The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.

NAME

HACKERS - Devel::PPPort internals for hackers

SYNOPSIS

So you probably want to hack Devel::PPPort?

Well, here's some information to get you started with what's lying around in this distribution.

DESCRIPTION

How to backport something

First, make sure that what you want to backport is documented. If it's worth backporting, it surely is worth documenting. Submit a documentation patch to https://github.com/Perl/perl5/issues if necessary. Also, Devel::PPPort cannot automatically generate proper information about the item without at least knowing its API prototype. It can get this from embed.fnc if the item is a function, but if it is a macro, there needs to be at least a =for apidoc line for Devel::PPPort to be able to figure things out on its own.

Next, figure out where to place your implementation. Look at all the files in parts/inc/ for one that fits what you're planning. If there isn't one, just start a new one and remember to include it from within PPPort_pm.PL. If you do create a new file, it's usually the best approach to just copy an existing file and use it as a template.

Each file holds all relevant data for implementing a certain part of the API:

  • A list of the provided API in the =provides section.

  • The optional =dontwarn section is used to suppress warnings about particular API elements. Don't use this unless you get such a warning, and be sure to think about using other other alternatives before resorting to adding something in this section.

  • The implementation to add to ppport.h in the =implementation section. See "Implementation Section Details".

  • The code required to add to PPPort.xs for testing the implementation. This code goes into the =xshead, =xsinit, =xsmisc, =xsboot and =xsubs section. Have a look at the template at the bottom of RealPPPort_xs.PL to see where the code ends up.

  • The tests in the =tests section. Remember not to use any fancy modules or syntax elements, as the test code needs to be able to run with Perl 5.003. (This is because Devel::PPPort itself will run all test files in the process of generating the information about when a feature came into existence.) This means, for example

    my isn't supported in for-loops
        for my $x (1, 2, 3) { }    # won't work with 5.003

    Instead declare $x just before the statement

    The postfix for statement modifier isn't supported
        foo for 1..2

    won't compile. Instead enclose foo in a loop.

    You can't use plain qr//

    Instead, wrap it in a string eval eval "qr//", and be sure it's skipped at execution time on perls earlier than 5.005

    As of version 3.56 of Devel::PPPort, the old Test style tests have been replaced with the more modern Test::More style, with some limitations. This means, for example, that is is finally available, as well as done_testing. You can pass the number of tests to skip, instead of having to have your own for loop.

    There is no like nor unlike (as those require qr which didn't exist in the earliest perls that Devel::PPPort runs on).

    skip doesn't do a last SKIP. (Perhaps it could, but that would mean converting all the skips in the existing tests.)

    The existing tests have been changed only as much as necessary so as to get things to work. But feel free to use the full functionality for any new tests you write.

    Here's a list of the supported functions:

     cmp_ok
     curr_test
     diag
     display
     done_testing
     eq_array
     eq_hash
     fail
     is
     isnt
     next_test
     note
     ok
     pass
     plan
     skip
     skip_all
     within

    These are copied from t/test.pl in the perl distribution. Not all of them have been tested back as far as Devel::PPPort supports. Bug reports welcome.

    It's fine to backport an element only as far as convenient and necessary. But remember that your test file will end up being called on all perl versions available, likely including ones earlier than your backport. That may mean that elements in the =xs sections will have to be #idef'd out so that the object will not get missing symbols when loaded.

    It also means you have to check for and skip tests that aren't relevant to this version. The recommended way to do this is like:

     if (ivers($]) < ivers(5.6)) {           # No quotes needed
        skip "reason", $count;
     }
     elsif (if (ivers($]) > ivers("5.5.4") {
         # Quotes needed for double-dotted versions prior to 5.6.0
        skip "other reason", $count;
     }
     else {
        do_the_test
     }

    ivers() is a function automatically made available to all .t files. It converts any reasonble expression of a version number into an integer, which can reliably be compared using numeric comparison operators, with the output of a second ivers() call on a different version number, like in the result above.

    It's worth emphasizing that, due to bugs in early perl parsing, if you use a version number containing two dots on a version befor 5.6.0, it has to be quoted.

In all sections, lines that begin with ## are completely ignored.

Implementation Section Details

You can implement API elements via C functions or macros, or simple variables. It is preferable to use a macro if feasible. Otherwise, the user must explicitly request that it get loaded, by defining a NEED_function (or variable) as described in ppport.h. If a function, foo is required, place its body in this =implementation section, like so:

 #if { NEED foo }

 char *
 foo(pTHX_ const U8 *arg1, const U32 arg2, ...)
 {
    ...
 }

 #endif

Similarly for a variable.

