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::*, AtomicRmwOp, Block, ConstantData, Endianness, ExternalName, FuncRef, Function, 13 LibCall, Opcode, SigRef, Signature, StackSlot, Type, 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/5468 527 (Opcode::Smulhi | Opcode::Umulhi, &[I8, I8]), 528 // https://github.com/bytecodealliance/wasmtime/issues/4756 529 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 530 // https://github.com/bytecodealliance/wasmtime/issues/5474 531 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 532 // https://github.com/bytecodealliance/wasmtime/issues/3370 533 ( 534 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 535 &[I128, I128] 536 ), 537 // https://github.com/bytecodealliance/wasmtime/issues/5107 538 (Opcode::Cls, &[I8], &[I8]), 539 (Opcode::Cls, &[I16], &[I16]), 540 (Opcode::Cls, &[I32], &[I32]), 541 (Opcode::Cls, &[I64], &[I64]), 542 (Opcode::Cls, &[I128], &[I128]), 543 // TODO 544 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 545 // https://github.com/bytecodealliance/wasmtime/issues/4897 546 // https://github.com/bytecodealliance/wasmtime/issues/4899 547 ( 548 Opcode::FcvtToUint 549 | Opcode::FcvtToUintSat 550 | Opcode::FcvtToSint 551 | Opcode::FcvtToSintSat, 552 &[F32 | F64], 553 &[I8 | I16 | I128] 554 ), 555 (Opcode::FcvtToUint | Opcode::FcvtToSint, &[F32X4], &[I32X4]), 556 ( 557 Opcode::FcvtToUint 558 | Opcode::FcvtToUintSat 559 | Opcode::FcvtToSint 560 | Opcode::FcvtToSintSat, 561 &[F64X2], 562 &[I64X2] 563 ), 564 // https://github.com/bytecodealliance/wasmtime/issues/4900 565 (Opcode::FcvtFromUint, &[I128], &[F32 | F64]), 566 // This has a lowering, but only when preceded by `uwiden_low`. 567 (Opcode::FcvtFromUint, &[I64X2], &[F64X2]), 568 // https://github.com/bytecodealliance/wasmtime/issues/4900 569 (Opcode::FcvtFromSint, &[I128], &[F32 | F64]), 570 (Opcode::FcvtFromSint, &[I64X2], &[F64X2]), 571 ( 572 Opcode::Umulhi | Opcode::Smulhi, 573 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 574 ), 575 ( 576 Opcode::UaddSat | Opcode::SaddSat | Opcode::UsubSat | Opcode::SsubSat, 577 &([I32X4, I32X4] | [I64X2, I64X2]) 578 ), 579 (Opcode::Fcopysign, &([F32X4, F32X4] | [F64X2, F64X2])), 580 (Opcode::Popcnt, &([I8X16] | [I16X8] | [I32X4] | [I64X2])), 581 ( 582 Opcode::Umax | Opcode::Smax | Opcode::Umin | Opcode::Smin, 583 &[I64X2, I64X2] 584 ), 585 // https://github.com/bytecodealliance/wasmtime/issues/6104 586 (Opcode::Bitcast, &[I128], &[_]), 587 (Opcode::Bitcast, &[_], &[I128]), 588 (Opcode::Uunarrow), 589 (Opcode::Snarrow | Opcode::Unarrow, &[I64X2, I64X2]), 590 (Opcode::SqmulRoundSat, &[I32X4, I32X4]), 591 // This Icmp is not implemented: #5529 592 (Opcode::Icmp, &[I64X2, I64X2]), 593 // IaddPairwise is implemented, but only for some types, and with some preceding ops. 594 (Opcode::IaddPairwise), 595 // Nothing wrong with this select. But we have an isle rule that can optimize it 596 // into a `min`/`max` instructions, which we don't have implemented yet. 597 (Opcode::Select, &[_, I128, I128]), 598 // These stack accesses can cause segfaults if they are merged into an SSE instruction. 599 // See: #5922 600 ( 601 Opcode::StackStore, 602 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 603 ), 604 ( 605 Opcode::StackLoad, 606 &[], 607 &[I8X16 | I16X8 | I32X4 | I64X2 | F32X4 | F64X2] 608 ), 609 // TODO 610 ( 611 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 612 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 613 ), 614 ( 615 Opcode::Rotr | Opcode::Rotl, 616 &[I8X16 | I16X8 | I32X4 | I64X2, _] 617 ), 618 ) 619 } 620 621 Architecture::Aarch64(_) => { 622 exceptions!( 623 op, 624 args, 625 rets, 626 (Opcode::UmulOverflow | Opcode::SmulOverflow, &[I128, I128]), 627 // https://github.com/bytecodealliance/wasmtime/issues/4864 628 (Opcode::Udiv | Opcode::Sdiv, &[I128, I128]), 629 // https://github.com/bytecodealliance/wasmtime/issues/5472 630 (Opcode::Urem | Opcode::Srem, &[I128, I128]), 631 // https://github.com/bytecodealliance/wasmtime/issues/4313 632 ( 633 Opcode::Smin | Opcode::Umin | Opcode::Smax | Opcode::Umax, 634 &[I128, I128] 635 ), 636 // https://github.com/bytecodealliance/wasmtime/issues/4870 637 (Opcode::Bnot, &[F32 | F64]), 638 ( 639 Opcode::Band 640 | Opcode::Bor 641 | Opcode::Bxor 642 | Opcode::BandNot 643 | Opcode::BorNot 644 | Opcode::BxorNot, 645 &([F32, F32] | [F64, F64]) 646 ), 647 // https://github.com/bytecodealliance/wasmtime/issues/5198 648 (Opcode::Bitselect, &[I128, I128, I128]), 649 // https://github.com/bytecodealliance/wasmtime/issues/4934 650 ( 651 Opcode::FcvtToUint 652 | Opcode::FcvtToUintSat 653 | Opcode::FcvtToSint 654 | Opcode::FcvtToSintSat, 655 &[F32 | F64], 656 &[I128] 657 ), 658 // https://github.com/bytecodealliance/wasmtime/issues/4933 659 ( 660 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 661 &[I128], 662 &[F32 | F64] 663 ), 664 ( 665 Opcode::Umulhi | Opcode::Smulhi, 666 &([I8X16, I8X16] | [I16X8, I16X8] | [I32X4, I32X4] | [I64X2, I64X2]) 667 ), 668 (Opcode::Popcnt, &[I16X8 | I32X4 | I64X2]), 669 // Nothing wrong with this select. But we have an isle rule that can optimize it 670 // into a `min`/`max` instructions, which we don't have implemented yet. 671 (Opcode::Select, &[I8, I128, I128]), 672 // https://github.com/bytecodealliance/wasmtime/issues/6104 673 (Opcode::Bitcast, &[I128], &[_]), 674 (Opcode::Bitcast, &[_], &[I128]), 675 // TODO 676 ( 677 Opcode::Sshr | Opcode::Ushr | Opcode::Ishl, 678 &[I8X16 | I16X8 | I32X4 | I64X2, I128] 679 ), 680 ( 681 Opcode::Rotr | Opcode::Rotl, 682 &[I8X16 | I16X8 | I32X4 | I64X2, _] 683 ), 684 // TODO 685 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 686 (Opcode::VhighBits, &[F32X4 | F64X2]), 687 ) 688 } 689 690 Architecture::S390x => { 691 exceptions!( 692 op, 693 args, 694 rets, 695 (Opcode::UaddOverflow | Opcode::SaddOverflow), 696 (Opcode::UsubOverflow | Opcode::SsubOverflow), 697 (Opcode::UmulOverflow | Opcode::SmulOverflow), 698 ( 699 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 700 &[I128, I128] 701 ), 702 (Opcode::Bnot, &[F32 | F64]), 703 ( 704 Opcode::Band 705 | Opcode::Bor 706 | Opcode::Bxor 707 | Opcode::BandNot 708 | Opcode::BorNot 709 | Opcode::BxorNot, 710 &([F32, F32] | [F64, F64]) 711 ), 712 ( 713 Opcode::FcvtToUint 714 | Opcode::FcvtToUintSat 715 | Opcode::FcvtToSint 716 | Opcode::FcvtToSintSat, 717 &[F32 | F64], 718 &[I128] 719 ), 720 ( 721 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 722 &[I128], 723 &[F32 | F64] 724 ), 725 (Opcode::SsubSat | Opcode::SaddSat, &[I64X2, I64X2]), 726 // https://github.com/bytecodealliance/wasmtime/issues/6104 727 (Opcode::Bitcast, &[I128], &[_]), 728 (Opcode::Bitcast, &[_], &[I128]), 729 // TODO 730 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 731 ) 732 } 733 734 Architecture::Riscv64(_) => { 735 exceptions!