Jean-Christophe PLAGNIOL-VILLARD plagnioj@jcrosoft.com wrote:
On 17:40 Tue 03 Mar , Tom Evans wrote:
BUG IN ITEST
itest crashes ARM chips. I can "iread.b *82000000 == 0", but "iread.b *82000001 == 0" throws an alignment exception that kills the box. The code in itest.c is:
static long evalexp(char *s, int w) { long l, *p;
/* if the parameter starts with a * then assume is a pointer to the value we want */ if (s[0] == '*') { p = (long *)simple_strtoul(&s[1], NULL, 16); l = *p;
here you are support to use the good accessor readb/readw/readl
Thank you for your response.
readl() is defined in asm-arm/io.h as "__arch_getl(a)" which is defined as "(*(volatile unsigned int *)(a))" which doesn't handle misalignment and will still crash.
I think those functions are only meant for I/O reads.
The code then "selects the byte" with the following:
return (l & ((1 << (w * 8)) - 1));
That selects the LOWER 8/16/32 bits of the 32-bit read.
Which works on a 486, ARM variants with unaligned transfers, but NOT on a PPC or any other big-endian CPUs (about half of them supported by U-Boot).
On a big-endian unaligned supporting CPU I'd expect a byte read of address "0" to read "3" and a read of "1" to read "4" (in the next word!). Ditto for itest.w.
Whoever owns/maintains this code, and/or can test it on a big-endian target should fix these, taking hints from the cmd_mem.c code which does all of this in a portable manner. I don't have a big-endian target so I can't test the big-endian fix.
the arm does not allow you to do an non aligned access
I know that. The current itest code doesn't. The "indirect pointer feature" mustn't be used much if I'm the first one to find these problems.