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 types::*, AliasRegion, AtomicRmwOp, Block, ConstantData, Endianness, ExternalName, FuncRef, 13 Function, LibCall, Opcode, SigRef, Signature, StackSlot, UserExternalName, UserFuncName, Value, 14 }; 15 use cranelift::codegen::isa::CallConv; 16 use cranelift::frontend::{FunctionBuilder, FunctionBuilderContext, Switch, Variable}; 17 use cranelift::prelude::isa::OwnedTargetIsa; 18 use cranelift::prelude::{ 19 EntityRef, ExtFuncData, FloatCC, InstBuilder, IntCC, JumpTableData, MemFlags, StackSlotData, 20 StackSlotKind, 21 }; 22 use once_cell::sync::Lazy; 23 use std::collections::HashMap; 24 use std::ops::RangeInclusive; 25 use std::str::FromStr; 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, category) = fgen.stack_slot_with_size(type_size)?; 136 137 // `stack_load` doesen'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, category) = fgen.stack_slot_with_size(type_size)?; 163 164 // `stack_store` doesen'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 _ => {} 516 } 517 518 match triple.architecture { 519 Architecture::X86_64 => { 520 exceptions!( 521 op, 522 args, 523 rets, 524 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 525 (Opcode::Imul, &[I8X16, I8X16]), 526 // https://github.com/bytecodealliance/wasmtime/issues/4756 527 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 528 // https://github.com/bytecodealliance/wasmtime/issues/5474 529 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 530 // https://github.com/bytecodealliance/wasmtime/issues/3370 531 ( 532 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 533 &[I128, I128] 534 ), 535 // https://github.com/bytecodealliance/wasmtime/issues/5107 536 (Opcode::Cls, &[I8], &[I8]), 537 (Opcode::Cls, &[I16], &[I16]), 538 (Opcode::Cls, &[I32], &[I32]), 539 (Opcode::Cls, &[I64], &[I64]), 540 (Opcode::Cls, &[I128], &[I128]), 541 // TODO 542 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 543 // https://github.com/bytecodealliance/wasmtime/issues/4897 544 // https://github.com/bytecodealliance/wasmtime/issues/4899 545 ( 546 Opcode::FcvtToUint 547 | Opcode::FcvtToUintSat 548 | Opcode::FcvtToSint 549 | Opcode::FcvtToSintSat, 550 &[F32 | F64], 551 &[I8 | I16 | I128] 552 ), 553 (Opcode::FcvtToUint | Opcode::FcvtToSint, &[F32X4], &[I32X4]), 554 ( 555 Opcode::FcvtToUint 556 | Opcode::FcvtToUintSat 557 | Opcode::FcvtToSint 558 | Opcode::FcvtToSintSat, 559 &[F64X2], 560 &[I64X2] 561 ), 562 // https://github.com/bytecodealliance/wasmtime/issues/4900 563 (Opcode::FcvtFromUint, &[I128], &[F32 | F64]), 564 // This has a lowering, but only when preceded by `uwiden_low`. 565 (Opcode::FcvtFromUint, &[I64X2], &[F64X2]), 566 // https://github.com/bytecodealliance/wasmtime/issues/4900 567 (Opcode::FcvtFromSint, &[I128], &[F32 | F64]), 568 (Opcode::FcvtFromSint, &[I64X2], &[F64X2]), 569 ( 570 Opcode::Umulhi | Opcode::Smulhi, 571 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 572 ), 573 ( 574 Opcode::UaddSat | Opcode::SaddSat | Opcode::UsubSat | Opcode::SsubSat, 575 &([I32X4, I32X4] | [I64X2, I64X2]) 576 ), 577 (Opcode::Fcopysign, &([F32X4, F32X4] | [F64X2, F64X2])), 578 (Opcode::Popcnt, &([I8X16] | [I16X8] | [I32X4] | [I64X2])), 579 ( 580 Opcode::Umax | Opcode::Smax | Opcode::Umin | Opcode::Smin, 581 &[I64X2, I64X2] 582 ), 583 // https://github.com/bytecodealliance/wasmtime/issues/6104 584 (Opcode::Bitcast, &[I128], &[_]), 585 (Opcode::Bitcast, &[_], &[I128]), 586 (Opcode::Uunarrow), 587 (Opcode::Snarrow | Opcode::Unarrow, &[I64X2, I64X2]), 588 (Opcode::SqmulRoundSat, &[I32X4, I32X4]), 589 // This Icmp is not implemented: #5529 590 (Opcode::Icmp, &[I64X2, I64X2]), 591 // IaddPairwise is implemented, but only for some types, and with some preceding ops. 592 (Opcode::IaddPairwise), 593 // Nothing wrong with this select. But we have an isle rule that can optimize it 594 // into a `min`/`max` instructions, which we don't have implemented yet. 595 (Opcode::Select, &[_, I128, I128]), 596 // These stack accesses can cause segfaults if they are merged into an SSE instruction. 597 // See: #5922 598 ( 599 Opcode::StackStore, 600 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 601 ), 602 ( 603 Opcode::StackLoad, 604 &[], 605 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 606 ), 607 // TODO 608 ( 609 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 610 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 611 ), 612 ( 613 Opcode::Rotr | Opcode::Rotl, 614 &[I8X16 | I16X8 | I32X4 | I64X2, _] 615 ), 616 ) 617 } 618 619 Architecture::Aarch64(_) => { 620 exceptions!( 621 op, 622 args, 623 rets, 624 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 625 // https://github.com/bytecodealliance/wasmtime/issues/4864 626 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 627 // https://github.com/bytecodealliance/wasmtime/issues/5472 628 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 629 // https://github.com/bytecodealliance/wasmtime/issues/4313 630 ( 631 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 632 &[I128, I128] 633 ), 634 // https://github.com/bytecodealliance/wasmtime/issues/4870 635 (Opcode::Bnot, &[F32 | F64]), 636 ( 637 Opcode::Band 638 | Opcode::Bor 639 | Opcode::Bxor 640 | Opcode::BandNot 641 | Opcode::BorNot 642 | Opcode::BxorNot, 643 &([F32, F32] | [F64, F64]) 644 ), 645 // https://github.com/bytecodealliance/wasmtime/issues/5198 646 (Opcode::Bitselect, &[I128, I128, I128]), 647 // https://github.com/bytecodealliance/wasmtime/issues/4934 648 ( 649 Opcode::FcvtToUint 650 | Opcode::FcvtToUintSat 651 | Opcode::FcvtToSint 652 | Opcode::FcvtToSintSat, 653 &[F32 | F64], 654 &[I128] 655 ), 656 // https://github.com/bytecodealliance/wasmtime/issues/4933 657 ( 658 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 659 &[I128], 660 &[F32 | F64] 661 ), 662 ( 663 Opcode::Umulhi | Opcode::Smulhi, 664 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 665 ), 666 (Opcode::Popcnt, &[I16X8 | I32X4 | I64X2]), 667 // Nothing wrong with this select. But we have an isle rule that can optimize it 668 // into a `min`/`max` instructions, which we don't have implemented yet. 669 (Opcode::Select, &[I8, I128, I128]), 670 // https://github.com/bytecodealliance/wasmtime/issues/6104 671 (Opcode::Bitcast, &[I128], &[_]), 672 (Opcode::Bitcast, &[_], &[I128]), 673 // TODO 674 ( 675 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 676 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 677 ), 678 ( 679 Opcode::Rotr | Opcode::Rotl, 680 &[I8X16 | I16X8 | I32X4 | I64X2, _] 681 ), 682 // TODO 683 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 684 (Opcode::VhighBits, &[F32X4 | F64X2]), 685 ) 686 } 687 688 Architecture::S390x => { 689 exceptions!