Hi, this is a re-send of two patches that didn't get anybody's attention, when I sent them out last week, plus some additional fixes for rarer instances of the same problem that I've encountered since then.
The problem that these patches fix is simple: the exec_list iterations often cast sentinel nodes that are mere exec_nodes to ir_* types. This leads to crashes when running ubsan, because ubsan attempts to verify the type of the object. Since the target type has a virtual method table, it does the verification by calling into the dynamic_cast implementation, under the assumption that this virtual method table is there. However, since the object is really only an exec_node, which _doesn't_ have a virtual method table, ubsan crashes. So I've converted loops into forms that only do the cast once we are assured that we are not looking at a sentinel node. This looks a bit funny in the macros, because an additional outer for-loop is used to define a variable of a different type, but the compiler is able to remove that outer loop and the __flag variable entirely. Please review! Thanks, Nicolai -- src/compiler/glsl/ast_function.cpp | 4 +- .../glsl/link_uniform_initializers.cpp | 8 +- src/compiler/glsl/list.h | 147 ++++++++++++----- src/compiler/glsl/lower_jumps.cpp | 7 +- src/compiler/glsl/opt_dead_code_local.cpp | 7 +- src/compiler/glsl/opt_tree_grafting.cpp | 10 +- 6 files changed, 111 insertions(+), 72 deletions(-) _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
