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 (Architecture::Aarch64(_), FloatCC::UnorderedOrLessThan) => true, 203 (Architecture::Aarch64(_), FloatCC::UnorderedOrLessThanOrEqual) => true, 204 (Architecture::Aarch64(_), FloatCC::UnorderedOrGreaterThan) => true, 205 (Architecture::Aarch64(_), FloatCC::UnorderedOrGreaterThanOrEqual) => true, 206 207 // These are not implemented on x86_64, for vectors. 208 (Architecture::X86_64, FloatCC::UnorderedOrEqual | FloatCC::OrderedNotEqual) => { 209 args[0].is_vector() 210 } 211 _ => false, 212 }; 213 if unimplemented_cc { 214 return Err(arbitrary::Error::IncorrectFormat.into()); 215 } 216 217 builder.ins().fcmp(cc, lhs, rhs) 218 } else { 219 let cc = *fgen.u.choose(IntCC::all())?; 220 builder.ins().icmp(cc, lhs, rhs) 221 }; 222 223 let var = fgen.get_variable_of_type(rets[0])?; 224 builder.def_var(var, res); 225 226 Ok(()) 227 } 228 229 fn insert_const( 230 fgen: &mut FunctionGenerator, 231 builder: &mut FunctionBuilder, 232 _opcode: Opcode, 233 _args: &[Type], 234 rets: &[Type], 235 ) -> Result<()> { 236 let typevar = rets[0]; 237 let var = fgen.get_variable_of_type(typevar)?; 238 let val = fgen.generate_const(builder, typevar)?; 239 builder.def_var(var, val); 240 Ok(()) 241 } 242 243 fn insert_bitcast( 244 fgen: &mut FunctionGenerator, 245 builder: &mut FunctionBuilder, 246 args: &[Type], 247 rets: &[Type], 248 ) -> Result<()> { 249 let from_var = fgen.get_variable_of_type(args[0])?; 250 let from_val = builder.use_var(from_var); 251 252 let to_var = fgen.get_variable_of_type(rets[0])?; 253 254 // TODO: We can generate little/big endian flags here. 255 let mut memflags = MemFlags::new(); 256 257 // When bitcasting between vectors of different lane counts, we need to 258 // specify the endianness. 259 if args[0].lane_count() != rets[0].lane_count() { 260 memflags.set_endianness(Endianness::Little); 261 } 262 263 let res = builder.ins().bitcast(rets[0], memflags, from_val); 264 builder.def_var(to_var, res); 265 Ok(()) 266 } 267 268 fn insert_load_store( 269 fgen: &mut FunctionGenerator, 270 builder: &mut FunctionBuilder, 271 opcode: Opcode, 272 args: &[Type], 273 rets: &[Type], 274 ) -> Result<()> { 275 if opcode == Opcode::Bitcast { 276 return insert_bitcast(fgen, builder, args, rets); 277 } 278 279 let ctrl_type = *rets.first().or(args.first()).unwrap(); 280 let type_size = ctrl_type.bytes(); 281 282 let is_atomic = [Opcode::AtomicLoad, Opcode::AtomicStore].contains(&opcode); 283 let (address, flags, offset) = 284 fgen.generate_address_and_memflags(builder, type_size, is_atomic)?; 285 286 // The variable being loaded or stored into 287 let var = fgen.get_variable_of_type(ctrl_type)?; 288 289 match opcode.format() { 290 InstructionFormat::LoadNoOffset => { 291 let (inst, dfg) = builder 292 .ins() 293 .LoadNoOffset(opcode, ctrl_type, flags, address); 294 295 let new_val = dfg.first_result(inst); 296 builder.def_var(var, new_val); 297 } 298 InstructionFormat::StoreNoOffset => { 299 let val = builder.use_var(var); 300 301 builder 302 .ins() 303 .StoreNoOffset(opcode, ctrl_type, flags, val, address); 304 } 305 InstructionFormat::Store => { 306 let val = builder.use_var(var); 307 308 builder 309 .ins() 310 .Store(opcode, ctrl_type, flags, offset, val, address); 311 } 312 InstructionFormat::Load => { 313 let (inst, dfg) = builder 314 .ins() 315 .Load(opcode, ctrl_type, flags, offset, address); 316 317 let new_val = dfg.first_result(inst); 318 builder.def_var(var, new_val); 319 } 320 _ => unimplemented!(), 321 } 322 323 Ok(()) 324 } 325 326 fn insert_atomic_rmw( 327 fgen: &mut FunctionGenerator, 328 builder: &mut FunctionBuilder, 329 _: Opcode, 330 _: &[Type], 331 rets: &[Type], 332 ) -> Result<()> { 333 let ctrl_type = *rets.first().unwrap(); 334 let type_size = ctrl_type.bytes(); 335 336 let rmw_op = *fgen.u.choose(AtomicRmwOp::all())?; 337 338 let (address, flags, offset) = fgen.generate_address_and_memflags(builder, type_size, true)?; 339 340 // AtomicRMW does not directly support offsets, so add the offset to the address separately. 341 let address = builder.ins().iadd_imm(address, i64::from(offset)); 342 343 // Load and store target variables 344 let source_var = fgen.get_variable_of_type(ctrl_type)?; 345 let target_var = fgen.get_variable_of_type(ctrl_type)?; 346 347 let source_val = builder.use_var(source_var); 348 let new_val = builder 349 .ins() 350 .atomic_rmw(ctrl_type, flags, rmw_op, address, source_val); 351 352 builder.def_var(target_var, new_val); 353 Ok(()) 354 } 355 356 fn insert_atomic_cas( 357 fgen: &mut FunctionGenerator, 358 builder: &mut FunctionBuilder, 359 _: Opcode, 360 _: &[Type], 361 rets: &[Type], 362 ) -> Result<()> { 363 let ctrl_type = *rets.first().unwrap(); 364 let type_size = ctrl_type.bytes(); 365 366 let (address, flags, offset) = fgen.generate_address_and_memflags(builder, type_size, true)?; 367 368 // AtomicCas does not directly support offsets, so add the offset to the address separately. 369 let address = builder.ins().iadd_imm(address, i64::from(offset)); 370 371 // Source and Target variables 372 let expected_var = fgen.get_variable_of_type(ctrl_type)?; 373 let store_var = fgen.get_variable_of_type(ctrl_type)?; 374 let loaded_var = fgen.get_variable_of_type(ctrl_type)?; 375 376 let expected_val = builder.use_var(expected_var); 377 let store_val = builder.use_var(store_var); 378 let new_val = builder 379 .ins() 380 .atomic_cas(flags, address, expected_val, store_val); 381 382 builder.def_var(loaded_var, new_val); 383 Ok(()) 384 } 385 386 fn insert_shuffle( 387 fgen: &mut FunctionGenerator, 388 builder: &mut FunctionBuilder, 389 opcode: Opcode, 390 _: &[Type], 391 rets: &[Type], 392 ) -> Result<()> { 393 let ctrl_type = *rets.first().unwrap(); 394 395 let lhs = builder.use_var(fgen.get_variable_of_type(ctrl_type)?); 396 let rhs = builder.use_var(fgen.get_variable_of_type(ctrl_type)?); 397 398 let mask = { 399 let mut lanes = [0u8; 16]; 400 for lane in lanes.iter_mut() { 401 *lane = fgen.u.int_in_range(0..=31)?; 402 } 403 let lanes = ConstantData::from(lanes.as_ref()); 404 builder.func.dfg.immediates.push(lanes) 405 }; 406 407 // This function is called for any `InstructionFormat::Shuffle`. Which today is just 408 // `shuffle`, but lets assert that, just to be sure we don't accidentally insert 409 // something else. 410 assert_eq!(opcode, Opcode::Shuffle); 411 let res = builder.ins().shuffle(lhs, rhs, mask); 412 413 let target_var = fgen.get_variable_of_type(ctrl_type)?; 414 builder.def_var(target_var, res); 415 416 Ok(()) 417 } 418 419 fn insert_ins_ext_lane( 420 fgen: &mut FunctionGenerator, 421 builder: &mut FunctionBuilder, 422 opcode: Opcode, 423 args: &[Type], 424 rets: &[Type], 425 ) -> Result<()> { 426 let vector_type = *args.first().unwrap(); 427 let ret_type = *rets.first().unwrap(); 428 429 let lhs = builder.use_var(fgen.get_variable_of_type(vector_type)?); 430 let max_lane = (vector_type.lane_count() as u8) - 1; 431 let lane = fgen.u.int_in_range(0..=max_lane)?; 432 433 let res = match opcode { 434 Opcode::Insertlane => { 435 let rhs = builder.use_var(fgen.get_variable_of_type(args[1])?); 436 builder.ins().insertlane(lhs, rhs, lane) 437 } 438 Opcode::Extractlane => builder.ins().extractlane(lhs, lane), 439 _ => todo!