( 690 op, 691 args, 692 rets, 693 (Opcode::UaddOverflow | Opcode::SaddOverflow), 694 (Opcode::UsubOverflow | Opcode::SsubOverflow), 695 (Opcode::UmulOverflow | Opcode::SmulOverflow), 696 ( 697 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 698 &[I128, I128] 699 ), 700 (Opcode::Bnot, &[F32 | F64]), 701 ( 702 Opcode::Band 703 | Opcode::Bor 704 | Opcode::Bxor 705 | Opcode::BandNot 706 | Opcode::BorNot 707 | Opcode::BxorNot, 708 &([F32, F32] | [F64, F64]) 709 ), 710 ( 711 Opcode::FcvtToUint 712 | Opcode::FcvtToUintSat 713 | Opcode::FcvtToSint 714 | Opcode::FcvtToSintSat, 715 &[F32 | F64], 716 &[I128] 717 ), 718 ( 719 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 720 &[I128], 721 &[F32 | F64] 722 ), 723 (Opcode::SsubSat | Opcode::SaddSat, &[I64X2, I64X2]), 724 // https://github.com/bytecodealliance/wasmtime/issues/6104 725 (Opcode::Bitcast, &[I128], &[_]), 726 (Opcode::Bitcast, &[_], &[I128]), 727 // TODO 728 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 729 ) 730 } 731 732 Architecture::Riscv64(_) => { 733 exceptions!( 734 op, 735 args, 736 rets, 737 // TODO 738 (Opcode::UaddOverflow | Opcode::SaddOverflow), 739 (Opcode::UsubOverflow | Opcode::SsubOverflow), 740 (Opcode::UmulOverflow | Opcode::SmulOverflow), 741 // TODO 742 ( 743 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 744 &[I128, I128] 745 ), 746 // TODO 747 (Opcode::Iabs, &[I128]), 748 // TODO 749 (Opcode::Bitselect, &[I128, I128, I128]), 750 // https://github.com/bytecodealliance/wasmtime/issues/5528 751 ( 752 Opcode::FcvtToUint | Opcode::FcvtToSint, 753 [F32 | F64], 754 &[I128] 755 ), 756 ( 757 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat, 758 &[F32 | F64], 759 &[I128] 760 ), 761 // https://github.com/bytecodealliance/wasmtime/issues/5528 762 ( 763 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 764 &[I128], 765 &[F32 | F64] 766 ), 767 // https://github.com/bytecodealliance/wasmtime/issues/6104 768 (Opcode::Bitcast, &[I128], &[_]), 769 (Opcode::Bitcast, &[_], &[I128]), 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: Lazy<Vec<OpcodeSignature>> = Lazy::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 => false, 821 822 // Constants are generated outside of `generate_instructions` 823 Opcode::Iconst => false, 824 825 // TODO: extract_vector raises exceptions during return type generation becuase it 826 // uses dynamic vectors. 827 Opcode::ExtractVector => false, 828 829 _ => true, 830 } 831 }) 832 .flat_map(|op| { 833 let constraints = op.constraints(); 834 835 let ctrl_types = if let Some(ctrls) = constraints.ctrl_typeset() { 836 Vec::from_iter(types.iter().copied().filter(|ty| ctrls.contains(*ty))) 837 } else { 838 vec![INVALID] 839 }; 840 841 ctrl_types.into_iter().flat_map(move |ctrl_type| { 842 let rets = Vec::from_iter( 843 (0..constraints.num_fixed_results()) 844 .map(|i| constraints.result_type(i, ctrl_type)), 845 ); 846 847 // Cols is a vector whose length will match `num_fixed_value_arguments`, and whose 848 // elements will be vectors of types that are valid for that fixed argument 849 // position. 850 let mut cols = vec![]; 851 852 for i in 0..constraints.num_fixed_value_arguments() { 853 match constraints.value_argument_constraint(i, ctrl_type) { 854 ResolvedConstraint::Bound(ty) => cols.push(Vec::from([ty])), 855 ResolvedConstraint::Free(tys) => cols.push(Vec::from_iter( 856 types.iter().copied().filter(|ty| tys.contains(*ty)), 857 )), 858 } 859 } 860 861 // Generate the cartesian product of cols to produce a vector of argument lists, 862 // argss. The argss vector is seeded with the empty argument list, so there's an 863 // initial value to be extended in the loop below. 864 let mut argss = vec![vec![]]; 865 let mut cols = cols.as_slice(); 866 while let Some((col, rest)) = cols.split_last() { 867 cols = rest; 868 869 let mut next = vec![]; 870 for current in argss.iter() { 871 // Extend the front of each argument candidate with every type in `col`. 872 for ty in col { 873 let mut args = vec![*ty]; 874 args.extend_from_slice(¤t); 875 next.push(args); 876 } 877 } 878 879 let _ = std::mem::replace(&mut argss, next); 880 } 881 882 argss.into_iter().map(move |args| (*op, args, rets.clone())) 883 }) 884 }) 885 .filter(|(op, args, rets)| { 886 // These op/signature combinations need to be vetted 887 exceptions!( 888 op, 889 args.as_slice(), 890 rets.as_slice(), 891 (Opcode::Debugtrap), 892 (Opcode::Trap), 893 (Opcode::Trapz), 894 (Opcode::ResumableTrap), 895 (Opcode::Trapnz), 896 (Opcode::ResumableTrapnz), 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::Null), 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::IaddCin), 931 (Opcode::IaddCarry), 932 (Opcode::UaddOverflowTrap), 933 (Opcode::IsubBin), 934 (Opcode::IsubBorrow), 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::IsNull), 944 (Opcode::IsInvalid), 945 (Opcode::ScalarToVector), 946 (Opcode::X86Pmaddubsw), 947 (Opcode::X86Cvtt2dq), 948 (Opcode::Umulhi, &[I128, I128], &[I128]), 949 (Opcode::Smulhi, &[I128, I128], &[I128]), 950 // https://github.com/bytecodealliance/wasmtime/issues/6073 951 (Opcode::Iconcat, &[I32, I32], &[I64]), 952 (Opcode::Iconcat, &[I16, I16], &[I32]), 953 (Opcode::Iconcat, &[I8, I8], &[I16]), 954 // https://github.com/bytecodealliance/wasmtime/issues/6073 955 (Opcode::Isplit, &[I64], &[I32, I32]), 956 (Opcode::Isplit, &[I32], &[I16, I16]), 957 (Opcode::Isplit, &[I16], &[I8, I8]), 958 (Opcode::Fmin, &[F32X4, F32X4], &[F32X4]), 959 (Opcode::Fmin, &[F64X2, F64X2], &[F64X2]), 960 (Opcode::Fmax, &[F32X4, F32X4], &[F32X4]), 961 (Opcode::Fmax, &[F64X2, F64X2], &[F64X2]), 962 (Opcode::FcvtToUintSat, &[F32X4], &[I8]), 