( 736 op, 737 args, 738 rets, 739 // TODO 740 (Opcode::UaddOverflow | Opcode::SaddOverflow), 741 (Opcode::UsubOverflow | Opcode::SsubOverflow), 742 (Opcode::UmulOverflow | Opcode::SmulOverflow), 743 // TODO 744 ( 745 Opcode::Udiv | Opcode::Sdiv | Opcode::Urem | Opcode::Srem, 746 &[I128, I128] 747 ), 748 // TODO 749 (Opcode::Iabs, &[I128]), 750 // TODO 751 (Opcode::Bitselect, &[I128, I128, I128]), 752 // https://github.com/bytecodealliance/wasmtime/issues/5528 753 ( 754 Opcode::FcvtToUint | Opcode::FcvtToSint, 755 [F32 | F64], 756 &[I128] 757 ), 758 ( 759 Opcode::FcvtToUintSat | Opcode::FcvtToSintSat, 760 &[F32 | F64], 761 &[I128] 762 ), 763 // https://github.com/bytecodealliance/wasmtime/issues/5528 764 ( 765 Opcode::FcvtFromUint | Opcode::FcvtFromSint, 766 &[I128], 767 &[F32 | F64] 768 ), 769 // https://github.com/bytecodealliance/wasmtime/issues/6104 770 (Opcode::Bitcast, &[I128], &[_]), 771 (Opcode::Bitcast, &[_], &[I128]), 772 // TODO 773 ( 774 Opcode::SelectSpectreGuard, 775 &[_, _, _], 776 &[F32 | F64 | I8X16 | I16X8 | I32X4 | I64X2 | F64X2 | F32X4] 777 ), 778 // TODO 779 (Opcode::Bitselect, &[_, _, _], &[F32 | F64]), 780 ( 781 Opcode::Rotr | Opcode::Rotl, 782 &[I8X16 | I16X8 | I32X4 | I64X2, _] 783 ), 784 ) 785 } 786 787 _ => true, 788 } 789 } 790 791 type OpcodeSignature = (Opcode, Vec<Type>, Vec<Type>); 792 793 static OPCODE_SIGNATURES: Lazy<Vec<OpcodeSignature>> = Lazy::new(|| { 794 let types = &[ 795 I8, I16, I32, I64, I128, // Scalar Integers 796 F32, F64, // Scalar Floats 797 I8X16, I16X8, I32X4, I64X2, // SIMD Integers 798 F32X4, F64X2, // SIMD Floats 799 ]; 800 801 // When this env variable is passed, we only generate instructions for the opcodes listed in 802 // the comma-separated list. This is useful for debugging, as it allows us to focus on a few 803 // specific opcodes. 804 let allowed_opcodes = std::env::var("FUZZGEN_ALLOWED_OPS").ok().map(|s| { 805 s.split(',') 806 .map(|s| s.trim()) 807 .filter(|s| !s.is_empty()) 808 .map(|s| Opcode::from_str(s).expect("Unrecoginzed opcode")) 809 .collect::<Vec<_>>() 810 }); 811 812 Opcode::all() 813 .iter() 814 .filter(|op| { 815 match op { 816 // Control flow opcodes should not be generated through `generate_instructions`. 817 Opcode::BrTable 818 | Opcode::Brif 819 | Opcode::Jump 820 | Opcode::Return 821 | Opcode::ReturnCall 822 | Opcode::ReturnCallIndirect => false, 823 824 // Constants are generated outside of `generate_instructions` 825 Opcode::Iconst => false, 826 827 // TODO: extract_vector raises exceptions during return type generation becuase it 828 // uses dynamic vectors. 829 Opcode::ExtractVector => false, 830 831 _ => true, 832 } 833 }) 834 .flat_map(|op| { 835 let constraints = op.constraints(); 836 837 let ctrl_types = if let Some(ctrls) = constraints.ctrl_typeset() { 838 Vec::from_iter(types.iter().copied().filter(|ty| ctrls.contains(*ty))) 839 } else { 840 vec![INVALID] 841 }; 842 843 ctrl_types.into_iter().flat_map(move |ctrl_type| { 844 let rets = Vec::from_iter( 845 (0..constraints.num_fixed_results()) 846 .map(|i| constraints.result_type(i, ctrl_type)), 847 ); 848 849 // Cols is a vector whose length will match `num_fixed_value_arguments`, and whose 850 // elements will be vectors of types that are valid for that fixed argument 851 // position. 852 let mut cols = vec![]; 853 854 for i in 0..constraints.num_fixed_value_arguments() { 855 match constraints.value_argument_constraint(i, ctrl_type) { 856 ResolvedConstraint::Bound(ty) => cols.push(Vec::from([ty])), 857 ResolvedConstraint::Free(tys) => cols.push(Vec::from_iter( 858 types.iter().copied().filter(|ty| tys.contains(*ty)), 859 )), 860 } 861 } 862 863 // Generate the cartesian product of cols to produce a vector of argument lists, 864 // argss. The argss vector is seeded with the empty argument list, so there's an 865 // initial value to be extended in the loop below. 866 let mut argss = vec![vec![]]; 867 let mut cols = cols.as_slice(); 868 while let Some((col, rest)) = cols.split_last() { 869 cols = rest; 870 871 let mut next = vec![]; 872 for current in argss.iter() { 873 // Extend the front of each argument candidate with every type in `col`. 874 for ty in col { 875 let mut args = vec![*ty]; 876 args.extend_from_slice(¤t); 877 next.push(args); 878 } 879 } 880 881 let _ = std::mem::replace(&mut argss, next); 882 } 883 884 argss.into_iter().map(move |args| (*op, args, rets.clone())) 885 }) 886 }) 887 .filter(|(op, args, rets)| { 888 // These op/signature combinations need to be vetted 889 exceptions!( 890 op, 891 args.as_slice(), 892 rets.as_slice(), 893 (Opcode::Debugtrap), 894 (Opcode::Trap), 895 (Opcode::Trapz), 896 (Opcode::ResumableTrap), 897 (Opcode::Trapnz), 898 (Opcode::ResumableTrapnz), 899 (Opcode::CallIndirect, &[I32]), 900 (Opcode::FuncAddr), 901 (Opcode::X86Pshufb), 902 (Opcode::AvgRound), 903 (Opcode::Uload8x8), 904 (Opcode::Sload8x8), 905 (Opcode::Uload16x4), 906 (Opcode::Sload16x4), 907 (Opcode::Uload32x2), 908 (Opcode::Sload32x2), 909 (Opcode::StackAddr), 910 (Opcode::DynamicStackLoad), 911 (Opcode::DynamicStackStore), 912 (Opcode::DynamicStackAddr), 913 (Opcode::GlobalValue), 914 (Opcode::SymbolValue), 915 (Opcode::TlsValue), 916 (Opcode::GetPinnedReg), 917 (Opcode::SetPinnedReg), 918 (Opcode::GetFramePointer), 919 (Opcode::GetStackPointer), 920 (Opcode::GetReturnAddress), 921 (Opcode::TableAddr), 922 (Opcode::Null), 923 (Opcode::X86Blendv), 924 (Opcode::IcmpImm), 925 (Opcode::X86Pmulhrsw), 926 (Opcode::IaddImm), 927 (Opcode::ImulImm), 928 (Opcode::UdivImm), 929 (Opcode::SdivImm), 930 (Opcode::UremImm), 931 (Opcode::SremImm), 932 (Opcode::IrsubImm), 933 (Opcode::IaddCin), 934 (Opcode::IaddCarry), 935 (Opcode::UaddOverflowTrap), 936 (Opcode::IsubBin), 937 (Opcode::IsubBorrow), 938 (Opcode::BandImm), 939 (Opcode::BorImm), 940 (Opcode::BxorImm), 941 (Opcode::RotlImm), 942 (Opcode::RotrImm), 