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(); 122 123 insert_call_to_function(fgen, builder, opcode, &sig, sig_ref, func_ref) 124 } 125 126 fn insert_stack_load( 127 fgen: &mut FunctionGenerator, 128 builder: &mut FunctionBuilder, 129 _opcode: Opcode, 130 _args: &[Type], 131 rets: &[Type], 132 ) -> Result<()> { 133 let typevar = rets[0]; 134 let type_size = typevar.bytes(); 135 let (slot, slot_size, _align, category) = fgen.stack_slot_with_size(type_size)?; 136 137 // `stack_load` doesn't support setting MemFlags, and it does not set any 138 // alias analysis bits, so we can only emit it for `Other` slots. 139 if category != AACategory::Other { 140 return Err(arbitrary::Error::IncorrectFormat.into()); 141 } 142 143 let offset = fgen.u.int_in_range(0..=(slot_size - type_size))? as i32; 144 145 let val = builder.ins().stack_load(typevar, slot, offset); 146 let var = fgen.get_variable_of_type(typevar)?; 147 builder.def_var(var, val); 148 149 Ok(()) 150 } 151 152 fn insert_stack_store( 153 fgen: &mut FunctionGenerator, 154 builder: &mut FunctionBuilder, 155 _opcode: Opcode, 156 args: &[Type], 157 _rets: &[Type], 158 ) -> Result<()> { 159 let typevar = args[0]; 160 let type_size = typevar.bytes(); 161 162 let (slot, slot_size, _align, category) = fgen.stack_slot_with_size(type_size)?; 163 164 // `stack_store` doesn't support setting MemFlags, and it does not set any 165 // alias analysis bits, so we can only emit it for `Other` slots. 166 if category != AACategory::Other { 167 return Err(arbitrary::Error::IncorrectFormat.into()); 168 } 169 170 let offset = fgen.u.int_in_range(0..=(slot_size - type_size))? as i32; 171 172 let arg0 = fgen.get_variable_of_type(typevar)?; 173 let arg0 = builder.use_var(arg0); 174 175 builder.ins().stack_store(arg0, slot, offset); 176 Ok(()) 177 } 178 179 fn insert_cmp( 180 fgen: &mut FunctionGenerator, 181 builder: &mut FunctionBuilder, 182 opcode: Opcode, 183 args: &[Type], 184 rets: &[Type], 185 ) -> Result<()> { 186 let lhs = fgen.get_variable_of_type(args[0])?; 187 let lhs = builder.use_var(lhs); 188 189 let rhs = fgen.get_variable_of_type(args[1])?; 190 let rhs = builder.use_var(rhs); 191 192 let res = if opcode == Opcode::Fcmp { 193 let cc = *fgen.u.choose(FloatCC::all())?; 194 195 // We filter out condition codes that aren't supported by the target at 196 // this point after randomly choosing one, instead of randomly choosing a 197 // supported one, to avoid invalidating the corpus when these get implemented. 198 let unimplemented_cc = match (fgen.isa.triple().architecture, cc) { 199 // Some FloatCC's are not implemented on AArch64, see: 200 // https://github.com/bytecodealliance/wasmtime/issues/4850 201 (Architecture::Aarch64(_), FloatCC::OrderedNotEqual) => true, 202 (Architecture::Aarch64(_), FloatCC::UnorderedOrEqual) => true, 203 (Architecture::Aarch64(_), FloatCC::UnorderedOrLessThan) => true, 204 (Architecture::Aarch64(_), FloatCC::UnorderedOrLessThanOrEqual) => true, 205 (Architecture::Aarch64(_), FloatCC::UnorderedOrGreaterThan) => true, 206 (Architecture::Aarch64(_), FloatCC::UnorderedOrGreaterThanOrEqual) => true, 207 208 // These are not implemented on x86_64, for vectors. 209 (Architecture::X86_64, FloatCC::UnorderedOrEqual | FloatCC::OrderedNotEqual) => { 210 args[0].is_vector() 211 } 212 _ => false, 213 }; 214 if unimplemented_cc { 215 return Err(arbitrary::Error::IncorrectFormat.into()); 216 } 217 218 builder.ins().fcmp(cc, lhs, rhs) 219 } else { 220 let cc = *fgen.u.choose(IntCC::all())?; 221 builder.ins().icmp(cc, lhs, rhs) 222 }; 223 224 let var = fgen.get_variable_of_type(rets[0])?; 225 builder.def_var(var, res); 226 227 Ok(()) 228 } 229 230 fn insert_const( 231 fgen: &mut FunctionGenerator, 232 builder: &mut FunctionBuilder, 233 _opcode: Opcode, 234 _args: &[Type], 235 rets: &[Type], 236 ) -> Result<()> { 237 let typevar = rets[0]; 238 let var = fgen.get_variable_of_type(typevar)?; 239 let val = fgen.generate_const(builder, typevar)?; 240 builder.def_var(var, val); 241 Ok(()) 242 } 243 244 fn insert_bitcast( 245 fgen: &mut FunctionGenerator, 246 builder: &mut FunctionBuilder, 247 args: &[Type], 248 rets: &[Type], 249 ) -> Result<()> { 250 let from_var = fgen.get_variable_of_type(args[0])?; 251 let from_val = builder.use_var(from_var); 252 253 let to_var = fgen.get_variable_of_type(rets[0])?; 254 255 // TODO: We can generate little/big endian flags here. 256 let mut memflags = MemFlags::new(); 257 258 // When bitcasting between vectors of different lane counts, we need to 259 // specify the endianness. 260 if args[0].lane_count() != rets[0].lane_count() { 261 memflags.set_endianness(Endianness::Little); 262 } 263 264 let res = builder.ins().bitcast(rets[0], memflags, from_val); 265 builder.def_var(to_var, res); 266 Ok(()) 267 } 268 269 fn insert_load_store( 270 fgen: &mut FunctionGenerator, 271 builder: &mut FunctionBuilder, 272 opcode: Opcode, 273 args: &[Type], 274 rets: &[Type], 275 ) -> Result<()> { 276 if opcode == Opcode::Bitcast { 277 return insert_bitcast(fgen, builder, args, rets); 278 } 279 280 let ctrl_type = *rets.first().or(args.first()).unwrap(); 281 let type_size = ctrl_type.bytes(); 282 283 let is_atomic = [Opcode::AtomicLoad, Opcode::AtomicStore].contains(&opcode); 284 let (address, flags, offset) = 285 fgen.generate_address_and_memflags(builder, type_size, is_atomic)?; 286 287 // The variable being loaded or stored into 288 let var = fgen.get_variable_of_type(ctrl_type)?; 289 290 match opcode.format() { 291 InstructionFormat::LoadNoOffset => { 292 let (inst, dfg) = builder 293 .ins() 294 .LoadNoOffset(opcode, ctrl_type, flags, address); 295 296 let new_val = dfg.first_result(inst); 297 builder.def_var(var, new_val); 298 } 299 InstructionFormat::StoreNoOffset => { 300 let val = builder.use_var(var); 301 302 builder 303 .ins() 304 .StoreNoOffset(opcode, ctrl_type, flags, val, address); 305 } 306 InstructionFormat::Store => { 307 let val = builder.use_var(var); 308 309 builder 310 .ins() 311 .Store(opcode, ctrl_type, flags, offset, val, address); 312 } 313 InstructionFormat::Load => { 314 let (inst, dfg) = builder 315 .ins() 316 .Load(opcode, ctrl_type, flags, offset, address); 317 318 let new_val = dfg.first_result(inst); 319 builder.def_var(var, new_val); 320 } 321 _ => unimplemented!