Set keyword flags on all events so those sharing a provider with
telemetry do not fire unless tracing is enabled
## PR Checklist
* [x] Closes#10093
* [x] I work here
* [x] Tests passed
* [x] Documentation added in `til.h` about how keywords work and at the
only other site of keywords we define in the Host project tracing
files.
## Detailed Description of the Pull Request / Additional comments
I initially thought that we would need to split providers here to
accomplish this... but @DHowett helped me realize that might be a lot of
additional metadata and bloat binary size. So with help from a friend
from fundamentals, I realized that we could use Keywords to
differentiate here. We can no longer define 0 keywords as that
represents an any/all scenario. Every `TraceLoggingWrite` event now
needs a keyword. When our events have a keyword, they're not included in
any trace. Additionally, when we have an explicit keyword to check that
is different from the ones used for the telemetry pipeline, we can
ensure that we only do "hard work" to generate debug trace data when an
"ALL" type listener like TraceView or Windows Performance Recorder with
our profiles is listening to these providers for ALL keyworded events.
## Validation Steps Performed
- [x] - Built with full release build config to confirm performance is
worse than dev builds BECAUSE of the telemetry event collector camping
our provider and triggering full trace event generation on shared
providers.
- [x] - Built with full release build config to enable statistics
collection and validated trace event collection is excluded and trace
event short-circuits work with this change.
- [x] - Checked that TraceView still sees both telemetry and tracing
events
- [x] - Checked that WPR with our .wprp profile sees both telemetry and
tracing events
## References
* See also #8617
## PR Checklist
* [x] Supports #3075
* [x] I work here.
* [x] Manual test.
## Detailed Description of the Pull Request / Additional comments
### Window Title Generation
Every time the renderer checks the title, it's doing two bad things that
I've fixed:
1. It's assembling the prefix to the full title doing a concatenation.
No one ever gets just the prefix ever after it is set besides the
concat. So instead of storing prefix and the title, I store the
assembled prefix + title and the bare title.
2. A copy must be made because it was returning `std::wstring` instead
of `std::wstring&`. Now it returns the ref.
### Dirty Area Return
Every time the renderer checks the dirty area, which is sometimes
multiple times per pass (regular text printing, again for selection,
etc.), a vector is created off the heap to return the rectangles. The
consumers only ever iterate this data. Now we return a span over a
rectangle or rectangles that the engine must store itself.
1. For some renderers, it's always a constant 1 element. They update
that 1 element when dirty is queried and return it in the span with a
span size of 1.
2. For other renderers with more complex behavior, they're already
holding a cached vector of rectangles. Now it's effectively giving
out the ref to those in the span for iteration.
### Bitmap Runs
The `til::bitmap` used a `std::optional<std::vector<til::rectangle>>`
inside itself to cache its runs and would clear the optional when the
runs became invalidated. Unfortunately doing `.reset()` to clear the
optional will destroy the underlying vector and have it release its
memory. We know it's about to get reallocated again, so we're just going
to make it a `std::pmr::vector` and give it a memory pool.
The alternative solution here was to use a `bool` and
`std::vector<til::rectangle>` and just flag when the vector was invalid,
but that was honestly more code changes and I love excuses to try out
PMR now.
Also, instead of returning the ref to the vector... I'm just returning a
span now. Everyone just iterates it anyway, may as well not share the
implementation detail.
### UTF-8 conversions
When testing with Terminal and looking at the `conhost.exe`'s PTY
renderer, it spends a TON of allocation time on converting all the
UTF-16 stuff inside to UTF-8 before it sends it out the PTY. This was
because `ConvertToA` was allocating a string inside itself and returning
it just to have it freed after printing and looping back around again...
as a PTY does.
The change here is to use `til::u16u8` that accepts a buffer out
parameter so the caller can just hold onto it.
## Validation Steps Performed
- [x] `big.txt` in conhost.exe (GDI renderer)
- [x] `big.txt` in Terminal (DX, PTY renderer)
- [x] Ensure WDDM and BGFX build under Razzle with this change.
## PR Checklist
* [x] Closes perf itch.
* [x] I work here.
* [x] Manual perf test.
