longcat and rotating-rust display all sorts of random graphics artifacts during animation

[larsbergstrom]

I suspect that we are reusing graphics buffers in a very interesting way:

To repro, just ./servo ../src/test/html/summit-two.html
note that longcat has similar issues: ./servo ../src/test/html/summit-one.html

@mrobinson @zwarich ideas?

[mrobinson]

This looks a lot like the issue I'm bisecting. Hoping to have something concrete today.

[mrobinson]

Pretty sure this originates with the rust upgrade. @larsbergstrom, do you also see this failure with servo -c?

[larsbergstrom]

@mrobinson No, this bug only occurs with GPU rendering.

[mrobinson]

This does seem like a Rust bug of some kind. I'm seeing Rust structures corrupted as they pass the FFI barrier. For instance, this patch to rust-azure fixes the corruption for me:

diff --git a/azure_hl.rs b/azure_hl.rs
index 69c0670..f48f379 100644
--- a/azure_hl.rs
+++ b/azure_hl.rs
@@ -488,6 +488,7 @@ impl DrawTarget {
             let draw_options = match draw_options {
                 None => ptr::mut_null(),
                 Some(mut draw_options) => {
+                    println!("This shouldn't fix anything, but it does...");
                     let draw_options: *mut AzDrawOptions = &mut draw_options;
                     draw_options
                 }

The issue seems to be that the alpha value of AzDrawOptions is set to 0 and the compositor operator is not passed properly either.

[zwarich]

@mrobinson Can you easily detect this on the Azure C side? If so then we should be able to make a reduced test case by calling a C function with the same signature.

[mrobinson]

Configuring with --disable-optimize also seems to fix this issue.

[metajack]

This sounds exactly like rust-lang/rust#16366, but the workaround is even more bizarre. I wonder if a println! would fix the example in the Rust issue.