(), 440 }; 441 442 let target_var = fgen.get_variable_of_type(ret_type)?; 443 builder.def_var(target_var, res); 444 445 Ok(()) 446 } 447 448 type OpcodeInserter = fn( 449 fgen: &mut FunctionGenerator, 450 builder: &mut FunctionBuilder, 451 Opcode, 452 &[Type], 453 &[Type], 454 ) -> Result<()>; 455 456 macro_rules! exceptions { 457 ($op:expr, $args:expr, $rets:expr, $(($($cases:pat),*)),* $(,)?) => { 458 match ($op, $args, $rets) { 459 $( ($($cases,)* ..) => return false, )* 460 _ => true, 461 } 462 } 463 } 464 465 /// Returns true if we believe this `OpcodeSignature` should compile correctly 466 /// for the given target triple. We currently have a range of known issues 467 /// with specific lowerings on specific backends, and we don't want to get 468 /// fuzz bug reports for those. Over time our goal is to eliminate all of these 469 /// exceptions. 470 fn valid_for_target(triple: &Triple, op: Opcode, args: &[Type], rets: &[Type]) -> bool { 471 // Rule out invalid combinations that we don't yet have a good way of rejecting with the 472 // instruction DSL type constraints. 473 match op { 474 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat => { 475 assert_eq!(args.len(), 1); 476 assert_eq!(rets.len(), 1); 477 478 let arg = args[0]; 479 let ret = rets[0]; 480 481 // Vector arguments must produce vector results, and scalar arguments must produce 482 // scalar results. 483 if arg.is_vector() != ret.is_vector() { 484 return false; 485 } 486 487 if arg.is_vector() && ret.is_vector() { 488 // Vector conversions must have the same number of lanes, and the lanes must be the 489 // same bit-width. 490 if arg.lane_count() != ret.lane_count() { 491 return false; 492 } 493 494 if arg.lane_of().bits() != ret.lane_of().bits() { 495 return false; 496 } 497 } 498 } 499 500 Opcode::Bitcast => { 501 assert_eq!(args.len(), 1); 502 assert_eq!(rets.len(), 1); 503 504 let arg = args[0]; 505 let ret = rets[0]; 506 507 // The opcode generator still allows bitcasts between different sized types, but these 508 // are rejected in the verifier. 509 if arg.bits() != ret.bits() { 510 return false; 511 } 512 } 513 514 // This requires precise runtime integration so it's not supported at 515 // all in fuzzgen just yet. 516 Opcode::StackSwitch => return false, 517 518 _ => {} 519 } 520 521 match triple.architecture { 522 Architecture::X86_64 => { 523 exceptions!( 524 op, 525 args, 526 rets, 527 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 528 (Opcode::Imul, &[I8X16, I8X16]), 529 // https://github.com/bytecodealliance/wasmtime/issues/4756 530 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 531 // https://github.com/bytecodealliance/wasmtime/issues/5474 532 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 533 // https://github.com/bytecodealliance/wasmtime/issues/3370 534 ( 535 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 536 &[I128, I128] 537 ), 538 // https://github.com/bytecodealliance/wasmtime/issues/5107 539 (Opcode::Cls, &[I8], &[I8]), 540 (Opcode::Cls, &[I16], &[I16]), 541 (Opcode::Cls, &[I32], &[I32]), 542 (Opcode::Cls, &[I64], &[I64]), 543 (Opcode::Cls, &[I128], &[I128]), 544 // TODO 545 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 546 // https://github.com/bytecodealliance/wasmtime/issues/4897 547 // https://github.com/bytecodealliance/wasmtime/issues/4899 548 ( 549 Opcode::FcvtToUint 550 | Opcode::FcvtToUintSat 551 | Opcode::FcvtToSint 552 | Opcode::FcvtToSintSat, 553 &[F32 | F64], 554 &[I8 | I16 | I128] 555 ), 556 (Opcode::FcvtToUint | Opcode::FcvtToSint, &[F32X4], &[I32X4]), 557 ( 558 Opcode::FcvtToUint 559 | Opcode::FcvtToUintSat 560 | Opcode::FcvtToSint 561 | Opcode::FcvtToSintSat, 562 &[F64X2], 563 &[I64X2] 564 ), 565 // https://github.com/bytecodealliance/wasmtime/issues/4900 566 (Opcode::FcvtFromUint, &[I128], &[F32 | F64]), 567 // This has a lowering, but only when preceded by `uwiden_low`. 568 (Opcode::FcvtFromUint, &[I64X2], &[F64X2]), 569 // https://github.com/bytecodealliance/wasmtime/issues/4900 570 (Opcode::FcvtFromSint, &[I128], &[F32 | F64]), 571 (Opcode::FcvtFromSint, &[I64X2], &[F64X2]), 572 ( 573 Opcode::Umulhi | Opcode::Smulhi, 574 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 575 ), 576 ( 577 Opcode::UaddSat | Opcode::SaddSat | Opcode::UsubSat | Opcode::SsubSat, 578 &([I32X4, I32X4] | [I64X2, I64X2]) 579 ), 580 (Opcode::Fcopysign, &([F32X4, F32X4] | [F64X2, F64X2])), 581 (Opcode::Popcnt, &([I8X16] | [I16X8] | [I32X4] | [I64X2])), 582 ( 583 Opcode::Umax | Opcode::Smax | Opcode::Umin | Opcode::Smin, 584 &[I64X2, I64X2] 585 ), 586 // https://github.com/bytecodealliance/wasmtime/issues/6104 587 (Opcode::Bitcast, &[I128], &[_]), 588 (Opcode::Bitcast, &[_], &[I128]), 589 (Opcode::Uunarrow), 590 (Opcode::Snarrow | Opcode::Unarrow, &[I64X2, I64X2]), 591 (Opcode::SqmulRoundSat, &[I32X4, I32X4]), 592 // This Icmp is not implemented: #5529 593 (Opcode::Icmp, &[I64X2, I64X2]), 594 // IaddPairwise is implemented, but only for some types, and with some preceding ops. 595 (Opcode::IaddPairwise), 596 // Nothing wrong with this select. But we have an isle rule that can optimize it 597 // into a `min`/`max` instructions, which we don't have implemented yet. 598 (Opcode::Select, &[_, I128, I128]), 599 // These stack accesses can cause segfaults if they are merged into an SSE instruction. 600 // See: #5922 601 ( 602 Opcode::StackStore, 603 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 604 ), 605 ( 606 Opcode::StackLoad, 607 &[], 608 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 609 ), 610 // TODO 611 ( 612 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 613 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 614 ), 615 ( 616 Opcode::Rotr | Opcode::Rotl, 617 &[I8X16 | I16X8 | I32X4 | I64X2, _] 618 ), 619 ) 620 } 621 622 Architecture::Aarch64(_) => { 623 exceptions!( 624 op, 625 args, 626 rets, 627 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 628 // https://github.com/bytecodealliance/wasmtime/issues/4864 629 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 630 // https://github.com/bytecodealliance/wasmtime/issues/5472 631 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 632 // https://github.com/bytecodealliance/wasmtime/issues/4313 633 ( 634 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 635 &[I128, I128] 636 ), 637 // https://github.com/bytecodealliance/wasmtime/issues/4870 638 (Opcode::Bnot, &[F32 | F64]), 639 ( 640 Opcode::Band 641 | Opcode::Bor 642 | Opcode::Bxor 643 | Opcode::BandNot 644 | Opcode::BorNot 645 | Opcode::BxorNot, 646 &([F32, F32] | [F64, F64]) 647 ), 648 // https://github.com/bytecodealliance/wasmtime/issues/5198 649 (Opcode::Bitselect, &[I128, I128, I128]), 650 // https://github.com/bytecodealliance/wasmtime/issues/4934 651 ( 652 Opcode::FcvtToUint 653 | Opcode::FcvtToUintSat 654 | Opcode::FcvtToSint 655 | Opcode::FcvtToSintSat, 656 &[F32 | F64], 657 &[I128] 658 ), 659 // https://github.com/bytecodealliance/wasmtime/issues/4933 660 ( 661 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 662 &[I128], 663 &[F32 | F64] 664 ), 665 ( 666 Opcode::Umulhi | Opcode::Smulhi, 667 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 668 ), 669 (Opcode::Popcnt, &[I16X8 | I32X4 | I64X2]), 670 // Nothing wrong with this select. But we have an isle rule that can optimize it 671 // into a `min`/`max` instructions, which we don't have implemented yet. 672 (Opcode::Select, &[I8, I128, I128]), 673 // https://github.com/bytecodealliance/wasmtime/issues/6104 674 (Opcode::Bitcast, &[I128], &[_]), 675 (Opcode::Bitcast, &[_], &[I128]), 676 // TODO 677 ( 678 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 679 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 680 ), 681 ( 682 Opcode::Rotr | Opcode::Rotl, 683 &[I8X16 | I16X8 | I32X4 | I64X2, _] 684 ), 685 // TODO 686 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 687 (Opcode::VhighBits, &[F32X4 | F64X2]), 688 ) 689 } 690 691 Architecture::S390x => { 692 exceptions!