GLXscreenConfigs is badly named and a dumping ground for a lot of stuff.
This patch creates private screen structs for the dri drivers and moves
some of their fields over there.
There was confusion on both the size of message we can send, and on
what the URB destination offset means.
The remaining problems appear to be due to spilling of regs in the
fragment shader being broken.
Instead of using ir_call::callee, search for the signature in the
linked shader. This will allow resolving calls from functions
imported from other shaders. The ir_call::callee pointer in the
imported function will still reference a signature in the original shader.
The list of shaders to search needs to be provided as an explicit
parameter to support coming changes. At that point there is no reason
for it to be in the class. Also, fix some of the 'const' decorators.
Give ir_function::parameter_lists_match_exist similar treatment. Make
the parameters const, and propogate the constness as far as it will
trivially go.
It now works correctly when nodes are removed, as it was originally
intended to do; it no longer processes nodes added to the list before
the current node, nor those added immediately after the current node.
This matches the behavior of Linux's list_for_each_safe.
The node being processed may be removed from the list and put in a
different list. Not using the safe version caused list processing to
change streams after moving a node.
For the shader containing 'main', use the linked shader (i.e., the
clone of the original shader that contained main) as the source for
global instructions to move into main.
When faced with a constructor like 'ivec4(0, 2, 0, 0)', we would
manage to get a value of 2 instead of 0 for the first "0". Usually 2
characters past "0" would point at some junk and lex as 0 anyway.
Fixes glsl-octal and glsl-unused-varyings.
The Mac OS X SCons build failed on 32-bit CPUs starting with commit
2f6d47a7c8 during linking of graw-null.
The build succeeds though on a 64-bit CPU. See FDO bug 29117.
This was the compiler error.
scons: building associated VariantDir targets: build/darwin-x86-debug
Linking build/darwin-x86-debug/gallium/targets/graw-null/libgraw.dylib ...
Undefined symbols:
"_lp_swizzled_cbuf", referenced from:
_lp_swizzled_cbuf$non_lazy_ptr in libllvmpipe.a(lp_rast.os)
_lp_swizzled_cbuf$non_lazy_ptr in libllvmpipe.a(lp_rast_tri.os)
(maybe you meant: _lp_swizzled_cbuf$non_lazy_ptr)
"_lp_dummy_tile", referenced from:
_lp_dummy_tile$non_lazy_ptr in libllvmpipe.a(lp_rast.os)
_lp_dummy_tile$non_lazy_ptr in libllvmpipe.a(lp_rast_tri.os)
_lp_dummy_tile$non_lazy_ptr in libllvmpipe.a(lp_setup.os)
(maybe you meant: _lp_dummy_tile$non_lazy_ptr)
The patch adds -fno-common to all Mac OS X builds to work around this issue.
This is a big deal for debugging if nothing else ("what class is this
ir_instruction, really?"), but is also nice for avoiding building a
whole visitor or an if (node->as_whatever() || node->as_other_thing())
chain.
DRI2 events are sent to the X drawable ID used to create the DRI2 drawable,
not the GLX drawable ID. So when an event comes in, we need to look up
the __GLXDRIdrawable by its X drawable ID, which needs a new hash table.
