Re: Current FOText implementation + Refinement whitespace handling
On Thu, 3 Nov 2005 04:08 am, Andreas L Delmelle wrote: Hi all, (Manuel, I guess this is mostly directed to you, as you may already have been browsing the same classes...) Just wandering a bit through the FOText source code (follow-up on Manuel's recent thread on whitespace handling), and I stumbled upon the following suspicious little detail: FOText has a static member 'lastFOTextProcessed', which doesn't seem to get cleared/flushed anywhere. Actually I think I have the issue of white space handling during refinement which is implemented in the handleWhiteSpace method in Block under control. The two issues I identified before: a) the character iterator needs to indicate inline boundaries as they act as a fence with respect to white space collapse b) we need to be able to remove leading white space before a hard break have been solved. a) The char iterator returned by inline's now does return NUL characters for start, end which the white space handling function automatically interprets as interupting consecutive white space b) The LF look ahead function already present has been enhanced to indicate LF for the end of the block thereby allowing white-space-treatment to be applied to the leading white space before the block end I have also rewritten the handleWhiteSpace method to behave in sync with the current understanding. My testcases so far indicate that this is working consistent with my expectations, other XSL-FO implementations, and comparable HTML use cases. The only limitation I am aware off at the moment is that the suppress-at-line-break property is not supported. This is because the char iterator used does only return characters (in the pure Java sense) and not their related XSL-FO properties. At this point in time I consider support for the suppress-at-line-break property not as high priority. The only bug (at this level) I am aware off is the already documented failure of a delete on a fo:character. But you already suggested some possible solutions I will have a look at. In terms of completing white space handling it really now boils down to the handling of white-space-treatment around formatter generated breaks. This is currently being discussed in other e-mail threads and is quite complex (not logically but technically) because it interacts with the Knuth sequences / Knuth algorithm, the internal LM structures, as well as the proposed UNICODE compliant line breaking. The intention is quite clear, but the possible effects of the current implementation may turn out rather nasty. IIC, this is what the warning is about in the FOText javadoc as well as the TODO for that member variable. Rough guess: since the variable doesn't get cleared, it always contains a reference to a char array containing the last portion of accumulated text (or, more precisely, a FOText instance carrying that reference, as well as one to the previous FOText etc.) --even after the document has finished, into the next run if within the same JVM (+ possible multi-thread mayhem?) The TODO hints at a solution involving the page-sequence. I somehow feel that moving it to the block level would be enough... Logically, whitespace handling --which is one of the prime reasons of existence of this static variable-- deals with line-breaks, and start-block/end- block are implicit after- or before-eol. To follow up on that last sentence, the current refinement whitespace handling works roughly as follows: 1. Add all text and inline children to the block, until the first non- inline child is encountered (or the block ends) 2. Recursively iterate over *all* text nodes anywhere in the block up to here, converting/removing any superfluous whitespace in the process and (+/-) repeat the above for each uninterrupted sequence of text/ inline children in the block. Seems to work nicely, for the most part. Manuel already raised the issue of inappropriate inter-FO whitespace- collapsing, but I have another question. Given this algorithm, and knowing that the inlines do not do any whitespace-handling themselves, what happens in the following case: fo:block fo:inline fo:block fo:inline fo:block ... ? My current best guess is that the inner block's underlying character sequence will be 'recursively' iterated over three times (?) That would be two too many, since all whitespace will have been collapsed the first time around. The Block class has a flag which prevents multiple iterations I believe. I'm still chewing on some ideas to move part of this to InlineLevel, so that ultimately, we can do away with the recursion and let each level handle its own small part. The higher level then chains these small parts together with its own character content. One way to make this happen would be to overload Block.handleWhiteSpace() to deal with an InlineLevel parameter. This has the advantage of the whitespace-related properties being easily
Re: Current FOText implementation + Refinement whitespace handling
