|
src/mesa/glapi. Basically, the scripts that did simple things (like
gl_offsets.py) were simple, and the scripts that did more complicated things
(like glX_proto_send.py) were getting progressively more and more out of
control. So, I re-write the foundation classes on which everything is based.
One problem with the existing code is that the division between the GL API
database representation and the way the output code is generated was either
blury or nonexistant. The new code somewhat follows the
Model-View-Controller pattern, minus the Controller. There is a distinct
set of classes that model the API data, and there is a distinct set of
classes that generate code from that data.
One big change is in the class that represents GL functions (was glFunction,
is now gl_function). There used to be an instance of this calls for each
function and for each alias to that function. For example, there was an
instance for PointParameterivSGIS, PointParameterivEXT, PointParameterivARB,
and PointParameteriv. In the new code, there is one instance. Each
instance has a list of entrypoint names for the function. In the next
revision, this will allow a couple useful things. The script will be able
to verify that the parameters, return type, and GLX protocol for a function
and all it's aliases match.
It will also allow aliases to be represented in the XML more compactly.
Instead of repeating all the information, an alias can be listed as:
<function name="PointParameterivARB" alias="PointParameterivEXT"/>
Because the data representation was changed, the order that the alias
functions are processed by the scripts also changed. This accounts for at
least 2,700 of the ~3,600 lines of diffs in the generated code.
Most of the remaining ~900 lines of diffs are the result of bugs *fixed* by
the new scripts. The old scripts also generated code with some bugs in it.
These bugs were discovered while the new code was being written.
These changes were discussed on the mesa3d-dev mailing list back at the end
of May:
http://marc.theaimsgroup.com/?t=111714569000004&r=1&w=2
Xorg bug: 3197, 3208
|
|
now contains 3 static tables. The first table is a single, large string of
all the enum names. The second table is an array, sorted by enum name, of
indexes to the string table and the matching enum value. The extra string
table is used to eliminate relocs (and save space) in the compiled file.
The third table is an array, sorted by enum value, of indexes into the
second table.
The [name, enum] table contains all of the enums, but the table sorted by
enum-value does not. This table contains one entry per enum value. For
enum values that have multiple names (e.g., 0x84C0 has GL_TEXTURE0_ARB and
GL_TEXTURE0), only an index to the "best" name will appear in the table.
gl_enums.py gives precedence to "core" GL versions of names, followed by ARB
versions, followed by EXT versions, followed, finally, by vendor versions
(i.e., anything that doesn't fall into one of the previous categories). By
filtering the unneeded elements from this table, not only can we guarantee
determinism in the generated tables, but we save 364 elements in the table.
The optimizations outlined above reduced the size of the stripped enums.o
(on x86) from ~80KB to ~53KB.
The internal organization of gl_enums.py was also heavily modified.
Previously enums were stored in an unsorted list as [value, name] tuples
(basically). This list was then sorted, using a user-specified compare
function (i.e., VERY slow in most Python implementations) to generate a
table sorted by enum value. It was then sorted again, using another
user-specified compare function, to generate a table sorted by name.
Enums are now stored in a dictionary, called enum_table, with the enum value
as the key. Each dictionary element is a list of [name, priority] pairs.
The priority is determined as described above. The table sorted by enum
value is generated by sorting the keys of enum_table (i.e., very fast). The
tables sorted by name are generated by creating a list, called name_table,
of [name, enum value] pairs. This table can then be sorted by doing
name_table.sort() (i.e., very fast).
The result is a fair amount more Python code, but execution time was reduced
from ~14 seconds to ~2 seconds.
|
