Condition where backtrace buffer for profiling is full

[kimikage]

On Linux, the number of stored instruction pointers will be less than the specified maximum number.

julia> versioninfo()
Julia Version 1.4.1
Commit 381693d3df* (2020-04-14 17:20 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-8565U CPU @ 1.80GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-8.0.1 (ORCJIT, skylake)

julia> using Profile

julia> Profile.init(n = 100)

julia> Profile.clear()

julia> @profile rand(Float64, 100, 100);

julia> length(Profile.fetch())
99

In fact, the source code specifies that. However, I don't understand the meaning of the -1. The null sentinels are counted properly.

julia/src/signals-unix.c

Lines 700 to 706 in f12cde0

	// Mark the end of this block with 0
	bt_data_prof[bt_size_cur++].uintptr = 0;
	}
	if (bt_size_cur >= bt_size_max - 1) {
	// Buffer full: Delete the timer
	jl_profile_stop_timer();
	}

Although the behavior is safer in terms of memory access, the warning below is not displayed correctly.

julia/stdlib/Profile/src/Profile.jl

Lines 308 to 315 in 3af96bc

	function fetch()
	len = len_data()
	maxlen = maxlen_data()
	if (len == maxlen)
	@warn """The profile data buffer is full; profiling probably terminated
	before your program finished. To profile for longer runs, call
	`Profile.init()` with a larger buffer and/or larger delay."""
	end

I'm still studying how profiling works, so I am not confident in making a PR for this issue.

[kimikage]

Even if the last one stack is free, there is no point in filling it because it is used for the sentinel.

[kimikage]

However, we may need not only "1" but also JL_BT_MAX_ENTRY_SIZE of free space. If there is not enough room, the end of the buffer will be filled with the sentinels. This isn't dangerous, but I don't think it is the desired behavior.