r=ben, sr=bryner).
- Pass null scope objects into nsIScriptContext::Compile{EventHandler,Script}
when precompiling for brutal sharing. The JS engine does not need a non-null
object parameter for static scope. That parameter can be non-null only if it
helps the compiler "pre-bind" functions to the same scope object that they'll
be parented by when executing, but with brutal sharing, functions are
precompiled once and executed against many different scope objects.
- A problem in XUL FastLoad was that it would serialize master .xul document
out-of-line scripts (those included via script src= from non-overlay, "master"
.xul docs) far from the place in the FastLoad file where XUL prototype script
info was serialized for the <script src=> tag itself. I fixed that so that,
unless the out-of-line script was previously serialized (by a different src=
reference from another .xul file), the OOL script data immediately follows the
proto-script info in the FastLoad file.
This required adding a SerializeOutOfLine method to nsXULPrototypeScript, which
restores symmetry by matching the existing DeserializeOutOfLine(Script) (note:
I dropped the redundant "Script" from the end of the latter method's name. We
need SerializeOutOfLine to handle overlay OOL scripts. They are serialized by
nsXULDocument::OnStreamComplete, because that code knows the difference between
an overlay and a master doc. This removes all trace of FastLoad writing from
nsXULPrototypeScript::Compile -- FastLoad stuff didn't belong there, not only
because we now want to write master OOL scripts later, when walking the master
XUL prototype doc's element tree, but also for modularity reasons. The caller
knows about FastLoad, nsXULPrototypeScript::Compile does just what its name
implies.
There are 132 seeks with the patch, only 49 of which dump the underlying file
stream's buffer, vs. 133 seeks without the patch, 87 of which dump the buffer.
- Nit-picked some comments to fit in 80 columns, and made other cosmetic fixes.
- Implicated the nsXULDocument::mIsWritingFastLoad flag from the useXULCache
"is the XUL cache enabled?" flag, so other places that test mIsWritingFastLoad
don't have to query whether the cache is enabled.
- Added METERING synchronous meter-dumping to /tmp/bufstats, only ifdef
DEBUG_brendan, in netwerk/base/src/nsBufferedStreams.cpp.
- Added the deferred seek optimization from the first patch in bug 195010 to
nsFastLoadFile.cpp.
- Fixed nsFastLoadFileReader so it overrides readSegments as well as read.
This catches up with the interface extension made to nsIInputStream to add
readSegments. The nsFastLoadFileReader extends nsBinaryInputStream, which is
a concrete class, and overrides Read in order to snoop on *all* bytes read.
It does this in order to demultiplex documents interleaved when the FastLoad
file was written.
But since the readSegments move into nsIInputStream.idl, certain primitives
in nsBinaryStream.cpp, e.g., nsBinaryInputStream::ReadCString, have used
ReadSegments, not Read, to consume parts of the underlying stream (to read
the C string's chars, in that example), and the FastLoad file implementation
has not accounted for those bytes.
- Added a new method to nsIFastLoadFileControl and its impls: hasMuxedDocument.
This is needed when serializing master XUL doc OOL scripts, because we must
not serialize twice, and any OOL script that other XUL docs can include via
script src= could already be in the FastLoad mux.
/be
is never 0, and make the memset-to-zero required for these assertions be
#ifdef NS_DEBUG.
- Extend FastLoad file format to encode a singleton flag in the high bit of
nsFastLoadSharpObjectInfo.mWeakRefCnt. The updater code needs to test this
bit in order to deserialize a singleton that was not read by the updater's
reader during this FastLoad episode. Otherwise the updater is likely to
reserialize the singleton, leading to multiple entries in the object map
for the same object, and UMRs when loading from the writer's object map
enumeration (objvec) in nsFastLoadFileWriter::WriteFooter.
- Remove bogus assertion that worked only when an XPCOM data structure being
serialized was still alive at the time the writer closed. In general, there
is no relationship between the strong ref-counts in a FastLoad file and the
XPCOM refcnts in memory at close time.
- Bug 189832, r=jrgm, sr=ben, a=dbaron.
Patch to make FastLoad Service return the previous URI selected when selecting a new one, as well as fixing a few issues relating to dependencies on non-existent files. Patch from brendan, r=dbaron, sr=waterson
for me on optimized and debug gcc2.96, rh7.1.
- Better failure codes from nsXULPrototypeScript::Deserialize.
- Call nsXULDocument::AbortFastLoads after nsXULPrototypeScript::Serialize
failure, instead of just nulling the FastLoad service's output stream.
- Expose nsXULDocument::AbortFastLoads via nsIXULPrototypeCache, for use from
nsChromeProtocolHandler.cpp. AbortFastLoads flushes the XUL cache now, for
good measure.
- The needless "Current" adjective in nsIFastLoadService attribute and method
names is no more.
