Source listing handles missing binaries and addresses.

[ghemawat]

Some profiles can contain locations without an address, or reference
binaries/libraries that are not available. Treat these better when
producing source listings by using any source information present
directly in the profile.Location.

Details:

If a location does not have an address, create a synthetic address
that does not fall in a known mapping.

If a location references an object file that is not found, use any
file/line information baked directly in the location.

Tweaked generated html to clicking for source lines that have no
inlining/assembly to display.

Fixes #621.

[codecov-commenter]

Codecov Report

Merging #628 (3178125) into master (3a04a4d) will increase coverage by 0.26%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #628      +/-   ##
==========================================
+ Coverage   67.05%   67.32%   +0.26%     
==========================================
  Files          39       40       +1     
  Lines        7246     7295      +49     
==========================================
+ Hits         4859     4911      +52     
+ Misses       1950     1948       -2     
+ Partials      437      436       -1     

Impacted Files	Coverage Δ
...c/github.com/google/pprof/internal/report/synth.go	100.00% <0.00%> (ø)
.../github.com/google/pprof/internal/report/source.go	79.49% <0.00%> (+2.22%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3a04a4d...3178125. Read the comment docs.

[aalexand]

This is pre-existing and doesn't have to be changed (if it needs to be changed) in this CL, but: do we handle recursion correctly here? It feels that we'll double count the cumulative time if the same address is present multiple times in the stack.

[ghemawat]

I think you are right. Let me look at that in a separate change.

[raidancampbell]

Hate to butt in here, but is there a chance that this is related to #720?

I'm not familiar with this code, but a flavor of #720 manifested for me with the following conditions:

• using weblist instead of list
• providing the binary for disassembly

and had the following behavior:

• in a sample function that had 4 caller locations, the count was 5x what the expected value (e.g. that returned by list or weblist without the binary) was.
• the printed assembly was printed 5 times

As far as I can tell, that fits the bill for the code in question here: the call tree for this function is only invoked for web printing and not the CLI interactions