Currently the reader is stateless (read is a static call) but utilizes a stateful reader (and has a few hooks into compiler/runtime state for autoresolving keywords, etc). If the call into the reader at the top level calls a splicing reader conditional, then only the first one will be returned. The remaining forms are stranded in the pendingForms list and will be lost for subsequent reads.
Approach: Make top level reader conditional splicing an error:
1. Make top-level reader conditional splicing an error and throw an exception. (SELECTED)
2. Allow the caller to pass in a stateful collection to catch or return the pendingForms. This changes the effective calling API for the reader. You would only need to do this in the cases where reader conditionals were allowed/preserved.
3. Add a static (or threadlocal?) pendingForms attribute to the reader to capture the pendingForms across calls. A static field would have concurrency issues - anyone using the reader across threads would get cross-talk in this field. The pendingForms could be threadlocal which would probably achieve separation in the majority of cases, but also creates a number of lifecycle questions about those forms. When do they get cleared or reset? What happens if reading the same reader happens across threads? Another option would be an identity map keyed by reader instance - would need to be careful about lifecycle management and clean up, as it's basically a cache.
4. Add more state into the reader itself to capture the pendingForms. The reader interfaces and hierarchy would be affected. This would allow the reader to stop passing the pendingForms around inside but modifies the interface in other ways. Again, this would only be needed for the specific case where reader conditionals are allowed so other uses could continue to work as is?
5. If read is going to exit with pendingForms on the stack, they could be printed and pushed back onto the reader. This adds new read/print roundtrip requirements on things at the top level of reader conditionals that didn't exist before.
6. Wrap spliced forms at the top level in a `do`. This seems to violate the intention of splicing reader conditional to read as spliced since it is not the same as if those forms were placed separately in the input stream.
Might be too late since Rich already gave the OK but the proposed patch doesn't prevent single-element top level conditional splicing forms.
Is this intended?
New -2 patch catches reader conditional splice of 0 or 1 element.
Attached alternative patch that is less intrusive than clj-1706-2.patch
clj-1706-3.patch is identical to 0001-CLJ-1706-Make-top-level-reader-conditional-splicing-.patch but with one whitespace change reverted. Marking latest as screened.
Rich didn't like the dynvar in -3, so switching back to -2.