963 (Opcode::FcvtToUintSat, &[F64X2], &[I8]), 964 (Opcode::FcvtToUintSat, &[F32X4], &[I16]), 965 (Opcode::FcvtToUintSat, &[F64X2], &[I16]), 966 (Opcode::FcvtToUintSat, &[F32X4], &[I32]), 967 (Opcode::FcvtToUintSat, &[F64X2], &[I32]), 968 (Opcode::FcvtToUintSat, &[F32X4], &[I64]), 969 (Opcode::FcvtToUintSat, &[F64X2], &[I64]), 970 (Opcode::FcvtToUintSat, &[F32X4], &[I128]), 971 (Opcode::FcvtToUintSat, &[F64X2], &[I128]), 972 (Opcode::FcvtToUintSat, &[F32], &[I8X16]), 973 (Opcode::FcvtToUintSat, &[F64], &[I8X16]), 974 (Opcode::FcvtToUintSat, &[F32X4], &[I8X16]), 975 (Opcode::FcvtToUintSat, &[F64X2], &[I8X16]), 976 (Opcode::FcvtToUintSat, &[F32], &[I16X8]), 977 (Opcode::FcvtToUintSat, &[F64], &[I16X8]), 978 (Opcode::FcvtToUintSat, &[F32X4], &[I16X8]), 979 (Opcode::FcvtToUintSat, &[F64X2], &[I16X8]), 980 (Opcode::FcvtToUintSat, &[F32], &[I32X4]), 981 (Opcode::FcvtToUintSat, &[F64], &[I32X4]), 982 (Opcode::FcvtToUintSat, &[F64X2], &[I32X4]), 983 (Opcode::FcvtToUintSat, &[F32], &[I64X2]), 984 (Opcode::FcvtToUintSat, &[F64], &[I64X2]), 985 (Opcode::FcvtToUintSat, &[F32X4], &[I64X2]), 986 (Opcode::FcvtToSintSat, &[F32X4], &[I8]), 987 (Opcode::FcvtToSintSat, &[F64X2], &[I8]), 988 (Opcode::FcvtToSintSat, &[F32X4], &[I16]), 989 (Opcode::FcvtToSintSat, &[F64X2], &[I16]), 990 (Opcode::FcvtToSintSat, &[F32X4], &[I32]), 991 (Opcode::FcvtToSintSat, &[F64X2], &[I32]), 992 (Opcode::FcvtToSintSat, &[F32X4], &[I64]), 993 (Opcode::FcvtToSintSat, &[F64X2], &[I64]), 994 (Opcode::FcvtToSintSat, &[F32X4], &[I128]), 995 (Opcode::FcvtToSintSat, &[F64X2], &[I128]), 996 (Opcode::FcvtToSintSat, &[F32], &[I8X16]), 997 (Opcode::FcvtToSintSat, &[F64], &[I8X16]), 998 (Opcode::FcvtToSintSat, &[F32X4], &[I8X16]), 999 (Opcode::FcvtToSintSat, &[F64X2], &[I8X16]), 1000 (Opcode::FcvtToSintSat, &[F32], &[I16X8]), 1001 (Opcode::FcvtToSintSat, &[F64], &[I16X8]), 1002 (Opcode::FcvtToSintSat, &[F32X4], &[I16X8]), 1003 (Opcode::FcvtToSintSat, &[F64X2], &[I16X8]), 1004 (Opcode::FcvtToSintSat, &[F32], &[I32X4]), 1005 (Opcode::FcvtToSintSat, &[F64], &[I32X4]), 1006 (Opcode::FcvtToSintSat, &[F64X2], &[I32X4]), 1007 (Opcode::FcvtToSintSat, &[F32], &[I64X2]), 1008 (Opcode::FcvtToSintSat, &[F64], &[I64X2]), 1009 (Opcode::FcvtToSintSat, &[F32X4], &[I64X2]), 1010 (Opcode::FcvtFromUint, &[I8X16], &[F32]), 1011 (Opcode::FcvtFromUint, &[I16X8], &[F32]), 1012 (Opcode::FcvtFromUint, &[I32X4], &[F32]), 1013 (Opcode::FcvtFromUint, &[I64X2], &[F32]), 1014 (Opcode::FcvtFromUint, &[I8X16], &[F64]), 1015 (Opcode::FcvtFromUint, &[I16X8], &[F64]), 1016 (Opcode::FcvtFromUint, &[I32X4], &[F64]), 1017 (Opcode::FcvtFromUint, &[I64X2], &[F64]), 1018 (Opcode::FcvtFromUint, &[I8], &[F32X4]), 1019 (Opcode::FcvtFromUint, &[I16], &[F32X4]), 1020 (Opcode::FcvtFromUint, &[I32], &[F32X4]), 1021 (Opcode::FcvtFromUint, &[I64], &[F32X4]), 1022 (Opcode::FcvtFromUint, &[I128], &[F32X4]), 1023 (Opcode::FcvtFromUint, &[I8X16], &[F32X4]), 1024 (Opcode::FcvtFromUint, &[I16X8], &[F32X4]), 1025 (Opcode::FcvtFromUint, &[I64X2], &[F32X4]), 1026 (Opcode::FcvtFromUint, &[I8], &[F64X2]), 1027 (Opcode::FcvtFromUint, &[I16], &[F64X2]), 1028 (Opcode::FcvtFromUint, &[I32], &[F64X2]), 1029 (Opcode::FcvtFromUint, &[I64], &[F64X2]), 1030 (Opcode::FcvtFromUint, &[I128], &[F64X2]), 1031 (Opcode::FcvtFromUint, &[I8X16], &[F64X2]), 1032 (Opcode::FcvtFromUint, &[I16X8], &[F64X2]), 1033 (Opcode::FcvtFromUint, &[I32X4], &[F64X2]), 1034 (Opcode::FcvtFromSint, &[I8X16], &[F32]), 1035 (Opcode::FcvtFromSint, &[I16X8], &[F32]), 1036 (Opcode::FcvtFromSint, &[I32X4], &[F32]), 1037 (Opcode::FcvtFromSint, &[I64X2], &[F32]), 1038 (Opcode::FcvtFromSint, &[I8X16], &[F64]), 1039 (Opcode::FcvtFromSint, &[I16X8], &[F64]), 1040 (Opcode::FcvtFromSint, &[I32X4], &[F64]), 1041 (Opcode::FcvtFromSint, &[I64X2], &[F64]), 1042 (Opcode::FcvtFromSint, &[I8], &[F32X4]), 1043 (Opcode::FcvtFromSint, &[I16], &[F32X4]), 1044 (Opcode::FcvtFromSint, &[I32], &[F32X4]), 1045 (Opcode::FcvtFromSint, &[I64], &[F32X4]), 1046 (Opcode::FcvtFromSint, &[I128], &[F32X4]), 1047 (Opcode::FcvtFromSint, &[I8X16], &[F32X4]), 1048 (Opcode::FcvtFromSint, &[I16X8], &[F32X4]), 1049 (Opcode::FcvtFromSint, &[I64X2], &[F32X4]), 1050 (Opcode::FcvtFromSint, &[I8], &[F64X2]), 1051 (Opcode::FcvtFromSint, &[I16], &[F64X2]), 1052 (Opcode::FcvtFromSint, &[I32], &[F64X2]), 1053 (Opcode::FcvtFromSint, &[I64], &[F64X2]), 1054 (Opcode::FcvtFromSint, &[I128], &[F64X2]), 1055 (Opcode::FcvtFromSint, &[I8X16], &[F64X2]), 1056 (Opcode::FcvtFromSint, &[I16X8], &[F64X2]), 1057 (Opcode::FcvtFromSint, &[I32X4], &[F64X2]), 1058 ) 1059 }) 1060 .filter(|(op, ..)| { 1061 allowed_opcodes 1062 .as_ref() 1063 .map_or(true, |opcodes| opcodes.contains(op)) 1064 }) 1065 .collect() 1066 }); 1067 1068 fn inserter_for_format(fmt: InstructionFormat) -> OpcodeInserter { 1069 match fmt { 1070 InstructionFormat::AtomicCas => insert_atomic_cas, 1071 InstructionFormat::AtomicRmw => insert_atomic_rmw, 1072 InstructionFormat::Binary => insert_opcode, 1073 InstructionFormat::BinaryImm64 => todo!(), 1074 InstructionFormat::BinaryImm8 => insert_ins_ext_lane, 1075 InstructionFormat::Call => insert_call, 1076 InstructionFormat::CallIndirect => insert_call, 1077 InstructionFormat::CondTrap => todo!(), 1078 InstructionFormat::DynamicStackLoad => todo!(), 1079 InstructionFormat::DynamicStackStore => todo!(), 1080 InstructionFormat::FloatCompare => insert_cmp, 1081 InstructionFormat::FuncAddr => todo!(), 1082 InstructionFormat::IntAddTrap => todo!(), 1083 InstructionFormat::IntCompare => insert_cmp, 1084 InstructionFormat::IntCompareImm => todo!(), 1085 InstructionFormat::Load => insert_load_store, 1086 InstructionFormat::LoadNoOffset => insert_load_store, 1087 InstructionFormat::NullAry => insert_opcode, 1088 InstructionFormat::Shuffle => insert_shuffle, 1089 InstructionFormat::StackLoad => insert_stack_load, 1090 InstructionFormat::StackStore => insert_stack_store, 1091 InstructionFormat::Store => insert_load_store, 1092 InstructionFormat::StoreNoOffset => insert_load_store, 1093 InstructionFormat::Ternary => insert_opcode, 1094 InstructionFormat::TernaryImm8 => insert_ins_ext_lane, 1095 InstructionFormat::Trap => todo!