On 20/02/23 14:56 +0100, Landry Breuil wrote: > Le Tue, Feb 21, 2023 at 12:00:06AM +1100, Jonathan Gray a écrit : > > On Mon, Feb 20, 2023 at 11:31:23AM +0100, Antoine Jacoutot wrote: > > > On Mon, Feb 20, 2023 at 02:06:34PM +1100, Jonathan Gray wrote: > > > > On Sun, Feb 19, 2023 at 01:34:41PM +0100, Antoine Jacoutot wrote: > > > > > Hi. > > > > > > > > > > There seems to be a regression with mesa that makes gtk+4 application > > > > > very slow > > > > > to start. > > > > > By default the GSK renderer uses OpenGL. > > > > > As a workaround, you can temporarily use this to go back to the cairo > > > > > renderer > > > > > which makes gtk+4 applications fast again: > > > > > > > > > > export GSK_RENDERER=cairo > > > > > > > > What hardware is this on? Is there a Mesa or gtk bug for it? > > > > > > See dmesg below. > > > > > > > When did this behaviour start? Before the gtk update that recently > > > > went in? Does it occur with GSK_RENDERER=vulkan? > > > > GSK_RENDERER described in > > > > https://docs.gtk.org/gtk4/running.html#gsk_renderer > > > > > > It did not happen on my previous amd ryzen. > > > As soon as I switched to the new intel laptop, I was that bug (exact same > > > installation, I rsync'd everything). > > > > > > It does *not* appear with GSK_RENDERER=vulkan but that renderer is buggy > > > (not > > > built by default) and segfaults on a regular basis. > > > > > > > There is > > > > https://gitlab.freedesktop.org/mesa/mesa/-/issues/5113 > > > > which briefly touches on shader cache. We disable the shader > > > > cache to be able to uses pledge(2). > > > > > > Yes, that is the bug I was looking into. > > > > Here is a xenocara diff to enable the shader cache. > > It is created in ~/.cache/mesa_shader_cache/ > > > > On an Intel system (x250 with Broadwell) launching gtk4-demo results in > > a 1.6M cache. The time to a window appearing is noticeably shorter > > running it again after that. > > > > The various firefox and chromium ports will need to change > > unveil/pledge policies to use it. Chromium at least still runs but > > there are multiple warnings that directories can't be created.
This just needs an unveil of that directory in the gpu process and an flock pledge for chromium to use it. Seems okay to me. I think this should go in.
