Summary:
This patch ports xptcodegen.py over to the new perfecthash.py system, removing
some special-case code generators, and taking advantage of the easier-to-use
interface.
In addition, the code was changed to take advantage of the endianness
information from Part 2, allowing us to avoid having to perform endianness swaps
at runtime when hashing nsIDs.
Depends On D2616
Reviewers: froydnj!
Tags: #secure-revision
Bug #: 1479484
Differential Revision: https://phabricator.services.mozilla.com/D2618
The build system knows at build-backend time where to find each IDL
file; making xpidl-process.py rediscover this by requiring
xpidl-process.py to search through directories to find input IDL files
is silly. To rememdy this, we're going to modify things so full paths
are passed into the script. Those paths can then be used directly, with
no searching.
The tail end of the xpidl Makefile.in contains a line, generated for
every xpt file:
$(1): $(addsuffix .idl,$(addprefix $(dist_idl_dir)/,$($(basename $(notdir $(1)))_deps)))
This line, in context, is saying that the xpt file depends on all of its
input IDL files. But xpidl-process.py already generates this
information when we pass it --depsdir, which we do. So this code is
redundant with what we already generate, and it can be removed.
The previous patch required us to pass a single -I argument pointing at
$(DIST)/idl so IDL include statements would work correctly. This patch
lifts that limitation and explicitly points xpidl-process.py at the
locations of all the IDL source directories to search for included IDL
files. Invocations of xpidl-process.py no longer depend on IDL files
being copied to the objdir.
Building on the last patch, we can change the build process to pass in
the directories where the input IDL files can be found. It is
convenient to pass in just the relative source directory paths, to
encourage people to not look in the object directory and to make the
command lines slightly shorter.
xpidl-process.py still assumes that included IDL files can be found by
looking in a single directory. We add a single -I argument to the
invocation of xpidl-process.py to accommodate this short-sightedness.
The current IDL build setup assumes that all IDL files can be found in a
single directory. This setup requires that all IDL files be copied to a
single directory, which is suboptimal in terms of disk I/O and also
complicates things like generating IDL files at build time.
As a first step in moving away from this state of affairs,
xpidl-process.py needs to be taught that the input IDL files could
potentially be found in multiple directories. The current setup can
just specify $(DIST)/idl as the lone directory to examine. Future
patches will change this to examine multiple directories.
This patch contains the meat of the changes here. The following summarize the changes:
1. xptinfo.h is rewritten to expose the new interface for reading the XPT data,
The nsXPTInterfaceInfo object exposes methods with the same signatures as
the methods on nsIInterfaceInfo, to make converting code which used
nsIInterfaceInfo as easy as possible, even when those methods don't have
signatures which make a ton of sense anymore. There are also a few methods
which are unnecessary (they return `true` or similar), which should be
removed over time.
Members of the data structures are made private in order to prevent reading
them directly. Code should instead call the getter methods. This should make
it easier to change their memory representation in the future. Constructing
these structs is made possible by making the structs `friend class` with the
XPTConstruct class, which is implemented by the code generator, and is able
to access the private fields.
In addition, rather than using integers with flag constants, I opted for
using C++ bitfields to store individual flags, as I found it made it easier
to both write the code generator, and reason about the layouts of the types.
I was able to shave a byte off of each nsXPTParamInfo (4 bytes -> 3 bytes)
by shoving the flags into spare bits in the nsXPTType. Unfortunately there
was not enough room for the retval flag. Fortunately, we already depend in
our code on the retval parameter being the last parameter, so I worked
around this by removing the retval flag and instead having a `hasretval`
flag on the method itself.
2. An xptinfo.cpp file is added for out-of-line definitions of more complex
methods, and the internal implementation details of the perfect hash.
Notable is the handling of xptshim interfaces. As the type is uniform, a
flag is checked when trying to read constant information, and a different
table with pointers into webidl data structures is checked when the type is
determined to be a shim.
Ideally we could remove this once we remove the remaining consumers of the
existing shim interfaces.
3. A python code generator which takes in the json XPT files generated in the
previous part, and emits a xptdata.cpp file with the data structures. I did
my best to heavily comment the code.
This code uses the friend class trick to construct the private fields of the
structs, and avoid a dependency on the ordering of fields in xptinfo.h.
The sInterfaces array's order is determined by a generated perfect hash
which is also written into the binary. This should allow for fast lookups by
IID or name of interfaces in memory. The hash function used for the perfect
hash is a simple FNV hash, as they're pretty fast.
For perfect hashing of names, another table is created which contains
indexes into the sInterfaces table. Lookup by name is less common, and this
form of lookup should still be very fast.
4. The necessary Makefiles are updated to use the new code generator, and
generate the file correctly.
This patch handles the actual generation of the static data structures
used to represent XPT information. XPT files are generated in the same
way as they are now, but they are used only as an intermediate
representation to speed up incremental compilation rather than
something used by Firefox itself. Instead of linking XPTs into a
single big XPT file at packaging time, they are linked into a single
big C++ file at build time, that defines the various static consts in
XPTHeader.
In xpt.py, every data structure that can get written to disk gets an
additional code_gen() method that returns a representation of that
data structure as C++ source code. CodeGenData aggregates this
information together, handling deduplication and the final source code
generation.
The ctors are needed for XPTConstValue to statically initialize the
different union cases without resorting to designated initializers,
which are part of C99, not C++. Designated initializers appear to be
supported in C++ code by Clang and GCC, but not MSVC. The ctors must
be constexpr to ensure they are actually statically initialized so
they can be shared between Firefox processes.
