This is a follow-up to the email thread discussing [PATCH 1/4] Revert "glsl: Skip processing the first function's body in do_dead_functions().".
As mentioned in that thread, since the linker emits functions at the head of the final linked program, they come before any global variable declarations, thus those functions can't access any global variables without violating IR rules. This bug was typically papered over by the fact that (a) this scenario only arises when linking several shaders together, which doesn't happen very often, and (b) linking is followed by inlining, which usually (but not always) is able to merge everything into main() in one pass, yielding valid IR again. Although I wasn't able to find a set of shaders that crashes Mesa or produces incorrect pixels because of this bug, the situation seemed fragile enough to be worth fixing. So patch 1 in the series adds a validation step, to make sure that the IR output by the linker is valid, and patch 2 fixes the bug by changing the linker to emit functions at the tail of the final linked program rather than the head. These patches need to be applied after [PATCH 1/4] Revert "glsl: Skip processing the first function's body in do_dead_functions().", since reverting that patch makes do_dead_functions() independent of the ordering of functions in the IR. [PATCH 1/2] glsl: validate IR after linking (debug builds only) [PATCH 2/2] glsl: When linking, emit functions at the tail of the final linked program. _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
