The R Project SVN R

Notes for Windows Maintainers

=============================

1) fixed/h/config.h can be made automatically.  Some of this is

achieved by overrides in the configure script.

You will need a copy of sh.exe in /bin.  Then with the R sources in

/R/R-2.10.0, I used in /TEMP/R (any other directory will do, except the

sources).

To test trio set

LIBS='<RHOME>/src/extra/trio/libtrio.a'

sh /R/R-2.10.0/configure --build=i386-pc-mingw32 --with-readline=no --with-x=no

This ran fairly slowly, currently producing a src/include/config.h

that differs only in that

 - it does not find the declarations of siglongjmp/sigsetjmp (in

   psignal.[ch])

 - MinGW gcc 4.2.1 supports pthreads, but I left that undefined.

Also, watch out for versions of the header files.  For example,

<strings.h> was in mingw-runtime-1.2 and later, but not (long ago) in

the mingw-1.1 bundle.  These days we insist on a current MinGW.

BDR 2002-04-15, 2003-02-10, 2004-06-28, 2005-11-25, 2006-07-06, 2007-08-09, 

    2009-07-08

2) The Makefiles for building a distribution have been reorganized.  Now

all of the decisions about what files go into the distributables are made

in installer/Makefile; the decisions about which component of the setup

program each file goes into are made in installer/JRins.pl.

DJM 2003-02-25

3) Making Tcl/Tk.

See https://github.com/waddella/Tclbuild.

4) Linking to DLLs.

MinGW's ld.exe takes the internal name of the DLL as the object to

link to, and hence for DLLs which are to be linked to (rather than

loaded), the internal names need to be dllname.dll.  This was not being

done by the %.dll rule used in 2.2.1, which builds via a .def file

with first line

LIBRARY $*

[The example in ld.info is

LIBRARY "xyz.dll"

whereas that in the MSDN documentation

(http://msdn2.microsoft.com/en-us/library/d91k01sh.aspx) is

LIBRARY BTREE

so MinGW is inconsistent with MSDN here.]

It would seem that this should just be replaced by $*.dll, but there

is another problem: ld.exe rejects .def files whose LIBRARY name

contains more than one dot, and so this is unable to cope with

packages with a dot in their name.  (We already document that two or

more dots are not allowed.)  So we have had to treat separately the

DLLs which are designed to be linked to.  These are

R.dll : is special-cased, and as it wants to export entry points from

static libraries and exports variables, nothing else we tried worked.

Rblas.dll : has a .def file, and the name was changed to Rblas.dll

there.

Rlapack.dll : is simple, so gcc -shared with no .def file works.

Rproxy.dll : is special-cased.  (Linked against by rcom package.)

The other DLLs which were in R_HOME/bin, Rbitmap.dll and Rchtml.dll,

have been moved to R_HOME/modules.  They are now made directly with

gcc -shared.  Rchtml.dll needs an import library as you cannot link

directly to a .ocx.

However, whereas MSDN says there must be a LIBRARY statement, it seems

not to be required for ld.exe.  So the %.dll rule in MkRules as from

2.3.0 does not have a LIBRARY statement, which circumvents the 'at

most one dot' rule.

BDR 2006-02-15, 2006-02-24

5) Rdll.hide

AllDevicesKilled

RConsole

RFrame

Rf_runcmd

RgetMDIheight

RgetMDIwidth

RguiMDI

Ri18n_wcwidth

Riconv

Riconv_close

Riconv_open

Rwin_graphicsx

Rwin_graphicsy

UserBreak

locale2charset

optclosefile

optfile

optline

optopenfile

optread

for grDevices

R_deferred_default_method

R_do_MAKE_CLASS

R_do_new_object

R_do_slot

R_do_slot_assign

R_execMethod

R_primitive_generic

R_primitive_methods

R_set_prim_method

R_set_quick_method_check

R_set_standardGeneric_ptr

R_subassign3_dflt

do_set_prim_method

for methods

set_R_Tcldo

unset_R_Tcldo

Rf_wtransChar

for tcltk

R_fixbackslash

consolefn

freeConsoleData

freemenuitems

lzma_crc64

orderVector1

wgl_histadd etc

for utils

R_gl_tab_set

cmdlineoptions

getDLLVersion

gl_hist_init

gl_loadhistory

readconsolecfg

saveConsoleTitle

setupui

for rgui/rterm/Rscript

Rf_mbrtowc

Rf_strchr

localeCP

for graphapp.dll

setup_term_ui

for package Rserve

optif9

for package nlme

BDR 2007-08-20

6) Conversion to Unicode/UTF-8.