( 693 op, 694 args, 695 rets, 696 (Opcode::UaddOverflow | Opcode::SaddOverflow), 697 (Opcode::UsubOverflow | Opcode::SsubOverflow), 698 (Opcode::UmulOverflow | Opcode::SmulOverflow), 699 ( 700 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 701 &[I128, I128] 702 ), 703 (Opcode::Bnot, &[F32 | F64]), 704 ( 705 Opcode::Band 706 | Opcode::Bor 707 | Opcode::Bxor 708 | Opcode::BandNot 709 | Opcode::BorNot 710 | Opcode::BxorNot, 711 &([F32, F32] | [F64, F64]) 712 ), 713 ( 714 Opcode::FcvtToUint 715 | Opcode::FcvtToUintSat 716 | Opcode::FcvtToSint 717 | Opcode::FcvtToSintSat, 718 &[F32 | F64], 719 &[I128] 720 ), 721 ( 722 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 723 &[I128], 724 &[F32 | F64] 725 ), 726 (Opcode::SsubSat | Opcode::SaddSat, &[I64X2, I64X2]), 727 // https://github.com/bytecodealliance/wasmtime/issues/6104 728 (Opcode::Bitcast, &[I128], &[_]), 729 (Opcode::Bitcast, &[_], &[I128]), 730 // TODO 731 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 732 ) 733 } 734 735 Architecture::Riscv64(_) => { 736 exceptions!( 737 op, 738 args, 739 rets, 740 // TODO 741 (Opcode::UaddOverflow | Opcode::SaddOverflow), 742 (Opcode::UsubOverflow | Opcode::SsubOverflow), 743 (Opcode::UmulOverflow | Opcode::SmulOverflow), 744 // TODO 745 ( 746 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 747 &[I128, I128] 748 ), 749 // TODO 750 (Opcode::Iabs, &[I128]), 751 // TODO 752 (Opcode::Bitselect, &[I128, I128, I128]), 753 // https://github.com/bytecodealliance/wasmtime/issues/5528 754 ( 755 Opcode::FcvtToUint | Opcode::FcvtToSint, 756 [F32 | F64], 757 &[I128] 758 ), 759 ( 760 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat, 761 &[F32 | F64], 762 &[I128] 763 ), 764 // https://github.com/bytecodealliance/wasmtime/issues/5528 765 ( 766 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 767 &[I128], 768 &[F32 | F64] 769 ), 770 // TODO 771 ( 772 Opcode::SelectSpectreGuard, 773 &[_, _, _], 774 &[F32 | F64 | I8X16 | I16X8 | I32X4 | I64X2 | F64X2 | F32X4] 775 ), 776 // TODO 777 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 778 ( 779 Opcode::Rotr | Opcode::Rotl, 780 &[I8X16 | I16X8 | I32X4 | I64X2, _] 781 ), 782 ) 783 } 784 785 _ => true, 786 } 787 } 788 789 type OpcodeSignature = (Opcode, Vec<Type>, Vec<Type>); 790 791 static OPCODE_SIGNATURES: LazyLock<Vec<OpcodeSignature>> = LazyLock::new(|| { 792 let types = &[ 793 I8, I16, I32, I64, I128, // Scalar Integers 794 F32, F64, // Scalar Floats 795 I8X16, I16X8, I32X4, I64X2, // SIMD Integers 796 F32X4, F64X2, // SIMD Floats 797 ]; 798 799 // When this env variable is passed, we only generate instructions for the opcodes listed in 800 // the comma-separated list. This is useful for debugging, as it allows us to focus on a few 801 // specific opcodes. 802 let allowed_opcodes = std::env::var("FUZZGEN_ALLOWED_OPS").ok().map(|s| { 803 s.split(',') 804 .map(|s| s.trim()) 805 .filter(|s| !s.is_empty()) 806 .map(|s| Opcode::from_str(s).expect("Unrecoginzed opcode")) 807 .collect::<Vec<_>>() 808 }); 809 810 Opcode::all() 811 .iter() 812 .filter(|op| { 813 match op { 814 // Control flow opcodes should not be generated through `generate_instructions`. 815 Opcode::BrTable 816 | Opcode::Brif 817 | Opcode::Jump 818 | Opcode::Return 819 | Opcode::ReturnCall 820 | Opcode::ReturnCallIndirect 821 | Opcode::TryCall 822 | Opcode::TryCallIndirect => false, 823 824 // Constants are generated outside of `generate_instructions` 825 Opcode::Iconst => false, 826 827 // TODO: extract_vector raises exceptions during return type generation because it 828 // uses dynamic vectors. 829 Opcode::ExtractVector => false, 830 831 _ => true, 832 } 833 }) 834 .flat_map(|op| { 835 let constraints = op.constraints(); 836 837 let ctrl_types = if let Some(ctrls) = constraints.ctrl_typeset() { 838 Vec::from_iter(types.iter().copied().filter(|ty| ctrls.contains(*ty))) 839 } else { 840 vec![INVALID] 841 }; 842 843 ctrl_types.into_iter().flat_map(move |ctrl_type| { 844 let rets = Vec::from_iter( 845 (0..constraints.num_fixed_results()) 846 .map(|i| constraints.result_type(i, ctrl_type)), 847 ); 848 849 // Cols is a vector whose length will match `num_fixed_value_arguments`, and whose 850 // elements will be vectors of types that are valid for that fixed argument 851 // position. 852 let mut cols = vec![]; 853 854 for i in 0..constraints.num_fixed_value_arguments() { 855 match constraints.value_argument_constraint(i, ctrl_type) { 856 ResolvedConstraint::Bound(ty) => cols.push(Vec::from([ty])), 857 ResolvedConstraint::Free(tys) => cols.push(Vec::from_iter( 858 types.iter().copied().filter(|ty| tys.contains(*ty)), 859 )), 860 } 861 } 862 863 // Generate the cartesian product of cols to produce a vector of argument lists, 864 // argss. The argss vector is seeded with the empty argument list, so there's an 865 // initial value to be extended in the loop below. 866 let mut argss = vec![vec![]]; 867 let mut cols = cols.as_slice(); 868 while let Some((col, rest)) = cols.split_last() { 869 cols = rest; 870 871 let mut next = vec![]; 872 for current in argss.iter() { 873 // Extend the front of each argument candidate with every type in `col`. 874 for ty in col { 875 let mut args = vec![*ty]; 876 args.extend_from_slice(¤t); 877 next.push(args); 878 } 879 } 880 881 let _ = std::mem::replace(&mut argss, next); 882 } 883 884 argss.into_iter().map(move |args| (*op, args, rets.clone())) 885 }) 886 }) 887 .filter(|(op, args, rets)| { 888 // These op/signature combinations need to be vetted 889 exceptions!( 890 op, 891 args.as_slice(), 892 rets.as_slice(), 893 (Opcode::Debugtrap), 894 (Opcode::Trap), 895 (Opcode::Trapz), 896 (Opcode::Trapnz), 897 (Opcode::CallIndirect, &[I32]), 898 (Opcode::FuncAddr), 899 (Opcode::X86Pshufb), 900 (Opcode::AvgRound), 901 (Opcode::Uload8x8), 902 (Opcode::Sload8x8), 903 (Opcode::Uload16x4), 904 (Opcode::Sload16x4), 905 (Opcode::Uload32x2), 906 (Opcode::Sload32x2), 907 (Opcode::StackAddr), 908 (Opcode::DynamicStackLoad), 909 (Opcode::DynamicStackStore), 910 (Opcode::DynamicStackAddr), 911 (Opcode::GlobalValue), 912 (Opcode::SymbolValue), 913 (Opcode::TlsValue), 914 (Opcode::GetPinnedReg), 915 (Opcode::SetPinnedReg), 916 (Opcode::GetFramePointer), 917 (Opcode::GetStackPointer), 918 (Opcode::GetReturnAddress), 919 (Opcode::X86Blendv), 920 (Opcode::IcmpImm), 921 (Opcode::X86Pmulhrsw), 922 (Opcode::IaddImm), 923 (Opcode::ImulImm), 924 (Opcode::UdivImm), 925 (Opcode::SdivImm), 926 (Opcode::UremImm), 927 (Opcode::SremImm), 928 (Opcode::IrsubImm), 929 (Opcode::UaddOverflowCin), 930 (Opcode::SaddOverflowCin), 931 (Opcode::UaddOverflowTrap), 932 (Opcode::UsubOverflowBin), 933 (Opcode::SsubOverflowBin), 934 (Opcode::BandImm), 935 (Opcode::BorImm), 936 (Opcode::BxorImm), 937 (Opcode::RotlImm), 938 (Opcode::RotrImm), 939 (Opcode::IshlImm), 940 (Opcode::UshrImm), 941 (Opcode::SshrImm), 942 (Opcode::ScalarToVector), 943 (Opcode::X86Pmaddubsw), 944 (Opcode::X86Cvtt2dq), 945 (Opcode::Umulhi, &[I128, I128], &[I128]), 946 (Opcode::Smulhi, &[I128, I128], &[I128]), 947 // https://github.com/bytecodealliance/wasmtime/issues/6073 948 (Opcode::Iconcat, &[I32, I32], &[I64]), 949 (Opcode::Iconcat, &[I16, I16], &[I32]), 950 (Opcode::Iconcat, &[I8, I8], &[I16]), 951 // https://github.com/bytecodealliance/wasmtime/issues/6073 952 (Opcode::Isplit, &[I64], &[I32, I32]), 953 (Opcode::Isplit, &[I32], &[I16, I16]), 954 (Opcode::Isplit, &[I16], &[I8, I8]), 955 (Opcode::Fmin, &[F32X4, F32X4], &[F32X4]), 956 (Opcode::Fmin, &[F64X2, F64X2], &[F64X2]), 957 (Opcode::Fmax, &[F32X4, F32X4], &[F32X4]), 958 (Opcode::Fmax, &[F64X2, F64X2], &[F64X2]), 959 (Opcode::FcvtToUintSat, &[F32X4], &[I8]), 960 (Opcode::FcvtToUintSat, &[F64X2], &[I8]), 961 (Opcode::FcvtToUintSat, &[F32X4], &[I16]), 962 (Opcode::FcvtToUintSat, &[F64X2], &[I16]), 963 (Opcode::FcvtToUintSat, &[F32X4], &[I32]), 