The creation of the primary key does not work.  We will have to flag
externally stored columns and copy the externally stored part from
the old table.

row_build_index_for_mysql(): Rename to row_merge_build_indexes().
Move from row0mysql.c to row0merge.c.

Remove private declarations from row0merge.h.  Make many functions static
in row0merge.c.

cmp_rec_rec_simple(): A new comparison function.

dict_index_get_min_size(): New function.

OS_FILE_FROM_FD(fd): A macro for converting from int to os_file_t.

rec_convert_dtuple_to_rec_comp(): Make the interface lower-level.

rec_get_converted_size_comp(): Return also extra_size.

UT_SORT_FUNCTION_BODY(): Remove reference to an obsolete test program.

row_rec_to_index_entry_low(): New function.

row0merge.c: Implement merge sort based on file streams instead of
fixed-size blocks.  Sort the small blocks as arrays of dfield_t*,
because it is faster than invoking rec_get_offsets() for every
comparison.

in fast index creation.

innobase_create_index_def(), innobase_copy_index_def(): Copy to pre-allocated
space.

innobase_create_key_def(): Change the return type from merge_index_def_t**
to merge_index_def_t*.

innobase_create_index_def(): Add parameter new_primary.

innobase_copy_index_def(): Simplify the documented algorithm,
and try to implement it properly.

innodb-index.test: Replace CHECKSUM TABLE with something more stable and
useful.  The test passes on an older BitKeeper snapshot:

ChangeSet@1.2475.18.9, 2007-05-08 11:16:41+02:00, jbruehe@mysql.com +1 -0
  Raise version number after cloning 5.1.18-beta

But it fails on a newer one where the statement
	alter table t1 add primary key (a), add key (b(20));
results in fast index creation:

ChangeSet@1.2500.1.40, 2007-06-01 20:06:13+04:00, kostja@bodhi.(none) +2 -0
  Merge bodhi.(none):/opt/local/work/mysql-5.0-runtime
  into  bodhi.(none):/opt/local/work/mysql-5.1-runtime
  MERGE: 1.1810.2984.14

infimum and supremum.  This was broken in r1515.

row_merge_create_temporary_table(), row_merge_create_index():
Return the dictionary object.  On error, set trx->error_state.

row_merge_mark_prebuilt_obsolete(): Rename to row_prebuilt_table_obsolete()
and move from row0merge.c to row0mysql.c.  Remove the parameter trx.

row_merge_print_table(): Remove.

row_remove_indexes_for_mysql(): Rename to row_merge_drop_indexes(),
move from row0mysql.c to row0merge.c and make the return type void.

row_merge_remove_index(): Rename to row_merge_drop_index() and make
the return type void.

Make use of UT_BITS_IN_BYTES().

of UT_SORT_FUNCTION_BODY is best done by defining SORT_FUN and CMP_FUN as
macros when needed.  The solution of r1523 allows for only one extra parameter.

needed in row0merge.c for merge sorting the small blocks in main memory.
Pass CTX also to SORT_FUN.  Adjust all users.

node_ptr.

rec_get_converted_size_new(): Rename to rec_get_converted_size_comp().
Add const qualifier to dtuple.  Remove the assertion
dict_table_is_comp(index->table), because this function will be invoked
when populating the temporary files in fast index creation.

rec_convert_dtuple_to_rec_new(): Rename to rec_convert_dtuple_to_rec_comp().
Add parameter "extra".

rec_init_offsets_comp_ordinary(): New function.  Used in rec_init_offsets().

and change the return type from ibool to int (-1, 0, 1).  This comparison
function will be needed in fast index creation.

row_merge_read_clustered_index(): Simplify the loop that scans the
clustered index.

otherwise like INNODB_ZIP, but resets the cumulated counts.

Rename innodb_buddy to innodb_zip in some comments.

innobase_stat_zip_fill(): Add parameter "reset".  Document the race condition
with the compression statistics counters.

trx_undo_report_row_operation(), trx_undo_report_dict_operation():
Reduce the scope of some variables.  Move the return(DB_SUCCESS)
case inside the for loop.

overflow.

Replace row0merge.h with row0mysql.h.

