M2.x: one intermediate soft mask serves every drawn size (hayro-interpret patch 77)

The raster cache key now includes the drawn size, so a soft mask drawn at
several sizes was decoded in full once per size: issue16263's rotated and
scaled copies hit four sizes and took about 10 s. Sizes are rounded up to a
power of two, and every averaged-down size is made from one cached
intermediate mask at most 4096 on a side, so the full mask is decoded once.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Pq7HyDBXygTf7jcqdsjcfn

crates/render/src/raster/tests.rs

@@ -8200,7 +8200,8 @@ fn decoded_images_serve_every_size_they_are_drawn_at() {
 /// there (vendored hayro-interpret patch 77): a 2 × 2 image whose mask is 34,862 × 4,332, drawn 40
 /// times on one slide (pdf.js `issue16263.pdf`), was decoded and drawn at full size each time and
 /// the page took minutes. Drawn ten times at one size, the mask is now decoded once and handed
-/// over at that size; a mask no larger than twice its drawn size keeps its own.
+/// over at that size rounded up to a power of two; a mask no larger than twice its drawn size
+/// keeps its own.
 #[test]
 fn a_soft_mask_far_larger_than_where_it_is_drawn_is_shrunk_once() {
     use hayro::hayro_interpret::font::Glyph;
@@ -8262,9 +8263,12 @@ fn a_soft_mask_far_larger_than_where_it_is_drawn_is_shrunk_once() {
         (device.0, cache.decoded_image_decodes())
     };
     let (alphas, decodes) = draws(1000);
-    assert_eq!(alphas, vec![(20, 20); 10], "averaged down to the size it is drawn at");
-    // One decode for the image and one for its averaged-down mask, shared by all ten draws.
-    assert_eq!(decodes, 2, "decoded once");
+    // Drawn 20 pixels square, averaged down to that rounded up to a power of two.
+    assert_eq!(alphas, vec![(32, 32); 10], "averaged down to the size it is drawn at");
+    // One decode for the image and one for its mask (here no larger than the intermediate the
+    // averaged-down sizes are made from, so that is the mask itself), plus the averaged-down mask,
+    // all shared by the ten draws.
+    assert_eq!(decodes, 3, "decoded once");
     let (alphas, _) = draws(40);
     assert_eq!(alphas, vec![(40, 40); 10], "no more than twice the drawn size: kept as it is");
 }
@@ -8273,7 +8277,7 @@ fn a_soft_mask_far_larger_than_where_it_is_drawn_is_shrunk_once() {
 /// dpi): its `mask` (`width`, `height`, 8-bit samples) is larger than the 2 × 1 image, so it may
 /// be averaged down (vendored hayro-interpret patch 77). Each mask is given as its size and
 /// samples.
-fn soft_masks_handed(width: u32, height: u32, mask: &[u8], sizes: &[(u32, u32)]) -> Vec<(u32, u32, Vec<u8>)> {
+fn soft_masks_handed(width: u32, height: u32, mask: &[u8], sizes: &[(u32, u32)]) -> (Vec<(u32, u32, Vec<u8>)>, usize) {
     use hayro::hayro_interpret::font::Glyph;
     use hayro::hayro_interpret::hayro_syntax::Pdf;
     use hayro::hayro_interpret::{
@@ -8325,24 +8329,42 @@ fn soft_masks_handed(width: u32, height: u32, mask: &[u8], sizes: &[(u32, u32)])
     let mut ctx = Context::new(initial, Rect::new(0.0, 0.0, 400.0, 400.0), &cache, page.xref(), InterpreterSettings::default());
     let mut device = Draws::default();
     interpret_page(page, &mut ctx, &mut device);
-    device.0
+    (device.0, cache.decoded_image_decodes())
 }
 
 /// An averaged-down soft mask (vendored hayro-interpret patch 77):
 /// - takes each source sample into the target sample its centre falls in: seven columns
-///   `0 0 255 255 255 255 255` drawn three wide give `0 255 255` (by left edges, `85 255 255`);
+///   `0 0 0 255 0 0 0`, drawn three wide and so averaged to four, give `0 0 127 0` (by left edges,
+///   `0 127 0 0`);
 /// - is never sized past the samples a truncated stream decodes: a mask that says it is 100 rows
 ///   but holds 10, drawn 20 high, is handed no rows that no sample covers;
-/// - is cached per drawn size: the same image drawn 20 and then 200 points wide is handed its
-///   mask at 20 and then at its own 100, not the first draw's 20 again.
+/// - is cached per drawn size, rounded up to a power of two: the same image drawn 20 and then 200
+///   points wide is handed its mask at 32 and then at its own 100, not the first draw's 32 again,
+///   and drawn 20 and 21 wide it is handed one mask of 32 both times;
+/// - is made, at every size, from one intermediate mask at most 4096 on a side, so a mask wider
+///   than that, drawn at four sizes, is decoded in full once and not four times.