964 (Opcode::FcvtToUintSat, &[F64X2], &[I32]), 965 (Opcode::FcvtToUintSat, &[F32X4], &[I64]), 966 (Opcode::FcvtToUintSat, &[F64X2], &[I64]), 967 (Opcode::FcvtToUintSat, &[F32X4], &[I128]), 968 (Opcode::FcvtToUintSat, &[F64X2], &[I128]), 969 (Opcode::FcvtToUintSat, &[F32], &[I8X16]), 970 (Opcode::FcvtToUintSat, &[F64], &[I8X16]), 971 (Opcode::FcvtToUintSat, &[F32X4], &[I8X16]), 972 (Opcode::FcvtToUintSat, &[F64X2], &[I8X16]), 973 (Opcode::FcvtToUintSat, &[F32], &[I16X8]), 974 (Opcode::FcvtToUintSat, &[F64], &[I16X8]), 975 (Opcode::FcvtToUintSat, &[F32X4], &[I16X8]), 976 (Opcode::FcvtToUintSat, &[F64X2], &[I16X8]), 977 (Opcode::FcvtToUintSat, &[F32], &[I32X4]), 978 (Opcode::FcvtToUintSat, &[F64], &[I32X4]), 979 (Opcode::FcvtToUintSat, &[F64X2], &[I32X4]), 980 (Opcode::FcvtToUintSat, &[F32], &[I64X2]), 981 (Opcode::FcvtToUintSat, &[F64], &[I64X2]), 982 (Opcode::FcvtToUintSat, &[F32X4], &[I64X2]), 983 (Opcode::FcvtToSintSat, &[F32X4], &[I8]), 984 (Opcode::FcvtToSintSat, &[F64X2], &[I8]), 985 (Opcode::FcvtToSintSat, &[F32X4], &[I16]), 986 (Opcode::FcvtToSintSat, &[F64X2], &[I16]), 987 (Opcode::FcvtToSintSat, &[F32X4], &[I32]), 988 (Opcode::FcvtToSintSat, &[F64X2], &[I32]), 989 (Opcode::FcvtToSintSat, &[F32X4], &[I64]), 990 (Opcode::FcvtToSintSat, &[F64X2], &[I64]), 991 (Opcode::FcvtToSintSat, &[F32X4], &[I128]), 992 (Opcode::FcvtToSintSat, &[F64X2], &[I128]), 993 (Opcode::FcvtToSintSat, &[F32], &[I8X16]), 994 (Opcode::FcvtToSintSat, &[F64], &[I8X16]), 995 (Opcode::FcvtToSintSat, &[F32X4], &[I8X16]), 996 (Opcode::FcvtToSintSat, &[F64X2], &[I8X16]), 997 (Opcode::FcvtToSintSat, &[F32], &[I16X8]), 998 (Opcode::FcvtToSintSat, &[F64], &[I16X8]), 999 (Opcode::FcvtToSintSat, &[F32X4], &[I16X8]), 1000 (Opcode::FcvtToSintSat, &[F64X2], &[I16X8]), 1001 (Opcode::FcvtToSintSat, &[F32], &[I32X4]), 1002 (Opcode::FcvtToSintSat, &[F64], &[I32X4]), 1003 (Opcode::FcvtToSintSat, &[F64X2], &[I32X4]), 1004 (Opcode::FcvtToSintSat, &[F32], &[I64X2]), 1005 (Opcode::FcvtToSintSat, &[F64], &[I64X2]), 1006 (Opcode::FcvtToSintSat, &[F32X4], &[I64X2]), 1007 (Opcode::FcvtFromUint, &[I8X16], &[F32]), 1008 (Opcode::FcvtFromUint, &[I16X8], &[F32]), 1009 (Opcode::FcvtFromUint, &[I32X4], &[F32]), 1010 (Opcode::FcvtFromUint, &[I64X2], &[F32]), 1011 (Opcode::FcvtFromUint, &[I8X16], &[F64]), 1012 (Opcode::FcvtFromUint, &[I16X8], &[F64]), 1013 (Opcode::FcvtFromUint, &[I32X4], &[F64]), 1014 (Opcode::FcvtFromUint, &[I64X2], &[F64]), 1015 (Opcode::FcvtFromUint, &[I8], &[F32X4]), 1016 (Opcode::FcvtFromUint, &[I16], &[F32X4]), 1017 (Opcode::FcvtFromUint, &[I32], &[F32X4]), 1018 (Opcode::FcvtFromUint, &[I64], &[F32X4]), 1019 (Opcode::FcvtFromUint, &[I128], &[F32X4]), 1020 (Opcode::FcvtFromUint, &[I8X16], &[F32X4]), 1021 (Opcode::FcvtFromUint, &[I16X8], &[F32X4]), 1022 (Opcode::FcvtFromUint, &[I64X2], &[F32X4]), 1023 (Opcode::FcvtFromUint, &[I8], &[F64X2]), 1024 (Opcode::FcvtFromUint, &[I16], &[F64X2]), 1025 (Opcode::FcvtFromUint, &[I32], &[F64X2]), 1026 (Opcode::FcvtFromUint, &[I64], &[F64X2]), 1027 (Opcode::FcvtFromUint, &[I128], &[F64X2]), 1028 (Opcode::FcvtFromUint, &[I8X16], &[F64X2]), 1029 (Opcode::FcvtFromUint, &[I16X8], &[F64X2]), 1030 (Opcode::FcvtFromUint, &[I32X4], &[F64X2]), 1031 (Opcode::FcvtFromSint, &[I8X16], &[F32]), 1032 (Opcode::FcvtFromSint, &[I16X8], &[F32]), 1033 (Opcode::FcvtFromSint, &[I32X4], &[F32]), 1034 (Opcode::FcvtFromSint, &[I64X2], &[F32]), 1035 (Opcode::FcvtFromSint, &[I8X16], &[F64]), 1036 (Opcode::FcvtFromSint, &[I16X8], &[F64]), 1037 (Opcode::FcvtFromSint, &[I32X4], &[F64]), 1038 (Opcode::FcvtFromSint, &[I64X2], &[F64]), 1039 (Opcode::FcvtFromSint, &[I8], &[F32X4]), 1040 (Opcode::FcvtFromSint, &[I16], &[F32X4]), 1041 (Opcode::FcvtFromSint, &[I32], &[F32X4]), 1042 (Opcode::FcvtFromSint, &[I64], &[F32X4]), 1043 (Opcode::FcvtFromSint, &[I128], &[F32X4]), 1044 (Opcode::FcvtFromSint, &[I8X16], &[F32X4]), 1045 (Opcode::FcvtFromSint, &[I16X8], &[F32X4]), 1046 (Opcode::FcvtFromSint, &[I64X2], &[F32X4]), 1047 (Opcode::FcvtFromSint, &[I8], &[F64X2]), 1048 (Opcode::FcvtFromSint, &[I16], &[F64X2]), 1049 (Opcode::FcvtFromSint, &[I32], &[F64X2]), 1050 (Opcode::FcvtFromSint, &[I64], &[F64X2]), 1051 (Opcode::FcvtFromSint, &[I128], &[F64X2]), 1052 (Opcode::FcvtFromSint, &[I8X16], &[F64X2]), 1053 (Opcode::FcvtFromSint, &[I16X8], &[F64X2]), 1054 (Opcode::FcvtFromSint, &[I32X4], &[F64X2]), 1055 // Only supported on x64 with a feature at this time, so 128-bit 1056 // atomics are not suitable to fuzz yet. 1057 (Opcode::AtomicRmw, _, &[I128]), 1058 (Opcode::AtomicCas, _, &[I128]), 1059 (Opcode::AtomicLoad, _, &[I128]), 1060 (Opcode::AtomicStore, &[I128, _], _), 1061 (Opcode::SequencePoint), 1062 ) 1063 }) 1064 .filter(|(op, ..)| { 1065 allowed_opcodes 1066 .as_ref() 1067 .map_or(true, |opcodes| opcodes.contains(op)) 1068 }) 1069 .collect() 1070 }); 1071 1072 fn inserter_for_format(fmt: InstructionFormat) -> OpcodeInserter { 1073 match fmt { 1074 InstructionFormat::AtomicCas => insert_atomic_cas, 1075 InstructionFormat::AtomicRmw => insert_atomic_rmw, 1076 InstructionFormat::Binary => insert_opcode, 1077 InstructionFormat::BinaryImm64 => todo!(), 1078 InstructionFormat::BinaryImm8 => insert_ins_ext_lane, 1079 InstructionFormat::Call => insert_call, 1080 InstructionFormat::CallIndirect => insert_call, 1081 InstructionFormat::CondTrap => todo!(), 1082 InstructionFormat::DynamicStackLoad => todo!(), 1083 InstructionFormat::DynamicStackStore => todo!(), 1084 InstructionFormat::FloatCompare => insert_cmp, 1085 InstructionFormat::FuncAddr => todo!(), 1086 InstructionFormat::IntAddTrap => todo!(), 1087 InstructionFormat::IntCompare => insert_cmp, 1088 InstructionFormat::IntCompareImm => todo!(), 1089 InstructionFormat::Load => insert_load_store, 1090 InstructionFormat::LoadNoOffset => insert_load_store, 1091 InstructionFormat::NullAry => insert_opcode, 1092 InstructionFormat::Shuffle => insert_shuffle, 1093 InstructionFormat::StackLoad => insert_stack_load, 1094 InstructionFormat::StackStore => insert_stack_store, 1095 InstructionFormat::Store => insert_load_store, 1096 InstructionFormat::StoreNoOffset => insert_load_store, 1097 InstructionFormat::Ternary => insert_opcode, 1098 InstructionFormat::TernaryImm8 => insert_ins_ext_lane, 1099 InstructionFormat::Trap => todo!(), 1100 InstructionFormat::Unary => insert_opcode, 1101 InstructionFormat::UnaryConst => insert_const, 1102 InstructionFormat::UnaryGlobalValue => todo!(), 1103 InstructionFormat::UnaryIeee16 => insert_const, 1104 InstructionFormat::UnaryIeee32 => insert_const, 1105 InstructionFormat::UnaryIeee64 => insert_const, 1106 InstructionFormat::UnaryImm => insert_const, 1107 InstructionFormat::ExceptionHandlerAddress => insert_const, 1108 1109 InstructionFormat::BranchTable 1110 | InstructionFormat::Brif 1111 | InstructionFormat::Jump 1112 | InstructionFormat::MultiAry 1113 | InstructionFormat::TryCall 1114 | InstructionFormat::TryCallIndirect => { 1115 panic!("Control-flow instructions should be handled by 'insert_terminator': {fmt:?}") 1116 } 1117 } 1118 } 1119 1120 pub struct FunctionGenerator<'r, 'data> 1121 where 1122 'data: 'r, 1123 { 1124 u: &'r mut Unstructured<'data>, 1125 config: &'r Config, 1126 resources: Resources, 1127 isa: OwnedTargetIsa, 1128 name: UserFuncName, 1129 signature: Signature, 1130 } 1131 1132 #[derive(Debug, Clone)] 1133 enum BlockTerminator { 1134 Return, 1135 Jump(Block), 1136 Br(Block, Block), 1137 BrTable(Block, Vec<Block>), 1138 Switch(Type, Block, HashMap<u128, Block>), 1139 TailCall(FuncRef), 1140 TailCallIndirect(FuncRef), 1141 } 1142 1143 #[derive(Debug, Clone)] 1144 enum BlockTerminatorKind { 1145 Return, 1146 Jump, 1147 Br, 1148 BrTable, 1149 Switch, 1150 TailCall, 1151 TailCallIndirect, 1152 } 1153 1154 /// Alias Analysis Category 1155 /// 1156 /// Our alias analysis pass supports 4 categories of accesses to distinguish 1157 /// different regions. The "Other" region is the general case, and is the default 1158 /// Although they have highly suggestive names there is no difference between any 1159 /// of the categories. 