(), 322 } 323 324 Ok(()) 325 } 326 327 fn insert_atomic_rmw( 328 fgen: &mut FunctionGenerator, 329 builder: &mut FunctionBuilder, 330 _: Opcode, 331 _: &[Type], 332 rets: &[Type], 333 ) -> Result<()> { 334 let ctrl_type = *rets.first().unwrap(); 335 let type_size = ctrl_type.bytes(); 336 337 let rmw_op = *fgen.u.choose(AtomicRmwOp::all())?; 338 339 let (address, flags, offset) = fgen.generate_address_and_memflags(builder, type_size, true)?; 340 341 // AtomicRMW does not directly support offsets, so add the offset to the address separately. 342 let address = builder.ins().iadd_imm(address, i64::from(offset)); 343 344 // Load and store target variables 345 let source_var = fgen.get_variable_of_type(ctrl_type)?; 346 let target_var = fgen.get_variable_of_type(ctrl_type)?; 347 348 let source_val = builder.use_var(source_var); 349 let new_val = builder 350 .ins() 351 .atomic_rmw(ctrl_type, flags, rmw_op, address, source_val); 352 353 builder.def_var(target_var, new_val); 354 Ok(()) 355 } 356 357 fn insert_atomic_cas( 358 fgen: &mut FunctionGenerator, 359 builder: &mut FunctionBuilder, 360 _: Opcode, 361 _: &[Type], 362 rets: &[Type], 363 ) -> Result<()> { 364 let ctrl_type = *rets.first().unwrap(); 365 let type_size = ctrl_type.bytes(); 366 367 let (address, flags, offset) = fgen.generate_address_and_memflags(builder, type_size, true)?; 368 369 // AtomicCas does not directly support offsets, so add the offset to the address separately. 370 let address = builder.ins().iadd_imm(address, i64::from(offset)); 371 372 // Source and Target variables 373 let expected_var = fgen.get_variable_of_type(ctrl_type)?; 374 let store_var = fgen.get_variable_of_type(ctrl_type)?; 375 let loaded_var = fgen.get_variable_of_type(ctrl_type)?; 376 377 let expected_val = builder.use_var(expected_var); 378 let store_val = builder.use_var(store_var); 379 let new_val = builder 380 .ins() 381 .atomic_cas(flags, address, expected_val, store_val); 382 383 builder.def_var(loaded_var, new_val); 384 Ok(()) 385 } 386 387 fn insert_shuffle( 388 fgen: &mut FunctionGenerator, 389 builder: &mut FunctionBuilder, 390 opcode: Opcode, 391 _: &[Type], 392 rets: &[Type], 393 ) -> Result<()> { 394 let ctrl_type = *rets.first().unwrap(); 395 396 let lhs = builder.use_var(fgen.get_variable_of_type(ctrl_type)?); 397 let rhs = builder.use_var(fgen.get_variable_of_type(ctrl_type)?); 398 399 let mask = { 400 let mut lanes = [0u8; 16]; 401 for lane in lanes.iter_mut() { 402 *lane = fgen.u.int_in_range(0..=31)?; 403 } 404 let lanes = ConstantData::from(lanes.as_ref()); 405 builder.func.dfg.immediates.push(lanes) 406 }; 407 408 // This function is called for any `InstructionFormat::Shuffle`. Which today is just 409 // `shuffle`, but lets assert that, just to be sure we don't accidentally insert 410 // something else. 411 assert_eq!(opcode, Opcode::Shuffle); 412 let res = builder.ins().shuffle(lhs, rhs, mask); 413 414 let target_var = fgen.get_variable_of_type(ctrl_type)?; 415 builder.def_var(target_var, res); 416 417 Ok(()) 418 } 419 420 fn insert_ins_ext_lane( 421 fgen: &mut FunctionGenerator, 422 builder: &mut FunctionBuilder, 423 opcode: Opcode, 424 args: &[Type], 425 rets: &[Type], 426 ) -> Result<()> { 427 let vector_type = *args.first().unwrap(); 428 let ret_type = *rets.first().unwrap(); 429 430 let lhs = builder.use_var(fgen.get_variable_of_type(vector_type)?); 431 let max_lane = (vector_type.lane_count() as u8) - 1; 432 let lane = fgen.u.int_in_range(0..=max_lane)?; 433 434 let res = match opcode { 435 Opcode::Insertlane => { 436 let rhs = builder.use_var(fgen.get_variable_of_type(args[1])?); 437 builder.ins().insertlane(lhs, rhs, lane) 438 } 439 Opcode::Extractlane => builder.ins().extractlane(lhs, lane), 440 _ => todo!(), 441 }; 442 443 let target_var = fgen.get_variable_of_type(ret_type)?; 444 builder.def_var(target_var, res); 445 446 Ok(()) 447 } 448 449 type OpcodeInserter = fn( 450 fgen: &mut FunctionGenerator, 451 builder: &mut FunctionBuilder, 452 Opcode, 453 &[Type], 454 &[Type], 455 ) -> Result<()>; 456 457 macro_rules! exceptions { 458 ($op:expr, $args:expr, $rets:expr, $(($($cases:pat),*)),* $(,)?) => { 459 match ($op, $args, $rets) { 460 $( ($($cases,)* ..) => return false, )* 461 _ => true, 462 } 463 } 464 } 465 466 /// Returns true if we believe this `OpcodeSignature` should compile correctly 467 /// for the given target triple. We currently have a range of known issues 468 /// with specific lowerings on specific backends, and we don't want to get 469 /// fuzz bug reports for those. Over time our goal is to eliminate all of these 470 /// exceptions. 471 fn valid_for_target(triple: &Triple, op: Opcode, args: &[Type], rets: &[Type]) -> bool { 472 // Rule out invalid combinations that we don't yet have a good way of rejecting with the 473 // instruction DSL type constraints. 474 match op { 475 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat => { 476 assert_eq!(args.len(), 1); 477 assert_eq!(rets.len(), 1); 478 479 let arg = args[0]; 480 let ret = rets[0]; 481 482 // Vector arguments must produce vector results, and scalar arguments must produce 483 // scalar results. 484 if arg.is_vector() != ret.is_vector() { 485 return false; 486 } 487 488 if arg.is_vector() && ret.is_vector() { 489 // Vector conversions must have the same number of lanes, and the lanes must be the 490 // same bit-width. 491 if arg.lane_count() != ret.lane_count() { 492 return false; 493 } 494 495 if arg.lane_of().bits() != ret.lane_of().bits() { 496 return false; 497 } 498 } 499 } 500 501 Opcode::Bitcast => { 502 assert_eq!(args.len(), 1); 503 assert_eq!(rets.len(), 1); 504 505 let arg = args[0]; 506 let ret = rets[0]; 507 508 // The opcode generator still allows bitcasts between different sized types, but these 509 // are rejected in the verifier. 510 if arg.bits() != ret.bits() { 511 return false; 512 } 513 } 514 515 // This requires precise runtime integration so it's not supported at 516 // all in fuzzgen just yet. 517 Opcode::StackSwitch => return false, 518 519 _ => {} 520 } 521 522 match triple.architecture { 523 Architecture::X86_64 => { 524 exceptions!