* [x] Documentation irrelevant.
* [x] Schema irrelevant.
* [x] Am core contributor.
## Detailed Description of the Pull Request / Additional comments
Passes the bitmap by ref into the tracing function instead of making a copy on the way in. It's only read anyway for tracing (if enabled) so the copy was a pointless oversight.
## Validation Steps Performed
- Observed WPR trace before and after with `time cat big.txt` in WSL.
Now that the Terminal is doing a better job of actually marking which
lines were and were not wrapped, we're not always copying lines as
"wrapped" when they should be. We're more correctly marking lines as not
wrapped, when previously we'd leave them marked wrapped.
The real problem is here in the `ScrollFrame` method - we'd manually
newline the cursor to make the terminal's viewport shift down to a new
line. If we had to scroll the viewport for a _wrapped_ line, this would
cause the Terminal to mark that line as broken, because conpty would
emit an extra `\n` that didn't actually exist.
This more correctly implements `ScrollFrame`. Now, well move where we
"thought" the cursor was, so when we get to the next `PaintBufferLine`,
if the cursor needs to newline for the next line, it'll newline, but if
we're in the middle of a wrapped line, we'll just keep printing the
wrapped line.
A couple follow up bugs were found to be caused by the same bad logic.
See #5039 and #5161 for more details on the investigations there.
## References
* #4741 RwR, which probably made this worse
* #5122, which I branched off of
* #1245, #357 - a pair of other conpty wrapped lines bugs
* #5228 - A followup issue for this PR
## PR Checklist
* [x] Closes#5113
* [x] Closes#5180 (by fixing DECRST 25)
* [x] Closes#5039
* [x] Closes#5161 (by ensuring we only `removeSpaces` on the actual
bottom line)
* [x] I work here
* [x] Tests added/passed
* [n/a] Requires documentation to be updated
## Validation Steps Performed
* Checked the cases from #1245, #357 to validate that they still work
* Added more and more tests for these scenarios, and then I added MORE
tests
* The entire team played with this in selfhost builds
Correct scrolling invalidation region for tmux in pty w/ bitmap
Add tracing for circling and scrolling operations. Fix improper
invalidation within AdjustCursorPosition routine in the subsection about
scrolling down at the bottom with a set of margins enabled.
## References
- Introduced with #5024
## Detailed Description of the Pull Request / Additional comments
- This occurs when there is a scroll region restriction applied and a
newline operation is performed to attempt to spin the contents of just
the scroll region. This is a frequent behavior of tmux.
- Right now, the Terminal doesn't support any sort of "scroll content"
operation, so what happens here generally speaking is that the PTY in
the ConHost will repaint everything when this happens.
- The PTY when doing `AdjustCursorPosition` with a scroll region
restriction would do the following things:
1. Slide literally everything in the direction it needed to go to take
advantage of rotating the circular buffer. (This would force a
repaint in PTY as the PTY always forces repaint when the buffer
circles.)
2. Copy the lines that weren't supposed to move back to where they were
supposed to go.
3. Backfill the "revealed" region that encompasses what was supposed to
be the newline.
- The invalidations for the three operations above were:
1. Invalidate the number of rows of the delta at the top of the buffer
(this part was wrong)
2. Invalidate the lines that got copied back into position (probably
unnecessary, but OK)
3. Invalidate the revealed/filled-with-spaces line (this is good).
- When we were using a simple single rectangle for invalidation, the
union of the top row of the buffer from 1 and the bottom row of the
buffer from 2 (and 3 was irrelevant as it was already unioned it)
resulted in repainting the entire buffer and all was good.
- When we switched to a bitmap, it dutifully only repainted the top line
and the bottom two lines as the middle ones weren't a consequence of
intersect.
- The logic was wrong. We shouldn't be invalidating rows-from-the-top
for the amount of the delta. The 1 part should be invalidating
everything BUT the lines that were invalidated in parts 2 and 3.
(Arguably part 2 shouldn't be happening at all, but I'm not optimizing
for that right now.)
- So this solves it by restoring an entire screen repaint for this sort
of slide data operation by giving the correct number of invalidated
lines to the bitmap.
