LLVMMemoryCopyFillLowering: in-bounds `memory.fill`/`memory.copy` trap on a 4 GiB memory

Open
#9,257 2 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
3/5
Estimated time
1-2 days
Newbie friendliness
62/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
cpp, wasm
Domain
compilers

Research direction

The bug is in src/passes/LLVMMemoryCopyFillLowering.cpp: the generated __memory_fill bounds check (around lines 213-225) and __memory_copy (around lines 123-130) compute the memory size in bytes with i32 arithmetic, which wraps to 0 at 65536 pages. Reproduce with the WAT in the issue using wasm-opt --llvm-memory-copy-fill-lowering, then run the output to confirm the unreachable trap. Done when the bounds check is done in i64 and the 65536-page module no longer traps, with the existing bulk-memory lowering tests still passing.

Written by the indexing model from the issue text.

Description

Repro

(module
 (memory 65536)
 (func (export "f")
  (memory.fill (i32.const 0) (i32.const 0) (i32.const 1))))
$ wasm-opt in.wat --enable-bulk-memory --llvm-memory-copy-fill-lowering -o out.wasm
$ wasmtime run --invoke f in.wat
$ wasmtime run --invoke f out.wasm
Error: failed to run main module `out.wasm`

Caused by:
    0: failed to invoke `f`
    1: error while executing at wasm backtrace:
    0:     0x44 - <unknown>!<wasm function 1>
    1:     0x32 - <unknown>!<wasm function 0>
    2: wasm trap: wasm `unreachable` instruction executed

f returns normally. After --llvm-memory-copy-fill-lowering, f traps (unreachable).

In pseudocode, the generated __memory_fill checks:

if (dst + size > memory.size << 16) trap    // i32: 65536 << 16 = 0

With 65535 pages it does not trap.

Cause

The memory size in bytes is computed in i32, which wraps to 0 for 65536 pages:

https://github.com/WebAssembly/binaryen/blob/1d7748aa4ef48ac93f90712c2437ef88c6441389/src/passes/LLVMMemoryCopyFillLowering.cpp#L213-L225

__memory_copy has the same computation:

https://github.com/WebAssembly/binaryen/blob/1d7748aa4ef48ac93f90712c2437ef88c6441389/src/passes/LLVMMemoryCopyFillLowering.cpp#L123-L130

The pass header excludes pointer overflow, but this access does not overflow; only the computed memory size does.

Possible fix

Do the bounds check in i64.

Question

I'm not familiar with how this pass is used in practice. It seems to work as a polyfill for targets without bulk memory, so I'm not sure whether a module with a 65536-page memory actually reaches it.

Dominant language
WebAssembly
Stars
8.7k
Forks
913
Avg merge
2d 14h
Merged PRs (30d)
92

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from WebAssembly/binaryen

All issues in WebAssembly/binaryen

Similar issues

More Compilers issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.