This is the mail archive of the cygwin-apps mailing list for the Cygwin project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [ITP] mingw-w64


On 7/1/2010 00:36, Charles Wilson wrote:
On 6/29/2010 1:13 PM, JonY wrote:
On 6/30/2010 00:10, Charles Wilson wrote:
Now, I thought you wanted to use the w64 prefix as a "project origin"
indicator, and assumed that "-mingw64-" would be the "target bitdepth"
indicator. However, given "w64-mingw64-pthreads-devel32" and
"w32-mingw64-pthreads-headers32" I'm not sure that's the case. It seems
NOW that you want to use 'mingw64' as the "project origin" indicator,
and "w64"/"w32" as the target bit depth (and I'm sure the trailing '32'
in these two anomalies is unnecessary).

So, I think I'd put the project origin first, followed by the bit depth.


Yes, I was using w64/w32 as bitness indicator, as with some releases made with the buildbot on sf, mingw64 to differentiate it from mingw.org.

Ah. Well, in that case I would think that the origin should come first, rather than the bitdepth. Our existing mingw-runtime package is from (effectively) mingw.org, after all (*).

More below.

(*) Well, both mingw.org mingwrt and cygwin's mingw-runtime come from a
common repository: the winsup/mingw directory on sourceware's CVS
server. [note that the "mingwrt" and "w32api" parts of "mingw.org" are
hosted by sourceware; all the rest of "mingw.org" including other source
code repos are hosted by sourceFORGE. The reason for the split is
historical.]


I understand, they do use sourceware's combined configure tree.


I don't know much about sf's buildbot; I assume that if you can get a
cygport to DTRT on your home PC without human intervention except for
kicking off the build, then you can convince the buildbot to do it for you?


Right now, human intervention still needed. cygport messes up the target dll locations by moving them around and trying to fix libtool files. Its also using cygwin strip(1) to strip 64bit dlls, it fails but doesn't lead to the cygport halting.


The trailing 32/64 is to indicate which toolchain the pthreads implibs
are for, it is too possible to setup a 32bit default multilib setup, the
current toolchain is defaulting to 64bit. So w64*devel32 means 64bit
implib for 32bit default toolchain that has yet to bet setup.

For the current multilib toolchain example, users would want w32*devel64
and w64*devel64 pthreads packages, 32bit and 64bit implib for 64bit
default toolchain (the tarballs have the same installation path to the
64bit libdir).


Hmm. So, big picture, we have possibly three different mingw-ish
compilers, and you're currently attempting to shepherd the first one,
while being mindful of future issues related to simultaneous
installation of both of the first two:

(1) mingw64-derived, multilib, default 64bit
(2) mingw64-derived, multilib, default 32bit
(3) mingw.org-derived, non-multilib, only 32bit

In the first two cases, they each must deliver some components in both
32 and 64 bit form, while some components (within a single case 1 or 2)
are shared. However, NO components are shared BETWEEN any of the three
cases, not even between (1) and (2). (**)

(**) except for a possible sharing of target runtime DLLs with the same
target bitdepth, but that's a small matter, discussed at the end of this
email.

Correct so far?


Correct, I am attempting the first.


Now, SOME of the components that cannot be shared between (1) and (2)
will be IDENTICAL in content, EXCEPT that they get installed into
different locations. This would include, for instance, the runtime DLLs
for, say, 32bit libgcc1 (UNLESS the 'share the target runtimes between
(1) and (2)' approach is taken, but assume for the moment that we do
not; wait until the end of this email for that discussion).


So, we need a naming scheme that accounts for: 1) origin 2) default bit depth of the tool chain 3) bit depth of the target

I think I would drop 'w64' since that's abiguous. "windows 64? is that
tool chain default, so I've got to use -m32 if I want 32bit, or is that
the actual bit depth of the contents, like "import libraries for linking
64bit target executables?" (Aside: I pity the organizational nightmare
that you have to deal with over at mingw64.org, because you've got yet
another set of 32/64 bifurcations: what is the bit depth of the
toolchain EXE's themselves: is ld.exe a 32bit or 64bit executable...at
least cygwin is just...cygwin.)