1160 /// 1161 /// We assign each stack slot a category when we first generate them, and then 1162 /// ensure that all accesses to that stack slot are correctly tagged. We already 1163 /// ensure that memory accesses never cross stack slots, so there is no risk 1164 /// of a memory access being tagged with the wrong category. 1165 #[derive(Debug, PartialEq, Clone, Copy)] 1166 enum AACategory { 1167 Other, 1168 Heap, 1169 Table, 1170 VmCtx, 1171 } 1172 1173 impl AACategory { 1174 pub fn all() -> &'static [Self] { 1175 &[ 1176 AACategory::Other, 1177 AACategory::Heap, 1178 AACategory::Table, 1179 AACategory::VmCtx, 1180 ] 1181 } 1182 1183 pub fn update_memflags(&self, flags: &mut MemFlags) { 1184 flags.set_alias_region(match self { 1185 AACategory::Other => None, 1186 AACategory::Heap => Some(AliasRegion::Heap), 1187 AACategory::Table => Some(AliasRegion::Table), 1188 AACategory::VmCtx => Some(AliasRegion::Vmctx), 1189 }) 1190 } 1191 } 1192 1193 pub type StackAlignment = StackSize; 1194 1195 #[derive(Default)] 1196 struct Resources { 1197 vars: HashMap<Type, Vec<Variable>>, 1198 blocks: Vec<(Block, BlockSignature)>, 1199 blocks_without_params: Vec<Block>, 1200 block_terminators: Vec<BlockTerminator>, 1201 func_refs: Vec<(Signature, SigRef, FuncRef)>, 1202 /// This field is required to be sorted by stack slot size at all times. 1203 /// We use this invariant when searching for stack slots with a given size. 1204 /// See [FunctionGenerator::stack_slot_with_size] 1205 stack_slots: Vec<(StackSlot, StackSize, StackAlignment, AACategory)>, 1206 usercalls: Vec<(UserExternalName, Signature)>, 1207 libcalls: Vec<LibCall>, 1208 } 1209 1210 impl Resources { 1211 /// Partitions blocks at `block`. Only blocks that can be targeted by branches are considered. 1212 /// 1213 /// The first slice includes all blocks up to and including `block`. 1214 /// The second slice includes all remaining blocks. 1215 fn partition_target_blocks( 1216 &self, 1217 block: Block, 1218 ) -> (&[(Block, BlockSignature)], &[(Block, BlockSignature)]) { 1219 // Blocks are stored in-order and have no gaps, this means that we can simply index them by 1220 // their number. We also need to exclude the entry block since it isn't a valid target. 1221 let target_blocks = &self.blocks[1..]; 1222 target_blocks.split_at(block.as_u32() as usize) 1223 } 1224 1225 /// Returns blocks forward of `block`. Only blocks that can be targeted by branches are considered. 1226 fn forward_blocks(&self, block: Block) -> &[(Block, BlockSignature)] { 1227 let (_, forward_blocks) = self.partition_target_blocks(block); 1228 forward_blocks 1229 } 1230 1231 /// Generates a slice of `blocks_without_params` ahead of `block` 1232 fn forward_blocks_without_params(&self, block: Block) -> &[Block] { 1233 let partition_point = self.blocks_without_params.partition_point(|b| *b <= block); 1234 &self.blocks_without_params[partition_point..] 1235 } 1236 1237 /// Generates an iterator of all valid tail call targets. This includes all functions with both 1238 /// the `tail` calling convention and the same return values as the caller. 1239 fn tail_call_targets<'a>( 1240 &'a self, 1241 caller_sig: &'a Signature, 1242 ) -> impl Iterator<Item = &'a (Signature, SigRef, FuncRef)> { 1243 self.func_refs.iter().filter(|(sig, _, _)| { 1244 sig.call_conv == CallConv::Tail && sig.returns == caller_sig.returns 1245 }) 1246 } 1247 } 1248 1249 impl<'r, 'data> FunctionGenerator<'r, 'data> 1250 where 1251 'data: 'r, 1252 { 1253 pub fn new( 1254 u: &'r mut Unstructured<'data>, 1255 config: &'r Config, 1256 isa: OwnedTargetIsa, 1257 name: UserFuncName, 1258 signature: Signature, 1259 usercalls: Vec<(UserExternalName, Signature)>, 1260 libcalls: Vec<LibCall>, 1261 ) -> Self { 1262 Self { 1263 u, 1264 config, 1265 resources: Resources { 1266 usercalls, 1267 libcalls, 1268 ..Resources::default() 1269 }, 1270 isa, 1271 name, 1272 signature, 1273 } 1274 } 1275 1276 /// Generates a random value for config `param` 1277 fn param(&mut self, param: &RangeInclusive<usize>) -> Result<usize> { 1278 Ok(self.u.int_in_range(param.clone())?) 1279 } 1280 1281 fn system_callconv(&mut self) -> CallConv { 1282 // TODO: This currently only runs on linux, so this is the only choice 1283 // We should improve this once we generate flags and targets 1284 CallConv::SystemV 1285 } 1286 1287 /// Finds a stack slot with size of at least n bytes 1288 fn stack_slot_with_size( 1289 &mut self, 1290 n: u32, 1291 ) -> Result<(StackSlot, StackSize, StackAlignment, AACategory)> { 1292 let first = self 1293 .resources 1294 .stack_slots 1295 .partition_point(|&(_slot, size, _align, _category)| size < n); 1296 Ok(*self.u.choose(&self.resources.stack_slots[first..])?) 1297 } 1298 1299 /// Generates an address that should allow for a store or a load. 1300 /// 1301 /// Addresses aren't generated like other values. They are never stored in variables so that 1302 /// we don't run the risk of returning them from a function, which would make the fuzzer 1303 /// complain since they are different from the interpreter to the backend. 1304 /// 1305 /// `min_size`: Controls the amount of space that the address should have. 1306 /// 1307 /// `aligned`: When passed as true, the resulting address is guaranteed to be aligned 1308 /// on an 8 byte boundary. 1309 /// 1310 /// Returns a valid address and the maximum possible offset that still respects `min_size`. 1311 fn generate_load_store_address( 1312 &mut self, 1313 builder: &mut FunctionBuilder, 1314 min_size: u32, 1315 aligned: bool, 1316 ) -> Result<(Value, u32, AACategory)> { 1317 // TODO: Currently our only source of addresses is stack_addr, but we 1318 // should add global_value, symbol_value eventually 1319 let (addr, available_size, category) = { 1320 let (ss, slot_size, _align, category) = self.stack_slot_with_size(min_size)?; 1321 1322 // stack_slot_with_size guarantees that slot_size >= min_size 1323 let max_offset = slot_size - min_size; 1324 let offset = if aligned { 1325 self.u.int_in_range(0..=max_offset / min_size)? * min_size 1326 } else { 1327 self.u.int_in_range(0..=max_offset)? 1328 }; 1329 1330 let base_addr = builder.ins().stack_addr(I64, ss, offset as i32); 1331 let available_size = slot_size.saturating_sub(offset); 1332 (base_addr, available_size, category) 1333 }; 1334 1335 // TODO: Insert a bunch of amode opcodes here to modify the address! 1336 1337 // Now that we have an address and a size, we just choose a random offset to return to the 1338 // caller. Preserving min_size bytes. 1339 let max_offset = available_size.saturating_sub(min_size); 1340 Ok((addr, max_offset, category)) 1341 } 1342 1343 // Generates an address and memflags for a load or store. 1344 fn generate_address_and_memflags( 1345 &mut self, 1346 builder: &mut FunctionBuilder, 1347 min_size: u32, 1348 is_atomic: bool, 1349 ) -> Result<(Value, MemFlags, Offset32)> { 1350 // Should we generate an aligned address 1351 // Some backends have issues with unaligned atomics. 1352 // AArch64: https://github.com/bytecodealliance/wasmtime/issues/5483 1353 // RISCV: https://github.com/bytecodealliance/wasmtime/issues/5882 1354 let requires_aligned_atomics = matches!