I checkout the gcc-3.4 branch today and try to build. xgcc generated ICE while it try to build libobjc. I tried to change/remove the compile options. And I found the ICE was caused by '-march=athlon-xp'. If I change it to '-march=athlon', No ICE. Here is ICE msg: ========================================================================= /home2/namsh/rpm/build/gcc/obj-athlon-redhat-linux/gcc/xgcc -B/home2/namsh/rpm/build/gcc/obj-athlon-redhat-linux/gcc/ -B/usr/athlon-redhat-linux/bin/ -B/usr/athlon-redhat-linux/lib/ -isystem /usr/athlon-redhat-linux/include -isystem /usr/athlon-redhat-linux/sys-include -c -I. -I/home2/namsh/rpm/build/gcc/libobjc -O2 -W -Wall -Wwrite-strings -Wstrict-prototypes -DHAVE_GTHR_DEFAULT -DIN_GCC -DIN_TARGET_LIBS -I/home2/namsh/rpm/build/gcc/libobjc/objc -I/home2/namsh/rpm/build/gcc/libobjc/../gcc -I/home2/namsh/rpm/build/gcc/libobjc/../gcc/config -I../../gcc -I/home2/namsh/rpm/build/gcc/libobjc/../include /home2/namsh/rpm/build/gcc/libobjc/sendmsg.c -fPIC -DPIC -march=athlon-xp -o .libs/sendmsg.o /home2/namsh/rpm/build/gcc/libobjc/sendmsg.c:42:1: warning: "rtx" redefined In file included from /home2/namsh/rpm/build/gcc/libobjc/sendmsg.c:31: /home2/namsh/rpm/build/gcc/libobjc/../gcc/coretypes.h:58:1: warning: this is the location of the previous definition /home2/namsh/rpm/build/gcc/libobjc/sendmsg.c: In function `objc_msg_sendv': /home2/namsh/rpm/build/gcc/libobjc/sendmsg.c:269: error: insn does not satisfy its constraints: (insn 104 73 105 0 (set (mem:DI (plus:SI (reg/f:SI 6 bp) (const_int -176 [0xffffff50])) [0 S8 A8]) (reg:DI 21 xmm0)) 59 {*movdi_2} (nil) (nil)) /home2/namsh/rpm/build/gcc/libobjc/sendmsg.c:269: internal compiler error: in reload_cse_simplify_operands, at postreload.c:391 Please submit a full bug report, with preprocessed source if appropriate. See <URL:http://gcc.gnu.org/bugs.html> for instructions.
Could you please retry with a new checkout? Richard's patch for PR19102 probably fixed this: http://gcc.gnu.org/ml/gcc-patches/2004-12/msg01777.html
Note I would not doubt that the reduced testcase is: int *f(void) { return __builtin_apply (0,0,0); }
I confirmed that the "Richard's patch for PR19102" fixed this problem. Thanks.
So closing as fixed.