Remove redundant test trx->dict_operation_lock_mode == 0.
Fix typos in comments.

row_undo_ins(): Make the function more similar to the one in trunk.

exactly.  Rename the local variable "ulint level" to "ibool leaf".
Document that if the function returns DB_SUCCESS on a compressed page that
is covered by the insert buffer, the mini-transaction must be committed
before latching any further pages.  Verify that this is the case on all
execution paths.

the insert buffer bitmap.

ibuf_set_free_bits_func(): Never disable redo logging.

ibuf_update_free_bits_zip(): Remove.

btr_page_reorganize_low(), page_zip_reorganize(): Do not update the insert
buffer bitmap.  Instead, document that callers will have to take care of it,
and adapt the callers.

btr_compress(): On error, reset the insert buffer free bits.

btr_cur_insert_if_possible(): Do not modify the insert buffer bitmap.

btr_compress(), btr_cur_optimistic_insert(): On compressed pages,
reset the insert buffer bitmap.  Document why.

btr_cur_update_alloc_zip(): Document why it is necessary and sufficient
to reset the insert buffer free bits.

btr_cur_update_in_place(), btr_cur_optimistic_update(),
btr_cur_pessimistic_update(): Update the free bits in the same
mini-transaction.  Document that the mini-transaction must be
committed before latching any further pages.  Verify that this
is the case in all execution paths.

row_ins_sec_index_entry_by_modify(), row_ins_clust_index_entry_by_modify(),
row_undo_mod_clust_low(): Because these functions call
btr_cur_update_in_place(), btr_cur_optimistic_update(), or
btr_cur_pessimistic_update(), document that the mini-transaction must be
committed before latching any further pages.  Verify that this is the case
in all execution paths.

page_cur_tuple_insert(), page_cur_rec_insert(): Improve the documentation.
Note that the physical position of page_cursor may change.

page_cur_rec_insert(), page_cur_insert_rec_low(), page_cur_insert_rec_zip():
Add a const qualifier to the parameter rec.

Remove mysql.patch, because the configuration parameter interface has changed.

mtr_commit(): Do not acquire the log mutex if no log records are to be written.

ibuf_set_free_bits_func(): Do not clear mtr.modifications, because that would
prevent the dirty insert buffer bitmap block from being moved to the flush
list.  Instead, assume that mtr_commit() will not acquire log_sys->mutex
because redo logging is disabled.

buf_page_init_for_read(): I/O-fix the block before buf_buddy_alloc(), because
the newly allocated block could otherwise be reused by the compressed page.

the pages be decompressed.

buf_read_page_low(), buf_page_init_for_read(): Add parameter "unzip".

crash recovery is in progress.  This avoids a hang when
btr_parse_page_reorganize(), called from an I/O handler thread,
attempts to acquire log_sys->mutex while it is being held by
the main thread (the one that runs innobase_init()).  This change
was committed accidentally.  It may be unsafe to clear
mtr.modifications, because buf_page_release() at mtr_commit() may
forget to put modified pages to the flush list.

Cleanup: Remove the "type" parameter from many ibuf functions.
Let the caller check that !dict_index_is_clust().  This should avoid
function calls and register spilling.

ibuf_set_free_bits_func(), ibuf_set_free_bits(): Remove the parameter "type".

ibuf_reset_free_bits_with_type(): Rename to ibuf_reset_free_bits().
Remove the parameter "type".

ibuf_update_free_bits_if_full(), ibuf_update_free_bits_zip(),
ibuf_update_free_bits_low(), ibuf_update_free_bits_for_two_pages_low():
Remove the parameter "index".

log_write_up_to(): Replace if-else if-else with a switch block.

log_archive_do(): Remove an excess mutex_enter().  Replace if-else if-else
with a switch block.  (This function is not included unless UNIV_LOG_ARCHIVE
is defined.)

pass the compressed page size) with KEY_BLOCK_SIZE.

to the switch (row_merge_cmp(...)) to eliminate bogus warning about
uninitialized variable tmp.

was broken in one of the earlier cleanups.

row_merge_block_merge(): Remove the variable "selected".  Clear the record
count of new_block1 and new_block2.  Anticipate the right return values
from row_merge_cmp().

Add __attribute__((const)) or __attribute__((pure)) to various
mach_read functions.

dtype_new_read_for_order_and_null_size(): Remove a statement
without effect: mach_read_from_2(buf + 4).

in r1445.  They are needed after all, but apparently the tests do not
actually verify that the records are sorted.  (Also, a buffer overflow
will occur in index_merge_innodb.test if MERGE_BLOCK_SIZE is made smaller.)

Correct the language of some comments.  Eliminate the local variables
"num_of_merges" (renamed to "sorted" in one function).