[m-dev.] diff: Linux/PPC global reg variable changes

[Fergus Henderson]

On 18-Dec-1997, Tyson Dowd <trd at cs.mu.oz.au> wrote:
> On 18-Dec-1997, Fergus Henderson <fjh at cs.mu.oz.au> wrote:
> > 
> > Add support for gcc global registers on Linux/PPC.
> > 
> > runtime/mercury_regs.h:
> > 	Change `#if defined(_POWER)' to
> > 	`#if defined(_POWER) || defined(__powerpc)',
> > 	since gcc on Linux/PPC does not defined `_POWER'.
> > 
> > runtime/machdeps/rs6000_regs.h:
> > 	Fix a typo that broke the PPC port:
> > 	`restore_transient_reg_from_mem' should have been
> > 	`restore_transient_regs_from_mem'.
> > 			      ^
> > README.Linux-PPC:
> > 	Update to reflect recent developments: with egcs,
> > 	we now support global register variables.
> 
> You forgot the diff again! ;-)

Oops.  I'm getting good at that.

cvs diff runtime/machdeps/rs6000_regs.h runtime/mercury_regs.h README.Linux-PPC
Index: runtime/machdeps/rs6000_regs.h
===================================================================
RCS file: /home/mercury1/repository/mercury/runtime/machdeps/rs6000_regs.h,v
retrieving revision 1.6
diff -u -r1.6 rs6000_regs.h
--- rs6000_regs.h	1997/07/27 15:09:11	1.6
+++ rs6000_regs.h	1997/12/18 02:21:01
@@ -59,7 +59,7 @@
 )
 
 #define save_transient_regs_to_mem(save_area)		((void)0)
-#define restore_transient_reg_from_mem(save_area)	((void)0)
+#define restore_transient_regs_from_mem(save_area)	((void)0)
 
 #define	mr10	fake_reg[10]
 #define	mr11	fake_reg[11]
Index: runtime/mercury_regs.h
===================================================================
RCS file: /home/mercury1/repository/mercury/runtime/mercury_regs.h,v
retrieving revision 1.2
diff -u -r1.2 mercury_regs.h
--- mercury_regs.h	1997/11/23 07:21:33	1.2
+++ mercury_regs.h	1997/12/18 02:20:35
@@ -70,7 +70,7 @@
     #include "machdeps/alpha_regs.h"
   #elif defined(__hppa__)
     #include "machdeps/pa_regs.h"
-  #elif defined(_POWER)
+  #elif defined(_POWER) || defined(__powerpc__)
     #include "machdeps/rs6000_regs.h"
   #else
     #error "USE_GCC_GLOBAL_REGISTERS not yet supported on this machine."
Index: README.Linux-PPC
===================================================================
RCS file: /home/mercury1/repository/mercury/README.Linux-PPC,v
retrieving revision 1.1
diff -u -r1.1 README.Linux-PPC
--- README.Linux-PPC	1997/11/24 06:30:51	1.1
+++ README.Linux-PPC	1997/12/18 03:08:35
@@ -1,133 +1,74 @@
-Mercury should compile out-of-the-box on Linux for PPC.
+Mercury should build "out-of-the-box" on Linux for PPC.
 Thanks to Robert A. Abernathy for his help with this port.
 