It's obviously best to use a macro if at all feasible. Sometimes what once was implemented with a macro now requires a function; perhaps an edge case was overlooked. Doing so will cause the new ppport.h to not be drop-in compatible with the older version, and can hence cause breakage. This incompatiblity (while easily solved) really needs to be stressed in documentation.

__UNDEFINED__

If you add the line __UNDEFINED__ to the =provides section, you can use lines like this in the =implementation section:

  __UNDEFINED__ macro    some definition

to both define macro and indicate that it is provided by ppport.h. This replaces these =implementation section lines:

  #ifndef macro
  #  define macro    some definition
  #endif

besides automagically making it be considered to be provided. macro can have optional arguments and the definition can even span multiple lines, like in

  __UNDEFINED__ SvMAGIC_set(sv, val) \
                STMT_START { assert(SvTYPE(sv) >= SVt_PVMG); \
                (((XPVMG*) SvANY(sv))->xmg_magic = (val)); } \
                STMT_END

This usually makes the code more compact and readable.

But you should only use this on things that you plan to publicly provide. If something, such as a mnemonic for a constant needs to be defined but isn't really needed for the public at large to know about, you should use

 __UNDEF_NOT_PROVIDED__ macro   some definition

instead. To avoid name space conflicts, follow what's in "Helper macros", below.

Helper macros

If you need to define a helper macro which is not part of Devel::PPPort API and its usage is only for the definition of other Devel::PPPort macros, then use the D_PPP_ prefix for this macro name (e.g. D_PPP_SVPV_NOLEN_LP_ARG). This suppresses any warnings when a macro is defined which is not part of the Perl public API.

Version numbers

Version checking used to be tricky to get correct (besides being buggy in some perl versions). ivers() is used in the =tests section to overcome this. and constructs like the following in the C language sections.

  #if { VERSION < 5.9.3 }

You SHOULD be using this construct or the alternatives listed below for ALL version checks, and not come up with something on your own.

In this form, the version number can be either of the new form 5.x.y or the older form 5.00x_yy. Both are translated into the correct preprocessor statements. It is also possible to combine this with other statements:

  #if { VERSION >= 5.004 } && !defined(sv_vcatpvf)
    /* a */
  #elif { VERSION < 5.004_63 } && { VERSION != 5.004_05 }
    /* b */
  #endif

This not only works in the =implementation section, but also in the =xsubs, =xsinit, =xsmisc, =xshead and =xsboot sections.

Alternatively, you can use the forms now available in regular Perl:

  #if PERL_VERSION_EQ(5,9,3)
  #if PERL_VERSION_NE(5,9,3)
  #if PERL_VERSION_LT(5,9,3)
  #if PERL_VERSION_GT(5,9,3)
  #if PERL_VERSION_LE(5,9,3)
  #if PERL_VERSION_GE(5,9,3)

These forms have the advantage over the '{ VERSION ... }' form in that you may use the special value '*' for the final number to mean ALL possible values for it. Thus,

 #if PERL_VERSION_EQ(5,31,'*')

means all perls in the 5.31 series. And

 #if PERL_VERSION_NE(5,24,'*')

means all perls EXCEPT 5.24 ones. And

 #if PERL_VERSION_LE(5,9,'*')

is effectively

 #if PERL_VERSION_LT(5,10,0)
Hints

If you add a comment like so:

 /* Hint: PL_expect, PL_copline, PL_rsfp
    paragraphs of stuff you want to have shown when ppport.h outputs
    something about any one of PL_expect, PL_copline, or PL_rsfp
 */

Earlier versions of ppport.h required an asterisk at the beginning of every continuation line, or else the content would be silently dropped.

Warnings

A more serious caution can be displayed by instead saying

 /* Warning: PL_expect, PL_copline, PL_rsfp
    paragraphs of stuff you want to have shown when ppport.h outputs
    something about any one of PL_expect, PL_copline, or PL_rsfp
 */

Earlier versions of ppport.h required an asterisk at the beginning of every continuation line, or else the content would be silently dropped.

Replace

When ppport.h is run on a file(s), you can cause it to automatically flag occurrences of the constructs you specify, encouraging the author to replace them with different (presumably better) ones. These also are used in any suggested edits and generated patches.

There are three ways to do this

in-line comment

You can add a trailing comment like so:

 #define bar foo    /* Replace */
 __UNDEFINED__ bar foo  /* Replace */

These say that foo should be replaced by bar. NOT the other way around.

separate comment

For situations not amenable to the above, you can say

 /* Replace foo with bar */
define a replacement region

It you have several replacements, you can group them together like so:

 /* Replace: 1 */
 #define foo bar
 #define bat baz
 /* Replace: 0 */

These replace bar with foo; baz with bat. NOT the other way around.

