coverage: fix count vs emit discrepancy in coverage counter data writing
This patch revises the way coverage counter data writing takes place
to avoid problems where useful counter data (for user-written functions)
is skipped in favor of counter data from stdlib functions that are
executed "late in the game", during the counter writing process itself.
Reading counter values from a running "--coverpkg=all" program is an
inherently racy operation; while the the code that scans the coverage
counter segment is reading values, the program is still executing,
potentially updating those values, and updates can include execution
of previously un-executed functions. The existing counter data writing
code was using a two-pass model (initial sweep over the counter
segment to count live functions, second sweep to actually write data),
and wasn't properly accounting for the fact that the second pass could
see more functions than the first.
In the bug in question, the first pass discovered an initial set of
1240 functions, but by the time the second pass kicked in, several
additional new functions were also live. The second pass scanned the
counter segment again to write out exactly 1240 functions, but since
some of the counters for the newly executed functions were earlier in
the segment (due to linker layout quirks) than the user's selected
function, the sweep terminated before writing out counters for the
function of interest.
The fix rewrites the counter data file encoder to make a single sweep
over the counter segment instead of using a two-pass scheme.
Fixes #59563.
Change-Id: I5e908e226bb224adb90a2fb783013e52deb341da
Reviewed-on: https://go-review.googlesource.com/c/go/+/484535 Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Than McIntosh <thanm@google.com>