new | ask | show | jobs
gurkwart 20 days ago  [ - ]

Nice, I like the colored output tables. Started tinkering with a small profiling lib as well a while ago.

https://github.com/gurki/glimmer

It focuses on creating flamegraphs to view on e.g. https://www.speedscope.app/. I wanted to use std::stacktrace, but they are very costly to evaluate, even just lazily at exit. Eventually, I just tracked thread and call layer manually.

If I understand correctly, you're tracking your call stack manually as well using some graph structure on linear ids? Mind elaborating a bit on its functionality and performance? Also proper platform-independent function names were a pita. Any comments on how you addressed that?

dustbunny 20 days ago  [ - ]

Speed scope is awesome.

Ive been thinking about using speed scope as a reference to make a native viewer like that.

Sampling profilers (like perf) are just so much easier to use than source markup ones. Just feel like the tooling around perf is bad and that speedscope is part of the solution.

GeorgeHaldane 20 days ago  [ - ]

General rundown of the logic can be found in this comment on reddit: https://www.reddit.com/r/cpp/comments/1jy6ver/comment/mmze20...

About linear IDs: A call graph in general case is a tree of nodes, each node has a single parent and an arbitrary amount of children. Each node accumulates time spend in the "lower" branches. A neat property of the callgraph relative to a generic tree, is that every node can be associated with a callsite. For example, if a some function f() calls itself 3 recursively, there will be multiple nodes corresponding to it, but in terms of callsite there is still only one. So lets take some simple call graph as an example:

  Callgraph:         f() -> f() -> f() -> g()
                                       -> h()
  Node id:           0      1      2      3,4
Let's say f() has callsite id '0', g() has callsite id '1', h() has callsite id '2'. The callgraph will then consist of N=5 nodes with M=3 different callsites:

  Node id:         { 0   1   2   3   4 }
  Callsite id:     { 0   0   0   1   2 }
We can then encode all "prev."" nodes as a single N vector, and all "next" nodes as a MxN matrix, which has some kind of sentinel value (like -1) in places with no connection. For this example this will result in following:

  Node id:         { 0   1   2   3   4 }
  Prev. id:        { x   0   1   2   2 }
  Next id matrix:  [ 1   2   3   x   x ]
                   [ x   x   4   x   x ]
Every thread has a thread-local callgraph object that keeps track of all this graph traversal, it holds 'current_node_id'. Traversing backwards on the graph is a single array lookup:

  current_node_id = prev_node_ids[current_node_id];
Traversing forwards to an existing callgraph node is a lookup & branch:

  next_node_id = next_node_ids[callsite_id, current_node_id]
  if (next_node_id = x) create_node(next_node_id); // will be usually predicted
  else                  current_node_id = next_node_id;
New nodes can be created pretty cheaply too, but too verbose for a comment. The key to tracking the callsites and assigning them IDs is thread_local local variables generated by the macro:

https://github.com/DmitriBogdanov/UTL/blob/master/include/UT...

When callsite marker initializes (which only happens once), it gets a new ID. Timer then gets this 'callsite_id' an passes it to the forwards-traversal. The way we get function names is by simply remembering __FILE__, __func__, __LINE__ pointers in another array of the call graph, they get saved during the callsite marker initialization too. As far as performance goes everything we do is cheap & simple operations, at this point the main overhead is just from taking the timestamps.