( 1355 self.isa.triple().architecture, 1356 Architecture::Aarch64(_) | Architecture::Riscv64(_) 1357 ); 1358 let aligned = if is_atomic && requires_aligned_atomics { 1359 true 1360 } else if min_size > 8 { 1361 // TODO: We currently can't guarantee that a stack_slot will be aligned on a 16 byte 1362 // boundary. We don't have a way to specify alignment when creating stack slots, and 1363 // cranelift only guarantees 8 byte alignment between stack slots. 1364 // See: https://github.com/bytecodealliance/wasmtime/issues/5922#issuecomment-1457926624 1365 false 1366 } else { 1367 bool::arbitrary(self.u)? 1368 }; 1369 1370 let mut flags = MemFlags::new(); 1371 // Even if we picked an aligned address, we can always generate unaligned memflags 1372 if aligned && bool::arbitrary(self.u)? { 1373 flags.set_aligned(); 1374 } 1375 // If the address is aligned, then we know it won't trap 1376 if aligned && bool::arbitrary(self.u)? { 1377 flags.set_notrap(); 1378 } 1379 1380 let (address, max_offset, category) = 1381 self.generate_load_store_address(builder, min_size, aligned)?; 1382 1383 // Set the Alias Analysis bits on the memflags 1384 category.update_memflags(&mut flags); 1385 1386 // Pick an offset to pass into the load/store. 1387 let offset = if aligned { 1388 0 1389 } else { 1390 self.u.int_in_range(0..=max_offset)? as i32 1391 } 1392 .into(); 1393 1394 Ok((address, flags, offset)) 1395 } 1396 1397 /// Get a variable of type `ty` from the current function 1398 fn get_variable_of_type(&mut self, ty: Type) -> Result<Variable> { 1399 let opts = self.resources.vars.get(&ty).map_or(&[][..], Vec::as_slice); 1400 let var = self.u.choose(opts)?; 1401 Ok(*var) 1402 } 1403 1404 /// Generates an instruction(`iconst`/`fconst`/etc...) to introduce a constant value 1405 fn generate_const(&mut self, builder: &mut FunctionBuilder, ty: Type) -> Result<Value> { 1406 Ok(match self.u.datavalue(ty)? { 1407 DataValue::I8(i) => builder.ins().iconst(ty, i as u8 as i64), 1408 DataValue::I16(i) => builder.ins().iconst(ty, i as u16 as i64), 1409 DataValue::I32(i) => builder.ins().iconst(ty, i as u32 as i64), 1410 DataValue::I64(i) => builder.ins().iconst(ty, i), 1411 DataValue::I128(i) => { 1412 let hi = builder.ins().iconst(I64, (i >> 64) as i64); 1413 let lo = builder.ins().iconst(I64, i as i64); 1414 builder.ins().iconcat(lo, hi) 1415 } 1416 DataValue::F16(f) => builder.ins().f16const(f), 1417 DataValue::F32(f) => builder.ins().f32const(f), 1418 DataValue::F64(f) => builder.ins().f64const(f), 1419 DataValue::F128(f) => { 1420 let handle = builder.func.dfg.constants.insert(f.into()); 1421 builder.ins().f128const(handle) 1422 } 1423 DataValue::V128(bytes) => { 1424 let data = bytes.to_vec().into(); 1425 let handle = builder.func.dfg.constants.insert(data); 1426 builder.ins().vconst(ty, handle) 1427 } 1428 _ => unimplemented!(), 1429 }) 1430 } 1431 1432 /// Chooses a random block which can be targeted by a jump / branch. 1433 /// This means any block that is not the first block. 1434 fn generate_target_block(&mut self, source_block: Block) -> Result<Block> { 1435 // We try to mostly generate forward branches to avoid generating an excessive amount of 1436 // infinite loops. But they are still important, so give them a small chance of existing. 1437 let (backwards_blocks, forward_blocks) = 1438 self.resources.partition_target_blocks(source_block); 1439 let ratio = self.config.backwards_branch_ratio; 1440 let block_targets = if !backwards_blocks.is_empty() && self.u.ratio(ratio.0, ratio.1)? { 1441 backwards_blocks 1442 } else { 1443 forward_blocks 1444 }; 1445 assert!(!block_targets.is_empty()); 1446 1447 let (block, _) = self.u.choose(block_targets)?.clone(); 1448 Ok(block) 1449 } 1450 1451 fn generate_values_for_block( 1452 &mut self, 1453 builder: &mut FunctionBuilder, 1454 block: Block, 1455 ) -> Result<Vec<BlockArg>> { 1456 let (_, sig) = self.resources.blocks[block.as_u32() as usize].clone(); 1457 Ok(self 1458 .generate_values_for_signature(builder, sig.iter().copied())? 1459 .into_iter() 1460 .map(|val| BlockArg::Value(val)) 1461 .collect::<Vec<_>>()) 1462 } 1463 1464 fn generate_values_for_signature<I: Iterator<Item = Type>>( 1465 &mut self, 1466 builder: &mut FunctionBuilder, 1467 signature: I, 1468 ) -> Result<Vec<Value>> { 1469 signature 1470 .map(|ty| { 1471 let var = self.get_variable_of_type(ty)?; 1472 let val = builder.use_var(var); 1473 Ok(val) 1474 }) 1475 .collect() 1476 } 1477 1478 /// The terminator that we need to insert has already been picked ahead of time 1479 /// we just need to build the instructions for it 1480 fn insert_terminator( 1481 &mut self, 1482 builder: &mut FunctionBuilder, 1483 source_block: Block, 1484 ) -> Result<()> { 1485 let terminator = self.resources.block_terminators[source_block.as_u32() as usize].clone(); 1486 1487 match terminator { 1488 BlockTerminator::Return => { 1489 let types: Vec<Type> = { 1490 let rets = &builder.func.signature.returns; 1491 rets.iter().map(|p| p.value_type).collect() 1492 }; 1493 let vals = self.generate_values_for_signature(builder, types.into_iter())?; 1494 1495 builder.ins().return_(&vals[..]); 1496 } 1497 BlockTerminator::Jump(target) => { 1498 let args = self.generate_values_for_block(builder, target)?; 1499 builder.ins().jump(target, &args[..]); 1500 } 1501 BlockTerminator::Br(left, right) => { 1502 let left_args = self.generate_values_for_block(builder, left)?; 1503 let right_args = self.generate_values_for_block(builder, right)?; 1504 1505 let condbr_types = [I8, I16, I32, I64, I128]; 1506 let _type = *self.u.choose(&condbr_types[..])?; 1507 let val = builder.use_var(self.get_variable_of_type(_type)?); 1508 builder 1509 .ins() 1510 .brif(val, left, &left_args[..], right, &right_args[..]); 1511 } 1512 BlockTerminator::BrTable(default, targets) => { 1513 // Create jump tables on demand 1514 let mut jt = Vec::with_capacity(targets.len()); 1515 for block in targets { 1516 let args = self.generate_values_for_block(builder, block)?; 1517 jt.push(builder.func.dfg.block_call(block, &args)) 1518 } 1519 1520 let args = self.generate_values_for_block(builder, default)?; 1521 let jt_data = JumpTableData::new(builder.func.dfg.block_call(default, &args), &jt); 1522 let jt = builder.create_jump_table(jt_data); 1523 1524 // br_table only supports I32 1525 let val = builder.use_var(self.get_variable_of_type(I32)?); 1526 1527 builder.ins().br_table(val, jt); 1528 } 1529 BlockTerminator::Switch(_type, default, entries) => { 1530 let mut switch = Switch::new(); 1531 for (&entry, &block) in entries.iter() { 1532 switch.set_entry(entry, block); 1533 } 1534 1535 let switch_val = builder.use_var(self.get_variable_of_type(_type)?); 1536 1537 switch.emit(builder, switch_val, default); 1538 } 1539 BlockTerminator::TailCall(target) | BlockTerminator::TailCallIndirect(target) => { 1540 let (sig, sig_ref, func_ref) = self 1541 .resources 1542 .func_refs 1543 .iter() 1544 .find(|(_, _, f)| *f == target) 1545 .expect("Failed to find previously selected function") 1546 .clone(); 1547 1548 let opcode = match terminator { 1549 BlockTerminator::TailCall(_) => Opcode::ReturnCall, 1550 BlockTerminator::TailCallIndirect(_) => Opcode::ReturnCallIndirect, 1551 _ => unreachable!(), 1552 }; 1553 1554 insert_call_to_function(self, builder, opcode, &sig, sig_ref, func_ref)?; 1555 } 1556 } 1557 1558 Ok(()) 1559 } 1560 1561 /// Fills the current block with random instructions 1562 fn generate_instructions(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1563 for _ in 0..self.param(&self.config.instructions_per_block)? { 1564 let (op, args, rets) = self.u.choose(&OPCODE_SIGNATURES)?; 1565 1566 // We filter out instructions that aren't supported by the target at this point instead 1567 // of building a single vector of valid instructions at the beginning of function 1568 // generation, to avoid invalidating the corpus when instructions are enabled/disabled. 