 #[test]
 fn an_averaged_down_soft_mask_follows_its_samples_and_its_drawn_size() {
-    let handed = soft_masks_handed(7, 1, &[0, 0, 255, 255, 255, 255, 255], &[(3, 1)]);
-    assert_eq!(handed, vec![(3, 1, vec![0, 255, 255])], "centres, not left edges");
-    let handed = soft_masks_handed(100, 100, &[255; 100 * 10], &[(20, 20)]);
+    let (handed, _) = soft_masks_handed(7, 1, &[0, 0, 0, 255, 0, 0, 0], &[(3, 1)]);
+    assert_eq!(handed, vec![(4, 1, vec![0, 0, 127, 0])], "centres, not left edges");
+    let (handed, _) = soft_masks_handed(100, 100, &[255; 100 * 10], &[(20, 20)]);
     assert!(!handed.is_empty() && handed.iter().all(|(_, _, data)| data.iter().all(|&a| a == 255)), "no uncovered rows: {handed:?}");
-    let sizes: Vec<(u32, u32)> = soft_masks_handed(100, 100, &[255; 100 * 100], &[(20, 20), (200, 200)]).iter().map(|(w, h, _)| (*w, *h)).collect();
-    assert_eq!(sizes, vec![(20, 20), (100, 100)], "each draw at its own size");
+    let sizes = |draws: &[(u32, u32)]| -> Vec<(u32, u32)> {
+        soft_masks_handed(100, 100, &[255; 100 * 100], draws).0.iter().map(|(w, h, _)| (*w, *h)).collect()
+    };
+    assert_eq!(sizes(&[(20, 20), (200, 200)]), vec![(32, 32), (100, 100)], "each draw at its own size");
+    assert_eq!(sizes(&[(20, 20), (21, 21)]), vec![(32, 32), (32, 32)], "draws a pixel apart share one");
+    // Columns alternating 0 and 255, so every averaged-down sample is about half: within 8, as a
+    // sample covering 19 or 20 columns, say, holds 9 or 10 opaque ones.
+    let wide: Vec<u8> = (0..5000 * 2).map(|i| if i % 2 == 0 { 0 } else { 255 }).collect();
+    let (handed, decodes) = soft_masks_handed(5000, 2, &wide, &[(20, 1), (40, 1), (80, 1), (160, 1)]);
+    let sizes: Vec<(u32, u32)> = handed.iter().map(|(w, h, _)| (*w, *h)).collect();
+    assert_eq!(sizes, vec![(32, 1), (64, 1), (128, 1), (256, 1)]);
+    assert!(handed.iter().all(|(_, _, data)| data.iter().all(|&a| a.abs_diff(127) <= 8)), "{handed:?}");
+    // The image at each of the four sizes, the 4096-wide intermediate once, and the mask at each
+    // of the four sizes made from it. Without the intermediate there are 8, each mask size a full
+    // decode.
+    assert_eq!(decodes, 9, "the full mask decoded once");
 }
 
 // ICC profiles whose stream dictionaries are the same (vendored hayro-interpret patch 48).

vendor/README.md

@@ -39,7 +39,7 @@ Other resource categories retain their existing parent-chain lookup.
 
 | Crate | Version | Licence | Patches | Test |
 |---|---|---|---|---|
-| hayro-interpret | 0.7.0 | Apache-2.0 OR MIT | (1) select `/AP /N` appearance states via `/AS` (checkboxes, radios, stateful icons) and honour the NoView flag; (2) Type3 Unicode fallback to `/Encoding` glyph names, then printable ASCII; (3) `Type3Glyph::advance_width`; (4) `InterpreterSettings::ocg_overrides` for viewer layer toggles; (5) soft masks whose group lacks `/CS` were dropped (Chrome gradient text); (6) paint nesting cap `MAX_PAINT_NESTING` for tiling patterns and Type 3 glyphs (self-referencing content overflowed the stack); (7) CID width ranges in `/W` and `/W2` clamped to `MAX_CID` (a `0 4294967295 w` range hung); (8) `InterpreterSettings::hide_comments` skips markup annotations but keeps widgets and links (View ▸ Hide all comments); (9) `image_size_ok`: images and masks over 2^28 pixels are not decoded (a fuzzed `/W 4294967295` stencil mask allocated 4 GiB); (10) `MAX_NESTED_PAINTS`: nested paints (forms, Type 3 glyphs and tiling patterns inside one another) share a per-page budget of 50,000 (a glyph showing eight glyphs of its own font, sixteen deep, hung); (11) depends on skrifa 0.47 instead of 0.42 (upstream read-fonts fixed a Type 1 tokenizer that recursed through runs of numbers: a fuzzed font hung or overflowed the stack), with the one API change (`Charset::glyph_id` returns `Option`); (12) annotation appearances whose `/BBox` or `/Matrix` collapses to a line or a point are skipped (mapping them onto `/Rect` divided by zero and handed the device infinite or NaN matrices); (13) a Highlight annotation without a usable normal appearance is drawn from its `/QuadPoints`, `/C` and `/CA` with Multiply blending, as pdfcraft-annot draws highlights (it was not drawn at all); (14) simple-font `/Widths` (`read_widths`): entries only for codes 0–255 (one entry per code below `/FirstChar` was allocated: `4294967295` asked for 34 GB), an entry that isn't a number leaves its code missing instead of ending the array, and a `/LastChar` below `/FirstChar` gives no widths instead of discarding the font for a standard one; (15) the appearance box is mapped onto `/Rect` corner to corner: the translation now accounts for the scale (an appearance whose box didn't start at the origin and differed in size from `/Rect` was drawn shifted); (16) a clip set by `W`/`W*` takes effect