I've now managed to get my version to successfully run against around
7000 files (basically any ELF object in /usr/bin, /usr/lib, /kernel,
/platform/sun4u) with a few caveats (explained below). Success was
defined as producting the same output as the closed source
disassembler, with the following exceptions:
-- Any differences in locations that don't have a symbol+offset
are ignored (assumed to be data). For example:
Orig: 0x100bc88: 20 52 65 63 bn,a,pn %xcc,
+0x9958c <0x10a5214>
New: 0x100bc88: 20 52 65 63 0x20526563
-- A minor discrepency with some of the ASR names:
Orig: setvecint_tl1+0xa8: a9 80 40 00 mov %g1, %set_softint
New: setvecint_tl1+0xa8: a9 80 40 00 mov %g1, %softint_set
Orig: wr_clr_softint+0x4: ab 82 00 00 mov %o0, %clear_softint
New: wr_clr_softint+0x4: ab 82 00 00 mov %o0, %softint_clr
%softint_set and %softint_clr are given in the US2005 manual -- if
this needs to change, this is trivial to do.
-- The stda syntax is assumed to be:
stda %Frd, [%rs1 + %rs2] immasi i.e. stda %d0, [%l1 + %l2] 0x80
The closed library at least under certain circumstances does something like:
stda %Frd, [%rs1] %rs2, immasi -- i.e. stda %f0, [%l1] %l2, 0x80
-- The closed source version has a number of issues with floating
point registers. When it decodes a double register name, it still
prepends %f to the resulting number (i.e. %f0) instead of %d. As a
temporary workaround, I've done something similar (but easy to
change).
However, it also often uses single-precision register names in
instruction arguments when it should be double precision (you can tell
because double register values are encoded slightly differently -- a
value of '1' in a register field meant for doubles actually means %d32
for example). I've found this on fxtos, fstox, fdtoi, fitod,
fmul8x16, fmul8ulx16, fmul8sux16, fcmpgt32, fcmple16, fcmple32,
fcmpgt16, fcmpeq16, and fcmpeq32. In the script I used to compare the
outputs (included in the repo), I correct the closed source output and
then compare the results.
As a reminder, the source can be obtained from
http://hg.genunix.org/libdisasm.hg
In it, I included libdisasm.[ch] and libdisasm_impl.h so that one
didn't need a bunch of stuff to build the library. compare.pl is the
script used to compare the outputs, and files.txt is the list of files
I've been using for validation.
_______________________________________________
tools-discuss mailing list
tools-discuss@opensolaris.org