On Nov 3, 2005, at 10:41, Manuel Mall wrote: snip / The only limitation I am aware off at the moment is that the suppress-at-line-break property is not supported. This is because the char iterator used does only return characters (in the pure Java sense) and not their related XSL-FO properties. At this point in time I consider support for the suppress-at-line-break property not as high priority. Well, IMO it shouldn't be handled during refinement, exactly because the iterators don't have access to the FO's properties. I know it can be argued that in cases like: ... fo:character character=... suppress-at-line-break=true / /fo:block During refinement, it is already known that the character will be suppressed. If the character is whitespace and the iterator is modified to deal with that, it will be dropped, regardless of the value of suppress-at-line-break (supposing default values for all other props). But in the most general cases, I think suppress-at-line-break is best dealt with during layout, not refinement. Then again, modifying the OneCharIterator (or subclassing to FOCharIterator?) to correctly deal with fo:characters (removal/ replacing) is easy enough. So, it's _possible_ to let it access the FOs properties. It's only that I'm not sure whether it would be _necessary_. snip / My current best guess is that the inner block's underlying character sequence will be 'recursively' iterated over three times (?) That would be two too many, since all whitespace will have been collapsed the first time around. The Block class has a flag which prevents multiple iterations I believe. See my immediate correction :-) This became apparent just after I had clicked 'Send'... snip / As I said above, I believe I got this sorted out. May be I should do an early commit? Would be nice. So at least the codebases we're talking about are completely in sync. Cheers, Andreas
Current FOText implementation + Refinement whitespace handling
Hi all, (Manuel, I guess this is mostly directed to you, as you may already have been browsing the same classes...) Just wandering a bit through the FOText source code (follow-up on Manuel's recent thread on whitespace handling), and I stumbled upon the following suspicious little detail: FOText has a static member 'lastFOTextProcessed', which doesn't seem to get cleared/flushed anywhere. The intention is quite clear, but the possible effects of the current implementation may turn out rather nasty. IIC, this is what the warning is about in the FOText javadoc as well as the TODO for that member variable. Rough guess: since the variable doesn't get cleared, it always contains a reference to a char array containing the last portion of accumulated text (or, more precisely, a FOText instance carrying that reference, as well as one to the previous FOText etc.) --even after the document has finished, into the next run if within the same JVM (+ possible multi-thread mayhem?) The TODO hints at a solution involving the page-sequence. I somehow feel that moving it to the block level would be enough... Logically, whitespace handling --which is one of the prime reasons of existence of this static variable-- deals with line-breaks, and start-block/end- block are implicit after- or before-eol. To follow up on that last sentence, the current refinement whitespace handling works roughly as follows: 1. Add all text and inline children to the block, until the first non- inline child is encountered (or the block ends) 2. Recursively iterate over *all* text nodes anywhere in the block up to here, converting/removing any superfluous whitespace in the process and (+/-) repeat the above for each uninterrupted sequence of text/ inline children in the block. Seems to work nicely, for the most part. Manuel already raised the issue of inappropriate inter-FO whitespace- collapsing, but I have another question. Given this algorithm, and knowing that the inlines do not do any whitespace-handling themselves, what happens in the following case: fo:block fo:inline fo:block fo:inline fo:block ... ? My current best guess is that the inner block's underlying character sequence will be 'recursively' iterated over three times (?) That would be two too many, since all whitespace will have been collapsed the first time around. I'm still chewing on some ideas to move part of this to InlineLevel, so that ultimately, we can do away with the recursion and let each level handle its own small part. The higher level then chains these small parts together with its own character content. One way to make this happen would be to overload Block.handleWhiteSpace() to deal with an InlineLevel parameter. This has the advantage of the whitespace-related properties being easily available. The call to this overloaded method would be made from InlineLevel.endOfNode(). If you're still following, I'd use a CharIterator that iterates over regular characters, fo:characters (and possibly the first and last characters of any nested FO). This iterator can operate very easily on both inlines and blocks. I don't immediately see any need to iterate backwards, at least not during refinement. Big advantage here would precisely be that we can wait until Block.endOfNode() to deal with any white-space for the entire block (leading and trailing), the nested bits will already have performed their parts at that point, so it is done sooner and far more efficiently IIC (guaranteed only one pass per level, no matter how deep the nesting goes). Food for thought :-) Cheers, Andreas