[MaxFreedomPollard]

Slicing the time axis of a buffered TimeFunction with a negative bound selects the wrong time slots, both when reading and when writing. With time_order=2, the usual setup for a second order wave equation, on main:

import numpy as np
from devito import Grid, TimeFunction

grid = Grid(shape=(4,))
u = TimeFunction(name='u', grid=grid, time_order=2)
u.data[:] = np.arange(3)[:, None]

u.data[-1:, 0]  # [1., 2.], NumPy gives [2.]
u.data[-2:, 0]  # [2.], NumPy gives [1., 2.]
u.data[:-1, 0]  # [], NumPy gives [0., 1.]

u.data[:-1] = 7                 # writes nothing
u.data[-1:] = 9                 # also overwrites slot 1
u.data[-2:] = np.zeros((2, 4))  # ValueError: could not broadcast input array from shape (2,4) into shape (1,4)

The default time_order=1 is affected too: u.data[:-1] comes back empty there as well.

The time Dimension of a buffered TimeFunction is modulo-indexed, so Data._index_glb_to_loc maps each index through index_apply_modulo in devito/data/utils.py. For a negative slice bound that function returns -(idx.start % modulo), or -(idx.stop % (modulo + 1)) for the stop. Python's % already gives a nonnegative remainder for a negative operand, so with three slots a start of -1 becomes -2, a start of -2 becomes -1, and a stop of -1 becomes -3.

The fix wraps the magnitude of a negative bound and keeps its sign, i.e. -(-idx.start % modulo) and -(-idx.stop % (modulo + 1)). A negative bound within the buffer now selects the same entries as in NumPy. I checked every slice with a start in [-n, n) or None, a stop in [-n, n] or None, and a step of None, ±1 or ±2, for buffers of 1 to 7 slots. All of them now match NumPy. Before the fix, 84 of the 280 cases for three slots did not. A negative bound beyond the buffer wraps the same way an integer index does, so u.data[-4:] now starts at the slot u.data[-4] refers to. Nonnegative bounds are unchanged. So is any Data without modulo indexing, which covers every Function, saved TimeFunction and SparseTimeFunction.

Verified on macOS (Apple silicon, Python 3.12, DEVITO_ARCH=clang, DEVITO_LANGUAGE=C):

• The new tests/test_data.py::TestDataBasic::test_negative_slicing_modulo reads and writes through eight slices with negative bounds and compares the results against NumPy. Six of the eight fail on main and all pass with the fix.
• pytest -m "not parallel" tests/test_data.py: 97 passed, 2 skipped.
• tests/test_builtins.py, test_buffering.py, test_checkpointing.py, test_dimension.py, test_save.py, test_timestepping.py, test_operator.py and test_docstrings.py: 484 passed. The 11 failures are environmental and identical on main: 8 cannot download scipy's ascent image, and 3 need OpenMP, which Apple clang lacks.
• tests/test_sparse.py, test_interpolation.py and test_tensors.py: 268 passed.
• isort, ruff check --preview, flake8 and typos are clean.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79.78%. Comparing base (a9e3f31) to head (cc46aa1).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3086      +/-   ##
==========================================
- Coverage   84.27%   79.78%   -4.50%     
==========================================
  Files         258      258              
  Lines       57421    57432      +11     
  Branches     4911     4911              
==========================================
- Hits        48392    45822    -2570     
- Misses       8205    10729    +2524     
- Partials      824      881      +57     

Flag	Coverage Δ
pytest-gpu-aomp-amdgpuX	?
pytest-gpu-gcc-	?
pytest-gpu-icx-	?
pytest-gpu-nvc-nvidiaX	?

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.