Fwd: Re: PDD 4, v1.3 Perl's internal data types (Final version)
On Friday 06 July 2001 10:13 am, Dan Sugalski wrote: > I should point out that the internal representation of large numbers isn't > going to be huge strings of ASCII characters--we'll probably be an array > of 15-bit integers. (As Hong pointed out a while ago, doing that makes > handling multiplication reasonably simple. Might go to arrays of 31-bit > integers on 64-bit platforms) Though I might be misreading you here. (I > probably am) Actually, you *shouldn't* have to point that at. No, you weren't misreading me, and Yes, Virginia, I am a fucking idiot. I can't even think of what I may have been thinking of. We have talked about this before. I write the damn summaries, for crying out loud Arggh! -- Bryan C. Warnock [EMAIL PROTECTED]
Re: PDD 4, v1.3 Perl's internal data types (Final version)
Dan Sugalski wrote: > The C structure that represents a bigint is: > >struct bigint { > void *buffer; > UV length; > IV exponent; > UV flags; >} > > =begin question > > Should we scrap the buffer pointer and just tack the buffer on the end > of the structure? Saves a level of indirection, but means if we need > to make the buffer bigger we have to adjust anything pointing to it. > > =end question Absolutely not. Keep as much static-sized as possible, so you can trivially recycle it. Nobody much liked the suggestion of tracking precision at the lowest levels, but here I am repeating it anyway. > Perl has a single internal string form: > =item unused > > Filler. Here to make sure we're both exactly double the size of a > bigint/bigfloat header and to make sure we don't cross cache lines on > any modern processor. Is this explicitly guaranteed to remain unused, so that it may be safely used for arbitrary user-magic (as long as they don't step on each others toes) and semantic analysis flags, and so forth? Or would that kind of thing be better included into whatever is containing these guys -- along with reference counts and other details of additional systems which are not referred to w/in this document. > =item Class > > Class refers to a higher-level piece of perl data. Each class has its > own vtable, which is a class' distinguishing mark. Classes live one > step below the perl source level, and should not be confused with perl > packages. Does this imply that perl packages will continue to be called perl packages, even when they start getting introduced with a "class" keyword?
Re: PDD 4, v1.3 Perl's internal data types (Final version)
> "DS" == Dan Sugalski <[EMAIL PROTECTED]> writes: DS> We won't be using a char-based string math library--it'll all be DS> some internal binary format or other. (I can make a good argument DS> for it being done with a base 10 exponent rather than a base 2 DS> one. I can see doing it all in decimal rather than binary, but I DS> can't think of a processor newer than the 6502 that does BCD DS> math. (Well, OK, I think the System/3x0 processors do--I suppose DS> that counts)) there is very neat and fast trick, called excess 3, for doing bcd sum and difference math with a set of binary words. this means doing 8 bcd digits at a time with common 32 bit words or 16 digits on alphas/sparc and itaniums. keeping the number (at least the mantissa) in bcd would also mean simpler and faster conversions to strings. in any case, when the bigint/float design starts up, i will jump in with my excess 3 suggestions and help. uri -- Uri Guttman - [EMAIL PROTECTED] -- http://www.sysarch.com SYStems ARCHitecture and Stem Development -- http://www.stemsystems.com Learn Advanced Object Oriented Perl from Damian Conway - Boston, July 10-11 Class and Registration info: http://www.sysarch.com/perl/OOP_class.html
Re: PDD 4, v1.3 Perl's internal data types (Final version)
At 02:49 AM 7/6/2001 -0400, Uri Guttman wrote: >question: > >can you declare at the language level a scalar to be a bigint or bignum? I think Larry's planning on that, yep. For arrays and hashes at least, I expect, and I don't see why not for scalars too. >that means that native format is never used. the reason might be >something like a fixed point decimal value for money with 2 decimal >places. the bigint/float thingies imply decimal math and that also means >a decimal math library. this came up in #perl in a discussion about >bcd. i think a true decimal math package for this would be useful and >faster than a char string based one. We won't be using a char-based string math library--it'll all be some internal binary format or other. (I can make a good argument for it being done with a base 10 exponent rather than a base 2 one. I can see doing it all in decimal rather than binary, but I can't think of a processor newer than the 6502 that does BCD math. (Well, OK, I think the System/3x0 processors do--I suppose that counts)) Dan --"it's like this"--- Dan Sugalski even samurai [EMAIL PROTECTED] have teddy bears and even teddy bears get drunk
Re: PDD 4, v1.3 Perl's internal data types (Final version)