after whichever path-painting operator ends the path (`S`, `f`, `B`, … as well as `n`; ISO 32000-2 §8.5.4): it was applied only on `n`, so `re W* S` clipped nothing and the clip stayed pending for a later, unrelated `n`; (17) `InterpreterSettings::cancelled`: once set, interpretation stops before the next content operator (a render pool sets it for every worker when dropped, after an edit or on closing, and for one worker when the watchdog gives up on its render, so renders whose answers nobody can receive stop instead of running to the end); (18) `MAX_PAGE_CONTENT`: the content a page holds at once (its contents, and the forms, tiling patterns and Type 3 glyph procedures being painted) stays within 512 MiB (`decode_content`; once it runs out, `InterpreterWarning::ContentTruncated` is reported (once per page by pdfcraft-render) so a truncated page is not a clean render; each level gives its bytes back when it finishes): forms nest fifty deep and each level kept its content while the next ran, so a form painting itself around a stream inflating to 256 MiB would have held over 12 GB; and Type 0 sampled functions take at most `MAX_SAMPLE_ENTRIES` (2^21) table points and `MAX_SAMPLE_VALUES` (2^23) samples, read no further than `/Size` asks (256 MiB of 1-bit samples asked for over 8 GB); (19) mesh shadings (types 4–7) are sampled only inside the area being drawn (`ShadingPattern::encode_within`, which the vendored hayro renderer calls with that area plus a pixel), within 8 visits per pixel of it plus 4 per triangle (`TriangleSampler`), and patches are triangulated one at a time: points far off the page were sampled over their whole bounding box, up to 65536 × 65536 entries (a fuzzed 147-byte patch stream; a synthetic page passed 6 GB in six seconds), ten thousand overlapping triangles ran for minutes, and every patch became 722 triangles before any limit; meshes hold at most `MAX_MESH_PATCHES` (2^16) patches and `MAX_MESH_TRIANGLES` (2^20) triangles, a lattice needs two to 2^19 vertices per row (`/VerticesPerRow 0` pushed empty rows for ever; one huge row read the whole stream), and a page samples at most `MAX_PAGE_MESH_TRIANGLES` (2^25) mesh triangles in all (a mesh painted a thousand times was triangulated a thousand times); each cut logs a warning; (20) PDF function construction is limited to 64 levels and 10,000 total nodes with cycle detection; calculator parsing and evaluation share procedure depth/work budgets, integer division/bitshift/rotation/index are panic-free, and fixed-size stack errors propagate; (21) a Type 0 font drawn with a substitute that isn't CID-keyed (it isn't embedded, or its program doesn't load) draws `.nLine truncated
+| hayro-interpret | 0.7.0 | Apache-2.0 OR MIT | (1) select `/AP /N` appearance states via `/AS` (checkboxes, radios, stateful icons) and honour the NoView flag; (2) Type3 Unicode fallback to `/Encoding` glyph names, then printable ASCII; (3) `Type3Glyph::advance_width`; (4) `InterpreterSettings::ocg_overrides` for viewer layer toggles; (5) soft masks whose group lacks `/CS` were dropped (Chrome gradient text); (6) paint nesting cap `MAX_PAINT_NESTING` for tiling patterns and Type 3 glyphs (self-referencing content overflowed the stack); (7) CID width ranges in `/W` and `/W2` clamped to `MAX_CID` (a `0 4294967295 w` range hung); (8) `InterpreterSettings::hide_comments` skips markup annotations but keeps widgets and links (View ▸ Hide all comments); (9) `image_size_ok`: images and masks over 2^28 pixels are not decoded (a fuzzed `/W 4294967295` stencil mask allocated 4 GiB); (10) `MAX_NESTED_PAINTS`: nested paints (forms, Type 3 glyphs and tiling patterns inside one another) share a per-page budget of 50,000 (a glyph showing eight glyphs of its own font, sixteen deep, hung); (11) depends on skrifa 0.47 instead of 0.42 (upstream read-fonts fixed a Type 1 tokenizer that recursed through runs of numbers: a fuzzed font hung or overflowed the stack), with the one API change (`Charset::glyph_id` returns `Option`); (12) annotation appearances whose `/BBox` or `/Matrix` collapses to a line or a point are skipped (mapping them onto `/Rect` divided by zero and handed the device infinite or NaN matrices); (13) a Highlight annotation without a usable normal appearance is drawn from its `/QuadPoints`, `/C` and `/CA` with Multiply blending, as pdfcraft-annot draws highlights (it was not drawn at all); (14) simple-font `/Widths` (`read_widths`): entries only for codes 0–255 (one entry per code below `/FirstChar` was allocated: `4294967295` asked for 34 GB), an entry that isn't a number leaves its code missing instead of ending the array, and a `/LastChar` below `/FirstChar` gives no widths instead of discarding the font for a standard one; (15) the appearance box is mapped onto `/Rect` corner to corner: the translation now accounts