## Description The `Rv32BranchLessThan256Chip` uses the wrong opcode offset when creating its internal `BranchLessThanCoreChip`. It uses `Rv32LessThan256Opcode::CLASS_OFFSET` instead of `Rv32BranchLessThan256Opcode::CLASS_OFFSET`. This will cause the branch less than instruction to be decoded and executed incorrectly. This affects all 256-bit branch-less-than operations (BLT, BGE, BLTU, BGEU) leading to incorrect branch condition evaluation. ## Proof of Concept This part of the code from the `Int256::build` method shows the issue: https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/bigint/circuit/src/extension.rs#L184-L196 ```rust 184: let branch_less_than_chip = Rv32BranchLessThan256Chip::new( 185: Rv32HeapBranchAdapterChip::new( 186: execution_bus, 187: program_bus, 188: memory_bridge, 189: address_bits, 190: bitwise_lu_chip.clone(), 191: ), 192: BranchLessThanCoreChip::new( 193: bitwise_lu_chip.clone(), 194: Rv32LessThan256Opcode::CLASS_OFFSET, ❌ 195: ), 196: offline_memory.clone(), ``` This code will cause branch instructions to be decoded as regular less-than operations which will bypass control flow checks. Also, this isn't consistent with other patterns seen in other chip implementations where each chip uses its corresponding opcode's `CLASS_OFFSET`. For example: - [`BaseAluCoreChip`](https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/bigint/circuit/src/extension.rs#L144) uses `Rv32BaseAlu256Opcode::CLASS_OFFSET` - [`BranchEqualCoreChip`](https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/bigint/circuit/src/extension.rs#L176) uses `Rv32BranchEqual256Opcode::CLASS_OFFSET` If we take a look at [`core.rs`](https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/rv32im/circuit/src/branch_lt/core.rs) file, we can also see that: The `BranchLessThanCoreChip` is built to handle branch comparison operations with dedicated logic for: 1. Branch-specific opcode flags: ```rust 37: pub opcode_blt_flag: T, 38: pub opcode_bltu_flag: T, 39: pub opcode_bge_flag: T, 40: pub opcode_bgeu_flag: T, ``` 2. Branch-specific comparison logic that takees care of both signed and unsigned comparisons: ```rust 104: let lt = cols.opcode_blt_flag + cols.opcode_bltu_flag; 105: let ge = cols.opcode_bge_flag + cols.opcode_bgeu_flag; 106: let signed = cols.opcode_blt_flag + cols.opcode_bge_flag; ``` 3. Program counter are updated based on comparison results: ```rust 159: let to_pc = from_pc 160: + cols.cmp_result * cols.imm 161: + not(cols.cmp_result) * AB::Expr::from_canonical_u32(DEFAULT_PC_STEP); ``` Using `Rv32LessThan256Opcode::CLASS_OFFSET` instead of `Rv32BranchLessThan256Opcode::CLASS_OFFSET` means: - The opcode validation in `core.rs` will fail to match the actual instruction opcode - The branch-specific flags wont be set correctly - The PC updates will be incorrect, potentially causing the program to continue execution at the wrong address - The signed vs unsigned comparison handling will be broken The mismatch between the core chip's implementation and the opcode offset used to create it explains why branch conditions would be evaluated incorrectly. Also look at this [`execute_instruction`](https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/rv32im/circuit/src/branch_lt/core.rs#L217-L235) function: ```rust 217: impl, const NUM_LIMBS: usize, const LIMB_BITS: usize> 218: VmCoreChip for BranchLessThanCoreChip 219: where 220: I::Reads: Into<[[F; NUM_LIMBS]; 2]>, 221: I::Writes: Default, 222: { 223: type Record = BranchLessThanCoreRecord; 224: type Air = BranchLessThanCoreAir; 225: 226: #[allow(clippy::type_complexity)] 227: fn execute_instruction( 228: &self, 229: instruction: &Instruction, 230: from_pc: u32, 231: reads: I::Reads, 232: ) -> Result<(AdapterRuntimeContext, Self::Record)> { 233: let Instruction { opcode, c: imm, .. } = *instruction; 234: let blt_opcode = BranchLessThanOpcode::from_usize(opcode.local_opcode_idx(self.air.offset)); ``` The chip converts the opcode using `local_opcode_idx(self.air.offset)` to get the appropriate branch operation. Using the wrong offset means branch operations will be incorrectly indexed, executing the wrong comparison type. The control flow logic in `execute_instruction` that handles PC