Well, Windows is special. It has Cygwin, MinGW, Interix, mingw-w64 and etc, all possibly running on the same time, while Linux is just Linux. It takes some nerve to get them to cooperate :)


In fact, for complete clarity, I think I'd use:
-tc64- or -tc32- == toolchain default bitdepth 64 or 32
-m64- or -m32- == bit depth 64 or 32(use "m" as a mnemonic
for which "-mNN" option it corresponds to)

(And, returning to the aside, that 'frees up' the 'w64'/'w32' tag for
use in distinguishing the bit depth of native windows ld.exe/gcc.exe:
w64-gcc4 contains a binary gcc4.exe that runs as a 64bit native windows
application. Not that you guys need or want to use an elaborate package
naming scheme like it appears we need here. You don't have setup.exe to
worry about.)


If something is common to BOTH, then just drop that element. E.g. both the -m32 and -m64 mingw headers are provided by the same package, so that package would have only "mingw64-tc64-". OTOH, if you split the mingw import libs for the defaults-to-64bit toolchain between the "normal" ones (for that toolchain, e.g. 64bit) and the lib32/ ones, then you'd have "mingw64-tc64-m64" and "mingw64-tc64-m32"


Yes, this is a nice plan, clear and concise naming.



The, obviously, the mingw.org compiler -- if cygwin ever got one -- would prefix all of its elements with "mingw-" and no "toolchain bitdepth" or "target bitdepth" because, right now, it supports only 32bits. So everything would be -tc32-m32- which is rather silly.

[1] These two packages [1a] and [1b] contain a DLL that is identical,
but "lives" in a different location buried under the compiler's private
lib area:
mingw64-tc64-m64-libgcc1 : */lib/gcc/x86_64-w64-mingw32/4.6.0/
mingw64-tc32-m64-libgcc1 : */lib/gcc/i686-w64-mingw32/4.6.0/64/

If we had a /opt/mingw64/bin64 and /opt/mingw64/bin32/ "standard" then
maybe the DLLs would go there, and you'd consolidate these two DLL
packages into a single common package that just installs one copy of
this specific DLL into
/opt/mingw64/bin64/
But we don't yet, so for now "buried" -- e.g. two separate identical
copies in separate non-conflicting locations -- is good.

[2] Ditto
mingw64-tc64-m64-libstdc++6 :
*/lib/gcc/x86_64-w64-mingw32/4.6.0/libstdc++-6.dll
mingw64-tc32-m64-libstdc++6 :
*/lib/gcc/i686-w64-mingw32/4.6.0/64/libstdc++-6.dll
vs.
/opt/mingw64/bin64/

[3] Ditto, /opt/mingw64/bin32/
mingw64-tc64-m32-libgcc1 : */lib/gcc/x86_64-w64-mingw32/4.6.0/32/
mingw64-tc32-m32-libgcc1 : */lib/gcc/i686-w64-mingw32/4.6.0/
vs.
/opt/mingw64/bin32/

[4] Ditto
mingw64-tc64-m32-libstdc++6 :
*/lib/gcc/x86_64-w64-mingw32/4.6.0/32/libstdc++-6.dll
mingw64-tc32-m32-libstdc++6 :
*/lib/gcc/i686-w64-mingw32/4.6.0/libstdc++-6.dll
vs.
/opt/mingw64/bin32/


The pthreads naming is a bit confusing (to myself as well). I should
change it to something easier. Ideas welcome.

Well, if the ideas above make sense, then it should be pretty straightforward to extend them to libpthread2.


From the packaging point of view, having them buried in toolchain is a lot easier than trying to move them around at postinstall.


Which leads to...use a different prefix as the only viable choice, I
think (I'd go with /opt/mingw64/). Or asking to use a non-standard
--datarootdir -- and HOPE that the only conflicts we ever discover are
these /usr/share ones.



I'm also not too sure how to proceed, --datarootdir sounds good.