for the scale (an appearance whose box didn't start at the origin and differed in size from `/Rect` was drawn shifted); (16) a clip set by `W`/`W*` takes effect after whichever path-painting operator ends the path (`S`, `f`, `B`, … as well as `n`; ISO 32000-2 §8.5.4): it was applied only on `n`, so `re W* S` clipped nothing and the clip stayed pending for a later, unrelated `n`; (17) `InterpreterSettings::cancelled`: once set, interpretation stops before the next content operator (a render pool sets it for every worker when dropped, after an edit or on closing, and for one worker when the watchdog gives up on its render, so renders whose answers nobody can receive stop instead of running to the end); (18) `MAX_PAGE_CONTENT`: the content a page holds at once (its contents, and the forms, tiling patterns and Type 3 glyph procedures being painted) stays within 512 MiB (`decode_content`; once it runs out, `InterpreterWarning::ContentTruncated` is reported (once per page by pdfcraft-render) so a truncated page is not a clean render; each level gives its bytes back when it finishes): forms nest fifty deep and each level kept its content while the next ran, so a form painting itself around a stream inflating to 256 MiB would have held over 12 GB; and Type 0 sampled functions take at most `MAX_SAMPLE_ENTRIES` (2^21) table points and `MAX_SAMPLE_VALUES` (2^23) samples, read no further than `/Size` asks (256 MiB of 1-bit samples asked for over 8 GB); (19) mesh shadings (types 4–7) are sampled only inside the area being drawn (`ShadingPattern::encode_within`, which the vendored hayro renderer calls with that area plus a pixel), within 8 visits per pixel of it plus 4 per triangle (`TriangleSampler`), and patches are triangulated one at a time: points far off the page were sampled over their whole bounding box, up to 65536 × 65536 entries (a fuzzed 147-byte patch stream; a synthetic page passed 6 GB in six seconds), ten thousand overlapping triangles ran for minutes, and every patch became 722 triangles before any limit; meshes hold at most `MAX_MESH_PATCHES` (2^16) patches and `MAX_MESH_TRIANGLES` (2^20) triangles, a lattice needs two to 2^19 vertices per row (`/VerticesPerRow 0` pushed empty rows for ever; one huge row read the whole stream), and a page samples at most `MAX_PAGE_MESH_TRIANGLES` (2^25) mesh triangles in all (a mesh painted a thousand times was triangulated a thousand times); each cut logs a warning; (20) PDF function construction is limited to 64 levels and 10,000 total nodes with cycle detection; calculator parsing and evaluation share procedure depth/work budgets, integer division/bitshift/rotation/index are panic-free, and fixed-size stack errors propagate; (21) a Type 0 font drawn with a substitute that isn't CID-keyed (it isn't embedded, or its program doesn't load) draws `.nLine truncated
 | hayro | 0.7.1 | Apache-2.0 OR MIT | (1) `RenderSettings::{x_offset, y_offset}` to render one tile of a large page; (2) images over `MAX_IMAGE_PIXELS` (2^28) are skipped (a fuzzed `/W 4294967295` hung in resampling); (3) `tiling_cell_scale` keeps a tiling pattern's cell pixmap within 3000 px a side whatever its `/XStep`/`/YStep` (a fuzzed step far beyond `/BBox` allocated 2 GB); (4) stroke widths clamped to 4 × (canvas width + height) device pixels (a fuzzed `/LW 9223372036854775807` allocated 10 GB in stroke expansion); (5) `usable_dash_pattern`: a stroke whose dash pattern could cut it into more than `MAX_DASHES_PER_STROKE` (250,000) pieces, counting kurbo's restart of the pattern at every subpath and `ClosePath`, is drawn solid (a fuzzed `/D [[11] 0]` on a line from x = 92234775807 allocated over 5 GB, `[0.000001] 0 d` on a 20 pt line over 10 GB); (6) a dash pattern with a negative or non-finite entry, or a period that isn't positive, is drawn solid (`[-1 -1] 0 d` looped for ever in kurbo); (7) `render_size` and `render_into`: render into a buffer the caller allocates (PdfCraft renders into pixel-aligned memory that the GUI takes over as texture data); `render` is `render_into` on a new pixmap; (8) checked image source/target/RGBA, alpha-premultiplication scratch and intermediate sizes before resampling, finite transforms and exact source/mask buffers; sources wider than u16 may shrink before backend limits, with a 2^20 source-axis cap and conservative 128 MiB filter-planning estimate; an over-limit or failed resize keeps the original image resolution and geometry when the source fits the backend, otherwise skips it, with a logged warning; staging, resampling output/scratch, RGBA conversion and padding buffers use fallible reserves; (9) `may_paint_canvas`: a fill or stroke whose control-point box, grown by how far its paint reaches past the path (half the stroke width times the larger of the miter limit and √2) plus 0.5 user units and 2 pixels, misses