While materializing values in SET through the `OperationStack` we used
to use as a type the type of the value associate to the currently used
`BoundedValue`. This was wrong, this patch uses the type on the free
operand on the top of the `OperationsStack` to perform the required
computations.

Landing pads are basically the `catch` blocks in C++ `try`/`catch`
statements. So far we were missing them since they are encoded in a
particular way in a way similar to DWARF debugging information in the
`.eh_frame` and, more specifically, in the `.gcc_except_table` sections
of ELF programs.

This commit parses these sections so that the basic blocks associated to
landing pads are correctly identified. Personality functions are
detected too. A test is also introduced to assess the effectiveness of
our code.

Since we now support disjoint ranges in ORSA, let's test it. This commit
also introduces some license disclaimers in tests assembly files.

`OSRA:handleComparison` was too big and complex, it has been mostly
rewritten.

* Create `OSRA::identifyComparisonOperands` which expands the argument
  of the comparison in a list of possible values (constants or
  OSRs). The new way in which we handle possible operands also fixes a
  bug showing up in case a constant OSR was being compared with an LLVM
  constant, which was checked for being a tautology/contradiction,
  preventing the reaching definitions of the operand to be considered
  too.
* Squeeze more information from uge/ugt. Unsigned comparisons lead to
  two pieces information: the result of the comparison itself, and the
  fact the left-hand side is greather than or equal 0. This secondo
  information is precious, but we were not able to exploit it in the
  case the original comparison is already "greater than" or "greater
  than or equal". In fact, `x - 4 > 10` gives us `x >= 4` and `x > 14`,
  which boils down to `x > 14`.  This commit introduces a change that
  handles this case as `NOT x - 4 <= 10` leading to the negation of `x
  >= 4` and `x < 14` which is way more informative.
* Improve `OSRA::mergePredicate` and `OSRA::applyConstraints`
  interfaces.
* In case a comparison instructions leads to multiple constraints on the
  same `Value`, these constraints are now first or-merged together and
  then propagated. This change improves the quality of the analysis in
  certain situations.

This commit introduces radically changes the implementation of
`BoundedValue`: it no longer represents a single, contiguous range, but
an arbitrary number of ranges.

The bounds are now represented through a
`llvm::SmallVector<std::pair<uint64_t, uint64_t>, 3>`.

* Introduce the `BoundedValue::bounds()` method, which allows to iterate
  over all the ranges that a `BoundedValue` represents. The `bounds`
  method returns a `Bounds` object, which can be used as a range
  composed by `BoundsIterator`.
* All the methods dealing with the `BoundedValue`'s bounds have been
  rewritten.
* New debugging information: "bv-merge". Print all the computations
  performed by `BoundedValue::mergeImpl`.
* Drop dead code: `BoundedValue::setBound` and `isPositive`
* Introduce `BoundedValue::isRightOpen` and drop
  `BoundedValue::isSingleRange`

Some subtle bugs have been fixed in `OSRA::pathSensitiveMerge`:

* Do not alter the current `BoundedValue` if merging a component would
  lead to bottom.
* Do not deactivate a reacher in case an incoherent condition is met.

In our reaching definition analysis we used to consider all the loads
not reached by any store as definitions. However we forgot to actually
register them as such, with the result that two consecutive loads from
the same CSV would end up being two free loads.

In `ConditionNumberingPass` we used to consider as resetting the last
basic block possibly interested in a certain numbered
condition. However, what we really meant, was that its successors were
resetting basic blocks. This commit fixes this issue.

`OSR::constant()` used to forward the result of
`BoundedValue::constant()`, but this is wrong, since the factor and the
base value have to be considered too.

or-merging bottom with anything used to produce a bottom value, which is
wrong. The non-bottom value should be produced instead.

Constraints associated to a memory instruction are propagated to
reached loads. However, if a constraint on the same `Value` is already
present, the new constraint should be and-merged, not or-merged.

* Introduce some additional helpers
* Spread some `const`ness
* Improve documentation
* New debugging information: "osr-bv". Prints every update operation
  performed in `BVMap::update`.
* Remove dead code
* Whitespace fixes
* Some new TODOs
* Fix some typos in comments

This commit drops the original handcrafted implementation of
`BoundedValue` merging, in favor of an implementation based on Boost
intervals. The old implementation was the source of intermittend bugs,
using Boost should be a more reliable solution. Moreover, this commit
enables moves us towards supporting multiple ranges in `BoundedValues`.