(), 1096 InstructionFormat::Unary => insert_opcode, 1097 InstructionFormat::UnaryConst => insert_const, 1098 InstructionFormat::UnaryGlobalValue => todo!(), 1099 InstructionFormat::UnaryIeee32 => insert_const, 1100 InstructionFormat::UnaryIeee64 => insert_const, 1101 InstructionFormat::UnaryImm => insert_const, 1102 1103 InstructionFormat::BranchTable 1104 | InstructionFormat::Brif 1105 | InstructionFormat::Jump 1106 | InstructionFormat::MultiAry => { 1107 panic!( 1108 "Control-flow instructions should be handled by 'insert_terminator': {:?}", 1109 fmt 1110 ) 1111 } 1112 } 1113 } 1114 1115 pub struct FunctionGenerator<'r, 'data> 1116 where 1117 'data: 'r, 1118 { 1119 u: &'r mut Unstructured<'data>, 1120 config: &'r Config, 1121 resources: Resources, 1122 isa: OwnedTargetIsa, 1123 name: UserFuncName, 1124 signature: Signature, 1125 } 1126 1127 #[derive(Debug, Clone)] 1128 enum BlockTerminator { 1129 Return, 1130 Jump(Block), 1131 Br(Block, Block), 1132 BrTable(Block, Vec<Block>), 1133 Switch(Type, Block, HashMap<u128, Block>), 1134 TailCall(FuncRef), 1135 TailCallIndirect(FuncRef), 1136 } 1137 1138 #[derive(Debug, Clone)] 1139 enum BlockTerminatorKind { 1140 Return, 1141 Jump, 1142 Br, 1143 BrTable, 1144 Switch, 1145 TailCall, 1146 TailCallIndirect, 1147 } 1148 1149 /// Alias Analysis Category 1150 /// 1151 /// Our alias analysis pass supports 4 categories of accesses to distinguish 1152 /// different regions. The "Other" region is the general case, and is the default 1153 /// Although they have highly suggestive names there is no difference between any 1154 /// of the categories. 1155 /// 1156 /// We assign each stack slot a category when we first generate them, and then 1157 /// ensure that all accesses to that stack slot are correctly tagged. We already 1158 /// ensure that memory accesses never cross stack slots, so there is no risk 1159 /// of a memory access being tagged with the wrong category. 1160 #[derive(Debug, PartialEq, Clone, Copy)] 1161 enum AACategory { 1162 Other, 1163 Heap, 1164 Table, 1165 VmCtx, 1166 } 1167 1168 impl AACategory { 1169 pub fn all() -> &'static [Self] { 1170 &[ 1171 AACategory::Other, 1172 AACategory::Heap, 1173 AACategory::Table, 1174 AACategory::VmCtx, 1175 ] 1176 } 1177 1178 pub fn update_memflags(&self, flags: &mut MemFlags) { 1179 flags.set_alias_region(match self { 1180 AACategory::Other => None, 1181 AACategory::Heap => Some(AliasRegion::Heap), 1182 AACategory::Table => Some(AliasRegion::Table), 1183 AACategory::VmCtx => Some(AliasRegion::Vmctx), 1184 }) 1185 } 1186 } 1187 1188 #[derive(Default)] 1189 struct Resources { 1190 vars: HashMap<Type, Vec<Variable>>, 1191 blocks: Vec<(Block, BlockSignature)>, 1192 blocks_without_params: Vec<Block>, 1193 block_terminators: Vec<BlockTerminator>, 1194 func_refs: Vec<(Signature, SigRef, FuncRef)>, 1195 /// This field is required to be sorted by stack slot size at all times. 1196 /// We use this invariant when searching for stack slots with a given size. 1197 /// See [FunctionGenerator::stack_slot_with_size] 1198 stack_slots: Vec<(StackSlot, StackSize, AACategory)>, 1199 usercalls: Vec<(UserExternalName, Signature)>, 1200 libcalls: Vec<LibCall>, 1201 } 1202 1203 impl Resources { 1204 /// Partitions blocks at `block`. Only blocks that can be targeted by branches are considered. 1205 /// 1206 /// The first slice includes all blocks up to and including `block`. 1207 /// The second slice includes all remaining blocks. 1208 fn partition_target_blocks( 1209 &self, 1210 block: Block, 1211 ) -> (&[(Block, BlockSignature)], &[(Block, BlockSignature)]) { 1212 // Blocks are stored in-order and have no gaps, this means that we can simply index them by 1213 // their number. We also need to exclude the entry block since it isn't a valid target. 1214 let target_blocks = &self.blocks[1..]; 1215 target_blocks.split_at(block.as_u32() as usize) 1216 } 1217 1218 /// Returns blocks forward of `block`. Only blocks that can be targeted by branches are considered. 1219 fn forward_blocks(&self, block: Block) -> &[(Block, BlockSignature)] { 1220 let (_, forward_blocks) = self.partition_target_blocks(block); 1221 forward_blocks 1222 } 1223 1224 /// Generates a slice of `blocks_without_params` ahead of `block` 1225 fn forward_blocks_without_params(&self, block: Block) -> &[Block] { 1226 let partition_point = self.blocks_without_params.partition_point(|b| *b <= block); 1227 &self.blocks_without_params[partition_point..] 1228 } 1229 1230 /// Generates an iterator of all valid tail call targets. This includes all functions with both 1231 /// the `tail` calling convention and the same return values as the caller. 1232 fn tail_call_targets<'a>( 1233 &'a self, 1234 caller_sig: &'a Signature, 1235 ) -> impl Iterator<Item = &'a (Signature, SigRef, FuncRef)> { 1236 self.func_refs.iter().filter(|(sig, _, _)| { 1237 sig.call_conv == CallConv::Tail && sig.returns == caller_sig.returns 1238 }) 1239 } 1240 } 1241 1242 impl<'r, 'data> FunctionGenerator<'r, 'data> 1243 where 1244 'data: 'r, 1245 { 1246 pub fn new( 1247 u: &'r mut Unstructured<'data>, 1248 config: &'r Config, 1249 isa: OwnedTargetIsa, 1250 name: UserFuncName, 1251 signature: Signature, 1252 usercalls: Vec<(UserExternalName, Signature)>, 1253 libcalls: Vec<LibCall>, 1254 ) -> Self { 1255 Self { 1256 u, 1257 config, 1258 resources: Resources { 1259 usercalls, 1260 libcalls, 1261 ..Resources::default() 1262 }, 1263 isa, 1264 name, 1265 signature, 1266 } 1267 } 1268 1269 /// Generates a random value for config `param` 1270 fn param(&mut self, param: &RangeInclusive<usize>) -> Result<usize> { 1271 Ok(self.u.int_in_range(param.clone())?) 1272 } 1273 1274 fn system_callconv(&mut self) -> CallConv { 1275 // TODO: This currently only runs on linux, so this is the only choice 1276 // We should improve this once we generate flags and targets 1277 CallConv::SystemV 1278 } 1279 1280 /// Finds a stack slot with size of at least n bytes 1281 fn stack_slot_with_size(&mut self, n: u32) -> Result<(StackSlot, StackSize, AACategory)> { 1282 let first = self 1283 .resources 1284 .stack_slots 1285 .partition_point(|&(_slot, size, _category)| size < n); 1286 Ok(*self.u.choose(&self.resources.stack_slots[first..])?) 1287 } 1288 1289 /// Generates an address that should allow for a store or a load. 1290 /// 1291 /// Addresses aren't generated like other values. They are never stored in variables so that 1292 /// we don't run the risk of returning them from a function, which would make the fuzzer 1293 /// complain since they are different from the interpreter to the backend. 1294 /// 1295 /// `min_size`: Controls the amount of space that the address should have. 1296 /// 1297 /// `aligned`: When passed as true, the resulting address is guaranteed to be aligned 1298 /// on an 8 byte boundary. 1299 /// 1300 /// Returns a valid address and the maximum possible offset that still respects `min_size`. 1301 fn generate_load_store_address( 1302 &mut self, 1303 builder: &mut FunctionBuilder, 1304 min_size: u32, 1305 aligned: bool, 1306 ) -> Result<(Value, u32, AACategory)> { 1307 // TODO: Currently our only source of addresses is stack_addr, but we 1308 // should add global_value, symbol_value eventually 1309 let (addr, available_size, category) = { 1310 let (ss, slot_size, category) = self.stack_slot_with_size(min_size)?; 1311 1312 // stack_slot_with_size guarantees that slot_size >= min_size 1313 let max_offset = slot_size - min_size; 1314 let offset = if aligned { 1315 self.u.int_in_range(0..=max_offset / min_size)? * min_size 1316 } else { 1317 self.u.int_in_range(0..=max_offset)? 1318 }; 1319 1320 let base_addr = builder.ins().stack_addr(I64, ss, offset as i32); 1321 let available_size = slot_size.saturating_sub(offset); 1322 (base_addr, available_size, category) 1323 }; 1324 1325 // TODO: Insert a bunch of amode opcodes here to modify the address! 1326 1327 // Now that we have an address and a size, we just choose a random offset to return to the 1328 // caller. Preserving min_size bytes. 1329 let max_offset = available_size.saturating_sub(min_size); 1330 Ok((addr, max_offset, category)) 1331 } 1332 1333 // Generates an address and memflags for a load or store. 1334 fn generate_address_and_memflags( 1335 &mut self, 1336 builder: &mut FunctionBuilder, 1337 min_size: u32, 1338 is_atomic: bool, 1339 ) -> Result<(Value, MemFlags, Offset32)> { 1340 // Should we generate an aligned address 1341 // Some backends have issues with unaligned atomics. 1342 // AArch64: https://github.com/bytecodealliance/wasmtime/issues/5483 1343 // RISCV: https://github.com/bytecodealliance/wasmtime/issues/5882 1344 let requires_aligned_atomics = matches!( 1345 self.isa.triple().architecture, 1346 Architecture::Aarch64(_) | Architecture::Riscv64(_) 1347 ); 1348 let aligned = if is_atomic && requires_aligned_atomics { 1349 true 1350 } else if min_size > 8 { 1351 // TODO: We currently can't guarantee that a stack_slot will be aligned on a 16 byte 1352 // boundary. We don't have a way to specify alignment when creating stack slots, and 1353 // cranelift only guarantees 8 byte alignment between stack slots. 1354 // See: https://github.com/bytecodealliance/wasmtime/issues/5922#issuecomment-1457926624 1355 false 1356 } else { 1357 bool::arbitrary(self.u)? 1358 }; 1359 1360 let mut flags = MemFlags::new(); 1361 // Even if we picked an aligned address, we can always generate unaligned memflags 1362 if aligned && bool::arbitrary(self.u)? { 1363 flags.set_aligned(); 1364 } 1365 // If the address is aligned, then we know it won't trap 1366 if aligned && bool::arbitrary(self.u)? { 1367 flags.set_notrap(); 1368 } 1369 1370 let (address, max_offset, category) = 1371 self.generate_load_store_address(builder, min_size, aligned)?; 1372 1373 // Set the Alias Analysis bits on the memflags 1374 category.update_memflags(&mut flags); 1375 1376 // Pick an offset to pass into the load/store. 1377 let offset = if aligned { 1378 0 1379 } else { 1380 self.u.int_in_range(0..=max_offset)? as i32 1381 } 1382 .into(); 1383 1384 Ok((address, flags, offset)) 1385 } 1386 1387 /// Get a variable of type `ty` from the current function 1388 fn get_variable_of_type(&mut self, ty: Type) -> Result<Variable> { 1389 let opts = self.resources.vars.get(&ty).map_or(&[][..], Vec::as_slice); 1390 let var = self.u.choose(opts)?; 1391 Ok(*var) 1392 } 1393 1394 /// Generates an instruction(`iconst`/`fconst`/etc...) to introduce a constant value 1395 fn generate_const(&mut self, builder: &mut FunctionBuilder, ty: Type) -> Result<Value> { 1396 Ok(match self.u.datavalue(ty)? { 1397 DataValue::I8(i) => builder.ins().iconst(ty, i as u8 as i64), 1398 DataValue::I16(i) => builder.ins().iconst(ty, i as u16 as i64), 1399 DataValue::I32(i) => builder.ins().iconst(ty, i as u32 as i64), 1400 DataValue::I64(i) => builder.ins().iconst(ty, i), 1401 DataValue::I128(i) => { 1402 let hi = builder.ins().iconst(I64, (i >> 64) as i64); 1403 let lo = builder.ins().iconst(I64, i as i64); 1404 builder.ins().iconcat(lo, hi) 1405 } 1406 DataValue::F32(f) => builder.ins().f32const(f), 1407 DataValue::F64(f) => builder.ins().f64const(f), 1408 DataValue::V128(bytes) => { 1409 let data = bytes.to_vec().into(); 1410 let handle = builder.func.dfg.constants.insert(data); 1411 builder.ins().vconst(ty, handle) 1412 } 1413 _ => unimplemented!(), 1414 }) 1415 } 1416 1417 /// Chooses a random block which can be targeted by a jump / branch. 1418 /// This means any block that is not the first block. 1419 fn generate_target_block(&mut self, source_block: Block) -> Result<Block> { 1420 // We try to mostly generate forward branches to avoid generating an excessive amount of 1421 // infinite loops. But they are still important, so give them a small chance of existing. 1422 let (backwards_blocks, forward_blocks) = 1423 self.resources.partition_target_blocks(source_block); 1424 let ratio = self.config.backwards_branch_ratio; 1425 let block_targets = if !backwards_blocks.is_empty() && self.u.ratio(ratio.0, ratio.1)? { 1426 backwards_blocks 1427 } else { 1428 forward_blocks 1429 }; 1430 assert!