Re: Make depmode=cpp work with Microsoft Visual C++ on MSYS.

[Peter Rosin]

Den 2009-01-28 13:41, skrev Ralf Wildenhues:
> * Peter Rosin wrote on Wed, Jan 28, 2009 at 12:57:43PM CET:
> > Considering all of the above, how about the attached patch instead? (I
> > consider it tiny. Even if it touches a few more lines the changes are
> > trivial, much of it is copy-paste...)
>
> Let's not worry about the copyright implications of it yet.

Indeed.

> > Den 2009-01-28 09:12, skrev Ralf Wildenhues:
> > > It would be great to see how the Automake test suite fares with this
> > > compiler (see tests/README for how to enable verbose output), esp.
> > > the depcomp*.test tests.
> > Sorry, but I get this
> >
> > configure: error: Autoconf 2.61a-341 or better is required.
> >
> > and I have plain old 2.61. I'm not too keen on building autoconf and
> > possibly muddy up my install...
>
> You can install Autoconf below a different --prefix.

Yes, but MSYS is MSYS is MSYS is a bitch. I fail to get automake
to build even with autoconf-2.63 (in a new prefix, first in $PATH).

I get this when I type "make" (after ./configure CC="bla bla bla")

 cd .. && perllibdir="/home/peda/automake/git/automake/lib:./lib" 
"/home/peda/automake/git/automake/automake" --libdir=lib --gnu
/bin/sh: /home/peda/automake/git/automake/automake: No such file or directory
make: *** [Makefile.in] Error 1


> Which gets me to the next question: if you haven't run the Automake
> test suite, how have you tested this patch?  Does it rightly set
> depmode to msvisualcpp on Cygwin, and to msvcmsys when using MSYS?
> Does it produce correct dependencies?  In the end, I think it'd really
> be the easiest if you showed testsuite results, including verbose output
> for failures.  You need not take up work to fix all the failures you'll
> probably encounter.  (At least not yet  ;->  )

Since the patch is now limited to the depcomp program, I simple copied
that file into libtool and played with configure.
If I configure libtool like this, on MSYS

../configure --enable-dependency-tracking CC=".../path/to/compile cl" CFLAGS="-MD 
-Zi"

I get this output:

## ---------------------------------------------------- ##
## Configuring libtool (Build:1.3093 2009-01-27) 2.2.7a ##
## ---------------------------------------------------- ##

checking for a BSD-compatible install... /bin/install -c
checking whether build environment is sane... yes
checking for a thread-safe mkdir -p... /bin/mkdir -p
checking for gawk... gawk
checking whether make sets $(MAKE)... yes
checking whether subdir libobjs are useable... yes
checking for gcc... .../path/to/compile cl
checking for C compiler default output file name... conftest.exe
checking whether the C compiler works... yes
checking whether we are cross compiling... no
checking for suffix of executables... .exe
checking for suffix of object files... obj
checking whether we are using the GNU C compiler... no
checking whether .../path/to/compile cl accepts -g... yes
checking for .../path/to/compile cl option to accept ISO C89... none needed
checking for style of include used by make... GNU
checking dependency style of .../path/to/compile cl... msvcmsys
checking whether .../path/to/compile cl and cc understand -c and -o together... 
yes
*snip*


On Cygwin, with the same configure options:

## ---------------------------------------------------- ##
## Configuring libtool (Build:1.3093 2009-01-27) 2.2.7a ##
## ---------------------------------------------------- ##

checking for a BSD-compatible install... /usr/bin/install -c
checking whether build environment is sane... yes
checking for a thread-safe mkdir -p... /usr/bin/mkdir -p
checking for gawk... gawk
checking whether make sets $(MAKE)... yes
checking whether subdir libobjs are useable... yes
checking for gcc... .../path/to/compile cl
checking for C compiler default output file name... conftest.exe
checking whether the C compiler works... yes
checking whether we are cross compiling... no
checking for suffix of executables... .exe
checking for suffix of object files... obj
checking whether we are using the GNU C compiler... no
checking whether .../path/to/compile cl accepts -g... yes
checking for .../path/to/compile cl option to accept ISO C89... none needed
checking for style of include used by make... GNU
checking dependency style of .../path/to/compile cl... msvisualcpp
checking whether .../path/to/compile cl and cc understand -c and -o together... 
yes
*snip*

But I didn't test enough of course. Blast, it worked with depmode=cpp
so I just assumed depmode=msvisualcpp was good enough too. Not so.

depmode=msvisualcpp, when used with libtool clobbers the output file
in the second invocation (with -E). So, I need to copy-paste some code
from the depmode=cpp to filter out libtool (and -o $object) from
that second invocation.