Well, putting everything from mingw64 into /opt/mingw64 would also work. So the exe's (for both the default-to-64bit toolchain and the default-to-32bit toolchain) would go in /opt/mingw64/bin/ as well as the typical "cross-compiler" copies of those exe's in /opt/mingw64/x86_64-w64-mingw32/bin/ and /opt/mingw64/i686-w64-mingw32/bin/ As cygwin apps, these would likely depend, themselves, on DLLs in /usr/bin -- but that just means that /usr/bin/ must be in $PATH when launching these tools. But that's ALWAYS true for cygwin apps.

BUT, you'd have the added freedom of establishing also
/opt/mingw64/bin32/ and
/opt/mingw64/bin64/
specifically for the target DLLs (libgcc1_sjlj.dll, libstdc++6.dll,
libpthread2.dll, etc) for the two separate bitdepths. This would allow
you to consolidate the duplicate DLL packages [1a] == [1b], [2a] ==
[2b], [3a] == [3b], [4a] == [4b], etc, above.



OK, I will use --prefix=/opt/mingw64 and --with-sysroot=/opt/mingw64/{x86_64,i686}-w64-mingw32 to avoid stuff clashing with /usr/share.


Target DLLs will go in /opt/mingw64/bin{32,64}. Users will have a consistent place to look for them.

To consolidate the DLLs, only one copy for 32bit and 64bit will be packaged from one of the compilers and placed in /opt/mingw64/bin{32,64}. For example, packaging 32bit libgcc DLL from the 64bit toolchain, but leaving out the one from the 32bit toolchain since they're similar.

This leads to the next question, how do I add the /opt/mingw64/bin to the user's $PATH after installation?


As for the rest of your reply...


OK, ignore my silly comments about extra symlinks.

LIBGCC RUNTIME DLL
===================================
w64-mingw64-gcc4-libgcc1-4.6.20100619-1.tar.bz2

(First, no need to include 'gcc4' in the library package name).


I'm not sure how to do this with cygport. Any ideas?

In the cygport file, you can specify exactly the name of all sub-packages. Here's the relevant bit from gettext's cygport

PKG_NAMES="${PN} ${PN}-devel libgettextpo0 libasprintf0 libintl8"
PKG_HINTS="${PN} ${PN}-devel libgettextpo0 libasprintf0 libintl8"


No idea why I haven't thought of removing ${PN} for the subpackages.


The contents of the different packages in this case are controlled by
the following files, created by the .cygwin.patch:

gettext-0.17/CYGWIN-PATCHES/gettext-devel.list | 36 ++
gettext-0.17/CYGWIN-PATCHES/gettext.list | 33 ++
gettext-0.17/CYGWIN-PATCHES/libasprintf0.list | 1
gettext-0.17/CYGWIN-PATCHES/libgettextpo0.list | 3
gettext-0.17/CYGWIN-PATCHES/libintl8.list | 1

as long as the *.list name matches the value in PKG_NAMES, it will be
use. A more common approach is (from ncurses cygport):

PKG_NAMES="${PN} lib${PN}${NCURSES_ABI} lib${PN}-devel ${PN}-demo terminfo"
PKG_HINTS="${PN} lib${PN}${NCURSES_ABI} lib${PN}-devel ${PN}-demo terminfo"

PKG_CONTENTS[0]="--exclude=usr/bin/*.dll --exclude=usr/bin/*-config
usr/bin usr/share/doc/${PN} usr/share/doc/Cygwin usr/share/man/man1
usr/share/man/man5 usr/share/man/man7"

PKG_CONTENTS[1]='usr/bin/*.dll'

PKG_CONTENTS[2]='--exclude=usr/lib/ncurses etc usr/bin/*-config usr/lib
usr/include usr/share/man/man3'

PKG_CONTENTS[3]='usr/lib/ncurses'

PKG_CONTENTS[4]='usr/share/terminfo usr/share/tabset'


Where the element of the PKG_CONTENTS array is used when creating the tarball for the corresponding PKG_NAME.


I am currently using this approach, this leads to long and complicated --exclude lines. Does the list file allow \n delimited list of files to include in the tarball release?



Ok, right now, I should make all the dlls package common to 32bit and 64bit defaulting toolchain, but I'm not too sure how to do it with cygport. Any ideas?