943 (Opcode::IshlImm), 944 (Opcode::UshrImm), 945 (Opcode::SshrImm), 946 (Opcode::IsNull), 947 (Opcode::IsInvalid), 948 (Opcode::ScalarToVector), 949 (Opcode::X86Pmaddubsw), 950 (Opcode::X86Cvtt2dq), 951 (Opcode::Umulhi, &[I128, I128], &[I128]), 952 (Opcode::Smulhi, &[I128, I128], &[I128]), 953 // https://github.com/bytecodealliance/wasmtime/issues/6073 954 (Opcode::Iconcat, &[I32, I32], &[I64]), 955 (Opcode::Iconcat, &[I16, I16], &[I32]), 956 (Opcode::Iconcat, &[I8, I8], &[I16]), 957 // https://github.com/bytecodealliance/wasmtime/issues/6073 958 (Opcode::Isplit, &[I64], &[I32, I32]), 959 (Opcode::Isplit, &[I32], &[I16, I16]), 960 (Opcode::Isplit, &[I16], &[I8, I8]), 961 (Opcode::Fmin, &[F32X4, F32X4], &[F32X4]), 962 (Opcode::Fmin, &[F64X2, F64X2], &[F64X2]), 963 (Opcode::Fmax, &[F32X4, F32X4], &[F32X4]), 964 (Opcode::Fmax, &[F64X2, F64X2], &[F64X2]), 965 (Opcode::FcvtToUintSat, &[F32X4], &[I8]), 966 (Opcode::FcvtToUintSat, &[F64X2], &[I8]), 967 (Opcode::FcvtToUintSat, &[F32X4], &[I16]), 968 (Opcode::FcvtToUintSat, &[F64X2], &[I16]), 969 (Opcode::FcvtToUintSat, &[F32X4], &[I32]), 970 (Opcode::FcvtToUintSat, &[F64X2], &[I32]), 971 (Opcode::FcvtToUintSat, &[F32X4], &[I64]), 972 (Opcode::FcvtToUintSat, &[F64X2], &[I64]), 973 (Opcode::FcvtToUintSat, &[F32X4], &[I128]), 974 (Opcode::FcvtToUintSat, &[F64X2], &[I128]), 975 (Opcode::FcvtToUintSat, &[F32], &[I8X16]), 976 (Opcode::FcvtToUintSat, &[F64], &[I8X16]), 977 (Opcode::FcvtToUintSat, &[F32X4], &[I8X16]), 978 (Opcode::FcvtToUintSat, &[F64X2], &[I8X16]), 979 (Opcode::FcvtToUintSat, &[F32], &[I16X8]), 980 (Opcode::FcvtToUintSat, &[F64], &[I16X8]), 981 (Opcode::FcvtToUintSat, &[F32X4], &[I16X8]), 982 (Opcode::FcvtToUintSat, &[F64X2], &[I16X8]), 983 (Opcode::FcvtToUintSat, &[F32], &[I32X4]), 984 (Opcode::FcvtToUintSat, &[F64], &[I32X4]), 985 (Opcode::FcvtToUintSat, &[F64X2], &[I32X4]), 986 (Opcode::FcvtToUintSat, &[F32], &[I64X2]), 987 (Opcode::FcvtToUintSat, &[F64], &[I64X2]), 988 (Opcode::FcvtToUintSat, &[F32X4], &[I64X2]), 989 (Opcode::FcvtToSintSat, &[F32X4], &[I8]), 990 (Opcode::FcvtToSintSat, &[F64X2], &[I8]), 991 (Opcode::FcvtToSintSat, &[F32X4], &[I16]), 992 (Opcode::FcvtToSintSat, &[F64X2], &[I16]), 993 (Opcode::FcvtToSintSat, &[F32X4], &[I32]), 994 (Opcode::FcvtToSintSat, &[F64X2], &[I32]), 995 (Opcode::FcvtToSintSat, &[F32X4], &[I64]), 996 (Opcode::FcvtToSintSat, &[F64X2], &[I64]), 997 (Opcode::FcvtToSintSat, &[F32X4], &[I128]), 998 (Opcode::FcvtToSintSat, &[F64X2], &[I128]), 999 (Opcode::FcvtToSintSat, &[F32], &[I8X16]), 1000 (Opcode::FcvtToSintSat, &[F64], &[I8X16]), 1001 (Opcode::FcvtToSintSat, &[F32X4], &[I8X16]), 1002 (Opcode::FcvtToSintSat, &[F64X2], &[I8X16]), 1003 (Opcode::FcvtToSintSat, &[F32], &[I16X8]), 1004 (Opcode::FcvtToSintSat, &[F64], &[I16X8]), 1005 (Opcode::FcvtToSintSat, &[F32X4], &[I16X8]), 1006 (Opcode::FcvtToSintSat, &[F64X2], &[I16X8]), 1007 (Opcode::FcvtToSintSat, &[F32], &[I32X4]), 1008 (Opcode::FcvtToSintSat, &[F64], &[I32X4]), 1009 (Opcode::FcvtToSintSat, &[F64X2], &[I32X4]), 1010 (Opcode::FcvtToSintSat, &[F32], &[I64X2]), 1011 (Opcode::FcvtToSintSat, &[F64], &[I64X2]), 1012 (Opcode::FcvtToSintSat, &[F32X4], &[I64X2]), 1013 (Opcode::FcvtFromUint, &[I8X16], &[F32]), 1014 (Opcode::FcvtFromUint, &[I16X8], &[F32]), 1015 (Opcode::FcvtFromUint, &[I32X4], &[F32]), 1016 (Opcode::FcvtFromUint, &[I64X2], &[F32]), 1017 (Opcode::FcvtFromUint, &[I8X16], &[F64]), 1018 (Opcode::FcvtFromUint, &[I16X8], &[F64]), 1019 (Opcode::FcvtFromUint, &[I32X4], &[F64]), 1020 (Opcode::FcvtFromUint, &[I64X2], &[F64]), 1021 (Opcode::FcvtFromUint, &[I8], &[F32X4]), 1022 (Opcode::FcvtFromUint, &[I16], &[F32X4]), 1023 (Opcode::FcvtFromUint, &[I32], &[F32X4]), 1024 (Opcode::FcvtFromUint, &[I64], &[F32X4]), 1025 (Opcode::FcvtFromUint, &[I128], &[F32X4]), 1026 (Opcode::FcvtFromUint, &[I8X16], &[F32X4]), 1027 (Opcode::FcvtFromUint, &[I16X8], &[F32X4]), 1028 (Opcode::FcvtFromUint, &[I64X2], &[F32X4]), 1029 (Opcode::FcvtFromUint, &[I8], &[F64X2]), 1030 (Opcode::FcvtFromUint, &[I16], &[F64X2]), 1031 (Opcode::FcvtFromUint, &[I32], &[F64X2]), 1032 (Opcode::FcvtFromUint, &[I64], &[F64X2]), 1033 (Opcode::FcvtFromUint, &[I128], &[F64X2]), 1034 (Opcode::FcvtFromUint, &[I8X16], &[F64X2]), 1035 (Opcode::FcvtFromUint, &[I16X8], &[F64X2]), 1036 (Opcode::FcvtFromUint, &[I32X4], &[F64X2]), 1037 (Opcode::FcvtFromSint, &[I8X16], &[F32]), 1038 (Opcode::FcvtFromSint, &[I16X8], &[F32]), 1039 (Opcode::FcvtFromSint, &[I32X4], &[F32]), 1040 (Opcode::FcvtFromSint, &[I64X2], &[F32]), 1041 (Opcode::FcvtFromSint, &[I8X16], &[F64]), 1042 (Opcode::FcvtFromSint, &[I16X8], &[F64]), 1043 (Opcode::FcvtFromSint, &[I32X4], &[F64]), 1044 (Opcode::FcvtFromSint, &[I64X2], &[F64]), 1045 (Opcode::FcvtFromSint, &[I8], &[F32X4]), 1046 (Opcode::FcvtFromSint, &[I16], &[F32X4]), 1047 (Opcode::FcvtFromSint, &[I32], &[F32X4]), 1048 (Opcode::FcvtFromSint, &[I64], &[F32X4]), 1049 (Opcode::FcvtFromSint, &[I128], &[F32X4]), 1050 (Opcode::FcvtFromSint, &[I8X16], &[F32X4]), 1051 (Opcode::FcvtFromSint, &[I16X8], &[F32X4]), 1052 (Opcode::FcvtFromSint, &[I64X2], &[F32X4]), 1053 (Opcode::FcvtFromSint, &[I8], &[F64X2]), 1054 (Opcode::FcvtFromSint, &[I16], &[F64X2]), 1055 (Opcode::FcvtFromSint, &[I32], &[F64X2]), 1056 (Opcode::FcvtFromSint, &[I64], &[F64X2]), 1057 (Opcode::FcvtFromSint, &[I128], &[F64X2]), 1058 (Opcode::FcvtFromSint, &[I8X16], &[F64X2]), 1059 (Opcode::FcvtFromSint, &[I16X8], &[F64X2]), 1060 (Opcode::FcvtFromSint, &[I32X4], &[F64X2]), 1061 ) 1062 }) 1063 .filter(|(op, ..)| { 1064 allowed_opcodes 1065 .as_ref() 1066 .map_or(true, |opcodes| opcodes.contains(op)) 1067 }) 1068 .collect() 1069 }); 1070 1071 fn inserter_for_format(fmt: InstructionFormat) -> OpcodeInserter { 1072 match fmt { 1073 InstructionFormat::AtomicCas => insert_atomic_cas, 1074 InstructionFormat::AtomicRmw => insert_atomic_rmw, 1075 InstructionFormat::Binary => insert_opcode, 1076 InstructionFormat::BinaryImm64 => todo!(), 1077 InstructionFormat::BinaryImm8 => insert_ins_ext_lane, 1078 InstructionFormat::Call => insert_call, 1079 InstructionFormat::CallIndirect => insert_call, 1080 InstructionFormat::CondTrap => todo!(), 1081 InstructionFormat::DynamicStackLoad => todo!(), 1082 InstructionFormat::DynamicStackStore => todo!(), 1083 InstructionFormat::FloatCompare => insert_cmp, 1084 InstructionFormat::FuncAddr => todo!