Hi Julian -
Can you send me the top-level config.log file? The code you're
running into problems with shouldn't be compiled on a G4-based
machine. It exists for Mac OS X 10.3 on the G5, which could emit a
strange mixed mode assembly that supported 64 bit operations in a 32
bit application. Since you're on a 32 bit platform, it definitely
won't do what we want.
Thanks,
Brian
On Mar 12, 2006, at 7:24 AM, Julian Seward wrote:
r9191 (approximately Fri 3 March 06) does not build on ppc32-linux
due to use of register names of the form "r9" etc in
opal/include/opal/sys/powerpc/atomic.h, in opal_atomic_cmpset_64().
You must use either names of the form "%%r9" or just plain "9",
or give gcc the flag -mregnames (iirc). The plain-digit convention
is what the rest of the assembly fragments in that file use.
My inelegant fix is below.
This is with SuSE 10.0 running on a Mac Mini. (glibc 2.3.5, gcc
4.0.2)
Config args were --with-memory-manager=none --disable-mpi-f77
J
Index: opal/include/opal/sys/powerpc/atomic.h
===================================================================
--- opal/include/opal/sys/powerpc/atomic.h (revision 9191)
+++ opal/include/opal/sys/powerpc/atomic.h (working copy)
@@ -259,17 +259,17 @@
* is very similar to the pure 64 bit version.
*/
__asm__ __volatile__ (
- "ld r4,%2 \n\t"
- "ld r5,%3 \n\t"
- "1: ldarx r9, 0, %1 \n\t"
- " cmpd 0, r9, r4 \n\t"
+ "ld %%r4,%2 \n\t"
+ "ld %%r5,%3 \n\t"
+ "1: ldarx %%r9, 0, %1 \n\t"
+ " cmpd 0, %%r9, %%r4 \n\t"
" bne- 2f \n\t"
- " stdcx. r5, 0, %1 \n\t"
+ " stdcx. %%r5, 0, %1 \n\t"
" bne- 1b \n\t"
"2: \n\t"
- "xor r5,r4,r9 \n\t"
- "subfic r9,r5,0 \n\t"
- "adde %0,r9,r5 \n\t"
+ "xor %%r5,%%r4,%%r9 \n\t"
+ "subfic %%r9,%%r5,0 \n\t"
+ "adde %0,%%r9,%%r5 \n\t"
: "=&r" (ret)
: "r"(addr),
"m"(oldval), "m"(newval)
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users