## Validation Steps Performed
- Manual validation with the steps described in #5104
- Automatic test `ConptyRoundtripTests::ScrollWithMargins`.
Closes#5104
## Summary of the Pull Request
Moves the ConPTY drawing mechanism (`VtRenderer`) to use the fine-grained `til::bitmap` individual-dirty-bit tracking mechanism instead of coarse-grained rectangle unions to improve drawing performance by dramatically reducing the total area redrawn.
## PR Checklist
* [x] Part of #778 and #1064
* [x] I work here
* [x] Tests added and updated.
* [x] I'm a core contributor
## Detailed Description of the Pull Request / Additional comments
- Converted `GetDirtyArea()` interface from `IRenderEngine` to use a vector of `til::rectangle` instead of the `SMALL_RECT` to banhammer inclusive rectangles.
- `VtEngine` now holds and operates on the `til::bitmap` for invalidation regions. All invalidation operation functions that used to be embedded inside `VtEngine` are deleted in favor of using the ones in `til::bitmap`.
- Updated `VtEngine` tracing to use new `til::bitmap` on trace and the new `to_string()` methods detailed below.
- Comparison operators for `til::bitmap` and complementary tests.
- Fixed an issue where the dirty rectangle shortcut in `til::bitmap` was set to 0,0,0,0 by default which means that `|=` on it with each `set()` operation was stretching the rectangle from 0,0. Now it's a `std::optional` so it has no value after just being cleared and will build from whatever the first invalidated rectangle is. Complementary tests added.
- Optional run caching for `til::bitmap` in the `runs()` method since both VT and DX renderers will likely want to generate the set of runs at the beginning of a frame and refer to them over and over through that frame. Saves the iteration and creation and caches inside `til::bitmap` where the chance of invalidation of the underlying data is known best. It is still possible to iterate manually with `begin()` and `end()` from the outside without caching, if desired. Complementary tests added.
- WEX templates added for `til::bitmap` and used in tests.
- `translate()` method for `til::bitmap` which will slide the dirty points in the direction specified by a `til::point` and optionally back-fill the uncovered area as dirty. Complementary tests added.
- Moves all string generation for `til` types `size`, `point`, `rectangle`, and `some` into a `to_string` method on each object such that it can be used in both ETW tracing scenarios AND in the TAEF templates uniformly. Adds a similar method for `bitmap`.
- Add tagging to `_bitmap_const_iterator` such that it appears as a valid **Input Iterator** to STL collections and can be used in a `std::vector` constructor as a range. Adds and cleans up operators on this iterator to match the theoretical requirements for an **Input Iterator**. Complementary tests added.
- Add loose operators to `til` which will allow some basic math operations (+, -, *, /) between `til::size` and `til::point` and vice versa. Complementary tests added. Complementary tests added.
- Adds operators to `til::rectangle` to allow scaling with basic math operations (+, -, *) versus `til::size` and translation with basic math operations (+, -) against `til::point`. Complementary tests added.
- In-place variants of some operations added to assorted `til` objects. Complementary tests added.
- Update VT tests to compare invalidation against the new map structure instead of raw rectangles where possible.
## Validation Steps Performed
- Wrote additional til Unit Tests for all additional operators and functions added to the project to support this operation
- Updated the existing VT renderer tests
- Ran perf check
This PR adds support for "Resize with Reflow" to the Terminal. In
conhost, `ResizeWithReflow` is the function that's responsible for
reflowing wrapped lines of text as the buffer gets resized. Now that
#4415 has merged, we can also implement this in the Terminal. Now, when
the Terminal is resized, it will reflow the lines of it's buffer in the
same way that conhost does. This means, the terminal will no longer chop
off the ends of lines as the buffer is too small to represent them.
As a happy side effect of this PR, it also fixed#3490. This was a bug
that plagued me during the investigation into this functionality. The
original #3490 PR, #4354, tried to fix this bug with some heavy conpty
changes. Turns out, that only made things worse, and far more
complicated. When I really got to thinking about it, I realized "conhost
can handle this right, why can't the Terminal?". Turns out, by adding
resize with reflow, I was also able to fix this at the same time.