(), 1085 InstructionFormat::IntAddTrap => todo!(), 1086 InstructionFormat::IntCompare => insert_cmp, 1087 InstructionFormat::IntCompareImm => todo!(), 1088 InstructionFormat::Load => insert_load_store, 1089 InstructionFormat::LoadNoOffset => insert_load_store, 1090 InstructionFormat::NullAry => insert_opcode, 1091 InstructionFormat::Shuffle => insert_shuffle, 1092 InstructionFormat::StackLoad => insert_stack_load, 1093 InstructionFormat::StackStore => insert_stack_store, 1094 InstructionFormat::Store => insert_load_store, 1095 InstructionFormat::StoreNoOffset => insert_load_store, 1096 InstructionFormat::TableAddr => todo!(), 1097 InstructionFormat::Ternary => insert_opcode, 1098 InstructionFormat::TernaryImm8 => insert_ins_ext_lane, 1099 InstructionFormat::Trap => todo!(), 1100 InstructionFormat::Unary => insert_opcode, 1101 InstructionFormat::UnaryConst => insert_const, 1102 InstructionFormat::UnaryGlobalValue => todo!(), 1103 InstructionFormat::UnaryIeee32 => insert_const, 1104 InstructionFormat::UnaryIeee64 => insert_const, 1105 InstructionFormat::UnaryImm => insert_const, 1106 1107 InstructionFormat::BranchTable 1108 | InstructionFormat::Brif 1109 | InstructionFormat::Jump 1110 | InstructionFormat::MultiAry => { 1111 panic!( 1112 "Control-flow instructions should be handled by 'insert_terminator': {:?}", 1113 fmt 1114 ) 1115 } 1116 } 1117 } 1118 1119 pub struct FunctionGenerator<'r, 'data> 1120 where 1121 'data: 'r, 1122 { 1123 u: &'r mut Unstructured<'data>, 1124 config: &'r Config, 1125 resources: Resources, 1126 isa: OwnedTargetIsa, 1127 name: UserFuncName, 1128 signature: Signature, 1129 } 1130 1131 #[derive(Debug, Clone)] 1132 enum BlockTerminator { 1133 Return, 1134 Jump(Block), 1135 Br(Block, Block), 1136 BrTable(Block, Vec<Block>), 1137 Switch(Type, Block, HashMap<u128, Block>), 1138 TailCall(FuncRef), 1139 TailCallIndirect(FuncRef), 1140 } 1141 1142 #[derive(Debug, Clone)] 1143 enum BlockTerminatorKind { 1144 Return, 1145 Jump, 1146 Br, 1147 BrTable, 1148 Switch, 1149 TailCall, 1150 TailCallIndirect, 1151 } 1152 1153 /// Alias Analysis Category 1154 /// 1155 /// Our alias analysis pass supports 4 categories of accesses to distinguish 1156 /// different regions. The "Other" region is the general case, and is the default 1157 /// Although they have highly suggestive names there is no difference between any 1158 /// of the categories. 1159 /// 1160 /// We assign each stack slot a category when we first generate them, and then 1161 /// ensure that all accesses to that stack slot are correctly tagged. We already 1162 /// ensure that memory accesses never cross stack slots, so there is no risk 1163 /// of a memory access being tagged with the wrong category. 1164 #[derive(Debug, PartialEq, Clone, Copy)] 1165 enum AACategory { 1166 Other, 1167 Heap, 1168 Table, 1169 VmCtx, 1170 } 1171 1172 impl AACategory { 1173 pub fn all() -> &'static [Self] { 1174 &[ 1175 AACategory::Other, 1176 AACategory::Heap, 1177 AACategory::Table, 1178 AACategory::VmCtx, 1179 ] 1180 } 1181 1182 pub fn update_memflags(&self, flags: &mut MemFlags) { 1183 match self { 1184 AACategory::Other => {} 1185 AACategory::Heap => flags.set_heap(), 1186 AACategory::Table => flags.set_table(), 1187 AACategory::VmCtx => flags.set_vmctx(), 1188 } 1189 } 1190 } 1191 1192 #[derive(Default)] 1193 struct Resources { 1194 vars: HashMap<Type, Vec<Variable>>, 1195 blocks: Vec<(Block, BlockSignature)>, 1196 blocks_without_params: Vec<Block>, 1197 block_terminators: Vec<BlockTerminator>, 1198 func_refs: Vec<(Signature, SigRef, FuncRef)>, 1199 /// This field is required to be sorted by stack slot size at all times. 1200 /// We use this invariant when searching for stack slots with a given size. 1201 /// See [FunctionGenerator::stack_slot_with_size] 1202 stack_slots: Vec<(StackSlot, StackSize, AACategory)>, 1203 usercalls: Vec<(UserExternalName, Signature)>, 1204 libcalls: Vec<LibCall>, 1205 } 1206 1207 impl Resources { 1208 /// Partitions blocks at `block`. Only blocks that can be targeted by branches are considered. 1209 /// 1210 /// The first slice includes all blocks up to and including `block`. 1211 /// The second slice includes all remaining blocks. 1212 fn partition_target_blocks( 1213 &self, 1214 block: Block, 1215 ) -> (&[(Block, BlockSignature)], &[(Block, BlockSignature)]) { 1216 // Blocks are stored in-order and have no gaps, this means that we can simply index them by 1217 // their number. We also need to exclude the entry block since it isn't a valid target. 1218 let target_blocks = &self.blocks[1..]; 1219 target_blocks.split_at(block.as_u32() as usize) 1220 } 1221 1222 /// Returns blocks forward of `block`. Only blocks that can be targeted by branches are considered. 1223 fn forward_blocks(&self, block: Block) -> &[(Block, BlockSignature)] { 1224 let (_, forward_blocks) = self.partition_target_blocks(block); 1225 forward_blocks 1226 } 1227 1228 /// Generates a slice of `blocks_without_params` ahead of `block` 1229 fn forward_blocks_without_params(&self, block: Block) -> &[Block] { 1230 let partition_point = self.blocks_without_params.partition_point(|b| *b <= block); 1231 &self.blocks_without_params[partition_point..] 1232 } 1233 1234 /// Generates an iterator of all valid tail call targets. This includes all functions with both 1235 /// the `tail` calling convention and the same return values as the caller. 1236 fn tail_call_targets<'a>( 1237 &'a self, 1238 caller_sig: &'a Signature, 1239 ) -> impl Iterator<Item = &'a (Signature, SigRef, FuncRef)> { 1240 self.func_refs.iter().filter(|(sig, _, _)| { 1241 sig.call_conv == CallConv::Tail && sig.returns == caller_sig.returns 1242 }) 1243 } 1244 } 1245 1246 impl<'r, 'data> FunctionGenerator<'r, 'data> 1247 where 1248 'data: 'r, 1249 { 1250 pub fn new( 1251 u: &'r mut Unstructured<'data>, 1252 config: &'r Config, 1253 isa: OwnedTargetIsa, 1254 name: UserFuncName, 1255 signature: Signature, 1256 usercalls: Vec<(UserExternalName, Signature)>, 1257 libcalls: Vec<LibCall>, 1258 ) -> Self { 1259 Self { 1260 u, 1261 config, 1262 resources: Resources { 1263 usercalls, 1264 libcalls, 1265 ..Resources::default() 1266 }, 1267 isa, 1268 name, 1269 signature, 1270 } 1271 } 1272 1273 /// Generates a random value for config `param` 1274 fn param(&mut self, param: &RangeInclusive<usize>) -> Result<usize> { 1275 Ok(self.u.int_in_range(param.clone())?) 