For that, we maintain an abstract store typing of all variables with LOCAL location (i.e., those that do not escape the function's own scope). We treat assignments as sequential effects that modify this store.
When control flow branches, we have to compute the disjunction of possible effects. To that end, we represent the store as a stack of effect sets, such that we can cheaply push and pop "local" effects when control flow has to branch.
In cases of non-local control transfer from an unknown source, we currently erase all knowledge about the store.
The 'switch' statement is still to come.
For a formulation of the typing rules, see:
https://docs.google.com/a/google.com/file/d/0B3wuXSv9YKuKeUNkVXZDemZ0Z1E
;)
R=jkummerow@chromium.org
BUG=
Review URL: https://codereview.chromium.org/19054006
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@15776 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Notes:
- For now, just adds the missing type info fields to the AST nodes directly.
I'd like to factor that out more nicely in a follow-up CL.
- All type feedback now is uniformly collected through AST nodes'
RecordTypeFeedback functions. At some point, this logic should be moved
out of ast.cc.
- The typing pass currently simulates the exact same conditions under
which feedback was collected in Hydrogen before. That also should be
made more generic in the future.
- Type information itself is unchanged. Making it more regular is
yet more future work.
Some additional cleanups:
- Lifted out nested class ObjectLiteral::Property, to enable forward declaration.
- Moved around some auxiliary enums.
R=svenpanne@chromium.org
BUG=
Review URL: https://codereview.chromium.org/14990014
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@14825 ce2b1a6d-e550-0410-aec6-3dcde31c8c00