At 07:10 PM 7/5/2001 -0400, Bryan C. Warnock wrote: >On Thursday 05 July 2001 02:11 pm, Dan Sugalski wrote: > > =begin question > > > > Should we scrap the buffer pointer and just tack the buffer on the end > > of the structure? Saves a level of indirection, but means if we need > > to make the buffer bigger we have to adjust anything pointing to it. > > > > =end question D'oh! I thought I'd chopped those question sections out! >This is probably silly to consider, and it may have been brought up before, >and, of course, I'm bringing it up *after* you've already closed it out, but >leaving it as a buffer pointer could make string->number->string conversions >almost as simple as a pointer copy. Yep, it could. Not likely, but possible. What is likely is that we'll see a bigint->bigfloat conversion that consists of swapping vtable pointers in the main PMC and (maybe) setting the flags in the number struct. >Taken to its extreme, you could run all conversions through bignum (or >bigfloat), if you decided not to rely on platform support for it. This >would give you a single piece of code to handle number detection and >manipulation, which means potential modularization (perhaps a third-party >lib), consistent results across platforms, and, well, just one piece of code >to tweak and maintain. (I think that that could also include bigocts, >bighexs, and bigbins, too.) Taken to its extreme extreme, there wouldn't >even really need to be a big* type - just big* code, a flag, and builtin >logic to treat a regular string as a big*. I should point out that the internal representation of large numbers isn't going to be huge strings of ASCII characters--we'll probably be an array of 15-bit integers. (As Hong pointed out a while ago, doing that makes handling multiplication reasonably simple. Might go to arrays of 31-bit integers on 64-bit platforms) Though I might be misreading you here. (I probably am) I think making bigint and bigfloat separate things is a reasonable performance win, but I might be wrong here. >Taken to its not-so-extreme-case, if you make an assumption (which may be >bad Bad BAD!) that most numeric work would have an implicit exponent of >10^0, then it *is* a simple pointer copy - at least until the big* decides >to normalize it to something else. That would cause the big* and the string >value to get out of sync, but you're going to need to address the conversion >back anyway. Assumptions are A Bad Thing. Explicit guarantees, however... If it turns out that an implicit exponent of 10^0 for bigints is warranted (and I don't see why it'd be a bad thing, though performance would probably be a little dodgy doing 10^45 + 10^46) then I'm all for doing that. Dan --"it's like this"--- Dan Sugalski even samurai [EMAIL PROTECTED] have teddy bears and even teddy bears get drunk
Re: PDD 4, v1.3 Perl's internal data types (Final version)
> "DS" == Dan Sugalski <[EMAIL PROTECTED]> writes: DS> 'Kay, here's the final version of this. DS>struct bigint { DS> void *buffer; DS> UV length; DS> IV exponent; DS> UV flags; DS>} DS> =begin question DS> Should we scrap the buffer pointer and just tack the buffer on the end DS> of the structure? Saves a level of indirection, but means if we need DS> to make the buffer bigger we have to adjust anything pointing to it. i think the indirection is good for that reason. tracking all refs to this structure is a lot of work for any resize operations. DS> =end question DS> The C pointer points to the buffer holding the actual s/num_// DS> and yes, this looks identical to the bigint structure. This isn't DS> accidental. Upgrading a bigint to a bignum should be quick. question: can you declare at the language level a scalar to be a bigint or bignum? that means that native format is never used. the reason might be something like a fixed point decimal value for money with 2 decimal places. the bigint/float thingies imply decimal math and that also means a decimal math library. this came up in #perl in a discussion about bcd. i think a true decimal math package for this would be useful and faster than a char string based one. uri -- Uri Guttman - [EMAIL PROTECTED] -- http://www.sysarch.com SYStems ARCHitecture and Stem Development -- http://www.stemsystems.com Learn Advanced Object Oriented Perl from Damian Conway - Boston, July 10-11 Class and Registration info: http://www.sysarch.com/perl/OOP_class.html
Re: PDD 4, v1.3 Perl's internal data types (Final version)
On Thursday 05 July 2001 02:11 pm, Dan Sugalski wrote: > =begin question > > Should we scrap the buffer pointer and just tack the buffer on the end > of the structure? Saves a level of indirection, but means if we need > to make the buffer bigger we have to adjust anything pointing to it. > > =end question This is probably silly to consider, and it may have been brought up before, and, of course, I'm bringing it up *after* you've already closed it out, but leaving it as a buffer pointer could make string->number->string conversions almost as simple as a pointer copy. Taken to its extreme, you could run all conversions through bignum (or bigfloat), if you decided not to rely on platform support for it. This would give you a single piece of code to handle number detection and manipulation, which means potential modularization (perhaps a third-party lib), consistent results across platforms, and, well, just one piece of code to tweak and maintain. (I think that that could also include bigocts, bighexs, and bigbins, too.) Taken to its extreme extreme, there wouldn't even really need to be a big* type - just big* code, a flag, and builtin logic to treat a regular string as a big*. With the buffer length specified, you may not even need to normalize an exponent string to a non-exponential string - you could simply set the buffer length short. (Of course, this could cause some resize problems, if the bignum tries to grow a string that is really longer than it thinks it is, if you follow.) Taken to its not-so-extreme-case, if you make an assumption (which may be bad Bad BAD!) that most numeric work would have an implicit exponent of 10^0, then it *is* a simple pointer copy - at least until the big* decides to normalize it to something else. That would cause the big* and the string value to get out of sync, but you're going to need to address the conversion back anyway. -- Bryan C. Warnock [EMAIL PROTECTED]
