We now have two tests: one tests behavior when we run out of space when
appending the node set (in which case the append fails), another one
tests behavior when we run out of space when filtering the node set (in
which case the set still contains redundant data).

Given an unsorted sequence, remove_duplicates would sort it using the
pointer value of attributes/nodes and then remove consecutive
duplicates.

This was problematic because it meant that the result of XPath queries
was dependent on the memory allocation pattern. While it's technically
incorrect to rely on the order, this results in easy to miss bugs.

This is particularly common when XPath queries use union operators -
although we also will call remove_duplicates in other cases.

This change reworks the code to use a hash set instead, using the same
hash function we use for compact storage. To make sure it performs well,
we allocate enough buckets for count * 1.5 (assuming all elements are
unique); since each bucket is a single pointer unlike xpath_node which
is two pointers, we need somewhere between size * 0.75 and size * 1.5
temporary storage.

The resulting filtering is stable - we remove elements that we have seen
before but we don't change the order - and is actually significantly
faster than sorting was.

With a large union operation, before this change it took ~56 ms per 100
query invocations to remove duplicates, and after this change it takes
~20ms.

Fixes #254.

This test is very sensitive to the particular implementation of union
aggregation; for now lets disable this.

We need a more robust way to test union allocation failures.

The behavior on Linux is very different between kernel versions, and it
triggers an unexpected OOM during sanitizer runs because somehow the
size is reported to be LONG_MAX. It's not clear that it helps us cover
any paths we don't cover otherwise - it would be nice to be able to test
failing to load a multi-gigabyte file on a 32-bit system, but we can't
do this easily atm anyway.

We had a few places in test code and library source where we used an
implicit float->double cast; while it should preserve the value exactly,
gcc/clang implement this warning to make sure uses of double are intentional.

This change also adds the warning to Makefile to make sure we don't
regress on this warning.

Fixes #243.

This change modifies the table entries for ctx_special_attr to treat TAB
character as special, which makes the output code escape it.

Before this change, trying to use TAB in an attribute value would output
it verbatim; during subsequent parsing, pugixml - and other compliant
parsers - would apply attribute-value normalization, turning the TAB
into a space and losing the original value.

Using 	 fixes this; if an input document has 	 in an attribute
value, that gets unescaped into \t during parsing and escaped back into
	 during output, which means we can now roundtrip values like this.

Fixes #242.

Intel compiler by default sets flush-to-zero flags which causes our
denorm test to produce 0.0. So make sure that denorms work on FPU before
testing the string output.

Fixes #218.

On some Debian systems it looks like we *can* open the current folder as
a file and read its contents, but parsing the result produces an empty
document. We now handle this case as well.

Fixes #225.

We need bad_alloc that lives in <new>, not <exception>.

pugixml.hpp wasn't including the <exception> header in this build but
test code needed it.

Several tests got the buffer size wrong when sizeof(char_t)>1, and one
test didn't meet the carefully tuned allocation criteria under compact
mode due to the hash table usage and had to be changed a bit.

Apparently at some point OSX behavior when reading /dev/tty switched
from "can't open the file" to "the file can be opened and 0 bytes can be
read from it" which generates a wrong error and doesn't exercise the
code path we care about.

This helps make sure our error handling logic works and is exercised.

This warning is new as of GCC 7 and highlights undefined behavior in the
preprocessor that ASAN detection was relying on.

These tests currently fail for compact mode because of ->reserve()
failing.

We now check that appending a child to a moved document performs no
allocations - this is already the case, but if we neglected to copy the
allocator state this test would fail.

Large test wasn't testing shared parent condition properly - add one
more level of hierarchy so that it works as expected.

Add a test that checks that static buffer pointer was moved correctly
by checking if offset_debug still works.

Make sure we have coverage for empty documents and for large documents
that trigger compact_shared_parent != root for some pages.

Verify that move doesn't allocate and that it preserves structures
required for tree memory management and append_buffer in tact.

These just verify that move ctor/assignment operator work as expected in
simple cases - there are a number of ways in which the internal
structure can be incorrect...

These new tests test that tellg() can fail when being called the second
time, which leads to seekable implementation failing.

These tests simulate various error conditions when reading data from
streams - seeks failing in seekable streams, underflow throwing an
exception causing read to set badbit, etc.

This change also adjusts memory thresholds to cause a reliable out of
memory during construction of a final buffer for non-seekable streams.

This fixes missing coverage in translate_table_generate and
xpath_node_set_raw::append.

Hiding using namespace in common.hpp is somewhat surprising so remove
common.hpp and move using namespace into all .cpp files that need it.

Most tests have `using namespace pugi` which makes explicit
qualifications unnecessary.

These functions were deprecated via comments in 1.5 but never got the
deprecated attribute; now is the time!

Using deprecated functions produces a warning; to silence it, this
change moves the relevant tests to a separate translation unit that has
deprecation disabled.

This requires moving the list of VS versions out of
autotest-appveyor.ps1 and into appveyor.yml.

Add memory allocation failure test for concact with a very large list
and make sure we have every single axis covered with and without a
predicate, with and without a previous step.

Apparently only narrow character streams had out of memory coverage -
fix that and also split this into a separate test.

Cover both char and wchar_t stream loading in a single run instead of
using pugi::char_t.

Cover more failure cases and simplify the streambuf implementation a
bit.

Rename partition to partition3 to resolve conflicts with std::partition.

Add more memory allocation failure tests.

Adjust the buffer size to be right on the edge of the overflow, make
sure we actually output " instead of ".