(!block_targets.is_empty()); 1431 1432 let (block, _) = self.u.choose(block_targets)?.clone(); 1433 Ok(block) 1434 } 1435 1436 fn generate_values_for_block( 1437 &mut self, 1438 builder: &mut FunctionBuilder, 1439 block: Block, 1440 ) -> Result<Vec<Value>> { 1441 let (_, sig) = self.resources.blocks[block.as_u32() as usize].clone(); 1442 self.generate_values_for_signature(builder, sig.iter().copied()) 1443 } 1444 1445 fn generate_values_for_signature<I: Iterator<Item = Type>>( 1446 &mut self, 1447 builder: &mut FunctionBuilder, 1448 signature: I, 1449 ) -> Result<Vec<Value>> { 1450 signature 1451 .map(|ty| { 1452 let var = self.get_variable_of_type(ty)?; 1453 let val = builder.use_var(var); 1454 Ok(val) 1455 }) 1456 .collect() 1457 } 1458 1459 /// The terminator that we need to insert has already been picked ahead of time 1460 /// we just need to build the instructions for it 1461 fn insert_terminator( 1462 &mut self, 1463 builder: &mut FunctionBuilder, 1464 source_block: Block, 1465 ) -> Result<()> { 1466 let terminator = self.resources.block_terminators[source_block.as_u32() as usize].clone(); 1467 1468 match terminator { 1469 BlockTerminator::Return => { 1470 let types: Vec<Type> = { 1471 let rets = &builder.func.signature.returns; 1472 rets.iter().map(|p| p.value_type).collect() 1473 }; 1474 let vals = self.generate_values_for_signature(builder, types.into_iter())?; 1475 1476 builder.ins().return_(&vals[..]); 1477 } 1478 BlockTerminator::Jump(target) => { 1479 let args = self.generate_values_for_block(builder, target)?; 1480 builder.ins().jump(target, &args[..]); 1481 } 1482 BlockTerminator::Br(left, right) => { 1483 let left_args = self.generate_values_for_block(builder, left)?; 1484 let right_args = self.generate_values_for_block(builder, right)?; 1485 1486 let condbr_types = [I8, I16, I32, I64, I128]; 1487 let _type = *self.u.choose(&condbr_types[..])?; 1488 let val = builder.use_var(self.get_variable_of_type(_type)?); 1489 builder 1490 .ins() 1491 .brif(val, left, &left_args[..], right, &right_args[..]); 1492 } 1493 BlockTerminator::BrTable(default, targets) => { 1494 // Create jump tables on demand 1495 let mut jt = Vec::with_capacity(targets.len()); 1496 for block in targets { 1497 let args = self.generate_values_for_block(builder, block)?; 1498 jt.push(builder.func.dfg.block_call(block, &args)) 1499 } 1500 1501 let args = self.generate_values_for_block(builder, default)?; 1502 let jt_data = JumpTableData::new(builder.func.dfg.block_call(default, &args), &jt); 1503 let jt = builder.create_jump_table(jt_data); 1504 1505 // br_table only supports I32 1506 let val = builder.use_var(self.get_variable_of_type(I32)?); 1507 1508 builder.ins().br_table(val, jt); 1509 } 1510 BlockTerminator::Switch(_type, default, entries) => { 1511 let mut switch = Switch::new(); 1512 for (&entry, &block) in entries.iter() { 1513 switch.set_entry(entry, block); 1514 } 1515 1516 let switch_val = builder.use_var(self.get_variable_of_type(_type)?); 1517 1518 switch.emit(builder, switch_val, default); 1519 } 1520 BlockTerminator::TailCall(target) | BlockTerminator::TailCallIndirect(target) => { 1521 let (sig, sig_ref, func_ref) = self 1522 .resources 1523 .func_refs 1524 .iter() 1525 .find(|(_, _, f)| *f == target) 1526 .expect("Failed to find previously selected function") 1527 .clone(); 1528 1529 let opcode = match terminator { 1530 BlockTerminator::TailCall(_) => Opcode::ReturnCall, 1531 BlockTerminator::TailCallIndirect(_) => Opcode::ReturnCallIndirect, 1532 _ => unreachable!(), 1533 }; 1534 1535 insert_call_to_function(self, builder, opcode, &sig, sig_ref, func_ref)?; 1536 } 1537 } 1538 1539 Ok(()) 1540 } 1541 1542 /// Fills the current block with random instructions 1543 fn generate_instructions(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1544 for _ in 0..self.param(&self.config.instructions_per_block)? { 1545 let (op, args, rets) = self.u.choose(&OPCODE_SIGNATURES)?; 1546 1547 // We filter out instructions that aren't supported by the target at this point instead 1548 // of building a single vector of valid instructions at the beginning of function 1549 // generation, to avoid invalidating the corpus when instructions are enabled/disabled. 1550 if !valid_for_target(&self.isa.triple(), *op, &args, &rets) { 1551 return Err(arbitrary::Error::IncorrectFormat.into()); 1552 } 1553 1554 let inserter = inserter_for_format(op.format()); 1555 inserter(self, builder, *op, &args, &rets)?; 1556 } 1557 1558 Ok(()) 1559 } 1560 1561 fn generate_funcrefs(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1562 let usercalls: Vec<(ExternalName, Signature)> = self 1563 .resources 1564 .usercalls 1565 .iter() 1566 .map(|(name, signature)| { 1567 let user_func_ref = builder.func.declare_imported_user_function(name.clone()); 1568 let name = ExternalName::User(user_func_ref); 1569 (name, signature.clone()) 1570 }) 1571 .collect(); 1572 1573 let lib_callconv = self.system_callconv(); 1574 let libcalls: Vec<(ExternalName, Signature)> = self 1575 .resources 1576 .libcalls 1577 .iter() 1578 .map(|libcall| { 1579 let pointer_type = Type::int_with_byte_size( 1580 self.isa.triple().pointer_width().unwrap().bytes().into(), 1581 ) 1582 .unwrap(); 1583 let signature = libcall.signature(lib_callconv, pointer_type); 1584 let name = ExternalName::LibCall(*libcall); 1585 (name, signature) 1586 }) 1587 .collect(); 1588 1589 for (name, signature) in usercalls.into_iter().chain(libcalls) { 1590 let sig_ref = builder.import_signature(signature.clone()); 1591 let func_ref = builder.import_function(ExtFuncData { 1592 name, 1593 signature: sig_ref, 1594 colocated: self.u.arbitrary()?, 1595 }); 1596 1597 self.resources 1598 .func_refs 1599 .push((signature, sig_ref, func_ref)); 1600 } 1601 1602 Ok(()) 1603 } 1604 1605 fn generate_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1606 for _ in 0..self.param(&self.config.static_stack_slots_per_function)? { 1607 let bytes = self.param(&self.config.static_stack_slot_size)? as u32; 1608 let ss_data = StackSlotData::new(StackSlotKind::ExplicitSlot, bytes, 0); 1609 let slot = builder.create_sized_stack_slot(ss_data); 1610 1611 // Generate one Alias Analysis Category for each slot 1612 let category = *self.u.choose(AACategory::all())?; 1613 1614 self.resources.stack_slots.push((slot, bytes, category)); 1615 } 1616 1617 self.resources 1618 .stack_slots 1619 .sort_unstable_by_key(|&(_slot, bytes, _category)| bytes); 1620 1621 Ok(()) 