[Bug 106533] r300_dri.so SIGSEGV in llvm_pipeline_generic under Cinnamon

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=106533

Roland Scheidegger <sroland at vmware.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |maraeo at gmail.com

--- Comment #4 from Roland Scheidegger <sroland at vmware.com> ---
You have the same tcl-less chipset?
So the assembly shows the segfault isn't due to alignment, but simply because
it's a null pointer.
This is also mesa 18.0? I'm just asking because according to the line numbers
(r300_render_stencilref.c:113) it's doing two-sided stencil emulation, which I
can't see how it could happen with the emulated clear (which definitely doesn't
enable two-sided stencil), so that looks a little bit suspect. Although maybe
the line numbers aren't quite accurate due to optimization...

You could try setting (with a debug build) DRAW_USE_LLVM="0" and see if this
fixes the crash - and if it still crashes it should be easier to figure out
what pointer is zero.

You could print out the shader (with a debug build) with
GALLIVM_DEBUG=tgsi,ir,asm to see what the assembly really might do. I think
though this is trying to read a vertex buffer (for position probably) which
just isn't there.

Or you could try setting (with a debug build) DRAW_USE_LLVM="0" and see if this
fixes the crash - and if it still crashes it should be easier to figure out
what pointer is zero.

I think I have an idea why it might fail, but someone more familiar with the
u_upload stuff and r300 would have to look at it:
r300_set_vertex_buffers_swtcl() would provide the vertex buffers to draw, and I
suspect it's always setting NULL buffer, because I think it's going to be not a
user buffer, but there won't be a malloced_buffer neither - the u_upload code
will call r300_buffer_create() but it won't alloc the malloced_buffer because
the PIPE_BIND_CUSTOM bit will always be set (because r300->stream_uploader is
the same as r300->uploader which will set that by default).

I'm not quite sure actually how it's supposed to work, maybe the logic in
r300_buffer_create should always use the malloced_buffer path when !has_tcl?

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/dri-devel/attachments/20180604/ac1412ed/attachment.html>