summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorNicholas Clark <nick@ccl4.org>2011-06-28 15:20:56 +0200
committerNicholas Clark <nick@ccl4.org>2011-07-01 14:05:41 +0200
commit72de20cdcc84ae35e0d8a55c7a92950fece19347 (patch)
treefde310b665c99a781df72f0aeab31f0e8ba8225e
parentb606cf7f37b8b46206c7f521b29167e037397a62 (diff)
downloadperl-72de20cdcc84ae35e0d8a55c7a92950fece19347.tar.gz
For shorter strings, store C<study>'s data as U8s or U16s, instead of U32s.
The assumption is that most studied strings are fairly short, hence the pain of the extra code is worth it, given the memory savings. 80 character string, 336 bytes as U8, down from 1344 as U32 800 character string, 2112 bytes as U16, down from 4224 as U32
-rw-r--r--ext/Devel-Peek/t/Peek.t57
-rw-r--r--pod/perldelta.pod6
-rw-r--r--pp.c61
-rw-r--r--regexec.c12
-rw-r--r--util.c85
5 files changed, 188 insertions, 33 deletions
diff --git a/ext/Devel-Peek/t/Peek.t b/ext/Devel-Peek/t/Peek.t
index 642d34c09b..d582a8f03c 100644
--- a/ext/Devel-Peek/t/Peek.t
+++ b/ext/Devel-Peek/t/Peek.t
@@ -874,9 +874,10 @@ unless ($Config{useithreads}) {
LEN = \d+
MAGIC = $ADDR
MG_VIRTUAL = &PL_vtbl_regexp
+ MG_PRIVATE = 1
MG_TYPE = PERL_MAGIC_study\\(G\\)
- MG_LEN = 1044
- MG_PTR = $ADDR "\\\\377\\\\377\\\\377\\\\377.*"
+ MG_LEN = 261
+ MG_PTR = $ADDR "\\\\377.*"
';
is(study beer, 1, "Our studies were successful");
@@ -903,10 +904,58 @@ unless ($Config{useithreads}) {
LEN = \d+
MAGIC = $ADDR
MG_VIRTUAL = &PL_vtbl_regexp
+ MG_PRIVATE = 1
MG_TYPE = PERL_MAGIC_study\\(G\\)
- MG_LEN = 1040
- MG_PTR = $ADDR "\\\\377\\\\377\\\\377\\\\377.*"
+ MG_LEN = 260
+ MG_PTR = $ADDR "\\\\377.*"
');
}
+{
+ my %z;
+ foreach (1, 254, 255, 65534, 65535) {
+ $z{$_} = "\0" x $_;
+ study $z{$_};
+ }
+ do_test('short studied representation', $z{1},
+'SV = PVMG\\($ADDR\\) at $ADDR
+ REFCNT = 1
+ FLAGS = \\(SMG,POK,pPOK,SCREAM\\)
+ IV = 0
+ NV = 0
+ PV = $ADDR "\\\\0"\\\0
+ CUR = 1
+ LEN = \d+
+ MAGIC = $ADDR
+ MG_VIRTUAL = &PL_vtbl_regexp
+ MG_PRIVATE = 1
+ MG_TYPE = PERL_MAGIC_study\\(G\\)
+ MG_LEN = 257
+ MG_PTR = $ADDR "\\\\0(?:\\\\377){256}"
+');
+
+ foreach ([254, 1], [255, 2], [65534, 2], [65535, 4]
+ ) {
+ my ($length, $bytes) = @$_;
+ my $quant = $length <= 32766 ? "{$length}" : '*';
+ do_test("studied representation for length $length", $z{$length},
+ sprintf
+'SV = PVMG\\($ADDR\\) at $ADDR
+ REFCNT = 1
+ FLAGS = \\(SMG,POK,pPOK,SCREAM\\)
+ IV = 0
+ NV = 0
+ PV = $ADDR "(?:\\\\0)%s"\\\0
+ CUR = %d
+ LEN = \d+
+ MAGIC = $ADDR
+ MG_VIRTUAL = &PL_vtbl_regexp
+ MG_PRIVATE = %d
+ MG_TYPE = PERL_MAGIC_study\\(G\\)
+ MG_LEN = %d
+ MG_PTR = $ADDR "\\\\0.*\\\\377"
+', $quant, $length, $bytes, (256 + $length) * $bytes);
+ }
+}
+
done_testing();
diff --git a/pod/perldelta.pod b/pod/perldelta.pod
index c65ee3e145..b06fc7a408 100644
--- a/pod/perldelta.pod
+++ b/pod/perldelta.pod
@@ -91,6 +91,12 @@ The implementation of C<s///r> makes one fewer copy of the scalar's value.
If a studied scalar is C<split> with a regex, the engine will now take
advantage of the C<study> data.
+=item *
+
+C<study> now uses considerably less memory for shorter strings. Strings shorter
+than 65535 characters use roughly half the memory than previously, strings
+shorter than 255 characters use roughly one quarter of the memory.
