diff options
author | Karl Williamson <khw@cpan.org> | 2016-05-13 11:32:44 -0600 |
---|---|---|
committer | Karl Williamson <khw@cpan.org> | 2016-05-24 10:28:38 -0600 |
commit | 17f41037d4817b6618a903e12aa1377ae078f66a (patch) | |
tree | d782a3e0de1110c764deb7dd62320f41b9319eb3 /locale.c | |
parent | 6ddd902ce7b4052f3d48e6dd638d08d705d1ee16 (diff) | |
download | perl-17f41037d4817b6618a903e12aa1377ae078f66a.tar.gz |
locale.c: Make locale collation predictions adaptive
We try to avoid calling strxfrm() more than needed by predicting its
needed buffer size. This generally works because the size of the
transformed string is roughly linear with the size of the input string.
But the key word here is "roughly". This commit changes things, so that
when we guess low, we change the coefficients in the equation to guess
higher the next time.
Diffstat (limited to 'locale.c')
-rw-r--r-- | locale.c | 60 |
1 files changed, 57 insertions, 3 deletions
@@ -1436,6 +1436,7 @@ Perl__mem_collxfrm(pTHX_ const char *input_string, STRLEN s_strlen = strlen(input_string); char *xbuf = NULL; STRLEN xAlloc; /* xalloc is a reserved word in VC */ + STRLEN length_in_chars; bool first_time = TRUE; /* Cleared after first loop iteration */ PERL_ARGS_ASSERT__MEM_COLLXFRM; @@ -1735,14 +1736,16 @@ Perl__mem_collxfrm(pTHX_ const char *input_string, } } + length_in_chars = (utf8) + ? utf8_length((U8 *) s, (U8 *) s + len) + : len; + /* The first element in the output is the collation id, used by * sv_collxfrm(); then comes the space for the transformed string. The * equation should give us a good estimate as to how much is needed */ xAlloc = COLLXFRM_HDR_LEN + PL_collxfrm_base - + (PL_collxfrm_mult * ((utf8) - ? utf8_length((U8 *) s, (U8 *) s + len) - : len)); + + (PL_collxfrm_mult * length_in_chars); Newx(xbuf, xAlloc, char); if (UNLIKELY(! xbuf)) goto bad; @@ -1759,6 +1762,57 @@ Perl__mem_collxfrm(pTHX_ const char *input_string, * was available, it means it successfully transformed the whole * string. */ if (*xlen < xAlloc - COLLXFRM_HDR_LEN) { + + /* If the first try didn't get it, it means our prediction was low. + * Modify the coefficients so that we predict a larger value in any + * future transformations */ + if (! first_time) { + STRLEN needed = *xlen + 1; /* +1 For trailing NUL */ + STRLEN computed_guess = PL_collxfrm_base + + (PL_collxfrm_mult * length_in_chars); + const STRLEN new_m = needed / length_in_chars; + + DEBUG_Lv(PerlIO_printf(Perl_debug_log, + "%s: %d: initial size of %"UVuf" bytes for a length " + "%"UVuf" string was insufficient, %"UVuf" needed\n", + __FILE__, __LINE__, + (UV) computed_guess, (UV) length_in_chars, (UV) needed)); + + /* If slope increased, use it, but discard this result for + * length 1 strings, as we can't be sure that it's a real slope + * change */ + if (length_in_chars > 1 && new_m > PL_collxfrm_mult) { +#ifdef DEBUGGING + STRLEN old_m = PL_collxfrm_mult; + STRLEN old_b = PL_collxfrm_base; +#endif + PL_collxfrm_mult = new_m; + PL_collxfrm_base = 1; /* +1 For trailing NUL */ + computed_guess = PL_collxfrm_base + + (PL_collxfrm_mult * length_in_chars); + if (computed_guess < needed) { + PL_collxfrm_base += needed - computed_guess; + } + + DEBUG_Lv(PerlIO_printf(Perl_debug_log, + "%s: %d: slope is now %"UVuf"; was %"UVuf", base " + "is now %"UVuf"; was %"UVuf"\n", + __FILE__, __LINE__, + (UV) PL_collxfrm_mult, (UV) old_m, + (UV) PL_collxfrm_base, (UV) old_b)); + } + else { /* Slope didn't change, but 'b' did */ + const STRLEN new_b = needed + - computed_guess + + PL_collxfrm_base; + DEBUG_Lv(PerlIO_printf(Perl_debug_log, + "%s: %d: base is now %"UVuf"; was %"UVuf"\n", + __FILE__, __LINE__, + (UV) new_b, (UV) PL_collxfrm_base)); + PL_collxfrm_base = new_b; + } + } + break; } |