-However, there are a few things that could be done to "tune" this port.
-Currently the Linux-PPC port does not yet support
-	- gcc global register variables
-	- shared libraries
-	- stack overflow detection
-
-Appended below is some information on how to add support for these.
-Intrepid hackers, read on ;-)
-
-> Robert A. Abernathy wrote:
-> > 
-> > You'll notice that shared libraries aren't supported on this
-> > machine yet.  The LinuxPPC on the machine is an ELF system
-> > and does support shared libs.  Most software I've tried to
-> > port that uses configure doesn't like to see this yet.
-> > I don't know in your case what the test is yet. 
-> > This certainly isn't crucial as far as I'm concerned.
-> > Just figured I'd let you know.
-> 
-> Thanks.
-> 
-> We don't try to auto-configure shared libraries;
-> it's not reliable, since often failures only exhibit
-> themselves on large programs, not simple test cases.
-> 
-> The code to handle shared libraries might be as simple as just
-> patching configure to tell it to assume that they work. 
-> (The way to do this is shown in the patch appended below.)
-> But more likely it will also require a little bit
-> of ABI-dependent code in runtime/goto.h, similar to the existing
-> stuff there for for __alpha or __sparc.
-> 
-> > checking whether we can use gcc labels and global registers... no
-> > checking whether we can use global registers without gcc labels... no
-> 
-> Actually it would probably be much more important to get gcc global
-> registers working; this would give you a big (50-100%) performance
-> boost.  This is also probably very easy.  You probably just need
-> to modify runtime/regs.h to change the line
-> 
-> 	#elif defined(_POWER)
-> 
-> to something that matches some symbol defined by gcc for Power PC,
-> e.g.
-> 
-> 	#elif defined(__powerpc__) || defined(_POWER)
-> 
-> We've already ported to rs6000-ibm-aix, so this architecture
-> has already been tested (RS/6000 is basically the same as PowerPC).
-> It should just be a matter of finding the right #define to enable it.
-> 
-> (Incidentally, if you're trying any of this stuff out, it is generally
-> quickest to start with a freshly untarred source distribution.)
-> 
-> > checking for sys/siginfo.h... no
-> > checking for ucontext.h... no
-> > checking for sys/ucontext.h... no
-> > checking for asm/sigcontext.h... yes
-> ...
-> > checking for `sigaction' field name... sa_handler
-> > checking for working `sigcontext_struct'... no
-> > checking for `siginfo_t'... no
-> 
-> The only other thing missing is stack overflow checking.
-> That requires a way of determining the fault address in a signal
-> handler, and there's no portable way of doing that.
-> We try a bunch of different ways, as shown above.
-> The method that is used on x86-linux is asm/sigcontext.h +
-> sigcontext_struct.  Unfortunately the sigcontext_struct field name
-> we use in this method is hard-coded as `cr2', so it won't work on
-> PPC-Linux.  There may be some way of getting the same information
-> on PPC-Linux.
-> 
-> Anyway, stack overflow checking is a lot less important than
-> the other two.  Global register variables would probably be
-> very easy and very beneficial, so it's very worthwhile.
-> Shared libraries would be useful, but might be a bit of work.
-> Stack overflow checking would be icing on the cake ;-)
-> 
-> Anyway, thanks for all your help, and I hope you enjoy using
-> Mercury.
-> 
-> Cheers,
-> 	Fergus.
-> 
-> -- 
-> Fergus Henderson <fjh at cs.mu.oz.au>   |  "I have always known that the pursuit
-> WWW: <http://www.cs.mu.oz.au/~fjh>   |  of excellence is a lethal habit"
-> PGP: finger fjh at 128.250.37.3         |     -- the last words of T. S. Garp.
-> 
-> P.S.  In the very latest beta-release, runtime/regs.h has been
-> renamed as runtime/mercury_regs.h, and similary goto.h is now
-> mercury_goto.h.
-> 
-> Index: regs.h
-> ===================================================================
-> RCS file: /home/staff/zs/imp/mercury/runtime/regs.h,v
-> retrieving revision 1.31
-> diff -u -u -r1.31 regs.h
-> --- regs.h	1997/08/23 22:34:17	1.31
-> +++ regs.h	1997/11/22 17:55:03