In OSRA we used to track `GlobalVariable`s and `AllocaInst` only,
despite the `MemoryAccess` infrastructure supported memory accesses of
the type register + constant. Enabling this, we're able to handle the
following x86-64 snippet found in the `omnetpp` SPEC benchmark:

    cmp    DWORD PTR [rbx+0x8],0x5
    ja     elsewhere
    mov    eax,DWORD PTR [rbx+0x8]

SET, when getting data from OSRA, used to check that the first and last
materialized address were within a certain range, under the assumption
that the last value would be greater that the first one. Turns out that
this is not always the case when in the operation stack we have a load
instruction. This commit improve the way such a situation is handled.

This commit changes the way instruction and basic block are purged when
re-translation is necessary. Specifically, the purge is now performed
through a post-order visit, which should prevent the removal of any
instructions still holding users.

This commit also introduces the `SubGraph` class, which is useful to be
able to navigate portions of a graph (e.g., a `Function`) in post-order
easily.

The main goal of this patch is to reduce the size of
`OSRAPass::runOnFunction()`. To do this we created the `OSRA` class
which handles everything `runOnFunction` was taking care of but without
the ugly lambdas nor being an endless function. Each class of
instruction is now handled by a dedicated function.

This also has the side effect of heavily reducing the amount of clutter
exposed by `OSRAPass` to its users.

Fix of another bug showing up only with LLVM in debug mode: splitting a
malformed basic block is not allowed, and we had a function call after a
`ret` instruction.

This commit should fix some bugs due to the fact that when we're
splitting a basic block we don't retranslate the basic block at the
split point but preserve the existing code. This lead to problems, in
particular in x86-64 where certain QEMU local variables were not
available. This change should fix it.

Basically, every time we split a basic block in
`JumpTargetManager::registerJT` we note down that the new basic block
must be purged, and in `JumpTargetManager::harvest` we perform the
purge. `harvest` has been chosen since it's a particularly quiet moment,
i.e., there should be no pending references/iterator to code we have to
delete.

This commit fixes a bug that appears only with debug builds of LLVM: in
RDA we were erasing a temporary common predecessor basic block before
removing the references to it in a `switch` statement.

If we need this again, we can do it in revamb-dump.

When performing a comparison, we try to attach its signedness to the
base OSR it's working on. However, this is not always
effective. Typically, even after this, the OSR remains with an unknown
signedness due to the fact that we don't have information about its
bounded value from all the predecessors, and therefore it goes to top.

`support.c` used to have a basic tracing system which would print on `stderr`
all the program counters. This commit reorganizes this system so that it has a
buffer, whose size can be specified through the `REVAMB_TRACE_BUFFER_SIZE`
environment variable, and it's possible to specify the output file
`REVAMB_TRACE_PATH`. The new tracing system also supports flushing the buffers
in case of clean exit of crash.

We also cleaned up `support.c`.

`support.c` used to be compiled using the system compiler and then
linked to the module generated by `revamb` as a separate translation
unit. This commit introduces a change that lets `clang` compile
`support.c`. This will allow us to make the CSV static, which should
enable more aggressive optimizations.

* Change the signature of the `root` function so that it accepts an
  argument: the initial value of the stack pointer, which the main is
  supposed to set up. QEMU now provides us with the offset of the stack
  pointer.
* Let the build system compile `support.c` for each supported
  architecture, both in normal and "tracing" mode.
* Remove the `--tracing` option, this is now handled by `support.c`, in
  particular depending on which version of `support.c` you link, you can
  have tracing enabled or not.
* In `support.c` drop global variables representing the stack pointer,
  we no longer need them.
* In `support.c` fix some warnings while handling the stack on 32-bit
  architectures.
* Extende the `translate` script to handle the new way we link the final
  binary and the tracing mechanism.

We used to jump directly to the program entry point (typically
`_start`), without initializing the program counter. We now initialize
the program counter and the let the program start with the dispatcher,
since it seems a safer solution.

This commit handles the situation where we have a function call with
more than one successor. This might be the case if there's an indirect
call but we're able to enumerate the possible targets statically.

This commit simply avoids an assertion, in the future we will register
all the possible destinations in the `function_call` marker.

We used to have a normalized address space formed only functions,
skipping all the holes. This was employed to identify "skipping
jumps". However, this method was ineffective, and we changed the
definition of skipping jump as a jump skipping over CFEPs that are
highly likely to be actual functions.

This commit removes the leftovers of the normalized address space
computation, which was also the cause of a bug in function detection.

This commit fixes a bug triggered by specifying the `--entry` switch:
the entry point would not be registered as a jump target, instead the
code would try to get the basic block associated to that address,
resulting in an assertion.

The manually specified entry point is now registered as a jump target
coming from global data.

Attach a `func.entry` metadata node to the entry basic block of a
function and rename `func` to `func.member.of`.

Introduce an option to prevent `revamb` from linking in all the QEMU
helpers. This is useful if the output doesn't need to be compiled, but
just analyzed.