While at it, make it finer-grained (63 bits of randomness)
function old new delta
evaluate 3303 3336 +33
.rodata 104107 104111 +4
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 2/0 up/down: 37/0) Total: 37 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Rework of the previous fix:
Can use operation attributes to disable arg evaluation instead of special-casing.
function old new delta
.rodata 104032 104036 +4
evaluate 3223 3215 -8
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 1/1 up/down: 4/-8) Total: -4 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
ptest() was using this idea already.
As far as I can see, this is safe. Ttestsuite passes.
One downside is that a temporary from e.g. printf invocation
won't be freed until the next printf call.
function old new delta
awk_printf 481 468 -13
as_regex 137 111 -26
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 0/2 up/down: 0/-39) Total: -39 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
It seems to be designed to reduce overhead of malloc's auxiliary data,
by allocating at least 64 variables as a block.
With "struct var" being about 20-32 bytes long (32/64 bits),
malloc overhead for one temporary indeed is high, ~33% more memory used
than needed.
function old new delta
evaluate 3137 3145 +8
modprobe_main 798 803 +5
exec_builtin 1414 1419 +5
awk_printf 476 481 +5
as_regex 132 137 +5
EMSG_INTERNAL_ERROR 15 - -15
nvfree 169 116 -53
nvalloc 145 - -145
------------------------------------------------------------------------------
(add/remove: 0/2 grow/shrink: 5/1 up/down: 28/-213) Total: -185 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
hash_find(): do not caclculate hash twice. Do not divide - can use
cheap multiply-by-8 shift.
nextword(): do not repeatedly increment in-memory value, do it in register,
then store final result.
hashwalk_init(): do not strlen() twice.
function old new delta
hash_search3 - 49 +49
hash_find 259 281 +22
nextword 19 16 -3
evaluate 3141 3137 -4
hash_search 54 28 -26
------------------------------------------------------------------------------
(add/remove: 1/0 grow/shrink: 1/3 up/down: 71/-33) Total: 38 bytes
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
We can free them after they are no longer needed.
(Currently, being a NOEXEC applet is much larger waste of memory
for the case of long-running awk script).
function old new delta
awk_main 831 827 -4
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
The same stored search pattern applies to both search ('/') and
search/replace (':s') operations.
A search/replace operation with an empty "find" string (':s//abc/')
should use the last stored search pattern, if available, and issue an
error message if there is none.
If the "find" string is not empty it should replace the stored search
pattern.
function old new delta
colon 3952 4024 +72
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 1/0 up/down: 72/0) Total: 72 bytes
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
With FEATURE_VI_REGEX_SEARCH enabled backward searches don't work.
This is problematic on distros that enable regexes, such as Tiny
Core Linux and Fedora.
When calling GNU re_search() with a negative range parameter
(indicating a backward search) the start offset must be set to
the end of the area being searched.
The return value of re_search() is the offset of the matched pattern
from the start of the area being searched. For a successful search
(positive return value) char_search() can return the pointer to
the start of the area plus the offset.
FEATURE_VI_REGEX_SEARCH isn't enabled by default but when it is:
function old new delta
char_search 256 247 -9
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-9) Total: -9 bytes
Signed-off-by: Andrey Dobrovolsky <andrey.dobrovolsky.odessa@gmail.com>
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Accepting nonsense like "--4", and even "-- -4" is confusing.
function old new delta
parse_expr 917 938 +21
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
If the motion command used to define the range of a change, yank or
delete fails the whole command should be rejected. BusyBox vi already
handled failed searches in these circumstances. Add some more cases:
- non-existent mark: d'x
- movement beyond end of file: c99999+ or 99999<<
This is implemented using a global variable which is set when a command
error is detected. Unlike the case of motion within a line it's
insufficient to check that the motion command doesn't move the cursor:
this fails to process 'LyL' correctly, for example, as the second 'L'
doesn't move the cursor.
function old new delta
indicate_error 75 82 +7
find_range 686 692 +6
do_cmd 4851 4852 +1
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 3/0 up/down: 14/0) Total: 14 bytes
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
In traditional vi and vim line motion commands ('+'/'-'/'j'/'k')
fail if the movement would exceed the bounds of the file. BusyBox vi
allowed such commands to succeed, leaving the cursor on the first or
last character of the file.
Make BusyBox vi work like vi/vim.
For the 'G'/'H'/'L' commands traditional vi treats an out of bounds
result as an error, vim doesn't. BusyBox vi behaves like vim, both
before and after this patch.
function old new delta
do_cmd 4785 4851 +66
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 1/0 up/down: 66/0) Total: 66 bytes
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
When ESC is entered to leave insert mode any autoindent should only
be removed if there's no content beyond the indent. This may be the
case if a line has been split by entering insert mode and then
entering a CR.
Add a check to ensure there's only a newline after the indent.
function old new delta
char_insert 912 929 +17
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 1/0 up/down: 17/0) Total: 17 bytes
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Second reference to a field reallocs/moves Fields[] array, but first ref
still tries to use the element where it was before move.
function old new delta
fsrealloc 94 106 +12
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
The default tabstop value should be set during early start up,
not reset for each new file.
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
When no line number is specified ':read' should place the inserted
text after the current line, not before.
This used to be correct but was broken when commit 0c42a6b07
(vi: fix empty line range regression) revealed a bug in commit
7a8ceb4eb (vi: changes to line addresses for colon commands).
function old new delta
colon 3960 3952 -8
------------------------------------------------------------------------------
(add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-8) Total: -8 bytes
Signed-off-by: Ron Yorston <rmy@pobox.com>
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
