gcc cross mips output sizes too big

I've been using crosstool-generated gcc-3.3.3 to build my mipsel targets, and I find that the file sizes are humungous. After stripping the objects, I get file sizes that are up to TWICE as large as similarly generated gcc-2.96 code. Do these sizes not look out of line?

bytes gcc-3.3.3 generated files
------- --------------------------------
1884160 /lib/libc-2.3.2.so
1257472 /lib/libstdc++.so.5.0.5
1048576 /usr/local/lib/libfreetype.so.1314.0.0.10
1036288 /bin/busybox
1011712 /bin/bash

bytes gcc-2.96 generated files
------- --------------------------------
1531904 /lib/libc-2.2.4.so
532480 /lib/libstdc++-3-libc6.2-2-2.10.0.so
495616 /usr/local/lib/libfreetype.so.1314.0.0.10
462848 /bin/busybox
847872 /bin/bash

Compiling with -Os helps, but doesn't get close to the file sizes I saw with my older cross-compiler. It seems strange to me that moving to a newer compiler will reduce the number of programs I can ship in my little NAND flash.

Is there some way to generate a more efficient cross-compiler?

Thanks,
Dave

It's been a month and I still haven't found any answers. Anyone here know why object files have grown significantly since 2.95.3?

I've been using crosstool-generated gcc-3.3.3 to build my mipsel targets, and I find that the file sizes are ginormous. After stripping the objects, I get file sizes that are up to TWICE as large as similarly generated gcc-2.96 code. Do these sizes not look out of line?

bytes gcc-3.3.3 generated files
------- --------------------------------
1884160 /lib/libc-2.3.2.so
1257472 /lib/libstdc++.so.5.0.5
1048576 /usr/local/lib/libfreetype.so.1314.0.0.10
1036288 /bin/busybox
1011712 /bin/bash

bytes gcc-2.96 generated files
------- --------------------------------
1531904 /lib/libc-2.2.4.so
532480 /lib/libstdc++-3-libc6.2-2-2.10.0.so
495616 /usr/local/lib/libfreetype.so.1314.0.0.10
462848 /bin/busybox
847872 /bin/bash

Compiling with -Os helps, but doesn't get close to the file sizes I saw with my older cross-compiler. It seems strange to me that moving to a newer compiler will reduce the number of programs I can ship in my little NAND flash.

Is there some way to generate a more efficient cross-compiler?

Thanks,
Dave