Files
revive/crates/integration/contracts/UnalignedMStoreBug.sol
T
98d818f958 The new Yul optimizer toolkit (#519)
This PR introduces powerful code size optimizations. This is achieved
with an additional IR layer in between Yul and LLVM IR. It works quite
well (see below for a comparison) because it captures optimization gains
that neither solc nor LLVM can realize.

The feature is gated behind `--newyork` or a corresponding standard JSON
input setting and **off by default**.

### How it works

We implemented an SSA IR modelled after or inspired by the [MLIR SCF
dialect](https://mlir.llvm.org/docs/Dialects/SCFDialect/). Yul is
translated into this IR and the IR implements mainly the following
optimizations:

- **Type narrowing**: narrows 256-bit values to I1..I160 where provable
(bidirectional + interprocedural), so LLVM emits native 32/64-bit ops
instead
    of software 256-bit arithmetic.
- **FMP range proof**: exposes the `mload(0x40) < heap_size` bound to
LLVM
IPSCCP, eliminating trailing overflow checks on FMP-derived offsets.
Single
    largest code-size contributor.
- **Heap analysis**: skips big-endian byte-swaps where native
little-endian
    layout is safe.
  - **mem_opt / keccak fusion+folding**, **compound outlining** (mapping
sload/sstore), **guard narrowing**, **inlining** tuned for PolkaVM, and
    **exact + fuzzy function dedup**.

### Results
The optimizations implemented with the newyork IR bring substantial code
size improvements.
  - Integration contracts: approximately **-20% to -50%** code size.
  - OpenZeppelin Wizard corpus: around **-47%** code size.

This table shows the code size improvements on various OpenZeppelin ERC
implementations, taken from their [contracts
wizard](https://wizard.openzeppelin.com/), with almost all features
enabled.

| Contract | Default: (bytes) | With `--newyork`: (bytes) | Reduction |
|---|---:|---:|---:|
| erc1155 | 59,206 | 30,649 | -48.2% |
| erc20 | 81,697 | 42,893 | -47.5% |
| erc721 | 92,052 | 49,423 | -46.3% |
| oz_gov | 141,861 | 81,159 | -42.8% |
| oz_rwa | 77,274 | 37,975 | -50.9% |
| oz_simple_erc20 | 29,950 | 16,554 | -44.7% |
| oz_stable | 82,117 | 39,032 | -52.5% |
| **Total** | **564,157** | **297,685** | **-47.2%** |

We believe this to be a reasonably representative dataset and conclude
that the newyork IR achieves drastic code size reductions on real world
dApp implementations.

### Methodology, security and runtime performance

Since the main contributors to these improvements are the FMP range
proofs and the type narrowing, we don't expect big runtime performance
regressions (it may even benefit runtime performance in some cases).

The PR also includes docs
(`book/src/developer_guide/newyork_{optimizer,ir}.md`) and many new
integration tests, including regression tests for known FMP soundness
traps.

The PR was authored collaboratively and reviewed extensively by the
`revive` maintainers, as well as Claude Opus, Claude Fable, Qwen,
Minimax and Deepseek LLMs, over a span of many months.

The `newyork` compiler pipeline introduces no new regressions over our
integration and differential test suites (containing over 30000 test
cases). This was achieved by careful manual reviews and many LLM bughunt
loops. Additionally, a final security review by Anthropic's Fable 5 LLM
found no remaining soundness issues. Nonetheless, like any brand new
software, please consider this an experimental feature upon release.

---------

Signed-off-by: xermicus <bigcyrill@hotmail.com>
Signed-off-by: xermicus <cyrill@parity.io>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: neo <me@90f9.org>
Co-authored-by: LJ <81748770+elle-j@users.noreply.github.com>
2026-06-17 12:35:16 +02:00

29 lines
1.4 KiB
Solidity

// SPDX-License-Identifier: MIT
pragma solidity ^0.8;
/// Soundness PoC: `mem_opt.rs` only updates `memory_state[word_offset]` when
/// recording an mstore, where `word_offset = static_offset / 32 * 32`. An
/// *unaligned* `mstore(0x70, v)` writes 32 bytes from byte 0x70 to 0x90, so
/// it overwrites the lower half of the word starting at 0x80. mem_opt only
/// touches `memory_state[0x60]` and leaves `memory_state[0x80]` intact.
///
/// A subsequent `mload(0x80)` finds the still-cached tracked entry at 0x80
/// (offset == 0x80, no aliasing check across overlap with 0x70) and is
/// forwarded to the *pre-overwrite* stored value. EVM reads the actual
/// memory which contains the lower half of the unaligned store plus the
/// untouched upper half of the original.
contract UnalignedMStoreBug {
function bug() external pure returns (bytes32 r) {
assembly {
// 32 byte sentinel at the aligned word 0x80.
mstore(0x80, 0xaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa)
// Unaligned write: 16 bytes of 0xbb at 0x70..0x80 and 16 bytes of
// 0xcc at 0x80..0x90 (overwriting the lower half of the sentinel).
mstore(0x70, 0xbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbcccccccccccccccccccccccccccccccc)
// mem_opt forwards the stale sentinel, missing the partial overwrite.
r := mload(0x80)
}
}
}