-> @@ -70,7 +70,7 @@
->      #include "machdeps/alpha_regs.h"
->    #elif defined(__hppa__)
->      #include "machdeps/pa_regs.h"
-> -  #elif defined(_POWER)
-> +  #elif defined(__powerpc__) || defined(_POWER)
->      #include "machdeps/rs6000_regs.h"
->    #else
->      #error "USE_GCC_GLOBAL_REGISTERS not yet supported on this machine."
-> 
-> Index: configure.in
-> ===================================================================
-> RCS file: /home/staff/zs/imp/mercury/configure.in,v
-> retrieving revision 1.117
-> diff -u -u -r1.117 configure.in
-> --- configure.in	1997/11/02 12:41:48	1.117
-> +++ configure.in	1997/11/22 17:37:07
-> @@ -1404,6 +1404,10 @@
->  		AC_MSG_RESULT(yes)
->  		EXT_FOR_SHARED_LIB=so
->  		;;
-> +	powerpc-*-linux|powerpc-*-linux-gnu)
-> +		AC_MSG_RESULT(yes)
-> +		EXT_FOR_SHARED_LIB=so
-> +		;;
->  	i?86-*-freebsd*)
->  		# From Cornelis van der Laan <nils at ims.uni-stuttgart.de>
->  		AC_MSG_RESULT(yes)
-> 
+However, gcc 2.7.2 does not yet support global register variables
+on this configuration, which means that the code generated by the
+Mercury compiler is much less efficient.  Perhaps a future release
+of gcc (2.8?) will incorporate this support.  But for the moment,
+we recommend that you use egcs (see <http://www.cygnus.com/egcs/>),
+which does have support for global register variables.
+
+In addition, there are a few things that could be done to "tune" this
+port.  Currently the Linux-PPC port does not yet support shared
+libraries, or mprotect()-based stack overflow detection.  These
+features are not necessary, but they would of course be nice to have.
+Appended below is some information about what would need to be done to
+add support for these features.  Intrepid hackers, read on ;-)
+
+----------------------------------------------------------------------
+
+> You'll notice that shared libraries aren't supported on this
+> machine yet.  The LinuxPPC on the machine is an ELF system
+> and does support shared libs.
+
+We don't try to auto-configure shared libraries;
+it's not reliable, since often failures only exhibit
+themselves on large programs, not simple test cases.
+
+The code to handle shared libraries might be as simple as just
+patching configure to tell it to assume that they work. 
+(The way to do this is shown in the patch appended below.)
+But more likely it will also require a little bit
+of ABI-dependent code in runtime/mercury_goto.h, similar to the
+existing stuff there for for __alpha or __sparc.
+
+> checking for sys/siginfo.h... no
+> checking for ucontext.h... no
+> checking for sys/ucontext.h... no
+> checking for asm/sigcontext.h... yes
+...
+> checking for `sigaction' field name... sa_handler
+> checking for working `sigcontext_struct'... no
+> checking for `siginfo_t'... no
+
+The only other thing missing is stack overflow checking.
+That requires a way of determining the fault address in a signal
+handler, and there's no portable way of doing that.
+We try a bunch of different ways, as shown above.
+The method that is used on x86-linux is asm/sigcontext.h +
+sigcontext_struct.  Unfortunately the sigcontext_struct field name
+we use in this method is hard-coded as `cr2', so it won't work on
+PPC-Linux.  There may be some way of getting the same information
+on PPC-Linux.  Try having a look at /usr/include/*/sigcontext.h.
+
+Index: configure.in
+===================================================================
+RCS file: /home/staff/zs/imp/mercury/configure.in,v
+retrieving revision 1.117
+diff -u -u -r1.117 configure.in
+--- configure.in	1997/11/02 12:41:48	1.117
++++ configure.in	1997/11/22 17:37:07
+@@ -1404,6 +1404,10 @@
+ 		AC_MSG_RESULT(yes)
+ 		EXT_FOR_SHARED_LIB=so
+ 		;;
++	powerpc-*-linux|powerpc-*-linux-gnu)
++		AC_MSG_RESULT(yes)
++		EXT_FOR_SHARED_LIB=so
++		;;
+ 	i?86-*-freebsd*)
+ 		# From Cornelis van der Laan <nils at ims.uni-stuttgart.de>
+ 		AC_MSG_RESULT(yes)
+
+
+
-- 
Fergus Henderson <fjh at cs.mu.oz.au>   |  "I have always known that the pursuit
WWW: <http://www.cs.mu.oz.au/~fjh>   |  of excellence is a lethal habit"
PGP: finger fjh at 128.250.37.3         |     -- the last words of T. S. Garp.