1276 } 1277 1278 fn system_callconv(&mut self) -> CallConv { 1279 // TODO: This currently only runs on linux, so this is the only choice 1280 // We should improve this once we generate flags and targets 1281 CallConv::SystemV 1282 } 1283 1284 /// Finds a stack slot with size of at least n bytes 1285 fn stack_slot_with_size(&mut self, n: u32) -> Result<(StackSlot, StackSize, AACategory)> { 1286 let first = self 1287 .resources 1288 .stack_slots 1289 .partition_point(|&(_slot, size, _category)| size < n); 1290 Ok(*self.u.choose(&self.resources.stack_slots[first..])?) 1291 } 1292 1293 /// Generates an address that should allow for a store or a load. 1294 /// 1295 /// Addresses aren't generated like other values. They are never stored in variables so that 1296 /// we don't run the risk of returning them from a function, which would make the fuzzer 1297 /// complain since they are different from the interpreter to the backend. 1298 /// 1299 /// `min_size`: Controls the amount of space that the address should have. 1300 /// 1301 /// `aligned`: When passed as true, the resulting address is guaranteed to be aligned 1302 /// on an 8 byte boundary. 1303 /// 1304 /// Returns a valid address and the maximum possible offset that still respects `min_size`. 1305 fn generate_load_store_address( 1306 &mut self, 1307 builder: &mut FunctionBuilder, 1308 min_size: u32, 1309 aligned: bool, 1310 ) -> Result<(Value, u32, AACategory)> { 1311 // TODO: Currently our only source of addresses is stack_addr, but we 1312 // should add global_value, symbol_value eventually 1313 let (addr, available_size, category) = { 1314 let (ss, slot_size, category) = self.stack_slot_with_size(min_size)?; 1315 1316 // stack_slot_with_size guarantees that slot_size >= min_size 1317 let max_offset = slot_size - min_size; 1318 let offset = if aligned { 1319 self.u.int_in_range(0..=max_offset / min_size)? * min_size 1320 } else { 1321 self.u.int_in_range(0..=max_offset)? 1322 }; 1323 1324 let base_addr = builder.ins().stack_addr(I64, ss, offset as i32); 1325 let available_size = slot_size.saturating_sub(offset); 1326 (base_addr, available_size, category) 1327 }; 1328 1329 // TODO: Insert a bunch of amode opcodes here to modify the address! 1330 1331 // Now that we have an address and a size, we just choose a random offset to return to the 1332 // caller. Preserving min_size bytes. 1333 let max_offset = available_size.saturating_sub(min_size); 1334 Ok((addr, max_offset, category)) 1335 } 1336 1337 // Generates an address and memflags for a load or store. 1338 fn generate_address_and_memflags( 1339 &mut self, 1340 builder: &mut FunctionBuilder, 1341 min_size: u32, 1342 is_atomic: bool, 1343 ) -> Result<(Value, MemFlags, Offset32)> { 1344 // Should we generate an aligned address 1345 // Some backends have issues with unaligned atomics. 1346 // AArch64: https://github.com/bytecodealliance/wasmtime/issues/5483 1347 // RISCV: https://github.com/bytecodealliance/wasmtime/issues/5882 1348 let requires_aligned_atomics = matches!( 1349 self.isa.triple().architecture, 1350 Architecture::Aarch64(_) | Architecture::Riscv64(_) 1351 ); 1352 let aligned = if is_atomic && requires_aligned_atomics { 1353 true 1354 } else if min_size > 8 { 1355 // TODO: We currently can't guarantee that a stack_slot will be aligned on a 16 byte 1356 // boundary. We don't have a way to specify alignment when creating stack slots, and 1357 // cranelift only guarantees 8 byte alignment between stack slots. 1358 // See: https://github.com/bytecodealliance/wasmtime/issues/5922#issuecomment-1457926624 1359 false 1360 } else { 1361 bool::arbitrary(self.u)? 1362 }; 1363 1364 let mut flags = MemFlags::new(); 1365 // Even if we picked an aligned address, we can always generate unaligned memflags 1366 if aligned && bool::arbitrary(self.u)? { 1367 flags.set_aligned(); 1368 } 1369 // If the address is aligned, then we know it won't trap 1370 if aligned && bool::arbitrary(self.u)? { 1371 flags.set_notrap(); 1372 } 1373 1374 let (address, max_offset, category) = 1375 self.generate_load_store_address(builder, min_size, aligned)?; 1376 1377 // Set the Alias Analysis bits on the memflags 1378 category.update_memflags(&mut flags); 1379 1380 // Pick an offset to pass into the load/store. 1381 let offset = if aligned { 1382 0 1383 } else { 1384 self.u.int_in_range(0..=max_offset)? as i32 1385 } 1386 .into(); 1387 1388 Ok((address, flags, offset)) 1389 } 1390 1391 /// Get a variable of type `ty` from the current function 1392 fn get_variable_of_type(&mut self, ty: Type) -> Result<Variable> { 1393 let opts = self.resources.vars.get(&ty).map_or(&[][..], Vec::as_slice); 1394 let var = self.u.choose(opts)?; 1395 Ok(*var) 1396 } 1397 1398 /// Generates an instruction(`iconst`/`fconst`/etc...) to introduce a constant value 1399 fn generate_const(&mut self, builder: &mut FunctionBuilder, ty: Type) -> Result<Value> { 1400 Ok(match self.u.datavalue(ty)? { 1401 DataValue::I8(i) => builder.ins().iconst(ty, i as u8 as i64), 1402 DataValue::I16(i) => builder.ins().iconst(ty, i as u16 as i64), 1403 DataValue::I32(i) => builder.ins().iconst(ty, i as u32 as i64), 1404 DataValue::I64(i) => builder.ins().iconst(ty, i), 1405 DataValue::I128(i) => { 1406 let hi = builder.ins().iconst(I64, (i >> 64) as i64); 1407 let lo = builder.ins().iconst(I64, i as i64); 1408 builder.ins().iconcat(lo, hi) 1409 } 1410 DataValue::F32(f) => builder.ins().f32const(f), 1411 DataValue::F64(f) => builder.ins().f64const(f), 1412 DataValue::V128(bytes) => { 1413 let data = bytes.to_vec().into(); 1414 let handle = builder.func.dfg.constants.insert(data); 1415 builder.ins().vconst(ty, handle) 1416 } 1417 _ => unimplemented!(), 1418 }) 1419 } 1420 1421 /// Chooses a random block which can be targeted by a jump / branch. 1422 /// This means any block that is not the first block. 1423 fn generate_target_block(&mut self, source_block: Block) -> Result<Block> { 1424 // We try to mostly generate forward branches to avoid generating an excessive amount of 1425 // infinite loops. But they are still important, so give them a small chance of existing. 1426 let (backwards_blocks, forward_blocks) = 1427 self.resources.partition_target_blocks(source_block); 1428 let ratio = self.config.backwards_branch_ratio; 1429 let block_targets = if !backwards_blocks.is_empty() && self.u.ratio(ratio.0, ratio.1)? { 1430 backwards_blocks 1431 } else { 1432 forward_blocks 1433 }; 1434 assert!(!block_targets.is_empty()); 1435 1436 let (block, _) = self.u.choose(block_targets)?.clone(); 1437 Ok(block) 1438 } 1439 1440 fn generate_values_for_block( 1441 &mut self, 1442 builder: &mut FunctionBuilder, 1443 block: Block, 1444 ) -> Result<Vec<Value>> { 1445 let (_, sig) = self.resources.blocks[block.as_u32() as usize].clone(); 1446 self.generate_values_for_signature(builder, sig.iter().copied()) 1447 } 1448 1449 fn generate_values_for_signature<I: Iterator<Item = Type>>( 1450 &mut self, 1451 builder: &mut FunctionBuilder, 1452 signature: I, 1453 ) -> Result<Vec<Value>> { 1454 signature 1455 .map(|ty| { 1456 let var = self.get_variable_of_type(ty)?; 1457 let val = builder.use_var(var); 1458 Ok(val) 1459 }) 1460 .collect() 1461 } 1462 1463 /// The terminator that we need to insert has already been picked ahead of time 1464 /// we just need to build the instructions for it 1465 fn insert_terminator( 1466 &mut self, 1467 builder: &mut FunctionBuilder, 1468 source_block: Block, 1469 ) -> Result<()> { 1470 let terminator = self.resources.block_terminators[source_block.as_u32() as usize].clone(); 1471 1472 match terminator { 1473 BlockTerminator::Return => { 1474 let types: Vec<Type> = { 1475 let rets = &builder.func.signature.returns; 1476 rets.iter().map(|p| p.value_type).collect() 1477 }; 1478 let vals = self.generate_values_for_signature(builder, types.into_iter())?; 1479 1480 builder.ins().return_(&vals[..]); 1481 } 1482 BlockTerminator::Jump(target) => { 1483 let args = self.generate_values_for_block(builder, target)?; 1484 builder.ins().jump(target, &args[..]); 1485 } 1486 BlockTerminator::Br(left, right) => { 1487 let left_args = self.generate_values_for_block(builder, left)?; 1488 let right_args = self.generate_values_for_block(builder, right)?; 1489 1490 let condbr_types = [I8, I16, I32, I64, I128]; 1491 let _type = *self.u.choose(&condbr_types[..])?; 1492 let val = builder.use_var(self.get_variable_of_type(_type)?); 1493 builder 1494 .ins() 1495 .brif(val, left, &left_args[..], right, &right_args[..]); 1496 } 1497 BlockTerminator::BrTable(default, targets) => { 1498 // Create jump tables on demand 1499 let mut jt = Vec::with_capacity(targets.len()); 1500 for block in targets { 1501 let args = self.generate_values_for_block(builder, block)?; 1502 jt.push(builder.func.dfg.block_call(block, &args)) 1503 } 1504 1505 let args = self.generate_values_for_block(builder, default)?; 1506 let jt_data = JumpTableData::new(builder.func.dfg.block_call(default, &args), &jt); 1507 let jt = builder.create_jump_table(jt_data); 1508 1509 // br_table only supports I32 1510 let val = builder.use_var(self.get_variable_of_type(I32)?); 1511 1512 builder.ins().br_table(val, jt); 1513 } 1514 BlockTerminator::Switch(_type, default, entries) => { 1515 let mut switch = Switch::new(); 1516 for (&entry, &block) in entries.iter() { 1517 switch.set_entry(entry, block); 1518 } 1519 1520 let switch_val = builder.use_var(self.get_variable_of_type(_type)?); 1521 1522 switch.emit(builder, switch_val, default); 1523 } 1524 BlockTerminator::TailCall(target) | BlockTerminator::TailCallIndirect(target) => { 1525 let (sig, sig_ref, func_ref) = self 1526 .resources 1527 .func_refs 1528 .iter() 1529 .find(|(_, _, f)| *f == target) 1530 .expect("Failed to find previously selected function") 1531 .clone(); 1532 1533 let opcode = match terminator { 1534 BlockTerminator::TailCall(_) => Opcode::ReturnCall, 1535 BlockTerminator::TailCallIndirect(_) => Opcode::ReturnCallIndirect, 1536 _ => unreachable!(), 1537 }; 1538 1539 insert_call_to_function(self, builder, opcode, &sig, sig_ref, func_ref)?; 1540 } 1541 } 1542 1543 Ok(()) 1544 } 1545 1546 /// Fills the current block with random instructions 1547 fn generate_instructions(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1548 for _ in 0..self.param(&self.config.instructions_per_block)? { 1549 let (op, args, rets) = self.u.choose(&OPCODE_SIGNATURES)?; 1550 1551 // We filter out instructions that aren't supported by the target at this point instead 1552 // of building a single vector of valid instructions at the beginning of function 1553 // generation, to avoid invalidating the corpus when instructions are enabled/disabled. 1554 if !valid_for_target(&self.isa.triple(), *op, &args, &rets) { 1555 return Err(arbitrary::Error::IncorrectFormat.into()); 1556 } 1557 1558 let inserter = inserter_for_format(op.format()); 1559 inserter(self, builder, *op, &args, &rets)?; 1560 } 1561 1562 Ok(()) 1563 } 1564 1565 fn generate_funcrefs(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1566 let usercalls: Vec<(ExternalName, Signature)> = self 1567 .resources 1568 .usercalls 1569 .iter() 1570 .map(|(name, signature)| { 1571 let user_func_ref = builder.func.declare_imported_user_function(name.clone()); 1572 let name = ExternalName::User(user_func_ref); 1573 (name, signature.clone()) 1574 }) 1575 .collect(); 1576 1577 let lib_callconv = self.system_callconv(); 1578 let libcalls: Vec<(ExternalName, Signature)> = self 1579 .resources 1580 .libcalls 1581 .iter() 1582 .map(|libcall| { 1583 let pointer_type = Type::int_with_byte_size( 1584 self.isa.triple().pointer_width().unwrap().bytes().into(), 1585 ) 1586 .unwrap(); 1587 let signature = libcall.signature(lib_callconv, pointer_type); 1588 let name = ExternalName::LibCall(*libcall); 1589 (name, signature) 1590 }) 1591 .collect(); 1592 1593 for (name, signature) in usercalls.into_iter().chain(libcalls) { 1594 let sig_ref = builder.import_signature(signature.clone()); 1595 let func_ref = builder.import_function(ExtFuncData { 1596 name, 1597 signature: sig_ref, 1598 colocated: self.u.arbitrary()?, 1599 }); 1600 1601 self.resources 1602 .func_refs 1603 .push((signature, sig_ref, func_ref)); 1604 } 1605 1606 Ok(()) 1607 } 1608 1609 fn generate_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1610 for _ in 0..self.param(&self.config.static_stack_slots_per_function)? { 1611 let bytes = self.param(&self.config.static_stack_slot_size)? as u32; 1612 let ss_data = StackSlotData::new(StackSlotKind::ExplicitSlot, bytes); 1613 let slot = builder.create_sized_stack_slot(ss_data); 1614 1615 // Generate one