BlendMode::Src overwrites unrelated pixels in affected rows

[Willow Chargin]

I have a program that uses raqote to render an image in "chunks" on
different threads, and then composite them all into a final image. By
"chunks" I mean disjoint rectangular regions that cover the canvas:
e.g., a 2×2 grid of quadrants.

This works fine with the default SrcOver blend mode, but I figured
that it would be faster to use Src since it's is a no-op
instead of some bit operations that compile down to 15 or so
instructions and probably prevent autovectorizing of the
copy loop.

I thought that blending with Src would produce the same result because
no two pixels in a chunk overlap. But in fact it seems that when drawing
an image into a DrawTarget with Src, all other pixels in the row
are cleared, set to transparency. Here is an example:

use raqote::{BlendMode, DrawOptions, DrawTarget, Image, PathBuilder, SolidSource, Source};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let rgb_source = |r, g, b| Source::Solid(SolidSource { r, g, b, a: 255 });
    let blue = rgb_source(34, 127, 190);
    let white = rgb_source(255, 255, 255);

    // Create a white circle on a blue field.
    let mut circle = DrawTarget::new(50, 50);
    circle.fill_rect(0.0, 0.0, 50.0, 50.0, &blue, &DrawOptions::new());
    let mut pb = PathBuilder::new();
    pb.arc(25.0, 25.0, 20.0, 0.0, std::f32::consts::TAU);
    circle.fill(&pb.finish(), &white, &DrawOptions::new());
    let circle_image = Image {
        width: circle.width(),
        height: circle.height(),
        data: circle.get_data(),
    };

    // Now, try to draw a 2x2 grid of those.
    let mut canvas = DrawTarget::new(100, 100);

    // Top row: two blue circles next to each other, default (SrcOver) compositing.
    canvas.draw_image_at(0.0, 0.0, &circle_image, &DrawOptions::new());
    canvas.draw_image_at(50.0, 0.0, &circle_image, &DrawOptions::new());

    // Bottom row: using Src compositing, since the pixels are disjoint and should cover the row
    // exactly. Should be the same, right?
    let mut opts_src = DrawOptions::new();
    opts_src.blend_mode = BlendMode::Src;
    canvas.draw_image_at(0.0, 50.0, &circle_image, &opts_src);
    canvas.draw_image_at(50.0, 50.0, &circle_image, &opts_src);

    canvas.write_png("canvas.png")?;
    Ok(())
}

At raqote v0.8.2, this produces the following image:

As you can see, drawing the bottom-right quadrant has caused the
bottom-left quadrant to be clobbered with transparency, even though it
shouldn't have been affected at all because it is not within the bounds
of the image that I asked to draw.

In addition to being a correctness issue, this suggests that raqote is
doing a lot more work compositing than it needs to! e.g., we would
expect drawing an M×N grid of fixed-size images onto a canvas to take
O(M·N) time, but it seems like it would actually take O(M·N²) time,
because each image that's drawn needs to rewrite the whole row.

I tried to track down what's happening, and this is my understanding:

• The data gets clobbered in ShaderBlendBlitter::blit_span, where
  blend_fn points to blend_row::<Src>;
• blend_row::<Src>(src, dst) effectively picks how much to blend
  based on the length of src, since dst is set to a whole suffix
  of the image data buffer;
• the instantiated value of src is &self.tmp[..], which is created
  in DrawTarget::choose_blitter as a vector with length width;
• and choose_blitter's unique call site, in DrawTarget::composite,
  unconditionally sets width to self.width, which does not at all
  incorporate the rect to which we're drawing.

Correspondingly, this patch seems to fix the problem for me:

diff --git a/src/draw_target.rs b/src/draw_target.rs
index 8206a5d..cc65771 100644
--- a/src/draw_target.rs
+++ b/src/draw_target.rs
@@ -979,7 +979,7 @@ impl<Backing : AsRef<[u32]> + AsMut<[u32]>> DrawTarget<Backing> {
         let shader = choose_shader(&ti, src, alpha, &mut shader_storage);
 
         let mut blitter_storage = ShaderBlitterStorage::None;
-        let blitter = DrawTarget::choose_blitter(mask, &self.clip_stack, &mut blitter_storage, shader, blend, dest, dest_bounds, self.width);
+        let blitter = DrawTarget::choose_blitter(mask, &self.clip_stack, &mut blitter_storage, shader, blend, dest, dest_bounds, self.width.min(rect.width()));
 
         match mask {
             Some(mask) => {

The raqote test suite continues to pass with this patch. I'm not
familiar enough to understand whether this is the right patch and good
in all cases, but I did try it out on the sweep-gradient example and
on images like this that my program generates; each one was
bitwise-identical before and after this patch.

What do you think—is this patch appropriate and sufficient, or is
something trickier needed?