Dependencies

ppport.h automatically gathers information as to what functions are dependent on what other things from inspecting the source, but if this is insufficient for you, you can add lines like the following:

 /* foo, bar depends on baz, bat */

Each of foo, bar depends on each of baz, bat.

Testing

After you have furnished your implementation, you need to test it.

Special Makefile targets

You can use

    make regen

to regenerate all of the autogenerated files. To get rid of all generated files (except for parts/todo/* and parts/base/*), use

    make purge_all

That's it.

To automatically test Devel::PPPort with lots of different Perl versions, you can use the soak script. Just pass it a list of all Perl binaries you want to test.

Regenerating ppport.h and PPPort.pm

Devel::PPPort keeps two directories of generated files, in parts/base and parts/todo. The files in each are named after Perl version numbers. When a function or macro came into existence is indicated by placing its name in the corresponding file in parts/base. The files in parts/todo are the same, except they indicate the earliest release that ppport.h supports the element. The delta is effectively what ppport.h buys you.

The generation process described in this section creates these files. It does so by examining as many perl versions as are available to it. It tries to make sure each element actually compiles, and it runs the test scripts you have furnished on every version.

Ideally, this should be done before every release that includes new backporting and/or when blead has added new public API. At a minimum, it should be done as the next major Perl release comes out.

The process isn't platform independent. It has currently been tested only under Linux, and it definitely requires at least gcc and the nm utility. The process used to be problematic, with random failures. But it has now been fixed to be reliable.

Before starting the regeneration, you need to have gathered certain data. (Options listed below apply to the tools that eventually will use the data, and which are described further below).

  • You will first need a whole bunch of different Perls, the more, the better, but only one per version tag (which one is random) will actually get used. dromedary has a sufficient set. They should all have the same Configure options with respect to what functions and macros are enabled. For example, they should all be threaded, or all non-threaded. A mixture will screw up the results. Similarly, they should all or none have quad math (at least as far back as that became available). You can use devel/buildperl.pl to build them.

    Previous maintainers of this module kept those perls in /tmp/perl/install/default, so most of the tools use this as a default, but you'll likely simply use the --install= option to specify where. This should be a path where a make install has been done, so has immediate subdirectories of /bin and /lib. /bin should contain the binaries. It will use all files in this directory whose names begin with perl5.

    Actually, not all the binaries need be in this directory. You can specify additional places to look since --install= takes a comma separated list of directories.

  • You also need a freshly built bleadperl. The --blead=path option should be used to specify it. (Some of the tools have a default of bleadperl-debug if this option is omitted.) Again, it needs the same Configure options as the earlier versions had. Using -DNO_MATHOMS will change the results, and probably should be avoided. True, these functions are allegedly on their way out, so it could be argued that they shouldn't be encouraged in any way; but some of these have been in limbo for many years, so should be documented.

  • And you will need updated API information. Copy the latest embed.fnc file from bleadperl to the parts directory and run devel/mkapidoc.pl to collect the remaining information in parts/apidoc.fnc. This needs to be done after the perl has been compiled, as there are generated files that feed it.

  • The final step before regenerating everything is to run devel/mkppport_fnc.pl to update the /parts/ppport.fnc file.

Having done this, run devel/regenerate which wraps the following steps (which you could instead do by hand, but it's easy to forget things):

  • It first does some sanity checking

  • Then it asks you if it's ok to remove all existing todo files in the parts/base and parts/todo directories. If you answer no, the process aborts.

    This is crtical to getting accurate results.

  • It builds the new baseline by running

        perl devel/mktodo --base

    in the root directory of the distribution.

    If there are warnings in blead, it will ask you to examine them, and to ok if it's all right to go ahead. If there are issues with blead, everything following could be wrong.

  • It builds the new todo files by running

        perl devel/mktodo

    in the root directory of the distribution.

  • Finally, it adds the remaining information by running

        perl Makefile.PL && make
        perl devel/scanprov --mode=write

How to build gobs of versions of Perl

Devel::PPPort supports Perl versions between 5.003 and bleadperl. To guarantee this support, its good to have as many versions as possible to test on. dromedary currently has many such versions.

There is a tool to build all the different versions and configurations. You can find it in devel/buildperl.pl. It can currently build the following Perl releases:

    5.003
    5.004 - 5.004_05
    5.005 - 5.005_04
    5.6.x
    5.7.x
    5.8.x
    5.9.x
    5.1x.x
    5.2x.x
    5.3x.x

Implementation

Knowing which parts of the API are not backwards compatible and probably need Devel::PPPort support is another problem that's not easy to deal with manually. If you run

    perl Makefile.PL --with-apicheck

a C file is generated by parts/apicheck.pl that is compiled and linked with Devel::PPPort. This C file has the purpose of using each of the public API functions/macros once.

