Reduces the amount of data written to temporary files.
The 3-level stack has to do a simple reduce after 4352 input files, for
a normal file sort this is 35GB of data or about 500 million records.
This needs about 50 open fd's - which should be ok.
Clearly the merge sort could process more input files in one go - speeding
up the sort, but at some point the number of input files would exceed
whatever limit was applied.
Split the merge sort so that fsort() can pass the 'FILE *' of the temporary
files to be merged into the merge code.
Don't rely on realloc() not moving the end address of a buffer!
Rework merge sort so that it sorts pointers to 'struct mfile' and only
copies about sort record descriptors.
No functional change intended.
This frees a byte value to use for 'end of key' (to correctly sort
short keys) while still having a weight assigned to the field sep.
(Unless -t is given, the field sep is in the field data.)
Do reverse sorts by writing the output file in reverse order (rather
than reversing the sort - apart from merges).
All key compares are now unweighted.
For 'sort -u' mark duplicates keys during the sort and don't write
to the output.
Use -S to mean a posix sort - where equal keys are sorted using the
raw record (rather than being kept in the original order).
For 'sort -f' (no keys) generate a key of the folded data (as for -n
-i and -d), simplifies the code and allows a 'posix' sort.
an array of 'RECHEADER *' and remove all the crappy stuff that backed up
by REC_DATA_OFFSET (etc).
Also change radix_sort() to return the number of elements, soon to be used
to drop duplicate keys (for sort -u).
- If we generate a key, it is always sortable using memcmp()
- If we are sorting the whole record, then a weight-table must be used
during compares.
- Major surgery to encoding of numbers to ensure unique keys for equal
numeric values. Reverse numerics are handled by inverting the sign.
- Case folding (-f) is handled when the sort keys are generated. No other
code has to care at all.
- Key uniqueness (-u) is done during merge for large datasets. It only
has to be done when writing the output file for small files.
Since the file is in key order this is simple!
Probably fixes all of: PR/27257 PR/25551 PR/22182 PR/31095 PR/30504
PR/36816 PR/37860 PR/39308
Also PR/18614 should no longer die, but a little more work needs to be
done on the merging for very large files.
Simplify logic for reading input records.
Do a merge sort whenever we have 16 partial sorted blocks.
The patient is breathing, but still carrying a lot of extra weight.
enlarge them as necessary to read records from merged files; the buffers
are allocated once per program run, so there shouldn't be any
performance difference
This makes sort(1) pass also regression 40B and should make it
fully arbitrary long record capable.
XXX the buffer array could probably be freed on end of fmerge() to save memory
for stable sort so that the records are not swapped arbitrarily - this makes
in-tree BSD sort(1) pass regression test 38
while here, do couple of cleanups, like s/16/MERGE_FNUM/ where appropriate,
making local stuff static and some intendation/code format changes
* get rid of union f_handle, replace by passing explicit int parameter
and (new) struct filelist
* add new typedefs gen_func_t and put_func_t and use where appropriate
end pointer has changed (full buffer is used instead of first DEFLLEN bytes)
the end pointer cannot be shared for crec and prec, we need to pass
different value in each case