Conhost does a little bit of math after reflowing to attempt to keep the
viewport in the same relative place after a reflow. By re-using that
logic in the Terminal, I was able to fix#3490.
I also included that big ole test from #3490, because everyone likes
adding 60 test cases in a PR.
## References
* #4200 - this scenario
* #405/#4415 - conpty emits wrapped lines, which was needed for this PR
* #4403 - delayed EOL wrapping via conpty, which was also needed for
this
* #4354 - we don't speak of this PR anymore
## PR Checklist
* [x] Closes#1465
* [x] Closes#3490
* [x] Closes#4771
* [x] Tests added/passed
## EDIT: Changes to this PR on 5 March 2020
I learned more since my original version of this PR. I wrote that in
January, and despite my notes that say it was totally working, it
_really_ wasn't.
Part of the hard problem, as mentioned in #3490, is that the Terminal
might request a resize to (W, H-1), and while conpty is preparing that
frame, or before the terminal has received that frame, the Terminal
resizes to (W, H-2). Now, there aren't enough lines in the terminal
buffer to catch all the lines that conpty is about to emit. When that
happens, lines get duplicated in the buffer. From a UX perspective, this
certainly looks a lot worse than a couple lost lines. It looks like
utter chaos.
So I've introduced a new mode to conpty to try and counteract this
behavior. This behavior I'm calling "quirky resize". The **TL;DR** of
quirky resize mode is that conpty won't emit the entire buffer on a
resize, and will trust that the terminal is prepared to reflow it's
buffer on it's own.
This will enable the quirky resize behavior for applications that are
prepared for it. The "quirky resize" is "don't `InvalidateAll` when the
terminal resizes". This is added as a quirk as to not regress other
terminal applications that aren't prepared for this behavior
(gnome-terminal, conhost in particular). For those kinds of terminals,
when the buffer is resized, it's just going to lose lines. That's what
currently happens for them.
When the quirk is enabled, conpty won't repaint the entire buffer. This
gets around the "duplicated lines" issue that requesting multiple
resizes in a row can cause. However, for these terminals that are
unprepared, the conpty cursor might end up in the wrong position after a
quirky resize.
The case in point is maximizing the terminal. For maximizing
(height->50) from a buffer that's 30 lines tall, with the cursor on
y=30, this is what happens:
* With the quirk disabled, conpty reprints the entire buffer. This is
60 lines that get printed. This ends up blowing away about 20 lines
of scrollback history, as the terminal app would have tried to keep
the text pinned to the bottom of the window. The term. app moved the
viewport up 20 lines, and then the 50 lines of conpty output (30
lines of text, and 20 blank lines at the bottom) overwrote the lines
from the scrollback. This is bad, but not immediately obvious, and
is **what currently happens**.
* With the quirk enabled, conpty doesn't emit any lines, but the
actual content of the window is still only in the top 30 lines.
However, the terminal app has still moved 20 lines down from the
scrollback back into the viewport. So the terminal's cursor is at
y=50 now, but conpty's is at 30. This means that the terminal and
conpty are out of sync, and there's not a good way of re-syncing
these. It's very possible (trivial in `powershell`) that the new
output will jump up to y=30 override the existing output in the
terminal buffer.
The Windows Terminal is already prepared for this quirky behavior, so it
doesn't keep the output at the bottom of the window. It shifts it's
viewport down to match what conpty things the buffer looks like.
What happens when we have passthrough mode and WT is like "I would like
quirky resize"? I guess things will just work fine, cause there won't be
a buffer behind the passthrough app that the terminal cares about. Sure,
in the passthrough case the Terminal could _not_ quirky resize, but the
quirky resize won't be wrong.
## Summary of the Pull Request
Changes how conpty emits text to preserve line-wrap state, and additionally adds rudimentary support to the Windows Terminal for wrapped lines.
## References
* Does _not_ fix (!) #3088, but that might be lower down in conhost. This makes wt behave like conhost, so at least there's that
* Still needs a proper deferred EOL wrap implementation in #780, which is left as a todo
* #4200 is the mega bucket with all this work
* MSFT:16485846 was the first attempt at this task, which caused the regression MSFT:18123777 so we backed it out.