+
=back
=head1 Modules and Pragmata
diff --git a/pp.c b/pp.c
index f177165a0f..98d64827e0 100644
--- a/pp.c
+++ b/pp.c
@@ -707,10 +707,11 @@ PP(pp_study)
{
dVAR; dSP; dPOPss;
register unsigned char *s;
- U32 *sfirst;
- U32 *snext;
+ char *sfirst_raw;
STRLEN len;
MAGIC *mg = SvMAGICAL(sv) ? mg_find(sv, PERL_MAGIC_study) : NULL;
+ U8 quanta;
+ STRLEN size;
if (mg && SvSCREAM(sv))
RETPUSHYES;
@@ -724,28 +725,64 @@ PP(pp_study)
stringification. Also refuse to study an FBM scalar, as this gives
more flexibility in SV flag usage. No real-world code would ever
end up studying an FBM scalar, so this isn't a real pessimisation.
+ Endemic use of I32 in Perl_screaminstr makes it hard to safely push
+ the study length limit from I32_MAX to U32_MAX - 1.
*/
RETPUSHNO;
}
- Newx(sfirst, 256 + len, U32);
+ if (len < 0xFF) {
+ quanta = 1;
+ } else if (len < 0xFFFF) {
+ quanta = 2;
+ } else
+ quanta = 4;
- if (!sfirst)
+ size = (256 + len) * quanta;
+ sfirst_raw = (char *)safemalloc(size);
+
+ if (!sfirst_raw)
DIE(aTHX_ "do_study: out of memory");
SvSCREAM_on(sv);
if (!mg)
mg = sv_magicext(sv, NULL, PERL_MAGIC_study, &PL_vtbl_regexp, NULL, 0);
- mg->mg_ptr = (char *) sfirst;
- mg->mg_len = (256 + len) * sizeof(U32);
+ mg->mg_ptr = sfirst_raw;
+ mg->mg_len = size;
+ mg->mg_private = quanta;
- snext = sfirst + 256;
- memset(sfirst, ~0, 256 * sizeof(U32));
+ memset(sfirst_raw, ~0, 256 * quanta);
- while (len-- > 0) {
- const U8 ch = s[len];
- snext[len] = sfirst[ch];
- sfirst[ch] = len;
+ /* The assumption here is that most studied strings are fairly short, hence
+ the pain of the extra code is worth it, given the memory savings.
+ 80 character string, 336 bytes as U8, down from 1344 as U32
+ 800 character string, 2112 bytes as U16, down from 4224 as U32
+ */
+
+ if (quanta == 1) {
+ U8 *const sfirst = (U8 *)sfirst_raw;
+ U8 *const snext = sfirst + 256;
+ while (len-- > 0) {
+ const U8 ch = s[len];
+ snext[len] = sfirst[ch];
+ sfirst[ch] = len;
+ }
+ } else if (quanta == 2) {
+ U16 *const sfirst = (U16 *)sfirst_raw;
+ U16 *const snext = sfirst + 256;
+ while (len-- > 0) {
+ const U8 ch = s[len];
+ snext[len] = sfirst[ch];
+ sfirst[ch] = len;
+ }
+ } else {
+ U32 *const sfirst = (U32 *)sfirst_raw;
+ U32 *const snext = sfirst + 256;
+ while (len-- > 0) {
+ const U8 ch = s[len];
+ snext[len] = sfirst[ch];
+ sfirst[ch] = len;
+ }
}
RETPUSHYES;
diff --git a/regexec.c b/regexec.c
index 516bf9556c..99ac5b381c 100644
--- a/regexec.c
+++ b/regexec.c
@@ -696,12 +696,22 @@ Perl_re_intuit_start(pTHX_ REGEXP * const rx, SV *sv, char *strpos,
I32 p = -1; /* Internal iterator of scream. */
I32 * const pp = data ? data->scream_pos : &p;
const MAGIC *mg;
+ bool found = FALSE;
assert(SvMAGICAL(sv));
mg = mg_find(sv, PERL_MAGIC_study);
assert(mg);
- if (((U32 *)mg->mg_ptr)[BmRARE(check)] != (U32)~0
+ if (mg->mg_private == 1) {
+ found = ((U8 *)mg->mg_ptr)[BmRARE(check)] != (U8)~0;
+ } else if (mg->mg_private == 2) {
+ found = ((U16 *)mg->mg_ptr)[BmRARE(check)] != (U16)~0;
+ } else {
+ assert (mg->mg_private == 4);
+ found = ((U32 *)mg->mg_ptr)[BmRARE(check)] != (U32)~0;
+ }
+
+ if (found
|| ( BmRARE(check) == '\n'
&& (BmPREVIOUS(check) == SvCUR(check) - 1)
&& SvTAIL(check) ))
diff --git a/util.c b/util.c
index 4dbd15e689..fcfeda9814 100644
--- a/util.c
+++ b/util.c
@@ -854,7 +854,7 @@ Perl_screaminstr(pTHX_ SV *bigstr, SV *littlestr, I32 start_shift, I32 end_shift
{
dVAR;
register const unsigned char *big;
- U32 pos;
+ U32 pos = 0; /* hush a gcc warning */
