I've tested u-boot 2009/03 with the last nios2 uClinux kernel and I've found 2 bugs (or things that can be changed to improve performance).
=> CFI "bug"
The first bug is related to the Common Flash Interface handling code: the write/clear/etc.. timeout is calculated assuming a 1kHz timer tick freq. and the expression used to scale the timeout leads to an incorrect timeou in case the tick frequency (integer) is less than 1kHz (e.g. 999Hz due to rounding like in our case with 83.33333Mhz clock). The integer division used in the cfi_flash.c routine can be improved like this:
662,663c662,665 < #if CONFIG_SYS_HZ != 1000 < tout *= CONFIG_SYS_HZ/1000; ---
#if CONFIG_SYS_HZ != 1000 unsigned long long ull; ull = tout*CONFIG_SYS_HZ + CONFIG_SYS_HZ/2; tout = ull/1000; /* Compute: tout *= CONFIG_SYS_HZ/1000; */
The new expression uses a long and an integer round trick to function properly even in case of CONFIG_SYS_HZ = 999.
=> Kernel decompression bug
The u-boot copies (sometime and in my test board!) the kernel image from somewhere to the execution address specified into the image header. After the copy, and some more work, the Nios2 bootm() procedure jumps to the entry point. The problem is that the bootm() procedure does not flush the data cache before jumping to the newly copied code, so the execution of the kernel head.S routine (that invalidates the cache and calls guzip) finds some invalid data into the dram memory (some data cache lines have been lost). The following code diff (just to flush data cache) fixes the problem in lib_nios2/bootm.c:
26a27
#include <asm/cache.h>
34c35,36 < ---
flush_dcache (0,CONFIG_SYS_DCACHE_SIZE ); flush_icache (0,CONFIG_SYS_ICACHE_SIZE);
Renato