( 525 op, 526 args, 527 rets, 528 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 529 (Opcode::Imul, &[I8X16, I8X16]), 530 // https://github.com/bytecodealliance/wasmtime/issues/4756 531 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 532 // https://github.com/bytecodealliance/wasmtime/issues/5474 533 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 534 // https://github.com/bytecodealliance/wasmtime/issues/3370 535 ( 536 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 537 &[I128, I128] 538 ), 539 // https://github.com/bytecodealliance/wasmtime/issues/5107 540 (Opcode::Cls, &[I8], &[I8]), 541 (Opcode::Cls, &[I16], &[I16]), 542 (Opcode::Cls, &[I32], &[I32]), 543 (Opcode::Cls, &[I64], &[I64]), 544 (Opcode::Cls, &[I128], &[I128]), 545 // TODO 546 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 547 // https://github.com/bytecodealliance/wasmtime/issues/4897 548 // https://github.com/bytecodealliance/wasmtime/issues/4899 549 ( 550 Opcode::FcvtToUint 551 | Opcode::FcvtToUintSat 552 | Opcode::FcvtToSint 553 | Opcode::FcvtToSintSat, 554 &[F32 | F64], 555 &[I8 | I16 | I128] 556 ), 557 (Opcode::FcvtToUint | Opcode::FcvtToSint, &[F32X4], &[I32X4]), 558 ( 559 Opcode::FcvtToUint 560 | Opcode::FcvtToUintSat 561 | Opcode::FcvtToSint 562 | Opcode::FcvtToSintSat, 563 &[F64X2], 564 &[I64X2] 565 ), 566 // https://github.com/bytecodealliance/wasmtime/issues/4900 567 (Opcode::FcvtFromUint, &[I128], &[F32 | F64]), 568 // This has a lowering, but only when preceded by `uwiden_low`. 569 (Opcode::FcvtFromUint, &[I64X2], &[F64X2]), 570 // https://github.com/bytecodealliance/wasmtime/issues/4900 571 (Opcode::FcvtFromSint, &[I128], &[F32 | F64]), 572 (Opcode::FcvtFromSint, &[I64X2], &[F64X2]), 573 ( 574 Opcode::Umulhi | Opcode::Smulhi, 575 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 576 ), 577 ( 578 Opcode::UaddSat | Opcode::SaddSat | Opcode::UsubSat | Opcode::SsubSat, 579 &([I32X4, I32X4] | [I64X2, I64X2]) 580 ), 581 (Opcode::Fcopysign, &([F32X4, F32X4] | [F64X2, F64X2])), 582 (Opcode::Popcnt, &([I8X16] | [I16X8] | [I32X4] | [I64X2])), 583 ( 584 Opcode::Umax | Opcode::Smax | Opcode::Umin | Opcode::Smin, 585 &[I64X2, I64X2] 586 ), 587 // https://github.com/bytecodealliance/wasmtime/issues/6104 588 (Opcode::Bitcast, &[I128], &[_]), 589 (Opcode::Bitcast, &[_], &[I128]), 590 (Opcode::Uunarrow), 591 (Opcode::Snarrow | Opcode::Unarrow, &[I64X2, I64X2]), 592 (Opcode::SqmulRoundSat, &[I32X4, I32X4]), 593 // This Icmp is not implemented: #5529 594 (Opcode::Icmp, &[I64X2, I64X2]), 595 // IaddPairwise is implemented, but only for some types, and with some preceding ops. 596 (Opcode::IaddPairwise), 597 // Nothing wrong with this select. But we have an isle rule that can optimize it 598 // into a `min`/`max` instructions, which we don't have implemented yet. 599 (Opcode::Select, &[_, I128, I128]), 600 // These stack accesses can cause segfaults if they are merged into an SSE instruction. 601 // See: #5922 602 ( 603 Opcode::StackStore, 604 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 605 ), 606 ( 607 Opcode::StackLoad, 608 &[], 609 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 610 ), 611 // TODO 612 ( 613 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 614 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 615 ), 616 ( 617 Opcode::Rotr | Opcode::Rotl, 618 &[I8X16 | I16X8 | I32X4 | I64X2, _] 619 ), 620 ) 621 } 622 623 Architecture::Aarch64(_) => { 624 exceptions!( 625 op, 626 args, 627 rets, 628 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 629 // https://github.com/bytecodealliance/wasmtime/issues/4864 630 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 631 // https://github.com/bytecodealliance/wasmtime/issues/5472 632 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 633 // https://github.com/bytecodealliance/wasmtime/issues/4313 634 ( 635 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 636 &[I128, I128] 637 ), 638 // https://github.com/bytecodealliance/wasmtime/issues/4870 639 (Opcode::Bnot, &[F32 | F64]), 640 ( 641 Opcode::Band 642 | Opcode::Bor 643 | Opcode::Bxor 644 | Opcode::BandNot 645 | Opcode::BorNot 646 | Opcode::BxorNot, 647 &([F32, F32] | [F64, F64]) 648 ), 649 // https://github.com/bytecodealliance/wasmtime/issues/5198 650 (Opcode::Bitselect, &[I128, I128, I128]), 651 // https://github.com/bytecodealliance/wasmtime/issues/4934 652 ( 653 Opcode::FcvtToUint 654 | Opcode::FcvtToUintSat 655 | Opcode::FcvtToSint 656 | Opcode::FcvtToSintSat, 657 &[F32 | F64], 658 &[I128] 659 ), 660 // https://github.com/bytecodealliance/wasmtime/issues/4933 661 ( 662 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 663 &[I128], 664 &[F32 | F64] 665 ), 666 ( 667 Opcode::Umulhi | Opcode::Smulhi, 668 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 669 ), 670 (Opcode::Popcnt, &[I16X8 | I32X4 | I64X2]), 671 // Nothing wrong with this select. But we have an isle rule that can optimize it 672 // into a `min`/`max` instructions, which we don't have implemented yet. 673 (Opcode::Select, &[I8, I128, I128]), 674 // https://github.com/bytecodealliance/wasmtime/issues/6104 675 (Opcode::Bitcast, &[I128], &[_]), 676 (Opcode::Bitcast, &[_], &[I128]), 677 // TODO 678 ( 679 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 680 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 681 ), 682 ( 683 Opcode::Rotr | Opcode::Rotl, 684 &[I8X16 | I16X8 | I32X4 | I64X2, _] 685 ), 686 // TODO 687 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 688 (Opcode::VhighBits, &[F32X4 | F64X2]), 689 ) 690 } 691 692 Architecture::S390x => { 693 exceptions!( 694 op, 695 args, 696 rets, 697 (Opcode::UaddOverflow | Opcode::SaddOverflow), 698 (Opcode::UsubOverflow | Opcode::SsubOverflow), 699 (Opcode::UmulOverflow | Opcode::SmulOverflow), 700 ( 701 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 702 &[I128, I128] 703 ), 704 (Opcode::Bnot, &[F32 | F64]), 705 ( 706 Opcode::Band 707 | Opcode::Bor 708 | Opcode::Bxor 709 | Opcode::BandNot 710 | Opcode::BorNot 711 | Opcode::BxorNot, 712 &([F32, F32] | [F64, F64]) 713 ), 714 ( 715 Opcode::FcvtToUint 716 | Opcode::FcvtToUintSat 717 | Opcode::FcvtToSint 718 | Opcode::FcvtToSintSat, 719 &[F32 | F64], 720 &[I128] 721 ), 722 ( 723 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 724 &[I128], 725 &[F32 | F64] 726 ), 727 (Opcode::SsubSat | Opcode::SaddSat, &[I64X2, I64X2]), 728 // https://github.com/bytecodealliance/wasmtime/issues/6104 729 (Opcode::Bitcast, &[I128], &[_]), 730 (Opcode::Bitcast, &[_], &[I128]), 731 // TODO 732 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 733 ) 734 } 735 736 Architecture::Riscv64(_) => { 737 exceptions!