the canvas is not drawn, so a tile no longer strokes and fills every path of its page (item 5 of #623); `set_skip_offscreen_paths` turns this off for the regression test; (10) `/Interpolate` governs the pre-resize too (#624): a minified image that does not request interpolation is pre-resized with nearest-neighbour (since (21), only inside Type 3 glyphs), planned without the Catmull-Rom filter limits (`image_resampling_size` takes the flag); a Type 3 glyph image is sampled bicubically only when interpolated; a minified image with a matching soft mask is nearest-sampled straight into premultiplied RGBA; (11) `may_paint_rect`: an image whose stream-size box, mapped to the device, misses the canvas is skipped before decoding (decoded data can only shrink against that box), so a tile of an image-heavy page no longer decodes, resamples and rasterizes every image on it; (12) preview renders (`InterpreterSettings::preview`: backdrops, thumbnails, grid cells, signature-drag backgrounds) skip images under `PREVIEW_MIN_IMAGE_AREA` device pixels, while final renders always draw them; (13) reading colours (#897): `RenderSettings::color_replacement` takes a `ColorReplacement` (text colour, background, `only_black`, `line_art`) that remaps solid text and line-art paint while rendering, glyph outlines counting as text and stencil-mask ink as line art; images, image alpha masks, soft masks and pattern cells keep their colours; `render_pdf` gains a `color_replacement: Option<ColorReplacement>` argument (pass `None` for true colours); the upstream-only `[dev-dependencies.log]` entry is removed from `Cargo.toml` (log is a normal dependency since the logging patch), so `cargo test -p hayro` runs this crate's unit tests; (14) a shrunk image that is interpolated (or, since (21), area-averaged) and whose axes lie along the device's (upright, mirrored or a quarter turn, as on a page under `/Rotate`) is snapped to whole device pixels: shrunk once to its rounded device rectangle (halves round up, so a tile snaps an edge where the whole page does) and drawn 1:1 from the rounded corner, instead of a ceil-sized raster resampled again at a fractional scale and offset (`snap_minified_image`; scans looked soft next to Acrobat, and until #5530 a turned one still did, up to 200 grey levels off the same page drawn upright). An image counts as axis-aligned when the terms the snap drops move its far edges by at most a hundredth of a device pixel (`SNAP_SKEW`), not by a tolerance on the coefficients, which let a slight skew of a very wide image through; and the snapped 1:1 transform is used only when the shrink produced the snapped size, so a failed resize keeps the full transform instead of drawing the source raster at its source size; (15) a soft mask of another size than its image is filtered when shrunk, whatever its `/Interpolate` (#1649): such a mask carries the shape on its own (Word saves text as a tiny black image with a glyph-bitmap mask), and nearest-neighbour dropped every stroLine truncated
 | hayro-jbig2 | 0.3.0 | Apache-2.0 OR MIT | (1) at most 2^28 pixels per bitmap and page bitmap, and 2^25 cells per halftone grid and grayscale image (each side could be 65535: 4.3 billion pixels decoded one by one, and a 4-byte value per cell); (2) at most 2^26 symbol pixels per symbol dictionary (a fuzzed dictionary decoded for minutes); (3) a symbol dictionary's height classes are capped at one per new symbol and its export runs at `2 × symbols + 1`, each plus `EXTRA_EMPTY_STEPS` (1024): an empty height class or a zero-length export run doesn't advance, and past the end of its data the arithmetic decoder can decode them for ever (one changed byte hung the renderer); (4) a segment header's long-form referred-to segment count is capped at `u16::MAX` and a larger count is refused before its retention or referred-to bytes are read: the 29-bit count went straight to `Vec::with_capacity`, a reservation of about 2 GiB whose failure aborts the process (from storytold/craft-libs#176); (5) what a page's segments hold for later ones (intermediate regions, pattern and symbol dictionaries with their headers and retained contexts, and Huffman tables) is at most 64 MiB, two square bitmaps of 16,384 a side (`MAX_RETAINED_WORDS`), and a dictionary's new and exported symbols are counted against it as they are made: each was bounded on its own, but a few dozen bytes an intermediate region held 33 MB each, and 8 MiB symbols exported again and again killed the process, so a few kilobytes could exhaust memory; (6) a segment may refer to at most 2^20 symbols, checked before any are collected: references may repeat, and 65,535 of them to one dictionary, in a 78 KB stream, asked for 17 GB; (7) a pattern dictionary keeps at most 16 MiB of shifted copies, which only speed up drawing, and none that wouldn't