ash: fix quadratic matching slowdown is ${v/*foo*/repl} (really bad one)
It is especially bad with patterns starting with "*". With ASH_OPTIMIZE_FOR_SIZE=y, only those are optimized, +few bytes: text data bss dec hex filename 836337 441 7564 844342 ce236 busybox_old 836341 441 7564 844346 ce23a busybox_unstripped With ASH_OPTIMIZE_FOR_SIZE off, we also optimize patterns _ending_ with "*", which costs about 80 bytes: text data bss dec hex filename 836656 441 7564 844661 ce375 busybox_old 836732 441 7564 844737 ce3c1 busybox_unstripped Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
This commit is contained in:
Denys Vlasenko
parent
6eaeb7737d
commit
b76356b28e
71
shell/ash.c
71
shell/ash.c
@ -6040,25 +6040,61 @@ scanleft(char *startp, char *rmesc, char *rmescend UNUSED_PARAM, char *str, int
|
||||
}
|
||||
|
||||
static char *
|
||||
scanright(char *startp, char *rmesc, char *rmescend, char *str, int quotes,
|
||||
int zero)
|
||||
scanright(char *startp, char *rmesc, char *rmescend, char *pattern, int quotes, int match_at_start)
|
||||
{
|
||||
#if !ENABLE_ASH_OPTIMIZE_FOR_SIZE
|
||||
int try2optimize = match_at_start;
|
||||
#endif
|
||||
int esc = 0;
|
||||
char *loc;
|
||||
char *loc2;
|
||||
|
||||
for (loc = str - 1, loc2 = rmescend; loc >= startp; loc2--) {
|
||||
/* If we called by "${v/pattern/repl}" or "${v//pattern/repl}":
|
||||
* startp="escaped_value_of_v" rmesc="raw_value_of_v"
|
||||
* rmescend=""(ptr to NUL in rmesc) pattern="pattern" quotes=match_at_start=1
|
||||
* Logic:
|
||||
* loc starts at NUL at the end of startp, loc2 starts at the end of rmesc,
|
||||
* and on each iteration they go back two/one char until they reach the beginning.
|
||||
* We try to find a match in "raw_value_of_v", "raw_value_of_", "raw_value_of" etc.
|
||||
*/
|
||||
/* TODO: document in what other circumstances we are called. */
|
||||
|
||||
for (loc = pattern - 1, loc2 = rmescend; loc >= startp; loc2--) {
|
||||
int match;
|
||||
char c = *loc2;
|
||||
const char *s = loc2;
|
||||
if (zero) {
|
||||
if (match_at_start) {
|
||||
*loc2 = '\0';
|
||||
s = rmesc;
|
||||
}
|
||||
match = pmatch(str, s);
|
||||
match = pmatch(pattern, s);
|
||||
//bb_error_msg("pmatch(pattern:'%s',s:'%s'):%d", pattern, s, match);
|
||||
*loc2 = c;
|
||||
if (match)
|
||||
return loc;
|
||||
#if !ENABLE_ASH_OPTIMIZE_FOR_SIZE
|
||||
if (try2optimize) {
|
||||
/* Maybe we can optimize this:
|
||||
* if pattern ends with unescaped *, we can avoid checking
|
||||
* shorter strings: if "foo*" doesnt match "raw_value_of_v",
|
||||
* it wont match truncated "raw_value_of_" strings too.
|
||||
*/
|
||||
unsigned plen = strlen(pattern);
|
||||
/* Does it end with "*"? */
|
||||
if (plen != 0 && pattern[--plen] == '*') {
|
||||
/* "xxxx*" is not escaped */
|
||||
/* "xxx\*" is escaped */
|
||||
/* "xx\\*" is not escaped */
|
||||
/* "x\\\*" is escaped */
|
||||
int slashes = 0;
|
||||
while (plen != 0 && pattern[--plen] == '\\')
|
||||
slashes++;
|
||||
if (!(slashes & 1))
|
||||
break; /* ends with unescaped "*" */
|
||||
}
|
||||
try2optimize = 0;
|
||||
}
|
||||
#endif
|
||||
loc--;
|
||||
if (quotes) {
|
||||
if (--esc < 0) {
|
||||
@ -6248,7 +6284,7 @@ subevalvar(char *p, char *str, int strloc, int subtype,
|
||||
|
||||
#if ENABLE_ASH_BASH_COMPAT
|
||||
if (subtype == VSREPLACE || subtype == VSREPLACEALL) {
|
||||
char *idx, *end, *restart_detect;
|
||||
char *idx, *end;
|
||||
|
||||
if (!repl) {
|
||||
repl = parse_sub_pattern(str, varflags & VSQUOTE);
|
||||
@ -6257,17 +6293,19 @@ subevalvar(char *p, char *str, int strloc, int subtype,
|
||||
}
|
||||
|
||||
/* If there's no pattern to match, return the expansion unmolested */
|
||||
if (*str == '\0')
|
||||
if (str[0] == '\0')
|
||||
return 0;
|
||||
|
||||
len = 0;
|
||||
idx = startp;
|
||||
end = str - 1;
|
||||
while (idx < end) {
|
||||
try_to_match:
|
||||
loc = scanright(idx, rmesc, rmescend, str, quotes, 1);
|
||||
if (!loc) {
|
||||
/* No match, advance */
|
||||
restart_detect = stackblock();
|
||||
char *restart_detect = stackblock();
|
||||
skip_matching:
|
||||
STPUTC(*idx, expdest);
|
||||
if (quotes && (unsigned char)*idx == CTLESC) {
|
||||
idx++;
|
||||
@ -6279,7 +6317,16 @@ subevalvar(char *p, char *str, int strloc, int subtype,
|
||||
idx++;
|
||||
len++;
|
||||
rmesc++;
|
||||
continue;
|
||||
/* continue; - prone to quadratic behavior, smarter code: */
|
||||
if (idx >= end)
|
||||
break;
|
||||
if (str[0] == '*') {
|
||||
/* Pattern is "*foo". If "*foo" does not match "long_string",
|
||||
* it would never match "ong_string" etc, no point in trying.
|
||||
*/
|
||||
goto skip_matching;
|
||||
}
|
||||
goto try_to_match;
|
||||
}
|
||||
|
||||
if (subtype == VSREPLACEALL) {
|
||||
@ -6294,7 +6341,7 @@ subevalvar(char *p, char *str, int strloc, int subtype,
|
||||
}
|
||||
|
||||
for (loc = repl; *loc; loc++) {
|
||||
restart_detect = stackblock();
|
||||
char *restart_detect = stackblock();
|
||||
STPUTC(*loc, expdest);
|
||||
if (stackblock() != restart_detect)
|
||||
goto restart;
|
||||
@ -6303,7 +6350,7 @@ subevalvar(char *p, char *str, int strloc, int subtype,
|
||||
|
||||
if (subtype == VSREPLACE) {
|
||||
while (*idx) {
|
||||
restart_detect = stackblock();
|
||||
char *restart_detect = stackblock();
|
||||
STPUTC(*idx, expdest);
|
||||
if (stackblock() != restart_detect)
|
||||
goto restart;
|
||||
@ -6332,7 +6379,7 @@ subevalvar(char *p, char *str, int strloc, int subtype,
|
||||
if (subtype < 0 || subtype > 7)
|
||||
abort();
|
||||
#endif
|
||||
/* zero = subtype == VSTRIMLEFT || subtype == VSTRIMLEFTMAX */
|
||||
/* zero = (subtype == VSTRIMLEFT || subtype == VSTRIMLEFTMAX) */
|
||||
zero = subtype >> 1;
|
||||
/* VSTRIMLEFT/VSTRIMRIGHTMAX -> scanleft */
|
||||
scan = (subtype & 1) ^ zero ? scanleft : scanright;
|
||||
|
||||
Loading…
Reference in New Issue
Block a user