( 738 op, 739 args, 740 rets, 741 // TODO 742 (Opcode::UaddOverflow | Opcode::SaddOverflow), 743 (Opcode::UsubOverflow | Opcode::SsubOverflow), 744 (Opcode::UmulOverflow | Opcode::SmulOverflow), 745 // TODO 746 ( 747 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 748 &[I128, I128] 749 ), 750 // TODO 751 (Opcode::Iabs, &[I128]), 752 // TODO 753 (Opcode::Bitselect, &[I128, I128, I128]), 754 // https://github.com/bytecodealliance/wasmtime/issues/5528 755 ( 756 Opcode::FcvtToUint | Opcode::FcvtToSint, 757 [F32 | F64], 758 &[I128] 759 ), 760 ( 761 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat, 762 &[F32 | F64], 763 &[I128] 764 ), 765 // https://github.com/bytecodealliance/wasmtime/issues/5528 766 ( 767 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 768 &[I128], 769 &[F32 | F64] 770 ), 771 // TODO 772 ( 773 Opcode::SelectSpectreGuard, 774 &[_, _, _], 775 &[F32 | F64 | I8X16 | I16X8 | I32X4 | I64X2 | F64X2 | F32X4] 776 ), 777 // TODO 778 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 779 ( 780 Opcode::Rotr | Opcode::Rotl, 781 &[I8X16 | I16X8 | I32X4 | I64X2, _] 782 ), 783 ) 784 } 785 786 _ => true, 787 } 788 } 789 790 type OpcodeSignature = (Opcode, Vec<Type>, Vec<Type>); 791 792 static OPCODE_SIGNATURES: LazyLock<Vec<OpcodeSignature>> = LazyLock::new(|| { 793 let types = &[ 794 I8, I16, I32, I64, I128, // Scalar Integers 795 F32, F64, // Scalar Floats 796 I8X16, I16X8, I32X4, I64X2, // SIMD Integers 797 F32X4, F64X2, // SIMD Floats 798 ]; 799 800 // When this env variable is passed, we only generate instructions for the opcodes listed in 801 // the comma-separated list. This is useful for debugging, as it allows us to focus on a few 802 // specific opcodes. 803 let allowed_opcodes = std::env::var("FUZZGEN_ALLOWED_OPS").ok().map(|s| { 804 s.split(',') 805 .map(|s| s.trim()) 806 .filter(|s| !s.is_empty()) 807 .map(|s| Opcode::from_str(s).expect("Unrecoginzed opcode")) 808 .collect::<Vec<_>>() 809 }); 810 811 Opcode::all() 812 .iter() 813 .filter(|op| { 814 match op { 815 // Control flow opcodes should not be generated through `generate_instructions`. 816 Opcode::BrTable 817 | Opcode::Brif 818 | Opcode::Jump 819 | Opcode::Return 820 | Opcode::ReturnCall 821 | Opcode::ReturnCallIndirect 822 | Opcode::TryCall 823 | Opcode::TryCallIndirect => false, 824 825 // Constants are generated outside of `generate_instructions` 826 Opcode::Iconst => false, 827 828 // TODO: extract_vector raises exceptions during return type generation because it 829 // uses dynamic vectors. 830 Opcode::ExtractVector => false, 831 832 _ => true, 833 } 834 }) 835 .flat_map(|op| { 836 let constraints = op.constraints(); 837 838 let ctrl_types = if let Some(ctrls) = constraints.ctrl_typeset() { 839 Vec::from_iter(types.iter().copied().filter(|ty| ctrls.contains(*ty))) 840 } else { 841 vec![INVALID] 842 }; 843 844 ctrl_types.into_iter().flat_map(move |ctrl_type| { 845 let rets = Vec::from_iter( 846 (0..constraints.num_fixed_results()) 847 .map(|i| constraints.result_type(i, ctrl_type)), 848 ); 849 850 // Cols is a vector whose length will match `num_fixed_value_arguments`, and whose 851 // elements will be vectors of types that are valid for that fixed argument 852 // position. 853 let mut cols = vec![]; 854 855 for i in 0..constraints.num_fixed_value_arguments() { 856 match constraints.value_argument_constraint(i, ctrl_type) { 857 ResolvedConstraint::Bound(ty) => cols.push(Vec::from([ty])), 858 ResolvedConstraint::Free(tys) => cols.push(Vec::from_iter( 859 types.iter().copied().filter(|ty| tys.contains(*ty)), 860 )), 861 } 862 } 863 864 // Generate the cartesian product of cols to produce a vector of argument lists, 865 // argss. The argss vector is seeded with the empty argument list, so there's an 866 // initial value to be extended in the loop below. 867 let mut argss = vec![vec![]]; 868 let mut cols = cols.as_slice(); 869 while let Some((col, rest)) = cols.split_last() { 870 cols = rest; 871 872 let mut next = vec![]; 873 for current in argss.iter() { 874 // Extend the front of each argument candidate with every type in `col`. 875 for ty in col { 876 let mut args = vec![*ty]; 877 args.extend_from_slice(¤t); 878 next.push(args); 879 } 880 } 881 882 let _ = std::mem::replace(&mut argss, next); 883 } 884 885 argss.into_iter().map(move |args| (*op, args, rets.clone())) 886 }) 887 }) 888 .filter(|(op, args, rets)| { 889 // These op/signature combinations need to be vetted 890 exceptions!( 891 op, 892 args.as_slice(), 893 rets.as_slice(), 894 (Opcode::Debugtrap), 895 (Opcode::Trap), 896 (Opcode::Trapz), 897 (Opcode::Trapnz), 898 (Opcode::CallIndirect, &[I32]), 899 (Opcode::FuncAddr), 900 (Opcode::X86Pshufb), 901 (Opcode::AvgRound), 902 (Opcode::Uload8x8), 903 (Opcode::Sload8x8), 904 (Opcode::Uload16x4), 905 (Opcode::Sload16x4), 906 (Opcode::Uload32x2), 907 (Opcode::Sload32x2), 908 (Opcode::StackAddr), 909 (Opcode::DynamicStackLoad), 910 (Opcode::DynamicStackStore), 911 (Opcode::DynamicStackAddr), 912 (Opcode::GlobalValue), 913 (Opcode::SymbolValue), 914 (Opcode::TlsValue), 915 (Opcode::GetPinnedReg), 916 (Opcode::SetPinnedReg), 917 (Opcode::GetFramePointer), 918 (Opcode::GetStackPointer), 919 (Opcode::GetReturnAddress), 920 (Opcode::X86Blendv), 921 (Opcode::IcmpImm), 922 (Opcode::X86Pmulhrsw), 923 (Opcode::IaddImm), 924 (Opcode::ImulImm), 925 (Opcode::UdivImm), 926 (Opcode::SdivImm), 927 (Opcode::UremImm), 928 (Opcode::SremImm), 929 (Opcode::IrsubImm), 930 (Opcode::UaddOverflowCin), 931 (Opcode::SaddOverflowCin), 932 (Opcode::UaddOverflowTrap), 933 (Opcode::UsubOverflowBin), 934 (Opcode::SsubOverflowBin), 935 (Opcode::BandImm), 936 (Opcode::BorImm), 937 (Opcode::BxorImm), 938 (Opcode::RotlImm), 939 (Opcode::RotrImm), 940 (Opcode::IshlImm), 941 (Opcode::UshrImm), 942 (Opcode::SshrImm), 943 (Opcode::ScalarToVector), 944 (Opcode::X86Pmaddubsw), 945 (Opcode::X86Cvtt2dq), 946 (Opcode::Umulhi, &[I128, I128], &[I128]), 947 (Opcode::Smulhi, &[I128, I128], &[I128]), 948 // https://github.com/bytecodealliance/wasmtime/issues/6073 949 (Opcode::Iconcat, &[I32, I32], &[I64]), 950 (Opcode::Iconcat, &[I16, I16], &[I32]), 951 (Opcode::Iconcat, &[I8, I8], &[I16]), 952 // https://github.com/bytecodealliance/wasmtime/issues/6073 953 (Opcode::Isplit, &[I64], &[I32, I32]), 