1622 } 1623 1624 /// Zero initializes the stack slot by inserting `stack_store`'s. 1625 fn initialize_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1626 let i8_zero = builder.ins().iconst(I8, 0); 1627 let i16_zero = builder.ins().iconst(I16, 0); 1628 let i32_zero = builder.ins().iconst(I32, 0); 1629 let i64_zero = builder.ins().iconst(I64, 0); 1630 let i128_zero = builder.ins().uextend(I128, i64_zero); 1631 1632 for &(slot, init_size, category) in self.resources.stack_slots.iter() { 1633 let mut size = init_size; 1634 1635 // Insert the largest available store for the remaining size. 1636 while size != 0 { 1637 let offset = (init_size - size) as i32; 1638 let (val, filled) = match size { 1639 sz if sz / 16 > 0 => (i128_zero, 16), 1640 sz if sz / 8 > 0 => (i64_zero, 8), 1641 sz if sz / 4 > 0 => (i32_zero, 4), 1642 sz if sz / 2 > 0 => (i16_zero, 2), 1643 _ => (i8_zero, 1), 1644 }; 1645 let addr = builder.ins().stack_addr(I64, slot, offset); 1646 1647 // Each stack slot has an associated category, that means we have to set the 1648 // correct memflags for it. So we can't use `stack_store` directly. 1649 let mut flags = MemFlags::new(); 1650 flags.set_notrap(); 1651 category.update_memflags(&mut flags); 1652 1653 builder.ins().store(flags, val, addr, 0); 1654 1655 size -= filled; 1656 } 1657 } 1658 Ok(()) 1659 } 1660 1661 /// Creates a random amount of blocks in this function 1662 fn generate_blocks(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1663 let extra_block_count = self.param(&self.config.blocks_per_function)?; 1664 1665 // We must always have at least one block, so we generate the "extra" blocks and add 1 for 1666 // the entry block. 1667 let block_count = 1 + extra_block_count; 1668 1669 // Blocks need to be sorted in ascending order 1670 self.resources.blocks = (0..block_count) 1671 .map(|i| { 1672 let is_entry = i == 0; 1673 let block = builder.create_block(); 1674 1675 // Optionally mark blocks that are not the entry block as cold 1676 if !is_entry { 1677 if bool::arbitrary(self.u)? { 1678 builder.set_cold_block(block); 1679 } 1680 } 1681 1682 // The first block has to have the function signature, but for the rest of them we generate 1683 // a random signature; 1684 if is_entry { 1685 builder.append_block_params_for_function_params(block); 1686 Ok(( 1687 block, 1688 self.signature.params.iter().map(|a| a.value_type).collect(), 1689 )) 1690 } else { 1691 let sig = self.generate_block_signature()?; 1692 sig.iter().for_each(|ty| { 1693 builder.append_block_param(block, *ty); 1694 }); 1695 Ok((block, sig)) 1696 } 1697 }) 1698 .collect::<Result<Vec<_>>>()?; 1699 1700 // Valid blocks for jump tables have to have no parameters in the signature, and must also 1701 // not be the first block. 1702 self.resources.blocks_without_params = self.resources.blocks[1..] 1703 .iter() 1704 .filter(|(_, sig)| sig.len() == 0) 1705 .map(|(b, _)| *b) 1706 .collect(); 1707 1708 // Compute the block CFG 1709 // 1710 // cranelift-frontend requires us to never generate unreachable blocks 1711 // To ensure this property we start by constructing a main "spine" of blocks. So block1 can 1712 // always jump to block2, and block2 can always jump to block3, etc... 1713 // 1714 // That is not a very interesting CFG, so we introduce variations on that, but always 1715 // ensuring that the property of pointing to the next block is maintained whatever the 1716 // branching mechanism we use. 1717 let blocks = self.resources.blocks.clone(); 1718 self.resources.block_terminators = blocks 1719 .iter() 1720 .map(|&(block, _)| { 1721 let next_block = Block::with_number(block.as_u32() + 1).unwrap(); 1722 let forward_blocks = self.resources.forward_blocks(block); 1723 let paramless_targets = self.resources.forward_blocks_without_params(block); 1724 let has_paramless_targets = !paramless_targets.is_empty(); 1725 let next_block_is_paramless = paramless_targets.contains(&next_block); 1726 1727 let mut valid_terminators = vec![]; 1728 1729 if forward_blocks.is_empty() { 1730 // Return is only valid on the last block. 1731 valid_terminators.push(BlockTerminatorKind::Return); 1732 } else { 1733 // If we have more than one block we can allow terminators that target blocks. 1734 // TODO: We could add some kind of BrReturn here, to explore edges where we 1735 // exit in the middle of the function 1736 valid_terminators.extend_from_slice(&[ 1737 BlockTerminatorKind::Jump, 1738 BlockTerminatorKind::Br, 1739 BlockTerminatorKind::BrTable, 1740 ]); 1741 } 1742 1743 // As the Switch interface only allows targeting blocks without params we need 1744 // to ensure that the next block has no params, since that one is guaranteed to be 1745 // picked in either case. 1746 if has_paramless_targets && next_block_is_paramless { 1747 valid_terminators.push(BlockTerminatorKind::Switch); 1748 } 1749 1750 // Tail Calls are a block terminator, so we should insert them as any other block 1751 // terminator. We should ensure that we can select at least one target before considering 1752 // them as candidate instructions. 1753 let has_tail_callees = self 1754 .resources 1755 .tail_call_targets(&self.signature) 1756 .next() 1757 .is_some(); 1758 let is_tail_caller = self.signature.call_conv == CallConv::Tail; 1759 1760 let supports_tail_calls = match self.isa.triple().architecture { 1761 Architecture::Aarch64(_) | Architecture::Riscv64(_) => true, 1762 // TODO: x64 currently requires frame pointers for tail calls. 1763 Architecture::X86_64 => self.isa.flags().preserve_frame_pointers(), 1764 // TODO: Other platforms do not support tail calls yet. 1765 _ => false, 1766 }; 1767 1768 if is_tail_caller && has_tail_callees && supports_tail_calls { 1769 valid_terminators.extend([ 1770 BlockTerminatorKind::TailCall, 1771 BlockTerminatorKind::TailCallIndirect, 1772 ]); 1773 } 1774 1775 let terminator = self.u.choose(&valid_terminators)?; 1776 1777 // Choose block targets for the terminators that we picked above 1778 Ok(match terminator { 1779 BlockTerminatorKind::Return => BlockTerminator::Return, 1780 BlockTerminatorKind::Jump => BlockTerminator::Jump(next_block), 1781 BlockTerminatorKind::Br => { 1782 BlockTerminator::Br(next_block, self.generate_target_block(block)?) 