[Notes are a work in progress.  This has been discussed since 2003.]

As from R 2.7.0, Rgui works internally in UCS-2.  Currently key strokes are

recorded as bytes, but that could easily be changed if we had testers with CJK

keyboards.  Input/output is converted to/from the current locale in

R_ReadConsole/R_WriteConsole.

The internal pager works in UCS-2.

History files are currently written in the locale's charset, for

compatibility with Rterm and past versions of R.  We could perhaps

alleviate this by using a BOM when writing in UCS-2, and detecting

that when loading history files.

Chris Jackson's script editor is still MBCS.  Quite a bit of work will

be needed to convert it, since e.g. GA_gettext() expects to work with

char *.

The only bar to converting Rterm to UCS-2 is the lack of CJK keyboards

to test on, but there would be little advantage in doing so.

6.1 Internal use of UTF-8

- Internal mbcs<->wchar/UCS translations would need to be modified to

  use UTF-8 and not the locale charset.  Several places

  (e.g. Riconv_open) already have support for this.  The

  gettext-runtime (src/extra/intl) does not.

*All* I/O would need to be translated.  This includes

- Rprintf and allies, error messaging.

  [This one is really tricky: we don't want to be writing UTF-8 files

   when sink() is in use, for example, and rterm/some embedded uses

   run in environments that probably do not support UCS-2.

   The current compromise is to encode UTF-8 character vector elements

   in EncodeString, at least when called from printing and from cat(),

   and if outputting to the Rgui console. This is done by surrounding

   them by 3-byte escape sequences starting with STX/ETX.]

- file names.  Since Windows NTFS can have file names not valid in the

  current locale, this needs to use wfopen and similar interfaces, as

  well as GetFullPathNameW ... also _w* versions of mkdir, rmdir,

  unlink, open, popen, stat, system.  Use with care, as not all file

  systems use UCS-2 file names.  [Done]

  This needs to include bzlib and zlib, as well as file/folder

  selection widgets. [Done]

- environment variables.  Both values and names, since _wputenv is

  needed to set name=value strings.  It would be better to use wmain

  and only ever use _wgetenv and _wputenv to avoid maintaining

  parallel environments.  [Done]

- graphics devices.  Calls to text() and points() (with pch as a

  string) need either to be translated or, preferably, the device

  adjusted to accept UTF-8 (and we would need a flag to know if it

  could).  Note that this impacts third-party devices.

  [Done for windows() family, postscript, pdf.]

- GetUserName, GetComputerName, [GS]etCurrentDirectory. [Done]

Also:

- There would be compatibility issues with saved workspaces, although

  few for those working in CP1252 locales.

- It is not clear what to do with strings passed to packages.  We

  could force conversion in .C and .Fortran, but most

  character-manipulation packages are likely to be using .Call.

BDR 2008-01-06

7) Changing the parser

By default the parser in gram.y is only processed in maintainer mode, 

which isn't supported in Windows builds.  Defining RUN_BISON will 

cause src/main/Makefile.win to process it (and correct for the erroneous

line number records caused by our renaming of the output).

8) Making iconv.dll

In the past this was done from the libiconv sources using VC++ 6, but

as from version 1.12 that is no longer supported.  It does not build

easily with MinGW (due to the overuse of libtool), but we used

setenv INSTALL cp

setenv INSTALL_DATA cp

sh configure --build=i386-pc-mingw32 --disable-nls

cd libcharset; make

cp include/localcharset.h ../lib

cd ../lib

make

gcc -shared  .libs/iconv.o .libs/localcharset.o .libs/relocatable.o \

  .libs/iconv-exports.o -export-all-symbols -o Riconv.dll

We now use a modified version of win-iconv, but the libiconv version

can still be used by replacing bin/Riconv.dll.

BDR 2009-02-08