> > But if you are looking for bugs to squash, one obvious problem is
> > that the standard
> > AC_PROG_CC
> > AM_PROG_CC_C_O
> > puts the dependency check before the -c -o check, which forces losers
> > like me to configure CC=".../path/to/compile cl" for the dependency
> > checker to work (as it compiles in subdirs with -c -o). Can some m4
> > magic be applied so that the dependency check is moved down or the
> > -c -o check moved up, that would be much appreciated!
>
> Can probably be done in principle, but would be hacky.  Not sure if that
> is the /first/ bug I'd be trying to fix.
>
> It may mean that you have to
>   export CC=".../path/to/compile cl"
> for some of the testsuite tests though.  Bummer.

Bummer. Exactly, being forced to feed various options to configure
should be avoided. I't bad enough as it is...

> Anyway, thank you for your work on this!

We'll see how far I get...

> > +2009-01-28  Peter Rosin  <address@hidden>
> > +
> > +       Add depmode=msvcmsys for Microsoft Visual C++ on MSYS.
> > +       * lib/depcomp: Add depmode=msvcmsys as a derivative of
> > +       depmode=msvisualcpp. msvcmsys uses echo instead of "cygpath -u"
> > +       to prevent any path translation.
> > +       Also transform any backslashes into forward slashes to make the
> > +       grep in depend.m4 match.
>
> > @@ -85,6 +85,13 @@ if test "$depmode" = dashXmstdout; then
> >     depmode=dashmstdout
> >  fi
> >  
> > +cygpath_u="cygpath -u"
> > +if test "$depmode" = msvcmsys; then
> > +   # This is just like msvisualcpp but w/o cygpath translation.
> > +   cygpath_u=echo
>
> Why not something involving 'CMD //C'?

Because MSVC spits out Win32 paths, no need to involve MSYS with those.

> > +   depmode=msvisualcpp
> > +fi
> > +
> >  case "$depmode" in
> >  gcc3)
> >  ## gcc 3 implements dependency tracking that does exactly what
> > @@ -556,15 +563,22 @@ msvisualcpp)
> >      esac
> >    done
> >    "$@" -E |
> > -  sed -n '/^#line [0-9][0-9]* "\([^"]*\)"/ s::echo "`cygpath -u \\"\1\\"`":p' | sort | uniq 
> > > "$tmpdepfile"
> > +  sed -n '/^#line [0-9][0-9]* "\([^"]*\)"/ s::echo "`$cygpath_u \\"\1\\"`":p' | sort | uniq 
> > > "$tmpdepfile"
>
> sort -u?

One fork less for each dep, thank you very much. But if it's forks
we're chasing, the echo in the echoed `subshell` isn't really
defendable when some form of simple echo would do...

> >    rm -f "$depfile"
> >    echo "$object : \\" > "$depfile"
> > -  . "$tmpdepfile" | sed 's% %\\ %g' | sed -n '/^\(.*\)$/ s:: \1 \\:p' >> 
> > "$depfile"
> > +  . "$tmpdepfile" | sed -e 's%\\%/%g' -e 's% %\\ %g' | sed -n '/^\(.*\)$/ s::        \1 \\:p' 
> > >> "$depfile"
>
> You can combine the two sed scripts.

One fork less for each dep, thank you very much.

> >    echo "  " >> "$depfile"
> > -  . "$tmpdepfile" | sed 's% %\\ %g' | sed -n '/^\(.*\)$/ s::\1\::p' >> 
> > "$depfile"
> > +  . "$tmpdepfile" | sed -e 's%\\%/%g' -e 's% %\\ %g' | sed -n '/^\(.*\)$/ s::\1\::p' >> 
> > "$depfile"
>
> Likewise.

One fork less for each dep, thank you very much.

> >    rm -f "$tmpdepfile"
> >    ;;

Ok, third iteration, taking care of all of the above except the
echo `echo file` thing.

