[PATCH] PR/target 30370 powerpc-unknown-eabispe can't build libgcc2

[PATCH] PR/target 30370 powerpc-unknown-eabispe can't build libgcc2

[Rask Ingemann Lambertsen]

   At some point after the GCC 4.1.1 release, libgcc2 stopped building for
powerpc-unknown-eabispe. It is currently broken on the GCC 4.1 and 4.2
branches while GCC 4.3 was fixed recently by the toplevel libgcc changes.

   The cause of the problem is a typo in the condition for using
128-bit long double, meaning that all powerpc targets will using it, where
the intention was that only GNU targets should use 128-bit long double. In
addition to fixing this, at Joseph's suggestion, gnu*spe targets will also
be exempt from building libgcc with -mlong-double-128 because it just isn't
supported yet. This might fix PR/target 30259 (ICE building libgcc2 on
powerpc-unknown-linux-gnuspe) also.

   I built and tested the GCC 4.2 branch with the patch. Testresults are at
<URL:http://gcc.gnu.org/ml/gcc-testresults/2007-01/msg00350.html>. I
investigated two of the execution failures, both of which turned out to be
'illegal instruction', so something isn't right with the simulator.

   (I tried to build Java, Fortran and ObjC too, but their support libraries
libgfortran, boehm-gc and libffi won't configure on targets like this one
which can't link an executable without additional options.)

   I'm building and testing on the GCC 4.1 branch and will report the
results tomorrow, I expect.

   Since this patch will also affect powerpc-unknown-linux-gnuspe, I'd like
someone to test it on that target also. Joseph?

ChangeLog
2007-01-08  Rask Ingemann Lambertsen  <rask@sygehus.dk>
	    Joseph S. Myers <joseph@codesourcery.com>

	PR target/30370
	* config/rs6000/t-ppccomm (TARGET_LIBGCC2_CFLAGS): Fix condition for
	using 128-bit long double when compiling libgcc2.

Index: config/rs6000/t-ppccomm
===================================================================
--- config/rs6000/t-ppccomm	(revision 120546)
+++ config/rs6000/t-ppccomm	(working copy)
@@ -11,7 +11,8 @@
 tramp.S: $(srcdir)/config/rs6000/tramp.asm
 	cat $(srcdir)/config/rs6000/tramp.asm > tramp.S
 
-ifneq (,$findstring gnu,$(target))
+ifneq (,$(findstring gnu,$(target)))
+ifeq (,$(findstring spe,$(target)))
 TARGET_LIBGCC2_CFLAGS += -specs=ldblspecs
 
 SHLIB_MAPFILES += $(srcdir)/config/rs6000/libgcc-ppc-glibc.ver
@@ -22,6 +23,7 @@
 ldblspecs: specs
 	sed -e '/cc1_options/{ n; s/$$/ %{!msoft-float:-mlong-double-128}/; }' < specs > $@
 endif
+endif
 
 # Switch synonyms
 MULTILIB_MATCHES_ENDIAN	= mlittle=mlittle-endian mbig=mbig-endian

-- 
Rask Ingemann Lambertsen

Re: [PATCH] PR/target 30370 powerpc-unknown-eabispe can't build libgcc2

[Joseph S. Myers]

On Mon, 8 Jan 2007, Rask Ingemann Lambertsen wrote:

>    Since this patch will also affect powerpc-unknown-linux-gnuspe, I'd like
> someone to test it on that target also. Joseph?

This patch works fine for powerpc-*-linux-gnuspe to produce a working 
libgcc and the expected test results.

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: [PATCH] PR/target 30370 powerpc-unknown-eabispe can't build libgcc2

[David Edelsohn]

	PR target/30370
	* config/rs6000/t-ppccomm (TARGET_LIBGCC2_CFLAGS): Fix condition
	for using 128-bit long double when compiling libgcc2.

Okay.

Thanks, David

Re: [PATCH] PR/target 30370 powerpc-unknown-eabispe can't build libgcc2

[Rask Ingemann Lambertsen]

On Mon, Jan 08, 2007 at 06:44:00PM -0500, David Edelsohn wrote:
> 	PR target/30370
> 	* config/rs6000/t-ppccomm (TARGET_LIBGCC2_CFLAGS): Fix condition
> 	for using 128-bit long double when compiling libgcc2.
> 
> Okay.

   The testresults for the 4.1 branch are in:
<URL:http://gcc.gnu.org/ml/gcc-testresults/2007-01/msg00390.html>

   Please commit the patch if it is OK. Thanks.

-- 
Rask Ingemann Lambertsen

Re: [PATCH] PR/target 30370 powerpc-unknown-eabispe can't build libgcc2

[Rask Ingemann Lambertsen]

On Mon, Jan 08, 2007 at 08:45:29PM +0100, Rask Ingemann Lambertsen wrote:
>    I built and tested the GCC 4.2 branch with the patch. Testresults are at
> <URL:http://gcc.gnu.org/ml/gcc-testresults/2007-01/msg00350.html>. I
> investigated two of the execution failures, both of which turned out to be
> 'illegal instruction', so something isn't right with the simulator.

   The problem is that psim doesn't support e500 double precision floating
point instructions.  :-(

-- 
Rask Ingemann Lambertsen