1569 if !valid_for_target(&self.isa.triple(), *op, &args, &rets) { 1570 return Err(arbitrary::Error::IncorrectFormat.into()); 1571 } 1572 1573 let inserter = inserter_for_format(op.format()); 1574 inserter(self, builder, *op, &args, &rets)?; 1575 } 1576 1577 Ok(()) 1578 } 1579 1580 fn generate_funcrefs(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1581 let usercalls: Vec<_> = self 1582 .resources 1583 .usercalls 1584 .iter() 1585 .map(|(name, signature)| { 1586 let user_func_ref = builder.func.declare_imported_user_function(name.clone()); 1587 let name = ExternalName::User(user_func_ref); 1588 (name, signature.clone()) 1589 }) 1590 .collect(); 1591 1592 let lib_callconv = self.system_callconv(); 1593 let libcalls: Vec<_> = self 1594 .resources 1595 .libcalls 1596 .iter() 1597 .map(|libcall| { 1598 let pointer_type = Type::int_with_byte_size( 1599 self.isa.triple().pointer_width().unwrap().bytes().into(), 1600 ) 1601 .unwrap(); 1602 let signature = libcall.signature(lib_callconv, pointer_type); 1603 let name = ExternalName::LibCall(*libcall); 1604 (name, signature) 1605 }) 1606 .collect(); 1607 1608 for (name, signature) in usercalls.into_iter().chain(libcalls) { 1609 let sig_ref = builder.import_signature(signature.clone()); 1610 let func_ref = builder.import_function(ExtFuncData { 1611 name, 1612 signature: sig_ref, 1613 1614 // Libcalls can't be colocated because they can be very far away 1615 // from allocated memory at runtime, and additionally at this 1616 // time cranelift-jit puts all functions in their own mmap so 1617 // they also cannot be colocated. 1618 colocated: false, 1619 patchable: false, 1620 }); 1621 1622 self.resources 1623 .func_refs 1624 .push((signature, sig_ref, func_ref)); 1625 } 1626 1627 Ok(()) 1628 } 1629 1630 fn generate_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1631 for _ in 0..self.param(&self.config.static_stack_slots_per_function)? { 1632 let bytes = self.param(&self.config.static_stack_slot_size)? as u32; 1633 let alignment = self.param(&self.config.stack_slot_alignment_log2)? as u8; 1634 let alignment_bytes = 1 << alignment; 1635 1636 let ss_data = StackSlotData::new(StackSlotKind::ExplicitSlot, bytes, alignment); 1637 let slot = builder.create_sized_stack_slot(ss_data); 1638 1639 // Generate one Alias Analysis Category for each slot 1640 let category = *self.u.choose(AACategory::all())?; 1641 1642 self.resources 1643 .stack_slots 1644 .push((slot, bytes, alignment_bytes, category)); 1645 } 1646 1647 self.resources 1648 .stack_slots 1649 .sort_unstable_by_key(|&(_slot, bytes, _align, _category)| bytes); 1650 1651 Ok(()) 1652 } 1653 1654 /// Zero initializes the stack slot by inserting `stack_store`'s. 1655 fn initialize_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1656 let i8_zero = builder.ins().iconst(I8, 0); 1657 let i16_zero = builder.ins().iconst(I16, 0); 1658 let i32_zero = builder.ins().iconst(I32, 0); 1659 let i64_zero = builder.ins().iconst(I64, 0); 1660 let i128_zero = builder.ins().uextend(I128, i64_zero); 1661 1662 for &(slot, init_size, _align, category) in self.resources.stack_slots.iter() { 1663 let mut size = init_size; 1664 1665 // Insert the largest available store for the remaining size. 1666 while size != 0 { 1667 let offset = (init_size - size) as i32; 1668 let (val, filled) = match size { 1669 sz if sz / 16 > 0 => (i128_zero, 16), 1670 sz if sz / 8 > 0 => (i64_zero, 8), 1671 sz if sz / 4 > 0 => (i32_zero, 4), 1672 sz if sz / 2 > 0 => (i16_zero, 2), 1673 _ => (i8_zero, 1), 1674 }; 1675 let addr = builder.ins().stack_addr(I64, slot, offset); 1676 1677 // Each stack slot has an associated category, that means we have to set the 1678 // correct memflags for it. So we can't use `stack_store` directly. 1679 let mut flags = MemFlags::new(); 1680 flags.set_notrap(); 1681 category.update_memflags(&mut flags); 1682 1683 builder.ins().store(flags, val, addr, 0); 1684 1685 size -= filled; 1686 } 1687 } 1688 Ok(()) 1689 } 1690 1691 /// Creates a random amount of blocks in this function 1692 fn generate_blocks(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1693 let extra_block_count = self.param(&self.config.blocks_per_function)?; 1694 1695 // We must always have at least one block, so we generate the "extra" blocks and add 1 for 1696 // the entry block. 1697 let block_count = 1 + extra_block_count; 1698 1699 // Blocks need to be sorted in ascending order 1700 self.resources.blocks = (0..block_count) 1701 .map(|i| { 1702 let is_entry = i == 0; 1703 let block = builder.create_block(); 1704 1705 // Optionally mark blocks that are not the entry block as cold 1706 if !is_entry { 1707 if bool::arbitrary(self.u)? { 1708 builder.set_cold_block(block); 1709 } 1710 } 1711 1712 // The first block has to have the function signature, but for the rest of them we generate 1713 // a random signature; 1714 if is_entry { 1715 builder.append_block_params_for_function_params(block); 1716 Ok(( 1717 block, 1718 self.signature.params.iter().map(|a| a.value_type).collect(), 1719 )) 1720 } else { 1721 let sig = self.generate_block_signature()?; 1722 sig.iter().for_each(|ty| { 1723 builder.append_block_param(block, *ty); 1724 }); 1725 Ok((block, sig)) 1726 } 1727 }) 1728 .collect::<Result<Vec<_>>>()?; 1729 1730 // Valid blocks for jump tables have to have no parameters in the signature, and must also 1731 // not be the first block. 1732 self.resources.blocks_without_params = self.resources.blocks[1..] 1733 .iter() 1734 .filter(|(_, sig)| sig.len() == 0) 1735 .map(|(b, _)| *b) 1736 .collect(); 1737 1738 // Compute the block CFG 1739 // 1740 // cranelift-frontend requires us to never generate unreachable blocks 1741 // To ensure this property we start by constructing a main "spine" of blocks. So block1 can 1742 // always jump to block2, and block2 can always jump to block3, etc... 1743 // 1744 // That is not a very interesting CFG, so we introduce variations on that, but always 1745 // ensuring that the property of pointing to the next block is maintained whatever the 1746 // branching mechanism we use. 1747 let blocks = self.resources.blocks.clone(); 1748 self.resources.block_terminators = blocks 1749 .iter() 1750 .map(|&(block, _)| { 1751 let next_block = Block::with_number(block.as_u32() + 1).unwrap(); 1752 let forward_blocks = self.resources.forward_blocks(block); 1753 let paramless_targets = self.resources.forward_blocks_without_params(block); 1754 let has_paramless_targets = !paramless_targets.is_empty(); 1755 let next_block_is_paramless = paramless_targets.contains(&next_block); 1756 1757 let mut valid_terminators = vec![]; 1758 1759 if forward_blocks.is_empty() { 1760 // Return is only valid on the last block. 1761 valid_terminators.push(BlockTerminatorKind::Return); 1762 } else { 1763 // If we have more than one block we can allow terminators that target blocks. 1764 // TODO: We could add some kind of BrReturn here, to explore edges where we 1765 // exit in the middle of the function 1766 valid_terminators.extend_from_slice(&[ 1767 BlockTerminatorKind::Jump, 1768 BlockTerminatorKind::Br, 1769 BlockTerminatorKind::BrTable, 1770 ]); 1771 } 1772 1773 // As the Switch interface only allows targeting blocks without params we need 1774 // to ensure that the next block has no params, since that one is guaranteed to be 1775 // picked in either case. 1776 if has_paramless_targets && next_block_is_paramless { 1777 valid_terminators.push(BlockTerminatorKind::Switch); 1778 } 1779 1780 // Tail Calls are a block terminator, so we should insert them as any other block 1781 // terminator. We should ensure that we can select at least one target before considering 1782 // them as candidate instructions. 