But comparing the resulting depmode=msvisualcpp with depmode=cpp,
maybe msvcmsys fit better as a tweeked cpp?
That would get rid of a whole bunch of forks for each dep.
(msvisualcpp is noticeably slower than cpp)

Cheers,
Peter

2009-01-29  Peter Rosin  <address@hidden>

        Add depmode=msvcmsys for Microsoft Visual C++ on MSYS.
        * lib/depcomp: Add depmode=msvcmsys as a derivative of
        depmode=msvisualcpp. msvcmsys uses echo instead of "cygpath -u"
        to prevent any path translation.
        Also transform any backslashes into forward slashes to make the
        grep in depend.m4 match.
        Fork fewer processes for depmode=msvisualcpp.
        Filter out libtool in the preprocessor invocation (as is done
        in depmode=cpp) for depmode=msvisualcpp.

diff --git a/ChangeLog b/ChangeLog
index 2463967..60084f5 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,15 @@
+2009-01-29  Peter Rosin  <address@hidden>
+
+       Add depmode=msvcmsys for Microsoft Visual C++ on MSYS.
+       * lib/depcomp: Add depmode=msvcmsys as a derivative of
+       depmode=msvisualcpp. msvcmsys uses echo instead of "cygpath -u"
+       to prevent any path translation.
+       Also transform any backslashes into forward slashes to make the
+       grep in depend.m4 match.
+       Fork fewer processes for depmode=msvisualcpp.
+       Filter out libtool in the preprocessor invocation (as is done
+       in depmode=cpp) for depmode=msvisualcpp.
+
 2009-01-23  Eric Blake  <address@hidden>
 
        Use no-arg macros via AC_REQUIRE for consistency.
diff --git a/lib/depcomp b/lib/depcomp
index 0f2bf43..ccb97d2 100755
--- a/lib/depcomp
+++ b/lib/depcomp
@@ -1,7 +1,7 @@
 #! /bin/sh
 # depcomp - compile a program generating dependencies as side-effects
 
-scriptversion=2007-03-29.01
+scriptversion=2009-01-29.01
 
 # Copyright (C) 1999, 2000, 2003, 2004, 2005, 2006, 2007 Free Software
 # Foundation, Inc.
@@ -85,6 +85,13 @@ if test "$depmode" = dashXmstdout; then
    depmode=dashmstdout
 fi
 
+cygpath_u="cygpath -u"
+if test "$depmode" = msvcmsys; then
+   # This is just like msvisualcpp but w/o cygpath translation.
+   cygpath_u=echo
+   depmode=msvisualcpp
+fi
+
 case "$depmode" in
 gcc3)
 ## gcc 3 implements dependency tracking that does exactly what
@@ -536,13 +543,27 @@ cpp)
 
 msvisualcpp)
   # Important note: in order to support this mode, a compiler *must*
-  # always write the preprocessed file to stdout, regardless of -o,
-  # because we must use -o when running libtool.
+  # always write the preprocessed file to stdout.
   "$@" || exit $?
+
+  # Remove the call to Libtool.
+  if test "$libtool" = yes; then
+    while test $1 != '--mode=compile'; do
+      shift
+    done
+    shift
+  fi
+
   IFS=" "
   for arg
   do
     case "$arg" in
+    -o)
+      shift
+      ;;
+    $object)
+      shift
+      ;;
     "-Gm"|"/Gm"|"-Gi"|"/Gi"|"-ZI"|"/ZI")
        set fnord "$@"
        shift
@@ -556,15 +577,22 @@ msvisualcpp)
     esac
   done
   "$@" -E |
-  sed -n '/^#line [0-9][0-9]* "\([^"]*\)"/ s::echo "`cygpath -u \\"\1\\"`":p' 
| sort | uniq > "$tmpdepfile"
+  sed -n '/^#line [0-9][0-9]* "\([^"]*\)"/ s::echo "`$cygpath_u \\"\1\\"`":p' 
| sort -u > "$tmpdepfile"
   rm -f "$depfile"
   echo "$object : \\" > "$depfile"
-  . "$tmpdepfile" | sed 's% %\\ %g' | sed -n '/^\(.*\)$/ s::   \1 \\:p' >> 
"$depfile"
+  . "$tmpdepfile" | sed -n -e 's%\\%/%g' -e 's% %\\ %g' -e '/^\(.*\)$/ s::     
\1 \\:p' >> "$depfile"
   echo "       " >> "$depfile"
-  . "$tmpdepfile" | sed 's% %\\ %g' | sed -n '/^\(.*\)$/ s::\1\::p' >> 
"$depfile"
+  . "$tmpdepfile" | sed -n -e 's%\\%/%g' -e 's% %\\ %g' -e '/^\(.*\)$/ 
s::\1\::p' >> "$depfile"
   rm -f "$tmpdepfile"
   ;;
 
+msvcmsys)
+  # This case exists only to let depend.m4 do its work.  It works by
+  # looking at the text of this script.  This case will never be run,
+  # since it is checked for above.
+  exit 1
+  ;;
+
 none)
   exec "$@"
   ;;