1783 } 1784 // TODO: Allow generating backwards branches here 1785 BlockTerminatorKind::BrTable => { 1786 // Make the default the next block, and then we don't have to worry 1787 // that we can reach it via the targets 1788 let default = next_block; 1789 1790 let target_count = self.param(&self.config.jump_table_entries)?; 1791 let targets = Result::from_iter( 1792 (0..target_count).map(|_| self.generate_target_block(block)), 1793 )?; 1794 1795 BlockTerminator::BrTable(default, targets) 1796 } 1797 BlockTerminatorKind::Switch => { 1798 // Make the default the next block, and then we don't have to worry 1799 // that we can reach it via the entries below 1800 let default_block = next_block; 1801 1802 let _type = *self.u.choose(&[I8, I16, I32, I64, I128][..])?; 1803 1804 // Build this into a HashMap since we cannot have duplicate entries. 1805 let mut entries = HashMap::new(); 1806 for _ in 0..self.param(&self.config.switch_cases)? { 1807 // The Switch API only allows for entries that are addressable by the index type 1808 // so we need to limit the range of values that we generate. 1809 let (ty_min, ty_max) = _type.bounds(false); 1810 let range_start = self.u.int_in_range(ty_min..=ty_max)?; 1811 1812 // We can either insert a contiguous range of blocks or a individual block 1813 // This is done because the Switch API specializes contiguous ranges. 1814 let range_size = if bool::arbitrary(self.u)? { 1815 1 1816 } else { 1817 self.param(&self.config.switch_max_range_size)? 1818 } as u128; 1819 1820 // Build the switch entries 1821 for i in 0..range_size { 1822 let index = range_start.wrapping_add(i) % ty_max; 1823 let block = *self 1824 .u 1825 .choose(self.resources.forward_blocks_without_params(block))?; 1826 1827 entries.insert(index, block); 1828 } 1829 } 1830 1831 BlockTerminator::Switch(_type, default_block, entries) 1832 } 1833 BlockTerminatorKind::TailCall => { 1834 let targets = self 1835 .resources 1836 .tail_call_targets(&self.signature) 1837 .collect::<Vec<_>>(); 1838 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1839 BlockTerminator::TailCall(*funcref) 1840 } 1841 BlockTerminatorKind::TailCallIndirect => { 1842 let targets = self 1843 .resources 1844 .tail_call_targets(&self.signature) 1845 .collect::<Vec<_>>(); 1846 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1847 BlockTerminator::TailCallIndirect(*funcref) 1848 } 1849 }) 1850 }) 1851 .collect::<Result<_>>()?; 1852 1853 Ok(()) 1854 } 1855 1856 fn generate_block_signature(&mut self) -> Result<BlockSignature> { 1857 let param_count = self.param(&self.config.block_signature_params)?; 1858 1859 let mut params = Vec::with_capacity(param_count); 1860 for _ in 0..param_count { 1861 params.push(self.u._type((&*self.isa).supports_simd())?); 1862 } 1863 Ok(params) 1864 } 1865 1866 fn build_variable_pool(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1867 let block = builder.current_block().unwrap(); 1868 1869 // Define variables for the function signature 1870 let mut vars: Vec<_> = builder 1871 .func 1872 .signature 1873 .params 1874 .iter() 1875 .map(|param| param.value_type) 1876 .zip(builder.block_params(block).iter().copied()) 1877 .collect(); 1878 1879 // Create a pool of vars that are going to be used in this function 1880 for _ in 0..self.param(&self.config.vars_per_function)? { 1881 let ty = self.u._type((&*self.isa).supports_simd())?; 1882 let value = self.generate_const(builder, ty)?; 1883 vars.push((ty, value)); 1884 } 1885 1886 for (id, (ty, value)) in vars.into_iter().enumerate() { 1887 let var = Variable::new(id); 1888 builder.declare_var(var, ty); 1889 builder.def_var(var, value); 1890 self.resources 1891 .vars 1892 .entry(ty) 1893 .or_insert_with(Vec::new) 1894 .push(var); 1895 } 1896 1897 Ok(()) 1898 } 1899 1900 /// We generate a function in multiple stages: 1901 /// 1902 /// * First we generate a random number of empty blocks 1903 /// * Then we generate a random pool of variables to be used throughout the function 1904 /// * We then visit each block and generate random instructions 1905 /// 1906 /// Because we generate all blocks and variables up front we already know everything that 1907 /// we need when generating instructions (i.e. jump targets / variables) 1908 pub fn generate(mut self) -> Result<Function> { 1909 let mut fn_builder_ctx = FunctionBuilderContext::new(); 1910 let mut func = Function::with_name_signature(self.name.clone(), self.signature.clone()); 1911 1912 let mut builder = FunctionBuilder::new(&mut func, &mut fn_builder_ctx); 1913 1914 // Build the function references before generating the block CFG since we store 1915 // function references in the CFG. 1916 self.generate_funcrefs(&mut builder)?; 1917 self.generate_blocks(&mut builder)?; 1918 1919 // Function preamble 1920 self.generate_stack_slots(&mut builder)?; 1921 1922 // Main instruction generation loop 1923 for (block, block_sig) in self.resources.blocks.clone().into_iter() { 1924 let is_block0 = block.as_u32() == 0; 1925 builder.switch_to_block(block); 1926 1927 if is_block0 { 1928 // The first block is special because we must create variables both for the 1929 // block signature and for the variable pool. Additionally, we must also define 1930 // initial values for all variables that are not the function signature. 1931 self.build_variable_pool(&mut builder)?; 1932 1933 // Stack slots have random bytes at the beginning of the function 1934 // initialize them to a constant value so that execution stays predictable. 1935 self.initialize_stack_slots(&mut builder)?; 1936 } else { 1937 // Define variables for the block params 1938 for (i, ty) in block_sig.iter().enumerate() { 1939 let var = self.get_variable_of_type(*ty)?; 1940 let block_param = builder.block_params(block)[i]; 1941 builder.def_var(var, block_param); 1942 } 1943 } 1944 1945 // Generate block instructions 1946 self.generate_instructions(&mut builder)?; 1947 1948 // Insert a terminator to safely exit the block 1949 self.insert_terminator(&mut builder, block)?; 1950 } 1951 1952 builder.seal_all_blocks(); 1953 builder.finalize(); 1954 1955 Ok(func) 1956 } 1957 } 1958