When thunks are evaluated, the thunk closure is replaced with an
`IndClosure`. To make sure we don't miss parts of the string, we also
expand `IndClosure` types.

515eb33d4fcc9382032a9e6b1757b9e2b00a145c changes the order and type of
these two fields. Therefore we have to update the decoding logic.

In particular we make sure to race all the appropiate threads against
each other, so that if any of them fail the tests just fail rather than
hang forever.

`viewport` is unreasonably inefficient for our usecase when the IOTree grows, as it
decides what to display once it rendered to full `IOTree`.
Especially while scrolling, this inefficiency became apparent.

Unfortunately, brick doesn't have any builtin way to improve this, see https://github.com/jtdaugherty/brick/issues/264

We take `drawListElement` as an inspiration, and only render up to
`heightOfBox * 2` elements of the `IOTree`, including expanded tree
nodes.
This speeds up the view noticeably, giving the user a responsive experience.

In particular, we decouple rendering metadata computation from the actual rendering to decide which rows to render.

We now display a "flat" cost centre stack rather than forcing the user
to expand it at each level.

We should never attempt to dereference a null pointer, any null pointers
should be captured at the point of creation.

These are no longer used

Matthew Pickering authored 1 year ago and Zubin committed 1 year ago

Inside an hBox if there are multiple greedy widgets they are allocated
equal space. This meant the commands and number display were allocated
the same number of columns which led to a lot of whitespace.

(cherry picked from commit 7a1fdab5)

Matthew Pickering authored 1 year ago and Zubin committed 1 year ago

(cherry picked from commit d0250893)

Matthew Pickering authored 1 year ago and Zubin committed 1 year ago

(cherry picked from commit 865dd6a4)

Matthew Pickering authored 1 year ago and Zubin committed 1 year ago

* Add find retainer by address, precisely find retainers of a specific
  closure address.
* Add ArrWords count analysis which performs a census of the ArrWords
  closures.
* Generalise display pane, to allow displaying trees of different types
  (not just ClosureDetails) (used by the arr words census)
* Allow commands to not have key bindings (and just appear in command
  picker)
* Cleanup: Remove "Analysis" modes such as dominator and size analysis.

(cherry picked from commit 349760d0)

Resolves #25

Query debugee for profiling mode and TNTC

Deserialize closures instead of relying on ghc-heap
The above is necessary to decode profiling headers and ghc-heap
does not

Additionally, deserialize info tables in haskell code instead of
relying on ghc-heap. This is also necessary to decode profiling
info tables properly.

Perhaps we should introduce a type for profiling info tables instead of
using StgInfoTable

Zubin authored 1 year ago and Matthew Pickering committed 1 year ago

1. When forking, don't repond to a pause request in the parent after the child has already responded

2. use waitpid() instead of wait() because wait can race with other wait() calls in the program