I also removed an unnecessary "union" in XPTConstDescriptor.
Together, these patches reduce the amount of memory reported by
xpti-working-set from about 860,000 bytes to about 200,000 bytes. The
remaining memory is used for xptiInterface and xptiTypelibGuts (which
are thin wrappers around the XPT interfaces and header) and hash
tables to speed up looking up interfaces by name or IID. That could
potentially be eliminated from dynamic allocations in follow up
work. These patches did not affect memory reporting because XPT arenas
are still used by the remaining XPTI data structures.
MozReview-Commit-ID: Jvi9ByCPa6H
--HG--
extra : rebase_source : a9e48e7026aab4ad1b7f97e50424adf4e3f4142f
Now that XPT files are not loaded from files at runtime, code for
packaging XPT files can be removed.
This means that a couple of test XPIDL interfaces will get shipped in
builds to users that weren't before, but I don't think that matters
much.
This also puts XPT files into the local objdir for the XPIDL makefile,
instead of dist/bin, because they are no longer part of the
distribution.
MozReview-Commit-ID: 7gWj8KWUun3
--HG--
extra : rebase_source : 65bac47c2cd1a20b3c675a01b44a25a1d2d3ab7a
This patch handles the actual generation of the static data structures
used to represent XPT information. XPT files are generated in the same
way as they are now, but they are used only as an intermediate
representation to speed up incremental compilation rather than
something used by Firefox itself. Instead of linking XPTs into a
single big XPT file at packaging time, they are linked into a single
big C++ file at build time, that defines the various static consts in
XPTHeader.
In xpt.py, every data structure that can get written to disk gets an
additional code_gen() method that returns a representation of that
data structure as C++ source code. CodeGenData aggregates this
information together, handling deduplication and the final source code
generation.
The ctors are needed for XPTConstValue to statically initialize the
different union cases without resorting to designated initializers,
which are part of C99, not C++. Designated initializers appear to be
supported in C++ code by Clang and GCC, but not MSVC. The ctors must
be constexpr to ensure they are actually statically initialized so
they can be shared between Firefox processes.
I also removed an unnecessary "union" in XPTConstDescriptor.
Together, these patches reduce the amount of memory reported by
xpti-working-set from about 860,000 bytes to about 200,000 bytes. The
remaining memory is used for xptiInterface and xptiTypelibGuts (which
are thin wrappers around the XPT interfaces and header) and hash
tables to speed up looking up interfaces by name or IID. That could
potentially be eliminated from dynamic allocations in follow up
work. These patches did not affect memory reporting because XPT arenas
are still used by the remaining XPTI data structures.
MozReview-Commit-ID: Jvi9ByCPa6H
--HG--
extra : rebase_source : 719dfbcb9f83235c0f1f0766270b7f127f9ab04e
Now that XPT files are not loaded from files at runtime, code for
packaging XPT files can be removed.
This means that a couple of test XPIDL interfaces will get shipped in
builds to users that weren't before, but I don't think that matters
much.
This also puts XPT files into the local objdir for the XPIDL makefile,
instead of dist/bin, because they are no longer part of the
distribution.
MozReview-Commit-ID: 7gWj8KWUun3
--HG--
extra : rebase_source : 6f7d4fd1d6cdea2c14866705a2dc972eb5f43382
By using $(REPORT_BUILD) instead of just echoing the filename, there is
no change during a regular build without REBUILD_CHECK, but specifying
REBUILD_CHECK in the environment will show which files triggered an .xpt
to rebuild. This is helpful in debugging why these files may be built
unnecessarily.
MozReview-Commit-ID: GGNaKAl02Ea
--HG--
extra : rebase_source : 098a81265deed9afc0b284d9863f18ebd74fda33
We can just generate xpidllex.py/xpidlyacc.py in the current directory
rather than one directory higher, and specify this directory as an
include path to xpidl-process.py
MozReview-Commit-ID: KLoGjudc4Y8
--HG--
extra : rebase_source : 8dda268c6490cdfb8b896de9da5b789208584193
The configure option has explicitly thrown an error for more than a year now,
and it happens that the remaining way to still forcefully use it has been
broken for more than 8 months.
xpt files don't have a dependency on backend files to avoid rebuilding all
of them when adding or removing new files. On incremental builds, some kind
of dependencies are required to ensure the xpt files are refreshed when
adding or removing new idls.
Generating the list of idl deps to generate an xpt from its dependency list
makes us give all _previous_ dependencies, inherited from the .deps makefiles.
This leads to removed files being listed on xpidl-process.py command line, and
the command subsequently failing.
Instead, use generated lists of idl dependencies. At the same time, lighten the
generated Makefile further by not emitting xpt dependencies on their containing
directory, and instead generating it from the $xpt_files list.
This brings down the Makefile size from 100k to 38k.
Now that the mozbuild backend knows about FINAL_TARGET, we are able to
install generated xpt files into their final location. This saves us
from copying xpt files into their final location on every build.
Original patch by gps, rebased and comments addressed by Ms2ger
--HG--
extra : transplant_source : %E2%DC%0F%E0%AD%C2%25%A1%B8%A9%FE%B0%8C%60%FF%CB%02G%25%E5
Nika Layzell
Nathan Froyd
Andrew McCreight
Dorel Luca
Mike Shal
Chris Manchester
Mike Hommey
Brian O'Keefe
Sylvestre Ledru
Gregory Szorc
Ryan VanderMeulen