954 (Opcode::Isplit, &[I32], &[I16, I16]), 955 (Opcode::Isplit, &[I16], &[I8, I8]), 956 (Opcode::Fmin, &[F32X4, F32X4], &[F32X4]), 957 (Opcode::Fmin, &[F64X2, F64X2], &[F64X2]), 958 (Opcode::Fmax, &[F32X4, F32X4], &[F32X4]), 959 (Opcode::Fmax, &[F64X2, F64X2], &[F64X2]), 960 (Opcode::FcvtToUintSat, &[F32X4], &[I8]), 961 (Opcode::FcvtToUintSat, &[F64X2], &[I8]), 962 (Opcode::FcvtToUintSat, &[F32X4], &[I16]), 963 (Opcode::FcvtToUintSat, &[F64X2], &[I16]), 964 (Opcode::FcvtToUintSat, &[F32X4], &[I32]), 965 (Opcode::FcvtToUintSat, &[F64X2], &[I32]), 966 (Opcode::FcvtToUintSat, &[F32X4], &[I64]), 967 (Opcode::FcvtToUintSat, &[F64X2], &[I64]), 968 (Opcode::FcvtToUintSat, &[F32X4], &[I128]), 969 (Opcode::FcvtToUintSat, &[F64X2], &[I128]), 970 (Opcode::FcvtToUintSat, &[F32], &[I8X16]), 971 (Opcode::FcvtToUintSat, &[F64], &[I8X16]), 972 (Opcode::FcvtToUintSat, &[F32X4], &[I8X16]), 973 (Opcode::FcvtToUintSat, &[F64X2], &[I8X16]), 974 (Opcode::FcvtToUintSat, &[F32], &[I16X8]), 975 (Opcode::FcvtToUintSat, &[F64], &[I16X8]), 976 (Opcode::FcvtToUintSat, &[F32X4], &[I16X8]), 977 (Opcode::FcvtToUintSat, &[F64X2], &[I16X8]), 978 (Opcode::FcvtToUintSat, &[F32], &[I32X4]), 979 (Opcode::FcvtToUintSat, &[F64], &[I32X4]), 980 (Opcode::FcvtToUintSat, &[F64X2], &[I32X4]), 981 (Opcode::FcvtToUintSat, &[F32], &[I64X2]), 982 (Opcode::FcvtToUintSat, &[F64], &[I64X2]), 983 (Opcode::FcvtToUintSat, &[F32X4], &[I64X2]), 984 (Opcode::FcvtToSintSat, &[F32X4], &[I8]), 985 (Opcode::FcvtToSintSat, &[F64X2], &[I8]), 986 (Opcode::FcvtToSintSat, &[F32X4], &[I16]), 987 (Opcode::FcvtToSintSat, &[F64X2], &[I16]), 988 (Opcode::FcvtToSintSat, &[F32X4], &[I32]), 989 (Opcode::FcvtToSintSat, &[F64X2], &[I32]), 990 (Opcode::FcvtToSintSat, &[F32X4], &[I64]), 991 (Opcode::FcvtToSintSat, &[F64X2], &[I64]), 992 (Opcode::FcvtToSintSat, &[F32X4], &[I128]), 993 (Opcode::FcvtToSintSat, &[F64X2], &[I128]), 994 (Opcode::FcvtToSintSat, &[F32], &[I8X16]), 995 (Opcode::FcvtToSintSat, &[F64], &[I8X16]), 996 (Opcode::FcvtToSintSat, &[F32X4], &[I8X16]), 997 (Opcode::FcvtToSintSat, &[F64X2], &[I8X16]), 998 (Opcode::FcvtToSintSat, &[F32], &[I16X8]), 999 (Opcode::FcvtToSintSat, &[F64], &[I16X8]), 1000 (Opcode::FcvtToSintSat, &[F32X4], &[I16X8]), 1001 (Opcode::FcvtToSintSat, &[F64X2], &[I16X8]), 1002 (Opcode::FcvtToSintSat, &[F32], &[I32X4]), 1003 (Opcode::FcvtToSintSat, &[F64], &[I32X4]), 1004 (Opcode::FcvtToSintSat, &[F64X2], &[I32X4]), 1005 (Opcode::FcvtToSintSat, &[F32], &[I64X2]), 1006 (Opcode::FcvtToSintSat, &[F64], &[I64X2]), 1007 (Opcode::FcvtToSintSat, &[F32X4], &[I64X2]), 1008 (Opcode::FcvtFromUint, &[I8X16], &[F32]), 1009 (Opcode::FcvtFromUint, &[I16X8], &[F32]), 1010 (Opcode::FcvtFromUint, &[I32X4], &[F32]), 1011 (Opcode::FcvtFromUint, &[I64X2], &[F32]), 1012 (Opcode::FcvtFromUint, &[I8X16], &[F64]), 1013 (Opcode::FcvtFromUint, &[I16X8], &[F64]), 1014 (Opcode::FcvtFromUint, &[I32X4], &[F64]), 1015 (Opcode::FcvtFromUint, &[I64X2], &[F64]), 1016 (Opcode::FcvtFromUint, &[I8], &[F32X4]), 1017 (Opcode::FcvtFromUint, &[I16], &[F32X4]), 1018 (Opcode::FcvtFromUint, &[I32], &[F32X4]), 1019 (Opcode::FcvtFromUint, &[I64], &[F32X4]), 1020 (Opcode::FcvtFromUint, &[I128], &[F32X4]), 1021 (Opcode::FcvtFromUint, &[I8X16], &[F32X4]), 1022 (Opcode::FcvtFromUint, &[I16X8], &[F32X4]), 1023 (Opcode::FcvtFromUint, &[I64X2], &[F32X4]), 1024 (Opcode::FcvtFromUint, &[I8], &[F64X2]), 1025 (Opcode::FcvtFromUint, &[I16], &[F64X2]), 1026 (Opcode::FcvtFromUint, &[I32], &[F64X2]), 1027 (Opcode::FcvtFromUint, &[I64], &[F64X2]), 1028 (Opcode::FcvtFromUint, &[I128], &[F64X2]), 1029 (Opcode::FcvtFromUint, &[I8X16], &[F64X2]), 1030 (Opcode::FcvtFromUint, &[I16X8], &[F64X2]), 1031 (Opcode::FcvtFromUint, &[I32X4], &[F64X2]), 1032 (Opcode::FcvtFromSint, &[I8X16], &[F32]), 1033 (Opcode::FcvtFromSint, &[I16X8], &[F32]), 1034 (Opcode::FcvtFromSint, &[I32X4], &[F32]), 1035 (Opcode::FcvtFromSint, &[I64X2], &[F32]), 1036 (Opcode::FcvtFromSint, &[I8X16], &[F64]), 1037 (Opcode::FcvtFromSint, &[I16X8], &[F64]), 1038 (Opcode::FcvtFromSint, &[I32X4], &[F64]), 1039 (Opcode::FcvtFromSint, &[I64X2], &[F64]), 1040 (Opcode::FcvtFromSint, &[I8], &[F32X4]), 1041 (Opcode::FcvtFromSint, &[I16], &[F32X4]), 1042 (Opcode::FcvtFromSint, &[I32], &[F32X4]), 1043 (Opcode::FcvtFromSint, &[I64], &[F32X4]), 1044 (Opcode::FcvtFromSint, &[I128], &[F32X4]), 1045 (Opcode::FcvtFromSint, &[I8X16], &[F32X4]), 1046 (Opcode::FcvtFromSint, &[I16X8], &[F32X4]), 1047 (Opcode::FcvtFromSint, &[I64X2], &[F32X4]), 1048 (Opcode::FcvtFromSint, &[I8], &[F64X2]), 1049 (Opcode::FcvtFromSint, &[I16], &[F64X2]), 1050 (Opcode::FcvtFromSint, &[I32], &[F64X2]), 1051 (Opcode::FcvtFromSint, &[I64], &[F64X2]), 1052 (Opcode::FcvtFromSint, &[I128], &[F64X2]), 1053 (Opcode::FcvtFromSint, &[I8X16], &[F64X2]), 1054 (Opcode::FcvtFromSint, &[I16X8], &[F64X2]), 1055 (Opcode::FcvtFromSint, &[I32X4], &[F64X2]), 1056 // Only supported on x64 with a feature at this time, so 128-bit 1057 // atomics are not suitable to fuzz yet. 1058 (Opcode::AtomicRmw, _, &[I128]), 1059 (Opcode::AtomicCas, _, &[I128]), 1060 (Opcode::AtomicLoad, _, &[I128]), 1061 (Opcode::AtomicStore, &[I128, _], _), 1062 (Opcode::SequencePoint), 1063 ) 1064 }) 1065 .filter(|(op, ..)| { 1066 allowed_opcodes 1067 .as_ref() 1068 .map_or(true, |opcodes| opcodes.contains(op)) 1069 }) 1070 .collect() 1071 }); 1072 1073 fn inserter_for_format(fmt: InstructionFormat) -> OpcodeInserter { 1074 match fmt { 1075 InstructionFormat::AtomicCas => insert_atomic_cas, 1076 InstructionFormat::AtomicRmw => insert_atomic_rmw, 1077 InstructionFormat::Binary => insert_opcode, 1078 InstructionFormat::BinaryImm64 => todo!(), 1079 InstructionFormat::BinaryImm8 => insert_ins_ext_lane, 1080 InstructionFormat::Call => insert_call, 1081 InstructionFormat::CallIndirect => insert_call, 1082 InstructionFormat::CondTrap => todo!(), 1083 InstructionFormat::DynamicStackLoad => todo!(), 1084 InstructionFormat::DynamicStackStore => todo!(), 1085 InstructionFormat::FloatCompare => insert_cmp, 1086 InstructionFormat::FuncAddr => todo!(), 1087 InstructionFormat::IntAddTrap => todo!(), 1088 InstructionFormat::IntCompare => insert_cmp, 1089 InstructionFormat::IntCompareImm => todo!