jump_table_entries)?; 1817 let targets = Result::from_iter( 1818 (0..target_count).map(|_| self.generate_target_block(block)), 1819 )?; 1820 1821 BlockTerminator::BrTable(default, targets) 1822 } 1823 BlockTerminatorKind::Switch => { 1824 // Make the default the next block, and then we don't have to worry 1825 // that we can reach it via the entries below 1826 let default_block = next_block; 1827 1828 let _type = *self.u.choose(&[I8, I16, I32, I64, I128][..])?; 1829 1830 // Build this into a HashMap since we cannot have duplicate entries. 1831 let mut entries = HashMap::new(); 1832 for _ in 0..self.param(&self.config.switch_cases)? { 1833 // The Switch API only allows for entries that are addressable by the index type 1834 // so we need to limit the range of values that we generate. 1835 let (ty_min, ty_max) = _type.bounds(false); 1836 let range_start = self.u.int_in_range(ty_min..=ty_max)?; 1837 1838 // We can either insert a contiguous range of blocks or a individual block 1839 // This is done because the Switch API specializes contiguous ranges. 1840 let range_size = if bool::arbitrary(self.u)? { 1841 1 1842 } else { 1843 self.param(&self.config.switch_max_range_size)? 1844 } as u128; 1845 1846 // Build the switch entries 1847 for i in 0..range_size { 1848 let index = range_start.wrapping_add(i) % ty_max; 1849 let block = *self 1850 .u 1851 .choose(self.resources.forward_blocks_without_params(block))?; 1852 1853 entries.insert(index, block); 1854 } 1855 } 1856 1857 BlockTerminator::Switch(_type, default_block, entries) 1858 } 1859 BlockTerminatorKind::TailCall => { 1860 let targets = self 1861 .resources 1862 .tail_call_targets(&self.signature) 1863 .collect::<Vec<_>>(); 1864 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1865 BlockTerminator::TailCall(*funcref) 1866 } 1867 BlockTerminatorKind::TailCallIndirect => { 1868 let targets = self 1869 .resources 1870 .tail_call_targets(&self.signature) 1871 .collect::<Vec<_>>(); 1872 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1873 BlockTerminator::TailCallIndirect(*funcref) 1874 } 1875 }) 1876 }) 1877 .collect::<Result<_>>()?; 1878 1879 Ok(()) 1880 } 1881 1882 fn generate_block_signature(&mut self) -> Result<BlockSignature> { 1883 let param_count = self.param(&self.config.block_signature_params)?; 1884 1885 let mut params = Vec::with_capacity(param_count); 1886 for _ in 0..param_count { 1887 params.push(self.u._type((&*self.isa).supports_simd())?); 1888 } 1889 Ok(params) 1890 } 1891 1892 fn build_variable_pool(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1893 let block = builder.current_block().unwrap(); 1894 1895 // Define variables for the function signature 1896 let mut vars: Vec<_> = builder 1897 .func 1898 .signature 1899 .params 1900 .iter() 1901 .map(|param| param.value_type) 1902 .zip(builder.block_params(block).iter().copied()) 1903 .collect(); 1904 1905 // Create a pool of vars that are going to be used in this function 1906 for _ in 0..self.param(&self.config.vars_per_function)? { 1907 let ty = self.u._type((&*self.isa).supports_simd())?; 1908 let value = self.generate_const(builder, ty)?; 1909 vars.push((ty, value)); 1910 } 1911 1912 for (ty, value) in vars.into_iter() { 1913 let var = builder.declare_var(ty); 1914 builder.def_var(var, value); 1915 1916 // Randomly declare variables as needing a stack map. 1917 // We limit these to only types that have fewer than 16 bytes 1918 // since the stack map mechanism does not support larger types. 1919 if ty.bytes() <= 16 && self.u.arbitrary()? { 1920 builder.declare_var_needs_stack_map(var); 1921 } 1922 1923 self.resources 1924 .vars 1925 .entry(ty) 1926 .or_insert_with(Vec::new) 1927 .push(var); 1928 } 1929 1930 Ok(()) 1931 } 1932 1933 /// We generate a function in multiple stages: 1934 /// 1935 /// * First we generate a random number of empty blocks 1936 /// * Then we generate a random pool of variables to be used throughout the function 1937 /// * We then visit each block and generate random instructions 1938 /// 1939 /// Because we generate all blocks and variables up front we already know everything that 1940 /// we need when generating instructions (i.e. jump targets / variables) 1941 pub fn generate(mut self) -> Result<Function> { 1942 let mut fn_builder_ctx = FunctionBuilderContext::new(); 1943 let mut func = Function::with_name_signature(self.name.clone(), self.signature.clone()); 1944 1945 let mut builder = FunctionBuilder::new(&mut func, &mut fn_builder_ctx); 1946 1947 // Build the function references before generating the block CFG since we store 1948 // function references in the CFG. 1949 self.generate_funcrefs(&mut builder)?; 1950 self.generate_blocks(&mut builder)?; 1951 1952 // Function preamble 1953 self.generate_stack_slots(&mut builder)?; 1954 1955 // Main instruction generation loop 1956 for (block, block_sig) in self.resources.blocks.clone().into_iter() { 1957 let is_block0 = block.as_u32() == 0; 1958 builder.switch_to_block(block); 1959 1960 if is_block0 { 1961 // The first block is special because we must create variables both for the 1962 // block signature and for the variable pool. Additionally, we must also define 1963 // initial values for all variables that are not the function signature. 1964 self.build_variable_pool(&mut builder)?; 1965 1966 // Stack slots have random bytes at the beginning of the function 1967 // initialize them to a constant value so that execution stays predictable. 1968 self.initialize_stack_slots(&mut builder)?; 1969 } else { 1970 // Define variables for the block params 1971 for (i, ty) in block_sig.iter().enumerate() { 1972 let var = self.get_variable_of_type(*ty)?; 1973 let block_param = builder.block_params(block)[i]; 1974 builder.def_var(var, block_param); 1975 } 1976 } 1977 1978 // Generate block instructions 1979 self.generate_instructions(&mut builder)?; 1980 1981 // Insert a terminator to safely exit the block 1982 self.insert_terminator(&mut builder, block)?; 1983 } 1984 1985 builder.seal_all_blocks(); 1986 builder.finalize(); 1987 1988 Ok(func) 1989 } 1990 } 1991