On Thu, 2009-10-22 at 23:17 +0100, Richard Sandiford wrote:
> David Daney <firstname.lastname@example.org> writes:
> > and here:
> > http://www.linux-mips.org/archives/linux-mips/2009-10/msg00290.html
> I'm not sure that the "search for a save of RA" thing is really a good idea.
> The last version of that seemed to be "assume that any register stores
> will be in a block that immediately precedes the move into RA", but even
> if that's true now, it might not be in future. And as Wu Zhangjin says,
> it doesn't cope with long calls, where the target address is loaded
> into a temporary register before the call.
-mlong-calls works with the current implementation of static function
tracer and function graph tracer for MIPS, just tried them, and module
support is supported by default with -mlong-calls, let's have a look at
the dumped code with -mlong-calls, only a few difference.
ffffffff80241520: 67bdff40 daddiu sp,sp,-192
ffffffff80241524: ffbe00b0 sd s8,176(sp)
ffffffff80241528: 03a0f02d move s8,sp
ffffffff8024152c: ffbf00b8 sd ra,184(sp)
ffffffff80241530: ffb700a8 sd s7,168(sp)
ffffffff80241534: ffb600a0 sd s6,160(sp)
ffffffff80241538: ffb50098 sd s5,152(sp)
ffffffff8024153c: ffb40090 sd s4,144(sp)
ffffffff80241540: ffb30088 sd s3,136(sp)
ffffffff80241544: ffb20080 sd s2,128(sp)
ffffffff80241548: ffb10078 sd s1,120(sp)
ffffffff8024154c: ffb00070 sd s0,112(sp)
ffffffff80241550: 3c038021 lui v1,0x8021
ffffffff80241554: 64631750 daddiu v1,v1,5968
ffffffff80241558: 03e0082d move at,ra
ffffffff8024155c: 0060f809 jalr v1
so, the only left job is making dynamic function tracer work with
-mlong-calls, I think it's not that complex, after using -mlong-calls,
we need to search "move at,ra; jalr v1" instead of "jal _mcount", and
also, some relative job need to do. will try to make it work next week.
> FWIW, I'd certainly be happy to make GCC pass an additional parameter
> to _mcount. The parameter could give the address of the return slot,
> or null for leaf functions. In almost all cases[*], there would be
> no overhead, since the move would go in the delay slot of the call.
> [*] Meaning when the frame is <=32k. ;) I'm guessing you never
> get anywhere near that, and if you did, the scan thing wouldn't
> work anyway.
> The new behaviour could be controlled by a command-line option,
> which would also give linux a cheap way of checking whether the
> feature is available.
I like your suggestion, and I have tried to make gcc do something like
this before your reply.
move at, ra
lui ra, 184 --> This is new
so, in a non-leaf function, the at register stored the stack offset of
the return address(range from 0 to PT_SIZE). in a leaf function, it is
the return address itself(at least bigger than PT_SIZE). we are easier
to distinguish them. and only a few lines of source code need to be
added for gcc.