(), 1090 InstructionFormat::Load => insert_load_store, 1091 InstructionFormat::LoadNoOffset => insert_load_store, 1092 InstructionFormat::NullAry => insert_opcode, 1093 InstructionFormat::Shuffle => insert_shuffle, 1094 InstructionFormat::StackLoad => insert_stack_load, 1095 InstructionFormat::StackStore => insert_stack_store, 1096 InstructionFormat::Store => insert_load_store, 1097 InstructionFormat::StoreNoOffset => insert_load_store, 1098 InstructionFormat::Ternary => insert_opcode, 1099 InstructionFormat::TernaryImm8 => insert_ins_ext_lane, 1100 InstructionFormat::Trap => todo!(), 1101 InstructionFormat::Unary => insert_opcode, 1102 InstructionFormat::UnaryConst => insert_const, 1103 InstructionFormat::UnaryGlobalValue => todo!(), 1104 InstructionFormat::UnaryIeee16 => insert_const, 1105 InstructionFormat::UnaryIeee32 => insert_const, 1106 InstructionFormat::UnaryIeee64 => insert_const, 1107 InstructionFormat::UnaryImm => insert_const, 1108 InstructionFormat::ExceptionHandlerAddress => insert_const, 1109 1110 InstructionFormat::BranchTable 1111 | InstructionFormat::Brif 1112 | InstructionFormat::Jump 1113 | InstructionFormat::MultiAry 1114 | InstructionFormat::TryCall 1115 | InstructionFormat::TryCallIndirect => { 1116 panic!("Control-flow instructions should be handled by 'insert_terminator': {fmt:?}") 1117 } 1118 } 1119 } 1120 1121 pub struct FunctionGenerator<'r, 'data> 1122 where 1123 'data: 'r, 1124 { 1125 u: &'r mut Unstructured<'data>, 1126 config: &'r Config, 1127 resources: Resources, 1128 isa: OwnedTargetIsa, 1129 name: UserFuncName, 1130 signature: Signature, 1131 } 1132 1133 #[derive(Debug, Clone)] 1134 enum BlockTerminator { 1135 Return, 1136 Jump(Block), 1137 Br(Block, Block), 1138 BrTable(Block, Vec<Block>), 1139 Switch(Type, Block, HashMap<u128, Block>), 1140 TailCall(FuncRef), 1141 TailCallIndirect(FuncRef), 1142 } 1143 1144 #[derive(Debug, Clone)] 1145 enum BlockTerminatorKind { 1146 Return, 1147 Jump, 1148 Br, 1149 BrTable, 1150 Switch, 1151 TailCall, 1152 TailCallIndirect, 1153 } 1154 1155 /// Alias Analysis Category 1156 /// 1157 /// Our alias analysis pass supports 4 categories of accesses to distinguish 1158 /// different regions. The "Other" region is the general case, and is the default 1159 /// Although they have highly suggestive names there is no difference between any 1160 /// of the categories. 1161 /// 1162 /// We assign each stack slot a category when we first generate them, and then 1163 /// ensure that all accesses to that stack slot are correctly tagged. We already 1164 /// ensure that memory accesses never cross stack slots, so there is no risk 1165 /// of a memory access being tagged with the wrong category. 1166 #[derive(Debug, PartialEq, Clone, Copy)] 1167 enum AACategory { 1168 Other, 1169 Heap, 1170 Table, 1171 VmCtx, 1172 } 1173 1174 impl AACategory { 1175 pub fn all() -> &'static [Self] { 1176 &[ 1177 AACategory::Other, 1178 AACategory::Heap, 1179 AACategory::Table, 1180 AACategory::VmCtx, 1181 ] 1182 } 1183 1184 pub fn update_memflags(&self, flags: &mut MemFlags) { 1185 flags.set_alias_region(match self { 1186 AACategory::Other => None, 1187 AACategory::Heap => Some(AliasRegion::Heap), 1188 AACategory::Table => Some(AliasRegion::Table), 1189 AACategory::VmCtx => Some(AliasRegion::Vmctx), 1190 }) 1191 } 1192 } 1193 1194 pub type StackAlignment = StackSize; 1195 1196 #[derive(Default)] 1197 struct Resources { 1198 vars: HashMap<Type, Vec<Variable>>, 1199 blocks: Vec<(Block, BlockSignature)>, 1200 blocks_without_params: Vec<Block>, 1201 block_terminators: Vec<BlockTerminator>, 1202 func_refs: Vec<(Signature, SigRef, FuncRef)>, 1203 /// This field is required to be sorted by stack slot size at all times. 1204 /// We use this invariant when searching for stack slots with a given size. 1205 /// See [FunctionGenerator::stack_slot_with_size] 1206 stack_slots: Vec<(StackSlot, StackSize, StackAlignment, AACategory)>, 1207 usercalls: Vec<(UserExternalName, Signature)>, 1208 libcalls: Vec<LibCall>, 1209 } 1210 1211 impl Resources { 1212 /// Partitions blocks at `block`. Only blocks that can be targeted by branches are considered. 1213 /// 1214 /// The first slice includes all blocks up to and including `block`. 1215 /// The second slice includes all remaining blocks. 1216 fn partition_target_blocks( 1217 &self, 1218 block: Block, 1219 ) -> (&[(Block, BlockSignature)], &[(Block, BlockSignature)]) { 1220 // Blocks are stored in-order and have no gaps, this means that we can simply index them by 1221 // their number. We also need to exclude the entry block since it isn't a valid target. 1222 let target_blocks = &self.blocks[1..]; 1223 target_blocks.split_at(block.as_u32() as usize) 1224 } 1225 1226 /// Returns blocks forward of `block`. Only blocks that can be targeted by branches are considered. 1227 fn forward_blocks(&self, block: Block) -> &[(Block, BlockSignature)] { 1228 let (_, forward_blocks) = self.partition_target_blocks(block); 1229 forward_blocks 1230 } 1231 1232 /// Generates a slice of `blocks_without_params` ahead of `block` 1233 fn forward_blocks_without_params(&self, block: Block) -> &[Block] { 1234 let partition_point = self.blocks_without_params.partition_point(|b| *b <= block); 1235 &self.blocks_without_params[partition_point..] 1236 } 1237 1238 /// Generates an iterator of all valid tail call targets. This includes all functions with both 1239 /// the `tail` calling convention and the same return values as the caller. 1240 fn tail_call_targets<'a>( 1241 &'a self, 1242 caller_sig: &'a Signature, 1243 ) -> impl Iterator<Item = &'a (Signature, SigRef, FuncRef)> { 1244 self.func_refs.iter().filter(|(sig, _, _)| { 1245 sig.call_conv == CallConv::Tail && sig.returns == caller_sig.returns 1246 }) 1247 } 1248 } 1249 1250 impl<'r, 'data> FunctionGenerator<'r, 'data> 1251 where 1252 'data: 'r, 1253 { 1254 pub fn new( 1255 u: &'r mut Unstructured<'data>, 1256 config: &'r Config, 1257 isa: OwnedTargetIsa, 1258 name: UserFuncName, 1259 signature: Signature, 1260 usercalls: Vec<(UserExternalName, Signature)>, 1261 libcalls: Vec<LibCall>, 1262 ) -> Self { 1263 Self { 1264 u, 1265 config, 1266 resources: Resources { 1267 usercalls, 1268 libcalls, 1269 ..Resources::default() 1270 }, 1271 isa, 1272 name, 1273 signature, 1274 } 1275 } 1276 1277 /// Generates a random value for config `param` 1278 fn param(&mut self, param: &RangeInclusive<usize>) -> Result<usize> { 1279 Ok(self.u.int_in_range(param.clone())?) 1280 } 1281 1282 fn system_callconv(&mut self) -> CallConv { 1283 // TODO: This currently only runs on linux, so this is the only choice 1284 // We should improve this once we generate flags and targets 1285 CallConv::SystemV 1286 } 1287 1288 /// Finds a stack slot with size of at least n bytes 1289 fn stack_slot_with_size( 1290 &mut self, 1291 n: u32, 1292 ) -> Result<(StackSlot, StackSize, StackAlignment, AACategory)> { 1293 let first = self 1294 .resources 1295 .stack_slots 1296 .partition_point(|&(_slot, size, _align, _category)| size < n); 1297 Ok(*self.u.choose(&self.resources.stack_slots[first..])?) 1298 } 1299 1300 /// Generates an address that should allow for a store or a load. 1301 /// 1302 /// Addresses aren't generated like other values. They are never stored in variables so that 1303 /// we don't run the risk of returning them from a function, which would make the fuzzer 1304 /// complain since they are different from the interpreter to the backend. 1305 /// 1306 /// `min_size`: Controls the amount of space that the address should have. 1307 /// 1308 /// `aligned`: When passed as true, the resulting address is guaranteed to be aligned 1309 /// on an 8 byte boundary. 1310 /// 1311 /// Returns a valid address and the maximum possible offset that still respects `min_size`. 1312 fn generate_load_store_address( 1313 &mut self, 1314 builder: &mut FunctionBuilder, 1315 min_size: u32, 1316 aligned: bool, 1317 ) -> Result<(Value, u32, AACategory)> { 1318 // TODO: Currently our only source of addresses is stack_addr, but we 1319 // should add global_value, symbol_value eventually 1320 let (addr, available_size, category) = { 1321 let (ss, slot_size, _align, category) = self.stack_slot_with_size(min_size)?; 1322 1323 // stack_slot_with_size guarantees that slot_size >= min_size 1324 let max_offset = slot_size - min_size; 1325 let offset = if aligned { 1326 self.u.int_in_range(0..=max_offset / min_size)? * min_size 1327 } else { 1328 self.u.int_in_range(0..=max_offset)? 1329 }; 1330 1331 let base_addr = builder.ins().stack_addr(I64, ss, offset as i32); 1332 let available_size = slot_size.saturating_sub(offset); 1333 (base_addr, available_size, category) 1334 }; 1335 1336 // TODO: Insert a bunch of amode opcodes here to modify the address! 1337 1338 // Now that we have an address and a size, we just choose a random offset to return to the 1339 // caller. Preserving min_size bytes. 1340 let max_offset = available_size.saturating_sub(min_size); 1341 Ok((addr, max_offset, category)) 1342 } 1343 1344 // Generates an address and memflags for a load or store. 1345 fn generate_address_and_memflags( 1346 &mut self, 1347 builder: &mut FunctionBuilder, 1348 min_size: u32, 1349 is_atomic: bool, 1350 ) -> Result<(Value, MemFlags, Offset32)> { 1351 // Should we generate an aligned address 1352 // Some backends have issues with unaligned atomics. 1353 // AArch64: https://github.com/bytecodealliance/wasmtime/issues/5483 1354 // RISCV: https://github.com/bytecodealliance/wasmtime/issues/5882 1355 let requires_aligned_atomics = matches!( 1356 self.isa.triple().architecture, 1357 Architecture::Aarch64(_) | Architecture::Riscv64(_) 1358 ); 1359 let aligned = if is_atomic && requires_aligned_atomics { 1360 true 1361 } else if min_size > 8 { 1362 // TODO: We currently can't guarantee that a stack_slot will be aligned on a 16 byte 1363 // boundary. We don't have a way to specify alignment when creating stack slots, and 1364 // cranelift only guarantees 8 byte alignment between stack slots. 1365 // See: https://github.com/bytecodealliance/wasmtime/issues/5922#issuecomment-1457926624 1366 false 1367 } else { 1368 bool::arbitrary(self.u)? 1369 }; 1370 1371 let mut flags = MemFlags::new(); 1372 // Even if we picked an aligned address, we can always generate unaligned memflags 1373 if aligned && bool::arbitrary(self.u)? { 1374 flags.set_aligned(); 1375 } 1376 // If the address is aligned, then we know it won't trap 1377 if aligned && bool::arbitrary(self.u)? { 1378 flags.set_notrap(); 1379 } 1380 1381 let (address, max_offset, category) = 1382 self.generate_load_store_address(builder, min_size, aligned)?; 1383 1384 // Set the Alias Analysis bits on the memflags 1385 category.update_memflags(&mut flags); 1386 1387 // Pick an offset to pass into the load/store. 1388 let offset = if aligned { 1389 0 1390 } else { 1391 self.u.int_in_range(0..=max_offset)? as i32 1392 } 1393 .into(); 1394 1395 Ok((address, flags, offset)) 1396 } 1397 1398 /// Get a variable of type `ty` from the current function 1399 fn get_variable_of_type(&mut self, ty: Type) -> Result<Variable> { 1400 let opts = self.resources.vars.get(&ty).map_or(&[][..], Vec::as_slice); 1401 let var = self.u.choose(opts)?; 1402 Ok(*var) 1403 } 1404 1405 /// Generates an instruction(`iconst`/`fconst`/etc...) to introduce a constant value 1406 fn generate_const(&mut self, builder: &mut FunctionBuilder, ty: Type) -> Result<Value> { 1407 Ok(match self.u.datavalue(ty)? { 1408 DataValue::I8(i) => builder.ins().iconst(ty, i as u8 as i64), 1409 DataValue::I16(i) => builder.ins().iconst(ty, i as u16 as i64), 1410 DataValue::I32(i) => builder.ins().iconst(ty, i as u32 as i64), 1411 DataValue::I64(i) => builder.ins().iconst(ty, i), 1412 DataValue::I128(i) => { 1413 let hi = builder.ins().iconst(I64, (i >> 64) as i64); 1414 let lo = builder.ins().iconst(I64, i as i64); 1415 builder.ins().iconcat(lo, hi) 1416 } 1417 DataValue::F16(f) => builder.ins().f16const(f), 1418 DataValue::F32(f) => builder.ins().f32const(f), 1419 DataValue::F64(f) => builder.ins().f64const(f), 1420 DataValue::F128(f) => { 1421 let handle = builder.func.dfg.constants.insert(f.into()); 1422 builder.ins().f128const(handle) 1423 } 1424 DataValue::V128(bytes) => { 1425 let data = bytes.to_vec().into(); 1426 let handle = builder.func.dfg.constants.insert(data); 1427 builder.ins().vconst(ty, handle) 1428 } 1429 _ => unimplemented!(), 1430 }) 1431 } 1432 1433 /// Chooses a random block which can be targeted by a jump / branch. 1434 /// This means any block that is not the first block. 1435 fn generate_target_block(&mut self, source_block: Block) -> Result<Block> { 1436 // We try to mostly generate forward branches to avoid generating an excessive amount of 1437 // infinite loops. But they are still important, so give them a small chance of existing. 