PDD 4, v1.3 Perl's internal data types (Final version)
'Kay, here's the final version of this. Cut here =head1 TITLE Perl's internal data types =head1 VERSION 1.3 =head2 CURRENT Maintainer: Dan Sugalski <[EMAIL PROTECTED]> Class: Internals PDD Number: 4 Version: 1.3 Status: Developing Last Modified: 02 July 2001 PDD Format: 1 Language: English =head2 HISTORY =over 4 =item Version 1.3, 2 July 2001 =item Version 1.2, 2 July 2001 =item Version 1.1, 2 March 2001 =item Version 1, 1 March 2001 =back =head1 CHANGES =item Version 1.3 Fixed some silly typos and dropped phrases. Took all the underscores out of the field names. =item Version 1.2 The string header format has changed some to allow for type tagging. The flags information for strings has changed as well. =item Version 1.1 INT and NUM are now concepts rather than data structures, as making them data structures was a Bad Idea. =item Version 1 None. First version =head1 ABSTRACT This PDD describes perl's known internal data types. =head1 DESCRIPTION This PDD details the primitive datatypes that the perl core knows how to deal with. These types are lower-level than what's presented to the perl programmer. =head1 IMPLEMENTATION =head2 Integer data types Integer data types are generically referred to as Cs. Cs are conceptual things, and there is no data structure that corresponds to them. =over 4 =item Platform-native integer These are whatever size native integer was chosen at perl configuration time. The C-level typedef C and C get you a platform-native signed and unsigned integer respectively. =item Arbitrary precision integers Big integers, or bigints, are arbitrary-length integer numbers. The only limit to the number of digits in a bigint is the lesser of the amount of memory available or the maximum value that can be represented by a C. This will generally allow at least 4 billion digits, which ought to be far more than enough for anyone. The C structure that represents a bigint is: struct bigint { void *buffer; UV length; IV exponent; UV flags; } =begin question Should we scrap the buffer pointer and just tack the buffer on the end of the structure? Saves a level of indirection, but means if we need to make the buffer bigger we have to adjust anything pointing to it. =end question The C pointer points to the buffer holding the actual number, C is the length of the buffer, C is the base 10 exponent for the number (so 2e4532 doesn't take up much space), and C are some flags for the bigint. BThe flags and exponent fields may be generally unused, but are in to make the base structure identical in size and field types to other structures. They may be removed before the first release of perl 6. =back =head2 Floating point data types Floating point data types are generically reffered to as Cs. Like Cs, Cs are a conceptual things, not a real data structure. =over 4 =item Platform native float These are whatever size float was chosen when perl was configured. The C level typedef C will get you one of these. =item Arbitrary precision decimal numbers Arbitrary precision decimal numbers, or bignums, can have any number of digits before and after the decimal point. They are represented by the structure: struct bignum { void *buffer; UV length; IV exponent; UV flags; } and yes, this looks identical to the bigint structure. This isn't accidental. Upgrading a bigint to a bignum should be quick. =for question Like the bigint structure, should we toss the data pointer and just tack the data on the end? =end question =back =head2 String data types Perl has a single internal string form: struct perl_string { void *buffer; UV allocated; UV bytes; UV flags; UV characters; UV encoding; UV type; UV unused; } The fields are: =over 4 =item buffer Pointer to the start of the string's data. =item allocated How many bytes are allocated in the buffer. =item bytes How many bytes are used in the buffer. =item flags Flags indicating whatever. Bits 0-15 are reserved for perl, bits 16-23 for the encoding/decoding code, and teh rest for the type code. =item characters How many characters are in the buffer. An optional cache field. =item encoding How the data is encoded, for example fixed 8-bit characters, utf-8, or utf-32. An index into the encoding/decoding function table. Note that this specifies encoding only--it's valid to encode EBCDIC characters with the utf-8 algorithm. Silly, but valid. =item type What sort of string data is in the buffer, for example ASCII, EBCDIC, or Unicode. Used to index into the table of string functions. =item unused Filler. Here to make sure we're both exactly double the size of a bigint/bigfloat header and to make sure we don't cross cache lines on any modern processor. =back =head1 ATTACHMENTS None =head1 REFERENCES The perl modules Mat