fit what the page has left (`MAX_SHIFTED_WORDS`); (8) a custom Huffman table has at most 2^16 lines, and a stream with its globals at most 2^20 segments and 2^22 references, checked as they are read: both grew before anything counted them, and PDFs of a kilobyte or so reached 5 GB and more behind nested Flate; (9) a page past the bounds of (1) is refused when the stream is opened: a caller sizes its output from the header before the page bitmap is made, and an Indexed image in a 743-byte PDF asked for 137 GB, which aborts; (10) a stream may ask for at most 2^30 units of decoding work, four of the largest pages, about 6 s to 17 s (`MAX_WORK`): decoding a bitmap costs a pixel each and 4 a row, copying one a word each and 1 a row, and each integer an item reads costs 38, what one takes at worst, so one with no pixels pays too, before it is decoded: each bitmap was bounded, but nothing bounded how many there were, and a 795-byte PDF took 265 s; (11) `Image::decode_within` spends from what a run of images may still do, starting with its page's pixels, and `charge_reading` charges reading a stream and its globals before anything is parsed, so an operation over a document can bound what `MAX_WORK` bounds for one image, and images that fail or decode nothing add up too: a file of thousands of images made to reach that limit took as many times as long; (12) adding to a location, a reference offset, a symbol's corner or a Huffman range line near the end of an `i32`, and negating a decoded integer of 2^31, overflowed: a panic wherever overflow is checked. The refinement context, a text region's corners, the integer decoder and the Huffman table wrap, as a release build did, and `Bitmap::combine` saturates, which comes to the same, as a location that overflows is past the end of any page. What is drawn doesn't change; (13) a bitmap filled with a default pixel of 1 (a page, or an intermediate text or halftone region) had 1s past its width in the last word of each row, and a refinement, which reads its reference a word at a time, read them as pixels where the standard has 0 (6.3.5.2): refining such a reference not a multiple of 32 wide decoded to other pixels than were coded, release builds too. `Bitmap::clear_padding` clears those bits when a bitmap is filled with 1s, as cutting symbols and patterns from a collective bitmap already did. Ghostscript reads the test's page and text streams as they decode now; (14) a region's location is unsigned (5.4.4), so one past `i32::MAX` is off the page, where the decoder took it for a signed one: a region there was drawn back onto the page from the left or the top (at 0xFFFFFFFC, four pixels left of it), and a refinement there, of the page or of a region at the same place (7.4.7.5), was refused as an overflow, and the page with it. `Bitmap::combine_region` keeps such a location past the page. A refinement of another region refines it where it is, as Table 35 has it (GRREFERENCEDX and GRREFERENCEDY 0), whatever the two locations say; a refinement of the page, see (16); (15) a refinement whose reference is offset right of it (GRREFERENCEDX above 0, which a refined text instance has from a wider bitmap, TaLine truncated
 | lopdf | 0.45.0 | MIT | (1) PNG predictor rows longer than the data are refused before their buffers are allocated (a fuzzed `/Columns 4294967295` allocated two 4 GiB rows), and empty frames return without row buffers; (2) predictor row widths use saturating arithmetic or reject overflow; (3) sub-byte TIFF Predictor 2 scratch is bounded by the samples present in the input instead of the declared `/Colors`, aligned with [lopdf#589](https://github.com/J-F-Liu/lopdf/pull/589); (4) the object lookup keeps its cycle history through compressed objects, requires an ordinary in-file container, and caps the chain of indirect lookups at 16 (a chain of new objects, each a `/Length` compressed in a container whose own `/Length` names the next, or a stream whose `/Length` names the next stream, overflowed the stack: a dev build's 2 MiB thread within 800 hops); (5) object-stream members are parsed within their own extent (up to the next distinct offset), an entry whose offset repeats an earlier entry's is skipped (so a repeated object number can resolve to its earlier entry), and offsets are processed in index order; (6) `find_latest_trailer`: a rebuild takes a `/Type /XRef` stream dictionary as the trailer when the file has no `trailer` keyword (the latest in the file wins, as a later `trailer` would), so a concatenated or truncated file keeps its /Info and /ID instead of failing to load; a candidate whose /Root is the catalog beats one whose /Root is another dictionary, and an in-file /Root must match its generation (#2365) | (1) `inspect::tests::png_predictor_rows_longer_than_the_data_are_refused_before_allocating`, `inspect::tests::empty_png_predictor_frames_do_not_allocate_rows`; (2) `inspect::tests::tiff_subbyte_predictor_row_width_overflow_is_refused`; (3) `inspect::tests::tiff_subbyte_predictor_scratch_is_bounded_by_available_samples`, `inspect::tests::tiff_subbyte_predictors_keep_components_rows_and_padding`; (4) `inspect::tests::compressed_object_self_container_does_not_recurse`, `inspect::tests::compressed_object_mutual_containers_do_not_recurse`, `inspect::tests::object_stream_length_cycle_does_not_recurse`, `inspect::tests::compressed_object_container_chain_does_not_recurse`, `inspect::tests::compressed_length_chain_does_not_recurse`, `inspect::tests::plain_stream_length_chain_does_not_recurse`, `inspect::tests::compressed_objects_and_indirect_stream_length_still_load`, `inspect::tests::compressed_length_in_a_container_with_an_indirect_length_still_loads`; (5) `inspect::tests::object_stream_member_extents_bound_owned_strings`, `inspect::tests::object_stream_member_extents_preserve_index_order_and_whitespace`; (6) `inspect::tests::rebuilt_file_keeps_the_info_of_its_cross_reference_stream`, `inspect::tests::rebuilt_file_prefers_a_trailer_whose_root_is_the_catalog` |

vendor/hayro-interpret/src/x_object.rs

@@ -627,9 +627,11 @@ impl<'a> ImageXObject<'a> {
             interpolate: self.interpolate,
             // PdfCraft patch (38): the same image drawn in a default colour space or without one.
             default_space: self.default_space,
-            // PdfCraft patch (32): see `decode_reads_target`.
+            // PdfCraft patch (32): see `decode_reads_target`; patch (77): see `has_larger_mask`.
             target: if self.decode_reads_target() {
                 target_dimension
+            } else if self.has_larger_mask() {
+                target_dimension.map(size_bucket)
             } else {
                 None
             },
@@ -659,20 +661,33 @@ impl<'a> ImageXObject<'a> {
         let Some((tw, th)) = shrink_to else {
             return self.decoded_mask(target_dimension);
         };
-        // Averaged down only on the sides that are larger than where they are drawn.
-        let to = (tw.min(self.width).max(1), th.min(self.height).max(1));
-        let shrink = || {
+        // Averaged down to the drawn size rounded up to a power of two (`size_bucket`), so draws a
+        // pixel apart share one mask and the renderer filters the rest, and only on the sides that
+        // are larger than that.
+        let (bw, bh) = size_bucket((tw, th));
+        let to = (bw.min(self.width).max(1), bh.min(self.height).max(1));
+        // Decoded in full and averaged down to `to`. A truncated stream decodes fewer rows than
+        // the dictionary says: never more target samples than decoded ones on a side.
+        let from_full = |to: (u32, u32)| {
             let full = decode_mask(self, target_dimension)?;
-            // A truncated stream decodes fewer rows than the dictionary says: never more target
-            // samples than decoded ones on a side.
-            let to = (to.0.min(full.luma.width).max(1), to.1.min(full.luma.height).max(1));
-            Some(DecodedMask { luma: shrink_luma(full.luma, to)? })
+            Some(DecodedMask { luma: shrink_luma_within(full.luma, to)? })
         };
         let Some(cache) = &self.decoded_cache else {
-            return shrink();
+            return from_full(to);
         };
-        let key = self.decode_key(cache, true, target_dimension).map(|key| DecodedImageKey { shrunk_to: Some(to), ..key });
-        cache.get_or_insert_with(key, MAX_DECODED_IMAGE_BYTES, decoded_mask_size, shrink)
+        let key = |to| self.decode_key(cache, true, target_dimension).map(|key| DecodedImageKey { shrunk_to: Some(to), ..key });
+        // One image drawn at several sizes (rotated and scaled copies on one slide) needs the
+        // mask at several sizes. Each comes from one intermediate averaged-down mask, cached at
+        // most `SOFT_MASK_INTERMEDIATE` on a side, so the full mask is decoded once, not once per
+        // size; a size larger than the intermediate comes from the full mask.
+        let between = (self.width.min(SOFT_MASK_INTERMEDIATE), self.height.min(SOFT_MASK_INTERMEDIATE));
+        if to == between || to.0 > between.0 || to.1 > between.1 {
+            return cache.get_or_insert_with(key(to), MAX_DECODED_IMAGE_BYTES, decoded_mask_size, || from_full(to));
+        }
+        cache.get_or_insert_with(key(to), MAX_DECODED_IMAGE_BYTES, decoded_mask_size, || {
+            let mid = cache.get_or_insert_with(key(between), MAX_DECODED_IMAGE_BYTES, decoded_mask_size, || from_full(between))?;
+            Some(DecodedMask { luma: shrink_luma_within(mid.luma, to)? })
+        })
     }
 
     /// PdfCraft patch (32): whether decoding this image depends on the size it is drawn at.
