> What about to compile the kernel with -mregparm=3? On linux-c-programming
> we made some tests and seems that using regparm(3) in the most of
> functions of bzip2 improve bzip2 of 6%.
<sigh> It's a wonderful idea that really cuts down on the cost of function
calls quite a bit, but alas, neither gcc nor egcs work even vaguely
reliably with the -mregparm option. Circa 2.1.24, I patched up the
assembly portions of the kernel to deal with the register parameter
passing, and it was able to boot fairly far, until it hit assembly that in
no way corresponded to the source... Unfortunately, incorrect code seems
to be occasionally emitted for function calls through function pointers --
which is a heavily used feature of the C language throughout the kernel.
Last time I reported the bug to the egcs people (gcc-bugs never
responded), I got back a reply to the tune of it being too complicated to
fix. Fwiw, I extracted a portion of the NE2000 driver that demonstrates
the bug (that's where my -mregparm'd kernel would Oops during booting).
-ben
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu