Document known issues

Address issue with one table being unpinned, and another table not being
unpinned.

Fixed calls break exception handling, and largely aren't needed anymore.
We were still "fixing" calls when it had no fallthrough instruction,
assuming the call is used for something else.  Except that sometimes a
"call _cxa_throw" ends a function/section and legitimately has no
fallthrough.  And this breaks the exception handling.

Testing this fix against the Zipr test suite to verify that it is OK
on a large suite of programs.

Was previously asserting isRegister() or isMemory(), but could also be
isConstant().

* Update 68-sled code to be keystone friendly
* Corrected spelling of getInstruction.  removed extraneous assert.
* Disable should-pin-immediately because it interfers with sleds.
* Remove assert from fix-calls about odd prefixing, just dont fix the call.  More of an issue with FII doing disassembly
* Update switch->code discovery path to only happen if switch table is in text segment
* fixed FII to do what the last commit said -- only disassembly when TABLE is in text.  also patched zipr to handle xbegin gracefully.
* Disable verbose setting
* Create lots of dbs instead of just a few

Update find-in-function method to be more vague about the leas it finds when precise-switch-statement finding is impossible

Issue was related to intermediate set of the compare register.
Leveraged backup_until's stopif argument to stop backup until sooner.

Note: as indicated in code comment, single register/register move
is common and could be used to improve accuracy.

1) FII was finding the wrong cmp for switch tables.
Removed the blind cmp and forced the cmp to use the proper
register or not find a cmp.  Also changed the limit for a table
size when no cmp is found from 2^32 to 255.

2) Fixed an issue with zipr dollop layout where the size
to layout 1 instruction + a trampoline was bigger than laying
out two+ full dollops (because both dollops were short), but
the dollops are disallowed from being coalesced.  The worst-size
detector needed to account for non-coalescable dollops.

Debugging code allows for one to binary-search the unpins.

But a .dynsym, .got, etc. entry can be duplicated in it's own section
or other sections.  Unpinning has to handle unpinning an instruction
and deal with updating all tables properly.  Bug was found in handling
of the unpin for .dynsym in this regard.  This patch fixes that bug.

In particulate, an alternate version of switch tables
was being mis-detected by FII.  In particular, switch
tables were assumed to be size=4, but we found size!=4
in sqlite (in particular, size=8).  And, I know windows
binaries have size=1,2.

Also fixed an issue where lea reg1, [reg2+reg3] assumed reg2/reg3
had values (i.e., reg2 is the base address of a switch table),
when reg3 might have the value.  Fixed the assumption by
tracing both reg2+reg3.

Some minor code cleanups.

The dynsym section is parsed by objdump/grep.  It's terrible and
needs to be implemented in C++ to reduce dependence on external tools
knowing what they are doing.

In this case, the terrible pattern matching with grep caused a symbol
defined in the binary to be missed, because `grep -v UND` was used
to determine whether a symbol "mattered".  1) symbol names might have
UND and get eliminated accidentally, but 2) (what really happened)
was that the symbol was defined in a section without a "link" field
in the dynsym section, thus objdump displayed the section name as UND
and thus the code skipped the symbol.  Don't use objdump to parse ELF
files!

Rida: Parsing of .plt.sec section added for x86/64 ubuntu 20 needed to be
updated to support x86/32 in ubuntu 20.

Fix-calls: Thunk checks aged poorly.  Revived/fixed some of them.

Shared libs need a reloc for a code pointer to live in rodata.
I found in ffmpeg that many spurious pins were occuring because constant
(particularly with relatively small values) data looked like code poitners
(because the code's load point is address 0 in a shared library).

To avoid these spuriosu pins, I extended the code pin algorithm
to check isDLL() (which includes shared objects) and for a reloc
before attempting to pin in the rodata.

A perf. optimization verifies that it _might_ result in a pin
before checking for a reloc, as checking for a reloc can be slow,
while checking for a possible pin location is fast.

Table size min was set to <=4 items, but readelf on ubuntu20
had a switch table with exactly 4 elements.  This caused
FII to miss the table size and subsequently include _many_ extra IBTs.
Ultimately, Zipr failed by trying to 68-sled an entire function due to
FII's failure.