The required information is derived from parts/embed.fnc (just a copy of bleadperl's embed.fnc), parts/apidoc.fnc (which is generated by devel/mkapidoc.pl and simply collects the rest of the apidoc entries spread over the Perl source code) and parts/ppport.fnc (which lists the API provided purely by Devel::PPPort, along with other elements that are tested only using ppport.h).

The generated C file (usually, apicheck.c) won't compile as-is with older perls. And even if it compiles, there's still a good chance of the dynamic linker failing at make test time. But that's on purpose!

We can use these failures to find changes in the API automatically. The Perl script devel/mktodo calls another script devel/mktodo.pl repeatedly to run Devel::PPPort on version after version of perl, in decreasing version order, so we start with blead and work backwards. The latter script generates an apicheck.c. It starts with the code that successfully worked in the previously tested Perl version, which should be the version one higher than the current one. Call the current one n, and the previous one n+1. The items that fail to compile in n, but did compile in n+1 must have become available in n+1. We run the Linux command nm to find those undefined symbols in n. We change apicheck.c to ignore (through #ifdef's) those and recompile, repeating until apicheck.c successfully compiles, the dynamic linker is happy, and make test runs on this version. Then we repeat the process for n-1, and so on. (Actually, this process may generate false positives, so by default each failing API call is checked again. If possible, this is done by generating an apicheck.c for just the one failing API.) Note that the make test is run using ppport.h during both passes.

Running devel/mktodo currently takes a couple hours on dromedary.

If you run it with the --nocheck option, it won't recheck the API calls that failed in the compilation stage and it'll take significantly less time. No one currently associated with maintaining this module understands under what circumstances it is safe to run with --nocheck.

By repeating the process over and over, we build up information on when every element first became supported. This information is stored in files in the parts/base directory, one file per version. The file for version n+1 is generated by running version n of perl.

We actually want a second piece of information, which is how much ppport.h buys you. What happens when regenerating is actually two entire runs through all the perls. The first is accomplished by calling devel/mktodo with the --base option. It automically will call devel/mktodo.pl with each version of perl, NOT using anything in ppport.h. When done the results indicate when each API element became available in stock perl, without using ppport.h.

And then the whole process is repeated, but this time ppport.h is included. The files are placed in parts/todo. Thus, at the end, we know when each element became available in modified perl, using ppport.h.

However, only the public API that is implemented as functions (and must appear in embed.fnc) plus macros whose calling sequence is documented can be checked this way. The final step in the process is calling devel/scanprov. It looks through the header files for when all the symbols provided by Devel::PPPort first became defined. It doesn't test the symbols or try to compile them, as it doesn't generally know the API, but it can tell that something exists in release n+1 but not n (by scanning the include files in the CORE directory of various Perl versions). (It does know if a macro has zero arguments or non-zero arguments, so it does get extra information from the zero argument ones.)

Files

Residing in parts/inc/ is the "heart" of Devel::PPPort. Each of the files implements a part of the supported API, along with hints, dependency information, XS code and tests. The files are in a POD-like format that is parsed using the functions in parts/ppptools.pl.

The scripts PPPort_pm.PL, RealPPPort_xs.PL and mktests.PL all use the information in parts/inc/ to generate the main module PPPort.pm, the XS code in RealPPPort.xs and various test files in t/.

You can get extra information from PPPort_pm.PL by setting the environment variable DPPP_CHECK_LEVEL to 1 or 2.

All of these files could be generated on the fly while building Devel::PPPort, but not having the tests in t/ will confuse TEST/harness in the core. Not having PPPort.pm will be bad for viewing the docs on search.cpan.org. So unfortunately, it's unavoidable to put some redundancy into the package.

Submitting Patches

If you've added some functionality to Devel::PPPort, please consider submitting a patch with your work to P5P by sending a pull request to

https://github.com/Dual-Life/Devel-PPPort/pulls.

When submitting patches, please only add the relevant changes and don't include the differences of the generated files. You can use the purge_all target to delete all autogenerated files.

Integrating into the Perl core

When integrating this module into the Perl core, be sure to remove the following files from the distribution. They are either not needed or generated on the fly when building this module in the core:

  MANIFEST
  META.yml
  PPPort.pm

BUGS

No known bugs.

COPYRIGHT

Version 3.x, Copyright (C) 2004-2020, Marcus Holland-Moritz and Perl 5 porters

Version 2.x, Copyright (C) 2001, Paul Marquess.

Version 1.x, Copyright (C) 1999, Kenneth Albanowski.

This program is free software; you can redistribute it and/or modify it under the same terms as Perl itself.

SEE ALSO

See ppport.h and devel/regenerate.