Alias Analysis Category for each slot 1616 let category = *self.u.choose(AACategory::all())?; 1617 1618 self.resources.stack_slots.push((slot, bytes, category)); 1619 } 1620 1621 self.resources 1622 .stack_slots 1623 .sort_unstable_by_key(|&(_slot, bytes, _category)| bytes); 1624 1625 Ok(()) 1626 } 1627 1628 /// Zero initializes the stack slot by inserting `stack_store`'s. 1629 fn initialize_stack_slots(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1630 let i8_zero = builder.ins().iconst(I8, 0); 1631 let i16_zero = builder.ins().iconst(I16, 0); 1632 let i32_zero = builder.ins().iconst(I32, 0); 1633 let i64_zero = builder.ins().iconst(I64, 0); 1634 let i128_zero = builder.ins().uextend(I128, i64_zero); 1635 1636 for &(slot, init_size, category) in self.resources.stack_slots.iter() { 1637 let mut size = init_size; 1638 1639 // Insert the largest available store for the remaining size. 1640 while size != 0 { 1641 let offset = (init_size - size) as i32; 1642 let (val, filled) = match size { 1643 sz if sz / 16 > 0 => (i128_zero, 16), 1644 sz if sz / 8 > 0 => (i64_zero, 8), 1645 sz if sz / 4 > 0 => (i32_zero, 4), 1646 sz if sz / 2 > 0 => (i16_zero, 2), 1647 _ => (i8_zero, 1), 1648 }; 1649 let addr = builder.ins().stack_addr(I64, slot, offset); 1650 1651 // Each stack slot has an associated category, that means we have to set the 1652 // correct memflags for it. So we can't use `stack_store` directly. 1653 let mut flags = MemFlags::new(); 1654 flags.set_notrap(); 1655 category.update_memflags(&mut flags); 1656 1657 builder.ins().store(flags, val, addr, 0); 1658 1659 size -= filled; 1660 } 1661 } 1662 Ok(()) 1663 } 1664 1665 /// Creates a random amount of blocks in this function 1666 fn generate_blocks(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1667 let extra_block_count = self.param(&self.config.blocks_per_function)?; 1668 1669 // We must always have at least one block, so we generate the "extra" blocks and add 1 for 1670 // the entry block. 1671 let block_count = 1 + extra_block_count; 1672 1673 // Blocks need to be sorted in ascending order 1674 self.resources.blocks = (0..block_count) 1675 .map(|i| { 1676 let is_entry = i == 0; 1677 let block = builder.create_block(); 1678 1679 // Optionally mark blocks that are not the entry block as cold 1680 if !is_entry { 1681 if bool::arbitrary(self.u)? { 1682 builder.set_cold_block(block); 1683 } 1684 } 1685 1686 // The first block has to have the function signature, but for the rest of them we generate 1687 // a random signature; 1688 if is_entry { 1689 builder.append_block_params_for_function_params(block); 1690 Ok(( 1691 block, 1692 self.signature.params.iter().map(|a| a.value_type).collect(), 1693 )) 1694 } else { 1695 let sig = self.generate_block_signature()?; 1696 sig.iter().for_each(|ty| { 1697 builder.append_block_param(block, *ty); 1698 }); 1699 Ok((block, sig)) 1700 } 1701 }) 1702 .collect::<Result<Vec<_>>>()?; 1703 1704 // Valid blocks for jump tables have to have no parameters in the signature, and must also 1705 // not be the first block. 1706 self.resources.blocks_without_params = self.resources.blocks[1..] 1707 .iter() 1708 .filter(|(_, sig)| sig.len() == 0) 1709 .map(|(b, _)| *b) 1710 .collect(); 1711 1712 // Compute the block CFG 1713 // 1714 // cranelift-frontend requires us to never generate unreachable blocks 1715 // To ensure this property we start by constructing a main "spine" of blocks. So block1 can 1716 // always jump to block2, and block2 can always jump to block3, etc... 1717 // 1718 // That is not a very interesting CFG, so we introduce variations on that, but always 1719 // ensuring that the property of pointing to the next block is maintained whatever the 1720 // branching mechanism we use. 1721 let blocks = self.resources.blocks.clone(); 1722 self.resources.block_terminators = blocks 1723 .iter() 1724 .map(|&(block, _)| { 1725 let next_block = Block::with_number(block.as_u32() + 1).unwrap(); 1726 let forward_blocks = self.resources.forward_blocks(block); 1727 let paramless_targets = self.resources.forward_blocks_without_params(block); 1728 let has_paramless_targets = !paramless_targets.is_empty(); 1729 let next_block_is_paramless = paramless_targets.contains(&next_block); 1730 1731 let mut valid_terminators = vec![]; 1732 1733 if forward_blocks.is_empty() { 1734 // Return is only valid on the last block. 1735 valid_terminators.push(BlockTerminatorKind::Return); 1736 } else { 1737 // If we have more than one block we can allow terminators that target blocks. 1738 // TODO: We could add some kind of BrReturn here, to explore edges where we 1739 // exit in the middle of the function 1740 valid_terminators.extend_from_slice(&[ 1741 BlockTerminatorKind::Jump, 1742 BlockTerminatorKind::Br, 1743 BlockTerminatorKind::BrTable, 1744 ]); 1745 } 1746 1747 // As the Switch interface only allows targeting blocks without params we need 1748 // to ensure that the next block has no params, since that one is guaranteed to be 1749 // picked in either case. 1750 if has_paramless_targets && next_block_is_paramless { 1751 valid_terminators.push(BlockTerminatorKind::Switch); 1752 } 1753 1754 // Tail Calls are a block terminator, so we should insert them as any other block 1755 // terminator. We should ensure that we can select at least one target before considering 1756 // them as candidate instructions. 1757 let has_tail_callees = self 1758 .resources 1759 .tail_call_targets(&self.signature) 1760 .next() 1761 .is_some(); 1762 let is_tail_caller = self.signature.call_conv == CallConv::Tail; 1763 1764 let supports_tail_calls = match self.isa.triple().architecture { 1765 Architecture::Aarch64(_) | Architecture::Riscv64(_) => true, 1766 // TODO: x64 currently requires frame pointers for tail calls. 1767 Architecture::X86_64 => self.isa.flags().preserve_frame_pointers(), 1768 // TODO: Other platforms do not support tail calls yet. 1769 _ => false, 1770 }; 1771 1772 if is_tail_caller && has_tail_callees && supports_tail_calls { 1773 valid_terminators.extend([ 1774 BlockTerminatorKind::TailCall, 1775 BlockTerminatorKind::TailCallIndirect, 1776 ]); 1777 } 1778 1779 let terminator = self.u.choose(&valid_terminators)?; 1780 1781 // Choose block targets for the terminators that we picked above 1782 Ok(match terminator { 1783 BlockTerminatorKind::Return => BlockTerminator::Return, 1784 BlockTerminatorKind::Jump => BlockTerminator::Jump(next_block), 1785 BlockTerminatorKind::Br => { 1786 BlockTerminator::Br(next_block, self.generate_target_block(block)?) 