1783 let has_tail_callees = self 1784 .resources 1785 .tail_call_targets(&self.signature) 1786 .next() 1787 .is_some(); 1788 let is_tail_caller = self.signature.call_conv == CallConv::Tail; 1789 1790 let supports_tail_calls = match self.isa.triple().architecture { 1791 Architecture::Aarch64(_) | Architecture::Riscv64(_) => true, 1792 // TODO: x64 currently requires frame pointers for tail calls. 1793 Architecture::X86_64 => self.isa.flags().preserve_frame_pointers(), 1794 // TODO: Other platforms do not support tail calls yet. 1795 _ => false, 1796 }; 1797 1798 if is_tail_caller && has_tail_callees && supports_tail_calls { 1799 valid_terminators.extend([ 1800 BlockTerminatorKind::TailCall, 1801 BlockTerminatorKind::TailCallIndirect, 1802 ]); 1803 } 1804 1805 let terminator = self.u.choose(&valid_terminators)?; 1806 1807 // Choose block targets for the terminators that we picked above 1808 Ok(match terminator { 1809 BlockTerminatorKind::Return => BlockTerminator::Return, 1810 BlockTerminatorKind::Jump => BlockTerminator::Jump(next_block), 1811 BlockTerminatorKind::Br => { 1812 BlockTerminator::Br(next_block, self.generate_target_block(block)?) 1813 } 1814 // TODO: Allow generating backwards branches here 1815 BlockTerminatorKind::BrTable => { 1816 // Make the default the next block, and then we don't have to worry 1817 // that we can reach it via the targets 1818 let default = next_block; 1819 1820 let target_count = self.param(&self.config.jump_table_entries)?; 1821 let targets = Result::from_iter( 1822 (0..target_count).map(|_| self.generate_target_block(block)), 1823 )?; 1824 1825 BlockTerminator::BrTable(default, targets) 1826 } 1827 BlockTerminatorKind::Switch => { 1828 // Make the default the next block, and then we don't have to worry 1829 // that we can reach it via the entries below 1830 let default_block = next_block; 1831 1832 let _type = *self.u.choose(&[I8, I16, I32, I64, I128][..])?; 1833 1834 // Build this into a HashMap since we cannot have duplicate entries. 1835 let mut entries = HashMap::new(); 1836 for _ in 0..self.param(&self.config.switch_cases)? { 1837 // The Switch API only allows for entries that are addressable by the index type 1838 // so we need to limit the range of values that we generate. 1839 let (ty_min, ty_max) = _type.bounds(false); 1840 let range_start = self.u.int_in_range(ty_min..=ty_max)?; 1841 1842 // We can either insert a contiguous range of blocks or a individual block 1843 // This is done because the Switch API specializes contiguous ranges. 1844 let range_size = if bool::arbitrary(self.u)? { 1845 1 1846 } else { 1847 self.param(&self.config.switch_max_range_size)? 1848 } as u128; 1849 1850 // Build the switch entries 1851 for i in 0..range_size { 1852 let index = range_start.wrapping_add(i) % ty_max; 1853 let block = *self 1854 .u 1855 .choose(self.resources.forward_blocks_without_params(block))?; 1856 1857 entries.insert(index, block); 1858 } 1859 } 1860 1861 BlockTerminator::Switch(_type, default_block, entries) 1862 } 1863 BlockTerminatorKind::TailCall => { 1864 let targets = self 1865 .resources 1866 .tail_call_targets(&self.signature) 1867 .collect::<Vec<_>>(); 1868 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1869 BlockTerminator::TailCall(*funcref) 1870 } 1871 BlockTerminatorKind::TailCallIndirect => { 1872 let targets = self 1873 .resources 1874 .tail_call_targets(&self.signature) 1875 .collect::<Vec<_>>(); 1876 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1877 BlockTerminator::TailCallIndirect(*funcref) 1878 } 1879 }) 1880 }) 1881 .collect::<Result<_>>()?; 1882 1883 Ok(()) 1884 } 1885 1886 fn generate_block_signature(&mut self) -> Result<BlockSignature> { 1887 let param_count = self.param(&self.config.block_signature_params)?; 1888 1889 let mut params = Vec::with_capacity(param_count); 1890 for _ in 0..param_count { 1891 params.push(self.u._type((&*self.isa).supports_simd())?); 1892 } 1893 Ok(params) 1894 } 1895 1896 fn build_variable_pool(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1897 let block = builder.current_block().unwrap(); 1898 1899 // Define variables for the function signature 1900 let mut vars: Vec<_> = builder 1901 .func 1902 .signature 1903 .params 1904 .iter() 1905 .map(|param| param.value_type) 1906 .zip(builder.block_params(block).iter().copied()) 1907 .collect(); 1908 1909 // Create a pool of vars that are going to be used in this function 1910 for _ in 0..self.param(&self.config.vars_per_function)? { 1911 let ty = self.u._type((&*self.isa).supports_simd())?; 1912 let value = self.generate_const(builder, ty)?; 1913 vars.push((ty, value)); 1914 } 1915 1916 for (ty, value) in vars.into_iter() { 1917 let var = builder.declare_var(ty); 1918 builder.def_var(var, value); 1919 1920 // Randomly declare variables as needing a stack map. 1921 // We limit these to only types that have fewer than 16 bytes 1922 // since the stack map mechanism does not support larger types. 1923 if ty.bytes() <= 16 && self.u.arbitrary()? { 1924 builder.declare_var_needs_stack_map(var); 1925 } 1926 1927 self.resources 1928 .vars 1929 .entry(ty) 1930 .or_insert_with(Vec::new) 1931 .push(var); 1932 } 1933 1934 Ok(()) 1935 } 1936 1937 /// We generate a function in multiple stages: 1938 /// 1939 /// * First we generate a random number of empty blocks 1940 /// * Then we generate a random pool of variables to be used throughout the function 1941 /// * We then visit each block and generate random instructions 1942 /// 1943 /// Because we generate all blocks and variables up front we already know everything that 1944 /// we need when generating instructions (i.e. jump targets / variables) 1945 pub fn generate(mut self) -> Result<Function> { 1946 let mut fn_builder_ctx = FunctionBuilderContext::new(); 1947 let mut func = Function::with_name_signature(self.name.clone(), self.signature.clone()); 1948 1949 let mut builder = FunctionBuilder::new(&mut func, &mut fn_builder_ctx); 1950 1951 // Build the function references before generating the block CFG since we store 1952 // function references in the CFG. 1953 self.generate_funcrefs(&mut builder)?; 1954 self.generate_blocks(&mut builder)?; 1955 1956 // Function preamble 1957 self.generate_stack_slots(&mut builder)?; 1958 1959 // Main instruction generation loop 1960 for (block, block_sig) in self.resources.blocks.clone().into_iter() { 1961 let is_block0 = block.as_u32() == 0; 1962 builder.switch_to_block(block); 1963 1964 if is_block0 { 1965 // The first block is special because we must create variables both for the 1966 // block signature and for the variable pool. Additionally, we must also define 1967 // initial values for all variables that are not the function signature. 1968 self.build_variable_pool(&mut builder)?; 1969 1970 // Stack slots have random bytes at the beginning of the function 1971 // initialize them to a constant value so that execution stays predictable. 1972 self.initialize_stack_slots(&mut builder)?; 1973 } else { 1974 // Define variables for the block params 1975 for (i, ty) in block_sig.iter().enumerate() { 1976 let var = self.get_variable_of_type(*ty)?; 1977 let block_param = builder.block_params(block)[i]; 1978 builder.def_var(var, block_param); 1979 } 1980 } 1981 1982 // Generate block instructions 1983 self.generate_instructions(&mut builder)?; 1984 1985 // Insert a terminator to safely exit the block 1986 self.insert_terminator(&mut builder, block)?; 1987 } 1988 1989 builder.seal_all_blocks(); 1990 builder.finalize(); 1991 1992 Ok(func) 1993 } 1994 } 1995