Well, I *assume* that you build the default-to-64bit gcc *completely* separately from the default-to-32bit? That is, there's a

something-something-gcc-VER-REL-src.tar.bz2 for the default-to-32bit
toolchain, and a something-something-gcc-VER-REL-src.tar.bz2 for the
default-to-64bit one? And each -src tarball contains a separate
something-something-gcc-VER-REL.cygport?


That was my first attempt, but I concluded its better to have them shared according to your replies. There will still be 2 sources (each for 32bit default and 64bit default), but the runtime target dlls won't get duplicated.


IF you wanted to ship only a single copy of the appropriate runtime DLL,
to be used by both toolchains, FIRST you need to decide where it should
go. If it's to be shared, then it can't go under
*/x86_64-w64-mingw32/* anywhere, nor
*/i686-w64-mingw32/*
because then they wouldn't be "sharing". One toolchain would have to
"know" about the other somehow. Ick.

However, you can't simply stick them in /usr/bin (or /opt/mingw64/bin,
whatever), because the true 64bit version and the 32bit version both
have the same name. Also Ick.


I tried fixing up libtool upstream to use "w64" and "w32" prefix instead of "lib" prefix like the cygwin "cyg" prefix, but maintainers weren't too welcoming of the idea.


That's why I suggest a "top-level" */bin32 and */bin64 specifically for
these target DLLs, so that they can be shared by both the default-to-64
and default-to-32 toolchains. However, cygwin package standards don't
allow additional top-level dirs under / or /usr. So... /bin32 and
/usr/bin32 are out.

Hence, the advantages of you putting everything under /opt/mingw64/.
You'd OWN that directory tree. Do whatever you like under there.

Once you make your decision about WHERE the target DLLs should go,
ensuring you only have one version is easy: both cygports will build and
create all the appropriate packages.

The default-to-64 cygport would create (for example) under the "rules"
above
mingw64-tc64-m64-libgcc1-VER-REL.tar.bz2
containing
/opt/mingw64/bin64/libgcc1_sjlj.dll

while the default-to-32 cygport would create (for example)
mingw64-tc32-m64-libgcc1-VER-REL.tar.bz2
ALSO containing
/opt/mingw64/bin64/libgcc1_sjlj.dll

Ideally, these two DLLs would be identical. Assuming that is the case,
then you rename one of those two tarballs
mingw64-m64-libgcc1-VER-REL.tar.bz2
(e.g. drop the -tc64-). Ditto for
mingw64-m32-libgcc1-VER-REL.tar.bz2
and all the other toolchain runtime libs (incl. libpthread2 ?)

As a final shortcut, you tweak both cygports to create
non-tc*-differentiated tarballs for the target runtime dlls to begin
with, so in the end, both cygports create

mingw64-m64-libgcc1-VER-REL.tar.bz2 <<< gen by -tc64- cygport
mingw64-m32-libgcc1-VER-REL.tar.bz2 <<< gen by -tc64- cygport
mingw64-m64-libgcc1-VER-REL.tar.bz2 <<< gen by -tc32- cygport
mingw64-m32-libgcc1-VER-REL.tar.bz2 <<< gen by -tc32- cygport

rather than

mingw64-tc64-m64-libgcc1-VER-REL.tar.bz2 <<< gen by -tc64- cygport
mingw64-tc64-m32-libgcc1-VER-REL.tar.bz2 <<< gen by -tc64- cygport
mingw64-tc32-m64-libgcc1-VER-REL.tar.bz2 <<< gen by -tc32- cygport
mingw64-tc32-m32-libgcc1-VER-REL.tar.bz2 <<< gen by -tc32- cygport

and you simply upload only one of each pair.


Setup won't fudge over similarly named packages? I thought of doing just: mingw64-m64-libgcc1-VER-REL.tar.bz2 <<< gen by -tc64- cygport mingw64-m32-libgcc1-VER-REL.tar.bz2 <<< gen by -tc64- cygport

and having them totally independent of the actual toolchain that generated them, from setup.exe's package dependency point of view.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]