We're currently always recomposing and representing the whole window when using the D3D11 compositor. This is wasteful. Let's not. This essentially is the same behavior as the Basic compositor.

Comment on attachment 8531752 [details] [diff] [review] Only recomposite and present the damaged area of the compositor Review of attachment 8531752 [details] [diff] [review]: ----------------------------------------------------------------- ::: gfx/layers/d3d11/CompositorD3D11.cpp @@ +889,5 @@ > + RefPtr<IDXGISwapChain1> chain; > + HRESULT hr = mSwapChain->QueryInterface((IDXGISwapChain1**)byRef(chain)); > + if (SUCCEEDED(hr) && chain) { > + DXGI_PRESENT_PARAMETERS params; > + memset(&params, 0, sizeof(DXGI_PRESENT_PARAMETERS)); Use mfbt/PodZero @@ +891,5 @@ > + if (SUCCEEDED(hr) && chain) { > + DXGI_PRESENT_PARAMETERS params; > + memset(&params, 0, sizeof(DXGI_PRESENT_PARAMETERS)); > + params.DirtyRectsCount = mInvalidRegion.GetNumRects(); > + std::vector<RECT> rects; rects.reserve(params.DirtyRectsCount)

Comment on attachment 8531752 [details] [diff] [review] Only recomposite and present the damaged area of the compositor Review of attachment 8531752 [details] [diff] [review]: ----------------------------------------------------------------- ::: gfx/layers/d3d11/CompositorD3D11.cpp @@ +621,5 @@ > + if (mCurrentRT == mDefaultRT) { > + scissor.left = std::max(int32_t(aClipRect.x), mCurrentClip.x); > + scissor.right = std::min(int32_t(aClipRect.XMost()), mCurrentClip.XMost()); > + scissor.top = std::max(int32_t(aClipRect.y), mCurrentClip.y); > + scissor.bottom = std::min(int32_t(aClipRect.YMost()), mCurrentClip.YMost()); I'd like this better if you intersected the rects and set from that. i.e. clipRect = aClipRect; if (mCurrentRT == mDefaultRT) { clipRect = intersect(aClipRect, mCurrentClip); } scissor.left = clipRect.x .... @@ +862,5 @@ > if (aRenderBoundsOut) { > *aRenderBoundsOut = Rect(0, 0, mSize.width, mSize.height); > } > > D3D11_RECT scissor; scissor is now unused?

I landed a trivial crosscompilation fixup: https://hg.mozilla.org/integration/mozilla-inbound/rev/16ea12664917

This should be backed out.

@:bas seems busy. @:jrmuizel], This should be backed out.

Yes, this should get backed out. I think we should also probably back out bug 1102499 as it depends on this for performance.

We set a Friday deadline for fixing the regressions, otherwise we'll back this out on aurora.