1438 let (backwards_blocks, forward_blocks) = 1439 self.resources.partition_target_blocks(source_block); 1440 let ratio = self.config.backwards_branch_ratio; 1441 let block_targets = if !backwards_blocks.is_empty() && self.u.ratio(ratio.0, ratio.1)? { 1442 backwards_blocks 1443 } else { 1444 forward_blocks 1445 }; 1446 assert!(!block_targets.is_empty()); 1447 1448 let (block, _) = self.u.choose(block_targets)?.clone(); 1449 Ok(block) 1450 } 1451 1452 fn generate_values_for_block( 1453 &mut self, 1454 builder: &mut FunctionBuilder, 1455 block: Block, 1456 ) -> Result<Vec<BlockArg>> { 1457 let (_, sig) = self.resources.blocks[block.as_u32() as usize].clone(); 1458 Ok(self 1459 .generate_values_for_signature(builder, sig.iter().copied())? 1460 .into_iter() 1461 .map(|val| BlockArg::Value(val)) 1462 .collect::<Vec<_>>()) 1463 } 1464 1465 fn generate_values_for_signature<I: Iterator<Item = Type>>( 1466 &mut self, 1467 builder: &mut FunctionBuilder, 1468 signature: I, 1469 ) -> Result<Vec<Value>> { 1470 signature 1471 .map(|ty| { 1472 let var = self.get_variable_of_type(ty)?; 1473 let val = builder.use_var(var); 1474 Ok(val) 1475 }) 1476 .collect() 1477 } 1478 1479 /// The terminator that we need to insert has already been picked ahead of time 1480 /// we just need to build the instructions for it 1481 fn insert_terminator( 1482 &mut self, 1483 builder: &mut FunctionBuilder, 1484 source_block: Block, 1485 ) -> Result<()> { 1486 let terminator = self.resources.block_terminators[source_block.as_u32() as usize].clone(); 1487 1488 match terminator { 1489 BlockTerminator::Return => { 1490 let types: Vec<Type> = { 1491 let rets = &builder.func.signature.returns; 1492 rets.iter().map(|p| p.value_type).collect() 1493 }; 1494 let vals = self.generate_values_for_signature(builder, types.into_iter())?; 1495 1496 builder.ins().return_(&vals[..]); 1497 } 1498 BlockTerminator::Jump(target) => { 1499 let args = self.generate_values_for_block(builder, target)?; 1500 builder.ins().jump(target, &args[..]); 1501 } 1502 BlockTerminator::Br(left, right) => { 1503 let left_args = self.generate_values_for_block(builder, left)?; 1504 let right_args = self.generate_values_for_block(builder, right)?; 1505 1506 let condbr_types = [I8, I16, I32, I64, I128]; 1507 let _type = *self.u.choose(&condbr_types[..])?; 1508 let val = builder.use_var(self.get_variable_of_type(_type)?); 1509 builder 1510 .ins() 1511 .brif(val, left, &left_args[..], right, &right_args[..]); 1512 } 1513 BlockTerminator::BrTable(default, targets) => { 1514 // Create jump tables on demand 1515 let mut jt = Vec::with_capacity(targets.len()); 1516 for block in targets { 1517 let args = self.generate_values_for_block(builder, block)?; 1518 jt.push(builder.func.dfg.block_call(block, &args)) 1519 } 1520 1521 let args = self.generate_values_for_block(builder, default)?; 1522 let jt_data = JumpTableData::new(builder.func.dfg.block_call(default, &args), &jt); 1523 let jt = builder.create_jump_table(jt_data); 1524 1525 // br_table only supports I32 1526 let val = builder.use_var(self.get_variable_of_type(I32)?); 1527 1528 builder.ins().br_table(val, jt); 1529 } 1530 BlockTerminator::Switch(_type, default, entries) => { 1531 let mut switch = Switch::new(); 1532 for (&entry, &block) in entries.iter() { 1533 switch.set_entry(entry, block); 1534 } 1535 1536 let switch_val = builder.use_var(self.get_variable_of_type(_type)?); 1537 1538 switch.emit(builder, switch_val, default); 1539 } 1540 BlockTerminator::TailCall(target) | BlockTerminator::TailCallIndirect(target) => { 1541 let (sig, sig_ref, func_ref) = self 1542 .resources 1543 .func_refs 1544 .iter() 1545 .find(|(_, _, f)| *f == target) 1546 .expect("Failed to find previously selected function") 1547 .clone(); 1548 1549 let opcode = match terminator { 1550 BlockTerminator::TailCall(_) => Opcode::ReturnCall, 1551 BlockTerminator::TailCallIndirect(_) => Opcode::ReturnCallIndirect, 1552 _ => unreachable!(), 1553 }; 1554 1555 insert_call_to_function(self, builder, opcode, &sig, sig_ref, func_ref)?; 1556 } 1557 } 1558 1559 Ok(()) 1560 } 1561 1562 /// Fills the current block with random instructions 1563 fn generate_instructions(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1564 for _ in 0..self.param(&self.config.instructions_per_block)? { 1565 let (op, args, rets) = self.u.choose(&OPCODE_SIGNATURES)?; 1566 1567 // We filter out instructions that aren't supported by the target at this point instead 1568 // of building a single vector of valid instructions at the beginning of function 1569 // generation, to avoid invalidating the corpus when instructions are enabled/disabled. 1570 if !valid_for_target(&self.isa.triple(), *op, &args, &rets) { 1571 return Err(arbitrary::Error::IncorrectFormat.into()); 1572 } 1573 1574 let inserter = inserter_for_format(op.format()); 1575 inserter(self, builder, *op, &args, &rets)?; 1576 } 1577 1578 Ok(()) 1579 } 1580 1581 fn generate_funcrefs(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1582 let usercalls: Vec<_> = self 1583 .resources 1584 .usercalls 1585 .iter() 1586 .map(|(name, signature)| { 1587 let user_func_ref = builder.func.declare_imported_user_function(name.clone()); 1588 let name = ExternalName::User(user_func_ref); 1589 (name, signature.clone()) 1590 }) 1591 .collect(); 1592 1593 let lib_callconv = self.system_callconv(); 1594 let libcalls: Vec<_> = self 1595 .resources 1596 .libcalls 1597 .iter() 1598 .map(|libcall| { 1599 let pointer_type = Type::int_with_byte_size( 1600 self.isa.triple().pointer_width().unwrap().bytes().into(), 1601 ) 1602 .unwrap(); 1603 let signature = libcall.signature(lib_callconv, pointer_type); 1604 let name = ExternalName::LibCall(*libcall); 1605 (name, signature) 1606 }) 1607 .collect(); 1608 1609 for (name, signature) in usercalls.into_iter().chain(libcalls) { 1610 let sig_ref = builder.import_signature(signature.clone()); 1611 let func_ref = builder.import_function(ExtFuncData { 1612 name, 1613 signature: sig_ref, 1614 1615 // Libcalls can't be colocated because they can be very far away 1616 // from allocated memory at runtime, and additionally at this 1617 // time cranelift-jit puts all functions in their own mmap so 1618 // they also cannot be colocated. 1619 colocated: 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