By default hide entries in the comment field that starts with #, and
add a "comments" command to show all comments.

Disassembly reports generated by pprof -disasm will now include line number
information as generated by objdump. This will make the generated assembly more
readable.

As part of this I've introduced a new assemblyInstruction struct. Previously
the code was reusing the graph.Node to represent assembly instructions but it
seems better to have a dedicated type for this.

Format numeric tag values using the value of the -unit flag.
Added a testcase for this behavior.

- Remove dead code, unnecessary conversions
- Consistently use a single space after period

When -mean is selected, currently pprof divides the sample value
by value[0], which is expected to be the number of samples. This
is intended to produce mean value per sample. These means cannot
be added. Instead, we should add the value and the number of samples
independently and perform the division at the end.

To do this we will create a separate function to get the number of samples,
and accumulate it independently from the sample value (weigth) and apply
the division after the accumulation is completed.

When generating callgrind format output, produce cost lines at
instruction granularity. This allows visualizers supporting the
callgrind format to display instruction-level profiling information.

We also need to provide the object file (ob=) in order for tools to find
the object file to disassemble when displaying assembly.

We opportunistically group cost lines corressponding to the same
function together, reducing the number of superfluous description lines.
Subposition compression (relative position numbering) is also used to
reduce the output size.

This will allow merging nodes for multiple instantiations of
C++ templates. A graph option is introduced to preserve older
behavior.

When using the call_tree option and generating a graph for kcachegrind,
it will merge back nodes that are distinct on the tree, producing some
confusing results. Add a suffix so that these entries are kept separate.

This addresses the problem described in
http://yosefk.com/blog/how-profilers-lie-the-cases-of-gprof-and-kcachegrind.html ,
particularly the summary "Choosing a profiler is hard" section.

This is to keep the new TrimTree functionality from breaking any code
currently using the public interface. We do this by separating NodeSet
from nodePtrSet and creating different functions for each.

In a graph, NodeInfo maps one to one to the nodes, so it suffices to
just find the top NodeInfo s and only keep those nodes in the graph. In
a tree however, a single NodeInfo may map to many nodes. As of this
commit, a call to 'web 10' in pprof on a tree will return all the nodes
corresponding with the top 10 NodeInfo s.

It is no longer representative of how omissions are made. pprof uses
more than just the cumulative value to make omissions.

Before this change, the node count used in the label is the proposed amount
provided by the user. If some nodes were trimmed and the graph ended up with
less nodes than the user asked for, the report label will now reflect this.

Currently pprof will look for source files only on the current directory
and its parents. This makes it hard to examine sources on jobs where
there are multiple source trees (eg from different libraries).

Add a variable to provide a search path for source files. It will default
to the cwd, so there will be no change in behavior by default.

This reverts commit 9c191ebbce.
The trimproto option is no longer needed, removing it to avoid code growth.

When generating a call tree, pprof was using a map to keep track of all the
inline nodes for a location. That is incorrect as it may cause inline functions
at different nesting levels to reuse the same node, causing the resulting graph
to not be a tree.

When creating a tree nodes with the same info may appear on multiple places
in the tree. Keeping one of them preserves them all, which may cause disconnected
nodes to remain. To ensure the resulting graph is a connected tree, do not include
children on any removed node, which is suitable for the normal tree refinement
(nodecount and nodefraction) but does not allow visual refinement, which may eliminate
intermediate nodes. Disable visual mode refinement for call_tree to avoid this issue.

Add new trimproto option that generates a new profile while removing
symbol information for functions below nodefraction. This reduces the
profile sizes significantly.

Separate implementation of graph and tree creation to speed it up.
Graph implementation maps upfront all locations to sequences of nodes,
tree implementation uses a per-parent map to keep track of a different
node per location per parent.

Also update the build instructions to use go get