updates based on comparison results will operate on incorrect opcodes. To demonstrate this, add this import to [`cargo.toml`](https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/rv32im/circuit/Cargo.toml) on line 19: ```rust openvm-bigint-transpiler = { workspace = true } ``` Add this import to [`tests.rs`](https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/rv32im/circuit/src/branch_lt/tests.rs) on line 15: ```rust use openvm_bigint_transpiler::Rv32LessThan256Opcode; ``` Then paste this in the same [`tests.rs`](https://github.com/openvm-org/openvm/blob/f1b484499b9c059d14949cdfaa648906757ca7aa/extensions/rv32im/circuit/src/branch_lt/tests.rs) file: ```rust #[test] #[should_panic(expected = "attempt to subtract with overflow")] fn rv32_branch_lt_wrong_offset_test() { let bitwise_bus = BitwiseOperationLookupBus::new(BITWISE_OP_LOOKUP_BUS); let bitwise_chip = SharedBitwiseOperationLookupChip::::new(bitwise_bus); let mut tester = VmChipTestBuilder::default(); // Create chip with wrong offset (using LessThan256 instead of BranchLessThan256) let mut chip_wrong_offset = Rv32BranchLessThanChip::::new( Rv32BranchAdapterChip::new( tester.execution_bus(), tester.program_bus(), tester.memory_bridge(), ), BranchLessThanCoreChip::new(bitwise_chip.clone(), Rv32LessThan256Opcode::CLASS_OFFSET), tester.offline_memory_mutex_arc(), ); let mut rng = create_seeded_rng(); // Test with values that should trigger a branch let a = [10, 0, 0, 0]; // Smaller value let b = [20, 0, 0, 0]; // Larger value let imm: i32 = 16; // Branch offset let rs1 = gen_pointer(&mut rng, 4); let rs2 = gen_pointer(&mut rng, 4); tester.write::(1, rs1, a.map(F::from_canonical_u32)); tester.write::(1, rs2, b.map(F::from_canonical_u32)); // This should fail due to incorrect opcode offset causing overflow let from_pc = rng.gen_range(imm.unsigned_abs()..(1 << (PC_BITS - 1))); tester.execute_with_pc( &mut chip_wrong_offset, &Instruction::from_isize( BranchLessThanOpcode::BLT.global_opcode(), rs1 as isize, rs2 as isize, imm as isize, 1, 1, ), from_pc, ); // We shouldn't reach this point panic!("Test should have failed with overflow error"); } ``` #### Logs ```shell Blocking waiting for file lock on build directory Compiling openvm-rv32im-circuit v1.0.0-rc.1 (/Users/rhaydden/openvm/extensions/rv32im/circuit) Finished `test` profile [optimized + debuginfo] target(s) in 15.75s Running unittests src/lib.rs (target/debug/deps/openvm_rv32im_circuit-9408da59e6f0f62e) running 1 test thread 'branch_lt::tests::rv32_branch_lt_wrong_offset_test' panicked at crates/toolchain/instructions/src/lib.rs:38:9: attempt to subtract with overflow note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace test branch_lt::tests::rv32_branch_lt_wrong_offset_test - should panic ... ok test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 173 filtered out; finished in 0.07s ``` Test loogs actually show that when `BranchLessThanOpcode::BLT.global_opcode()` is used in the instruction: - This gives us a global opcode for the BLT operation - The chip tries to map this global opcode to a local opcode using `opcode.local_opcode_idx(self.air.offset)` - But we providded `Rv32LessThan256Opcode::CLASS_OFFSET` instead of `Rv32BranchLessThan256Opcode::CLASS_OFFSET` The overflow happens because: - The global opcode for BLT is in the branch instruction range - When we subtract the wrong offset (`Rv32LessThan256Opcode::CLASS_OFFSET`), which is in a different range - We get an invalid result that causes arithmetic overflow ## Fix The `BranchLessThanCoreChip` within the `Rv32BranchLessThan256Chip` should make use of the opcode offset specifically for the branch less than instruction, which is `Rv32BranchLessThan256Opcode::CLASS_OFFSET`. Using the less than opcode offset will result in incorrect behavior. ```diff BranchLessThanCoreChip::new( bitwise_lu_chip.clone(), - Rv32LessThan256Opcode::CLASS_OFFSET, + Rv32BranchLessThan256Opcode::CLASS_OFFSET, ), offline_memory.clone(), ); ```

Contest	Position	Date	Payout
solaxy	1 / 789	June 2025	$13,951
openvm	5 / 327	January 2025	$4,524
hyperlend	5 / 260	November 2024	$2,708
Soon	8 / 452	December 2024	$1,624
ton-pool-contracts	7 / 208	January 2025	$947

Rhaydden

Top competitions