- Add a do_GetFastLoadService() helper, to use CID instead of contractid, and
to let the compiler consolidate the static inline CID.
- Add "nglayout.debug.checksum_xul_fastload_file" pref so people can do without
the checksum verification step when reading a FastLoad file.
- Verify the FastLoad file checksum, by default. Also, cache it in the FastLoad
service so we don't recompute it when re-opening the FastLoad file as mailnews
and other top-levels start up. Fill the checksum cache in EndFastLoad, when
the last pseudo-concurrent top-level finishes loading.
My hope to compute the checksum while writing the FastLoad file ran afoul of
misordered writes. The old code to checksum the in-memory nsFastLoadHeader
also was broken on little endian platforms. Now all checksumming is done via
a separate read pass over the complete file, save for the header's checksum
field, which is summed as if it contained zero.
- Track and check FastLoad file dependencies. This required groveling with a
bunch of Necko interfaces in nsChromeProtocolHandler::NewChannel -- read it
and weep. Dependency checking, as well as checksum access and computation,
use better-factored nsIFastLoad{File,Read,Write}Control interfaces.
- nsBufferedStream::Seek wasn't flushing the buffer when seeking backward
within the buffer, but it must, because mCursor bounds the amount to write
if the buffer contains the end of file.
- Add an unbufferedStream readonly attribute to nsIStreamBufferAccess, so we
don't have to screw around with the bufferying layer when checksumming. Also
implement nsIStreamBufferAccess in nsBufferedOutputStream.
- nsISeekableOutputStream was bogus, based on a bad state I had put the
nsBufferedOutputStream code in on its way from being completely broken when
you seek backwards outside of the buffer. Removing this interface required
using nsIFastLoadFileIO in nsFastLoadFileWriter, and it also required careful
ordering of Close calls (the Reader must close after the Writer or Updater,
so that the Reader's underlying, unbuffered input stream can be read by
nsFastLoadFileWriter::Close to compute the checksum.
- Miscellaneous tab/indentation, comment typo, bracing, if( => if ( style,
nsnull vs. 0, useless variable elimination, tortured control flow,
AutoString instead of String, and gratuitous ; after nsISupportsUtils.h
macro call cleanups.
- Call nsXULDocument::AbortFastLoads after nsXULPrototypeScript::Serialize
failure, instead of just nulling the FastLoad service's output stream.
- Expose nsXULDocument::AbortFastLoads via nsIXULPrototypeCache, for use from
nsChromeProtocolHandler.cpp. AbortFastLoads flushes the XUL cache now, for
good measure.
- The needless "Current" adjective in nsIFastLoadService attribute and method
names is no more.
- Add a do_GetFastLoadService() helper, to use CID instead of contractid, and
to let the compiler consolidate the static inline CID.
- Add "nglayout.debug.checksum_xul_fastload_file" pref so people can do without
the checksum verification step when reading a FastLoad file.
- Verify the FastLoad file checksum, by default. Also, cache it in the FastLoad
service so we don't recompute it when re-opening the FastLoad file as mailnews
and other top-levels start up. Fill the checksum cache in EndFastLoad, when
the last pseudo-concurrent top-level finishes loading.
My hope to compute the checksum while writing the FastLoad file ran afoul of
misordered writes. The old code to checksum the in-memory nsFastLoadHeader
also was broken on little endian platforms. Now all checksumming is done via
a separate read pass over the complete file, save for the header's checksum
field, which is summed as if it contained zero.
- Track and check FastLoad file dependencies. This required groveling with a
bunch of Necko interfaces in nsChromeProtocolHandler::NewChannel -- read it
and weep. Dependency checking, as well as checksum access and computation,
use better-factored nsIFastLoad{File,Read,Write}Control interfaces.
- nsBufferedStream::Seek wasn't flushing the buffer when seeking backward
within the buffer, but it must, because mCursor bounds the amount to write
if the buffer contains the end of file.
- Add an unbufferedStream readonly attribute to nsIStreamBufferAccess, so we
don't have to screw around with the bufferying layer when checksumming. Also
implement nsIStreamBufferAccess in nsBufferedOutputStream.
- nsISeekableOutputStream was bogus, based on a bad state I had put the
nsBufferedOutputStream code in on its way from being completely broken when
you seek backwards outside of the buffer. Removing this interface required
using nsIFastLoadFileIO in nsFastLoadFileWriter, and it also required careful
ordering of Close calls (the Writer or Updater must close after the Reader,
so that the Reader's underlying, unbuffered input stream can be read by
nsFastLoadFileWriter::Close to compute the checksum.
- Miscellaneous tab/indentation, comment typo, bracing, if( => if ( style,
nsnull vs. 0, useless variable elimination, tortured control flow,
AutoString instead of String, and gratuitous ; after nsISupportsUtils.h
macro call cleanups.