1787 } 1788 // TODO: Allow generating backwards branches here 1789 BlockTerminatorKind::BrTable => { 1790 // Make the default the next block, and then we don't have to worry 1791 // that we can reach it via the targets 1792 let default = next_block; 1793 1794 let target_count = self.param(&self.config.jump_table_entries)?; 1795 let targets = Result::from_iter( 1796 (0..target_count).map(|_| self.generate_target_block(block)), 1797 )?; 1798 1799 BlockTerminator::BrTable(default, targets) 1800 } 1801 BlockTerminatorKind::Switch => { 1802 // Make the default the next block, and then we don't have to worry 1803 // that we can reach it via the entries below 1804 let default_block = next_block; 1805 1806 let _type = *self.u.choose(&[I8, I16, I32, I64, I128][..])?; 1807 1808 // Build this into a HashMap since we cannot have duplicate entries. 1809 let mut entries = HashMap::new(); 1810 for _ in 0..self.param(&self.config.switch_cases)? { 1811 // The Switch API only allows for entries that are addressable by the index type 1812 // so we need to limit the range of values that we generate. 1813 let (ty_min, ty_max) = _type.bounds(false); 1814 let range_start = self.u.int_in_range(ty_min..=ty_max)?; 1815 1816 // We can either insert a contiguous range of blocks or a individual block 1817 // This is done because the Switch API specializes contiguous ranges. 1818 let range_size = if bool::arbitrary(self.u)? { 1819 1 1820 } else { 1821 self.param(&self.config.switch_max_range_size)? 1822 } as u128; 1823 1824 // Build the switch entries 1825 for i in 0..range_size { 1826 let index = range_start.wrapping_add(i) % ty_max; 1827 let block = *self 1828 .u 1829 .choose(self.resources.forward_blocks_without_params(block))?; 1830 1831 entries.insert(index, block); 1832 } 1833 } 1834 1835 BlockTerminator::Switch(_type, default_block, entries) 1836 } 1837 BlockTerminatorKind::TailCall => { 1838 let targets = self 1839 .resources 1840 .tail_call_targets(&self.signature) 1841 .collect::<Vec<_>>(); 1842 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1843 BlockTerminator::TailCall(*funcref) 1844 } 1845 BlockTerminatorKind::TailCallIndirect => { 1846 let targets = self 1847 .resources 1848 .tail_call_targets(&self.signature) 1849 .collect::<Vec<_>>(); 1850 let (_, _, funcref) = *self.u.choose(&targets[..])?; 1851 BlockTerminator::TailCallIndirect(*funcref) 1852 } 1853 }) 1854 }) 1855 .collect::<Result<_>>()?; 1856 1857 Ok(()) 1858 } 1859 1860 fn generate_block_signature(&mut self) -> Result<BlockSignature> { 1861 let param_count = self.param(&self.config.block_signature_params)?; 1862 1863 let mut params = Vec::with_capacity(param_count); 1864 for _ in 0..param_count { 1865 params.push(self.u._type((&*self.isa).supports_simd())?); 1866 } 1867 Ok(params) 1868 } 1869 1870 fn build_variable_pool(&mut self, builder: &mut FunctionBuilder) -> Result<()> { 1871 let block = builder.current_block().unwrap(); 1872 1873 // Define variables for the function signature 1874 let mut vars: Vec<_> = builder 1875 .func 1876 .signature 1877 .params 1878 .iter() 1879 .map(|param| param.value_type) 1880 .zip(builder.block_params(block).iter().copied()) 1881 .collect(); 1882 1883 // Create a pool of vars that are going to be used in this function 1884 for _ in 0..self.param(&self.config.vars_per_function)? { 1885 let ty = self.u._type((&*self.isa).supports_simd())?; 1886 let value = self.generate_const(builder, ty)?; 1887 vars.push((ty, value)); 1888 } 1889 1890 for (id, (ty, value)) in vars.into_iter().enumerate() { 1891 let var = Variable::new(id); 1892 builder.declare_var(var, ty); 1893 builder.def_var(var, value); 1894 self.resources 1895 .vars 1896 .entry(ty) 1897 .or_insert_with(Vec::new) 1898 .push(var); 1899 } 1900 1901 Ok(()) 1902 } 1903 1904 /// We generate a function in multiple stages: 1905 /// 1906 /// * First we generate a random number of empty blocks 1907 /// * Then we generate a random pool of variables to be used throughout the function 1908 /// * We then visit each block and generate random instructions 1909 /// 1910 /// Because we generate all blocks and variables up front we already know everything that 1911 /// we need when generating instructions (i.e. jump targets / variables) 1912 pub fn generate(mut self) -> Result<Function> { 1913 let mut fn_builder_ctx = FunctionBuilderContext::new(); 1914 let mut func = Function::with_name_signature(self.name.clone(), self.signature.clone()); 1915 1916 let mut builder = FunctionBuilder::new(&mut func, &mut fn_builder_ctx); 1917 1918 // Build the function references before generating the block CFG since we store 1919 // function references in the CFG. 1920 self.generate_funcrefs(&mut builder)?; 1921 self.generate_blocks(&mut builder)?; 1922 1923 // Function preamble 1924 self.generate_stack_slots(&mut builder)?; 1925 1926 // Main instruction generation loop 1927 for (block, block_sig) in self.resources.blocks.clone().into_iter() { 1928 let is_block0 = block.as_u32() == 0; 1929 builder.switch_to_block(block); 1930 1931 if is_block0 { 1932 // The first block is special because we must create variables both for the 1933 // block signature and for the variable pool. Additionally, we must also define 1934 // initial values for all variables that are not the function signature. 1935 self.build_variable_pool(&mut builder)?; 1936 1937 // Stack slots have random bytes at the beginning of the function 1938 // initialize them to a constant value so that execution stays predictable. 1939 self.initialize_stack_slots(&mut builder)?; 1940 } else { 1941 // Define variables for the block params 1942 for (i, ty) in block_sig.iter().enumerate() { 1943 let var = self.get_variable_of_type(*ty)?; 1944 let block_param = builder.block_params(block)[i]; 1945 builder.def_var(var, block_param); 1946 } 1947 } 1948 1949 // Generate block instructions 1950 self.generate_instructions(&mut builder)?; 1951 1952 // Insert a terminator to safely exit the block 1953 self.insert_terminator(&mut builder, block)?; 1954 } 1955 1956 builder.seal_all_blocks(); 1957 builder.finalize(); 1958 1959 Ok(func) 1960 } 1961 } 1962