* #4403 - I made sure this worked with that PR before I even sent #4403
## PR Checklist
* [x] Closes#405
* [x] Closes#3367
* [x] I work here
* [x] Tests added/passed
* [n/a] Requires documentation to be updated
## Detailed Description of the Pull Request / Additional comments
I started with the following implementation:
When conpty is about to write the last column, note that we wrapped this line here. If the next character the vt renderer is told to paint get is supposed to be at the start of the following line, then we know that the previous line had wrapped, so we _won't_ emit the usual `\r\n` here, and we'll just continue emitting text.
However, this isn't _exactly_ right - if someone fills the row _exactly_ with text, the information that's available to the vt renderer isn't enough to know for sure if this line broke or not. It is possible for the client to write a full line of text, with a `\n` at the end, to manually break the line. So, I had to also add the `lineWrapped` param to the `IRenderEngine` interface, which is about half the files in this changelist.
## Validation Steps Performed
* Ran tests
* Checked how the Windows Terminal behaves with these changes
* Made sure that conhost/inception and gnome-terminal both act as you'd expect with wrapped lines from conpty
## Summary of the Pull Request
- Height adjustment of a glyph is now restricted to itself in the DX
renderer instead of applying to the entire run
- ConPTY compensates for drawing the right half of a fullwidth
character. The entire render base has this behavior restored now as
well.
## PR Checklist
* [x] Closes#2191
* [x] I work here
* [x] Tests added/passed
* [x] No doc
* [x] Am core contributor.
## Detailed Description of the Pull Request / Additional comments
Two issues:
1. On the DirectX renderer side, when confronted with shrinking a glyph,
the correction code would apply the shrunken size to the entire run, not
just the potentially individual glyph that needed to be reduced in size.
Unfortunately while adjusting the horizontal X width can be done for
each glyph in a run, the vertical Y height has to be adjusted for an
entire run. So the solution here was to split the individual glyph
needing shrinking out of the run into its own run so it can be shrunk.
2. On the ConPTY side, there was a long standing TODO that was never
completed to deal with a request to draw only the right half of a
two-column character. This meant that when encountering a request for
the right half only, we would transmit the entire full character to be
drawn, left and right halves, struck over the right half position. Now
we correct the cursor back a position (if space) and draw it out so the
right half is struck over where we believe the right half should be (and
the left half is updated as well as a consequence, which should be OK.)
The reason this happens right now is because despite VIM only updating
two cells in the buffer, the differential drawing calculation in the
ConPTY is very simplistic and intersects only rectangles. This means
from the top left most character drawn down to the row/col cursor count
indicator in vim's modeline are redrawn with each character typed. This
catches the line below the edited line in the typing and refreshes it.
But incorrectly.
We need to address making ConPTY smarter about what it draws
incrementally as it's clearly way too chatty. But I plan to do that with
some of the structures I will be creating to solve #778.
## Validation Steps Performed
- Ran the scenario listed in #2191 in vim in the Terminal
- Added unit tests similar to examples given around glyph/text mapping
in runs from Microsoft community page
## Summary of the Pull Request
- If no one is listening to the ETW provider for the VT Renderer for diagnostic purposes, do not spend time allocating/deleting/formatting strings for presentation in TraceLogging messages.
<!-- Please review the items on the PR checklist before submitting-->
## PR Checklist
* [x] Closes something I noticed while working on the renderer.
* [x] I work here.
* [x] Existing tests should pass
* [x] No doc
* [x] Am core contributor.
## Validation Steps Performed
WPR/WPA
Before: 321/3016 samples on hot path (10.64%)
After: 0/1266 samples on the same path (0%)
* static analysis fixes
* using C++ style casts
* explicit delete changed to reset(nullptr)
* fix for null apiMsg.OtherId during tracing in Compare()
* changed INVALID_ID macro to constexpr
* properly handle null ReplyMsg in ConsoleIoThread()
* Fixed wrong static_cast for State.InputBuffer
* compensate for null reply message to fix deref problem of ReplyMsg in srvinit.cpp by changing signature in DeviceComm.h
Michael Niksa
Mike Griese
adiviness
Eric Budai
Dustin Howett