@@ -682,16 +697,22 @@ impl<'a> ImageXObject<'a> {
     /// step and kept once per zoom level; without it, one decode serves them all.
     fn decode_reads_target(&self) -> bool {
         let dict = self.stream.dict();
-        // PdfCraft patch (77): a soft mask (or `/Mask` image) larger than the image may be
-        // averaged down to the size the image is drawn at (`decoded_soft_mask`), so the decoded
-        // raster, which holds it as its alpha, depends on that size too.
+        uses_jpx(&self.stream)
+            || dict.get::<Stream<'_>>(SMASK).is_some_and(|s| uses_jpx(&s))
+            || dict.get::<Stream<'_>>(MASK).is_some_and(|s| uses_jpx(&s))
+    }
+
+    /// PdfCraft patch (77): whether this image has a soft mask (or `/Mask` image) larger than
+    /// itself, which may be averaged down to the size the image is drawn at
+    /// (`decoded_soft_mask`): the decoded raster holds it as its alpha, so it depends on that size,
+    /// to the [`size_bucket`] it is averaged down to.
+    fn has_larger_mask(&self) -> bool {
+        let dict = self.stream.dict();
         let larger = |s: &Stream<'_>| {
             let d = s.dict();
             d.get::<u32>(WIDTH).is_some_and(|w| w > self.width) || d.get::<u32>(HEIGHT).is_some_and(|h| h > self.height)
         };
-        uses_jpx(&self.stream)
-            || dict.get::<Stream<'_>>(SMASK).is_some_and(|s| uses_jpx(&s) || larger(&s))
-            || dict.get::<Stream<'_>>(MASK).is_some_and(|s| uses_jpx(&s) || larger(&s))
+        dict.get::<Stream<'_>>(SMASK).is_some_and(|s| larger(&s)) || dict.get::<Stream<'_>>(MASK).is_some_and(|s| larger(&s))
     }
 
     pub(crate) fn width(&self) -> u32 {
@@ -722,6 +743,19 @@ const MAX_DECODED_IMAGE_BYTES: usize = 64 << 20;
 /// keeps its own resolution, which the renderer filters as it draws.
 const SOFT_MASK_SHRINK_FACTOR: u64 = 2;
 
+/// PdfCraft patch (77): the largest side of the intermediate mask that every smaller averaged-down
+/// size of one soft mask is made from (`ImageXObject::decoded_soft_mask`): 4096 × 4096 one-byte
+/// samples are 16 MiB, a quarter of the decoded-image budget, so it stays cached between draws.
+const SOFT_MASK_INTERMEDIATE: u32 = 4096;
+
+/// PdfCraft patch (77): the size a large soft mask is averaged down to for an image drawn at
+/// `(w, h)` device pixels: each side rounded up to a power of two. Draws of one image a pixel or so
+/// apart (scaled by 0.36 and 0.36167, say) share one averaged mask and one cache entry, and a mask
+/// is never averaged below where it is drawn, so the renderer shrinks the rest at most twofold.
+fn size_bucket((w, h): (u32, u32)) -> (u32, u32) {
+    (w.checked_next_power_of_two().unwrap_or(u32::MAX), h.checked_next_power_of_two().unwrap_or(u32::MAX))
+}
+
 /// PdfCraft patch (77): `luma` averaged down to `to` (no larger than it on either side): each
 /// target sample is the mean of the source samples whose centres fall in it. Source sample `x`
 /// of `sw`, centred at `x + ½`, falls in target sample `⌊(2x + 1)·tw / 2sw⌋`. The only buffers are
@@ -758,6 +792,17 @@ fn shrink_luma(luma: LumaData, to: (u32, u32)) -> Option<LumaData> {
     Some(LumaData { data, width: to.0, height: to.1, ..luma })
 }
 
+/// PdfCraft patch (77): `luma` averaged down to `to`, but never to more samples than it has on a
+/// side (a truncated stream decodes fewer rows than its dictionary says), and returned as it is
+/// when that is its own size.
+fn shrink_luma_within(luma: LumaData, to: (u32, u32)) -> Option<LumaData> {
+    let to = (to.0.min(luma.width).max(1), to.1.min(luma.height).max(1));
+    if to == (luma.width, luma.height) {
+        return Some(luma);
+    }
+    shrink_luma(luma, to)
+}
+
 /// PdfCraft patch: pixel bytes held by a decoded mask.
 fn decoded_mask_size(mask: &DecodedMask) -> usize {
     mask.luma.width as usize * mask.luma.height as usize