GitLab

Changes during post-3.3/pre-3.4 development refactored storage of most
configure-time information, including variable bindings and function
scopes.  All scopes (even short-lived) were kept persistently for
possible future debugging features, causing huge accumulated memory
usage.  This was mostly addressed by commit v3.4.1~4^2 (cmState: Avoid
accumulating snapshot storage for short-lived scopes, 2015-11-24).

Since then we still keep short-lived scopes when they are needed for a
backtrace.  This is because since commit v3.4.0-rc1~378^2
(cmListFileBacktrace: Implement in terms of cmState::Snapshot,
2015-05-29) backtraces have been lightweight objects that simply point
into the snapshot tree.  While the intention of this approach was to
avoid duplicating the call stack file path strings, the cost turned out
to be holding on to the entire call stack worth of scope snapshots,
which is much worse.

Furthermore, since commit v3.4.0-rc2~1^2 (cmIfCommand: Issue CMP0054
warning with appropriate context, 2015-10-20) all conditions used in
`if()` commands hold a backtrace for use in diagnostic messages.  Even
though the backtrace is short-lived it still causes the scope snapshot
to be kept.  This means that code like

    function(foo)
      if(0)
      endif()
    endfunction()

    foreach(i RANGE 1000000)
      foo()
    endforeach()

accumulates storage for the function call scope snapshots.

Fix this by partially reverting commit v3.4.0-rc1~378^2 and saving the
entire call stack during cmListFileBacktrace construction.  This way
we can avoid keeping short-lived scope snapshot storage in all cases.

The bottom of the call stack is always a long-lived snapshot and can be
saved for later use with cmOutputConverter.

We need to keep only certain snapshot types and their ancestors.
Also keep those needed for backtraces.

We enforce policy push/pop balance around any scope that pushes/pops a
snapshot.  Therefore a snapshot may never reference entries of
PolicyStack that were created in nested scopes.  Free storage of
short-lived policy stack entries when they are popped.

Logically the method pushes a nested scope on top of a given scope
because the "up" pointer sequence forms a stack independent of any
other branches of the tree.

Everywhere we use cmMakefile::ScopePushPop to manage variable scopes
also expects policy scopes to be balanced.  There is no place that we
use cmMakefile::PolicyPushPop without also using ScopePushPop.  Relieve
PolicyPushPop of responsibility for policy scope balance checks by
moving it to ScopePushPop.

Since commit v3.4.0-rc1~179^2~1 (cmState: Add a VariableScope snapshot
type, 2015-08-23) the snapshot stack may have a VariableScopeType entry.
Skip over these when constructing the call stack, just as we do for
policy scopes.  Otherwise we report the command causing the variable
scope to be entered twice (e.g. find_package while loading a package
version file).

Add new API for the subdirs command to cmState.

This fixes a regression introduced in commit f716460e (cmMakefile: Move
invokation to initialize snapshot., 2015-10-06).

Don't leave this as cmMakefile responsibility.

Don't leave this as cmMakefile responsibility.

Don't leave this as cmMakefile responsibility.

Don't leave this as cmMakefile responsibility.

This is the appropriate snapshot for future use.  Existing users
of this method are mostly just calling GetDirectory on the result,
but the progress tracking of the Makefile generator needs a consistent
snapshot to be used, so the end snapshot should be consistently used.

This is a change in behavior from CMake 3.3, but there is no semantic meaning
to empty entries in buildsystem properties.  This also restores behavior to
that of CMake 2.8.10.

Match the scopes currently used in cmMakefile for definitions.

For completeness mostly.

Refactoring in commit 6ed9c7e0 (cmState: Host buildsystem properties for
directories, 2015-07-18) broke include_directories(BEFORE).  Fix it and
add a test case.

Implement lexical scope checking in terms of the state stack instead
of barriers.

At generate time, variables and policies are determined from the end
state of the cmMakefile.

This was added at a time when I thought the cmState would report errors,
but that is not the direction it is going.  The ordering of the stack of
scopes is canonical and clients are required to maintain it.

Brad King authored Jul 23, 2015 and Stephen Kelly committed Jul 25, 2015

Delay use of the PositionType constructor until after SnapshotDataType
is fully defined.