I was looking at conhost/OpenConsole and noticed it was being pretty
inefficient with allocations due to some usages of std::deque and
std::vector that didn't need to be done quite that way.
So this uses std::vector for the TextBuffer's storage of ROW objects,
which allows one allocation to contiguously reserve space for all the
ROWs - on Desktop this is 9001 ROW objects which means it saves 9000
allocations that the std::deque would have done. Plus it has the
benefit of increasing locality of the ROW objects since deque is going
to chase pointers more often with its data structure.
Then, within each ROW there are CharRow and ATTR_ROW objects that use
std::vector today. This changes them to use Boost's small_vector, which
is a variation of vector that allows for the so-called "small string
optimization." Since we know the typical size of these vectors, we can
pre-reserve the right number of elements directly in the
CharRow/ATTR_ROW instances, avoiding any heap allocations at all for
constructing these objects.
There are a ton of variations on this "small_vector" concept out there
in the world - this one in Boost, LLVM has one called SmallVector,
Electronic Arts' STL has a small_vector, Facebook's folly library has
one...there are a silly number of these out there. But Boost seems like
it's by far the easiest to consume in terms of integration into this
repo, the CI/CD pipeline, licensing, and stuff like that, so I went with
the boost version.
In terms of numbers, I measured the startup path of OpenConsole.exe on
my dev box for Release x64 configuration. My box is an i7-6700k @ 4
Ghz, with 32 GB RAM, not that I think machine config matters much here:
| | Allocation count | Allocated bytes | CPU usage (ms) |
| ------ | ------------------- | ------------------ | -------------- |
| Before | 29,461 | 4,984,640 | 103 |
| After | 2,459 (-91%) | 4,853,931 (-2.6%) | 96 (-7%) |
Along the way, I also fixed a dynamic initializer I happened to spot in
the registry code, and updated some docs.
## Validation Steps Performed
- Ran "runut", "runft" and "runuia" locally and confirmed results are
the same as the main branch
- Profiled the before/after numbers in the Visual Studio profiler, for
the numbers shown in the table
Co-authored-by: Austin Lamb <austinl@microsoft.com>
This is an attempt to simplify the SGR (Select Graphic Rendition)
implementation in conhost, to cut down on the number of methods required
in the `ConGetSet` interface, and pave the way for future improvements
and bug fixes. It already fixes one bug that prevented SGR 0 from being
correctly applied when combined with meta attributes.
* This a first step towards fixing the conpty narrowing bugs in issue
#2661
* I'm hoping the simplification of `ConGetSet` will also help with
#3849.
* Some of the `TextAttribute` refactoring in this PR overlaps with
similar work in PR #1978.
## Detailed Description of the Pull Request / Additional comments
The main point of this PR was to simplify the
`AdaptDispatch::SetGraphicsRendition` implementation. So instead of
having it call a half a dozen methods in the `ConGetSet` API, depending
on what kinds of attributes needed to be set, there is now just one call
to get current attributes, and another call to set the new value. All
adjustments to the attributes are made in the `AdaptDispatch` class, in
a simple switch statement.
To help with this refactoring, I also made some change to the
`TextAttribute` class to make it easier to work with. This included
adding a set of methods for setting (and getting) the individual
attribute flags, instead of having the calling code being exposed to the
internal attribute structures and messing with bit manipulation. I've
tried to get rid of any methods that were directly setting legacy, meta,
and extended attributes.
Other than the fix to the `SGR 0` bug, the `AdaptDispatch` refactoring
mostly follows the behaviour of the original code. In particular, it
still maps the `SGR 38/48` indexed colors to RGB instead of retaining
the index, which is what we ultimately need it to do. Fixing that will
first require the color tables to be unified (issue #1223), which I'm
hoping to address in a followup PR.
But for now, mapping the indexed colors to RGB values required adding an
an additional `ConGetSet` API to lookup the color table entries. In the
future that won't be necessary, but the API will still be useful for
other color reporting operations that we may want to support. I've made
this API, and the existing setter, standardise on index values being in
the "Xterm" order, since that'll be essential for unifying the code with
the terminal adapter one day.
I should also point out one minor change to the `SGR 38/48` behavior,
which is that out-of-range RGB colors are now ignored rather than being
clamped, since that matches the way Xterm works.
## Validation Steps Performed
This refactoring has obviously required corresponding changes to the
unit tests, but most were just minor updates to use the new
`TextAttribute` methods without any real change in behavior. However,
the adapter tests did require significant changes to accommodate the new
`ConGetSet` API. The basic structure of the tests remain the same, but
the simpler API has meant fewer values needed to be checked in each test
case. I think they are all still covering the areas there were intended
to, though, and they are all still passing.
Other than getting the unit tests to work, I've also done a bunch of
manual testing of my own. I've made sure the color tests in Vttest all
still work as well as they used to. And I've confirmed that the test
case from issue #5341 is now working correctly.
Closes#5341
