Venom: an unrelated bounded-array function makes a four-byte runtime-sized call cost 34M gas

[AlbertoCentonze]

Version information

• Vyper: 0.5.0b2+commit.5764199a, master at 5764199a9ab64e13bc64d8f41371378db4fa6ab0 (checked September 5, 2026).
• Also observed at the port's pinned compiler, 8af5e83c10af4f065cc660fe20293dac11fdff83. The intervening commit changes only .github/workflows/codeql.yml.
• Python 3.13.14; macOS 26.6.2, arm64.
• Venom (--experimental-codegen), Cancun; bytecode metadata disabled.

Observed behavior

Changing a bound in an external function that is never called changes the cost of calling a separate length(Bytes[INF]) function with four bytes from roughly 2,100 gas to 34 million gas.

The complete Vyper contract is:

# pragma version 0.5.0b2
BOUND: constant(uint256) = 32
@external
@pure
def echo(values: DynArray[Bytes[BOUND], 64]) -> DynArray[Bytes[BOUND], 64]:
    return values

@external
@pure
def length(data: Bytes[INF]) -> uint256:
    return len(data)

Compile once with BOUND = 32 and once with BOUND = 32768. In both cases, deploy a fresh contract and call only length(hex"01020304"); do not call echo.

Fresh measurements against the master commit above:

Bound in unused echo function	O2 call gas	O3 call gas
32	2,126	2,132
32,768	34,050,123	34,050,129

All four calls return 4. These numbers are the gasleft() difference around the call, including the identical Solidity caller overhead; deployment is outside the measurement. Foundry's whole-test gas number is different and should not be substituted for the logged call gas.

Reproduction

git clone https://github.com/aave/aave-v4.git aave-v4-vyper-feedback
cd aave-v4-vyper-feedback
git checkout 1f7129042fe57e24bfc4c934be5088514cb93503
python3.13 -m venv .venv
.venv/bin/python -m pip install 'vyper @ git+https://github.com/vyperlang/vyper.git@5764199a9ab64e13bc64d8f41371378db4fa6ab0'

With Foundry available (measured using Forge 1.8.1 and Solidity 0.8.28):

.venv/bin/python scripts/check_vyper_memory_allocation.py

The self-contained runner compiles both variants with O2/O3, generates a temporary Solidity measurement harness, and executes four local tests. It needs no RPC, deployed protocol, or Aave test fixtures. Small source, large source, and retained pinned-compiler results. The current-master rerun produced exactly the same call-gas values.

Expected behavior / investigation direction

Executing the small function should not incur memory-expansion cost proportional to bounds in an unrelated, unexecuted external function. Investigate whether runtime allocation starts above a contract-wide static-memory maximum rather than the memory needed by the selected entry point. This is a hypothesis about the mechanism; the cross-function cost coupling is directly measured.

This looks related to the allocation design discussed in #4426 and #4856. The report isolates cross-function coupling, rather than the small extra buffer allocations described in #4103 and #4117.