Fix BMP save of a P image with an empty palette

[gaoflow]

Saving a P-mode image whose palette is empty produces a BMP that Pillow cannot reopen:

import io
from PIL import Image

im = Image.new("P", (8, 8))          # empty palette, no putpalette()
im.frombytes(bytes(range(64)))
buf = io.BytesIO()
im.save(buf, "BMP")
buf.seek(0)
Image.open(buf).load()               # OSError: image file is truncated (0 bytes not processed)

PNG, GIF and TIFF all round-trip the same image; only BMP fails, and it fails to read back its own output.

Cause

In BmpImagePlugin._save, the P branch computes colors = len(palette) // 4, which is 0 for an empty palette. The writer then sets biClrUsed = 0 and writes no color table, with the pixel-data offset at 14 + 40 = 54. On read, biClrUsed == 0 at 8 bpp is interpreted as 1 << 8 = 256 entries, so the reader advances the pixel offset by 256 * 4 bytes past a table that was never written — it reads past EOF and reports the file as truncated.

Fix

When the palette is empty, write a full 256-entry table (mirroring the existing L mode branch) so the header is self-consistent and the file reads back. An all-zero table is used rather than a grayscale ramp so the image reopens as P (a ramp would be detected as grayscale and downgraded to L). A populated palette is unaffected.

Tests

test_save_empty_palette parametrizes (1, 1), (7, 5), (8, 8), (16, 16), asserting the written header has a non-zero biClrUsed (pinning the root cause, not just the symptom) and that the image reopens as P with its pixel indices intact. test_save_empty_palette_round_trips_like_other_formats pins the cross-format invariant that BMP reopens the image like PNG/GIF/TIFF do. Both fail on main (OSError / biClrUsed == 0) and pass with the fix; the full Tests/test_file_bmp.py suite stays green (37 passed); ruff format, ruff check and mypy are clean.

[codspeed-hq]

Merging this PR will not alter performance

✅ 345 untouched benchmarks
⏩ 2 skipped benchmarks¹

---

_{Comparing gaoflow:fix-bmp-empty-palette (10b57cf) with main (abb8ba0)}

Footnotes

1. 2 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[radarhere]

You have assert that if you save a P mode image as a BMP, it should be re-opened as a P mode image.

BmpImagePlugin is actively trying to move to the simpler mode though.

Pillow/src/PIL/BmpImagePlugin.py

Lines 275 to 283 in fc7cec1

	# ----------------- Check if grayscale and ignore palette if so
	for ind, val in enumerate(indices):
	rgb = palette[ind * padding : ind * padding + 3]
	if rgb != o8(val) * 3:
	grayscale = False
	# ------- If all colors are gray, white or black, ditch palette
	if grayscale:
	self._mode = "1" if file_info["colors"] == 2 else "L"

I don't think that all image formats need to behave exactly the same. The PNG format actually has a 'Color type' that distinguishes whether the data is grayscale or palette. The BMP format doesn't have that information.

So for a P mode image with an empty palette, I would be inclined to save only 1 palette entry, so that the final file is as small as possible.

[gaoflow]

Fair point on both. Switched to a single palette entry — 122 bytes now instead of 1142 — and dropped the mode assertion and the cross-format test, since BMP has no way to record the grayscale/palette distinction and the reader is right to simplify. It reopens as L with the pixel indices intact.

The test now just pins biClrUsed == 1 and the round-trip.