register I32 previous;
register I32 first;
register const unsigned char *little;
@@ -862,9 +862,8 @@ Perl_screaminstr(pTHX_ SV *bigstr, SV *littlestr, I32 start_shift, I32 end_shift
register const unsigned char *littleend;
bool found = FALSE;
const MAGIC * mg;
- U32 *screamfirst;
- U32 *screamnext;
- U32 const nope = ~0;
+ const void *screamnext_raw = NULL; /* hush a gcc warning */
+ bool cant_find = FALSE; /* hush a gcc warning */
PERL_ARGS_ASSERT_SCREAMINSTR;
@@ -874,12 +873,37 @@ Perl_screaminstr(pTHX_ SV *bigstr, SV *littlestr, I32 start_shift, I32 end_shift
assert(SvTYPE(littlestr) == SVt_PVMG);
assert(SvVALID(littlestr));
- screamfirst = (U32 *)mg->mg_ptr;
- screamnext = screamfirst + 256;
+ if (mg->mg_private == 1) {
+ const U8 *const screamfirst = (U8 *)mg->mg_ptr;
+ const U8 *const screamnext = screamfirst + 256;
- pos = *old_posp == -1
- ? screamfirst[BmRARE(littlestr)] : screamnext[*old_posp];
- if (pos == nope) {
+ screamnext_raw = (const void *)screamnext;
+
+ pos = *old_posp == -1
+ ? screamfirst[BmRARE(littlestr)] : screamnext[*old_posp];
+ cant_find = pos == (U8)~0;
+ } else if (mg->mg_private == 2) {
+ const U16 *const screamfirst = (U16 *)mg->mg_ptr;
+ const U16 *const screamnext = screamfirst + 256;
+
+ screamnext_raw = (const void *)screamnext;
+
+ pos = *old_posp == -1
+ ? screamfirst[BmRARE(littlestr)] : screamnext[*old_posp];
+ cant_find = pos == (U16)~0;
+ } else if (mg->mg_private == 4) {
+ const U32 *const screamfirst = (U32 *)mg->mg_ptr;
+ const U32 *const screamnext = screamfirst + 256;
+
+ screamnext_raw = (const void *)screamnext;
+
+ pos = *old_posp == -1
+ ? screamfirst[BmRARE(littlestr)] : screamnext[*old_posp];
+ cant_find = pos == (U32)~0;
+ } else
+ Perl_croak(aTHX_ "panic: unknown study size %u", mg->mg_private);
+
+ if (cant_find) {
cant_find:
if ( BmRARE(littlestr) == '\n'
&& BmPREVIOUS(littlestr) == SvCUR(littlestr) - 1) {
@@ -910,13 +934,30 @@ Perl_screaminstr(pTHX_ SV *bigstr, SV *littlestr, I32 start_shift, I32 end_shift
#endif
return NULL;
}
- while ((I32)pos < previous + start_shift) {
- pos = screamnext[pos];
- if (pos == nope)
- goto cant_find;
+ if (mg->mg_private == 1) {
+ const U8 *const screamnext = (const U8 *const) screamnext_raw;
+ while ((I32)pos < previous + start_shift) {
+ pos = screamnext[pos];
+ if (pos == (U8)~0)
+ goto cant_find;
+ }
+ } else if (mg->mg_private == 2) {
+ const U16 *const screamnext = (const U16 *const) screamnext_raw;
+ while ((I32)pos < previous + start_shift) {
+ pos = screamnext[pos];
+ if (pos == (U16)~0)
+ goto cant_find;
+ }
+ } else if (mg->mg_private == 4) {
+ const U32 *const screamnext = (const U32 *const) screamnext_raw;
+ while ((I32)pos < previous + start_shift) {
+ pos = screamnext[pos];
+ if (pos == (U32)~0)
+ goto cant_find;
+ }
}
big -= previous;
- do {
+ while (1) {
if ((I32)pos >= stop_pos) break;
if (big[pos] == first) {
const unsigned char *s = little;
@@ -932,8 +973,20 @@ Perl_screaminstr(pTHX_ SV *bigstr, SV *littlestr, I32 start_shift, I32 end_shift
found = TRUE;
}
}
- pos = screamnext[pos];
- } while (pos != nope);
+ if (mg->mg_private == 1) {
+ pos = ((const U8 *const)screamnext_raw)[pos];
+ if (pos == (U8)~0)
+ break;
+ } else if (mg->mg_private == 2) {
+ pos = ((const U16 *const)screamnext_raw)[pos];
+ if (pos == (U16)~0)
+ break;
+ } else if (mg->mg_private == 4) {
+ pos = ((const U32 *const)screamnext_raw)[pos];
+ if (pos == (U32)~0)
+ break;
+ }
+ };
if (last && found)
return (char *)(big+(*old_posp));
check_tail: