Cherry-pick "LibRegex: Only search start of line if pattern begins with ^"

[nico]

LadybirdBrowser/ladybird#1570

[nico]

@alimpfard looks like this causes 54.298 RegexLibC(374): Failed test 'simple_notbol_noteol' in 0ms. Is that expected?

[alimpfard]

The tests should pass, I guess we don't have enough tests in ladybird (or this is a posix-specific issue)

[nico]

Want to take a look into what's going on?

[nico]

Here's a standalone demo:

#include <regex.h>
#include <stddef.h>
#include <stdio.h>

#define REG_NOERR 0

#define EXPECT_EQ(a, b) printf("%d\n", a)

int main()
{
    const char pattern[] = "^hello friends$";
    const char pattern2 []= "hello friends";
    regex_t regex, regex2;

    EXPECT_EQ(regcomp(&regex, pattern, REG_EXTENDED | REG_NOSUB | REG_ICASE), REG_NOERR);
    EXPECT_EQ(regcomp(&regex2, pattern2, REG_EXTENDED | REG_NOSUB | REG_ICASE), REG_NOERR);

    EXPECT_EQ(regexec(&regex, "hello friends", 0, NULL, 0), REG_NOERR);

    EXPECT_EQ(regexec(&regex, "hello friends", 0, NULL, REG_NOTBOL), REG_NOMATCH);
    EXPECT_EQ(regexec(&regex, "hello friends", 0, NULL, REG_NOTEOL), REG_NOMATCH);
    EXPECT_EQ(regexec(&regex, "hello friends", 0, NULL, REG_NOTBOL | REG_NOTEOL), REG_NOMATCH);

    EXPECT_EQ(regexec(&regex, "a hello friends b", 0, NULL, REG_NOTBOL), REG_NOMATCH);
    EXPECT_EQ(regexec(&regex, "a hello friends", 0, NULL, REG_NOTBOL), REG_NOMATCH);
    //EXPECT_EQ(regexec(&regex, "a hello friends", 0, NULL, REG_NOTBOL | REG_SEARCH), REG_NOERR);
    //EXPECT_EQ(regexec(&regex, "a hello friends b", 0, NULL, REG_NOTBOL | REG_SEARCH), REG_NOERR);

    EXPECT_EQ(regexec(&regex, "a hello friends b", 0, NULL, REG_NOTEOL), REG_NOMATCH);
    EXPECT_EQ(regexec(&regex, "hello friends b", 0, NULL, REG_NOTEOL), REG_NOMATCH);
    //EXPECT_EQ(regexec(&regex, "hello friends b", 0, NULL, REG_NOTEOL | REG_SEARCH), REG_NOERR);
    //EXPECT_EQ(regexec(&regex, "a hello friends b", 0, NULL, REG_NOTEOL | REG_SEARCH), REG_NOMATCH);

    EXPECT_EQ(regexec(&regex, "a hello friends b", 0, NULL, REG_NOTBOL | REG_NOTEOL), REG_NOMATCH);
    //EXPECT_EQ(regexec(&regex, "a hello friends b", 0, NULL, REG_NOTBOL | REG_NOTEOL | REG_SEARCH), REG_NOMATCH);

    EXPECT_EQ(regexec(&regex2, "hello friends", 0, NULL, REG_NOTBOL), REG_NOMATCH);
    EXPECT_EQ(regexec(&regex2, "hello friends", 0, NULL, REG_NOTEOL), REG_NOMATCH);

    regfree(&regex);
    regfree(&regex2);
}

The failing lines are the commented-out ones – REG_SEARCH is a serenity extension, so it doesn't work with my system libc.

I'm confused by those lines. We pass REG_NOTBOL, so the input string doesn't start at start-of-line, but independently of that the string shouldn't match, independent of REG_SEARCH, right?

So it looks correct to me that we don't return REG_NOERR there (…or am I misreading the test?). But why did we pass it before this optimization? The commit message sounds like this is a perf optimization, not a correctness fix.

[nico]

(…and we still return REG_NOTEOL in the REG_NOTEOL | REG_SEARCH case, which still seems wrong?)