This is the mail archive of the
fortran@gcc.gnu.org
mailing list for the GNU Fortran project.
Re: [Patch, Fortran] Experimental LEN_TRIM and TRIM patch
- From: Daniel Kraft <d at domob dot eu>
- To: Fortran List <fortran at gcc dot gnu dot org>
- Cc: gcc-patches <gcc-patches at gcc dot gnu dot org>
- Date: Thu, 12 Feb 2009 16:57:49 +0100
- Subject: Re: [Patch, Fortran] Experimental LEN_TRIM and TRIM patch
- References: <49933925.90909@domob.eu>
Slightly improved patch and ChangeLog.
Daniel Kraft wrote:
Hi,
here's an experimental patch to tune the LEN_TRIM and TRIM intrinsics.
In my own experience, I usually use TRIM (or LEN_TRIM) in combination
with strings allocated too large (because no dynamic length is
possible), so that there are usually lots of blanks at the end.
In this case (and for KIND=1 strings), instead of checking each and
every character if it is a blank, we can check them in chunks of 4 (or
8) by comparing them as unsigned long integers. That's what the patch
does, as well as add a test-case and re-use LEN_TRIM for TRIM instead of
duplicating the code there.
For the attached (extreme) test-case, the patch really gives nearly 4x
speed-up as expected for a best case; if the full string length is only
marginally larger than the trimmed one (like one trailing blank), the
patch slows down up to a factor of 2. But as soon as there are 4-5
trailing blanks, the new version is faster.
Comments welcome!
I guess there are configure-set macros telling me if unsigned long is 4
or 8 bytes on the machine, so we can make the blank_long initialization
work on 64-bit, too...
Yours,
Daniel
PS: BTW, while gfortran does not complain about CHARACTER(LEN=0) (don't
know if this is legal), for LEN=-1 one gets "Warning: CHARACTER variable
has *zero* length" which sounds quite a little confusing.
--
Done: Arc-Bar-Cav-Ran-Rog-Sam-Tou-Val-Wiz
To go: Hea-Kni-Mon-Pri
Index: libgfortran/intrinsics/string_intrinsics_inc.c
===================================================================
--- libgfortran/intrinsics/string_intrinsics_inc.c (revision 144104)
+++ libgfortran/intrinsics/string_intrinsics_inc.c (working copy)
@@ -165,15 +165,7 @@ void
string_trim (gfc_charlen_type *len, CHARTYPE **dest, gfc_charlen_type slen,
const CHARTYPE *src)
{
- gfc_charlen_type i;
-
- /* Determine length of result string. */
- for (i = slen - 1; i >= 0; i--)
- {
- if (src[i] != ' ')
- break;
- }
- *len = i + 1;
+ *len = string_len_trim (slen, src);
if (*len == 0)
*dest = &zero_length_string;
@@ -193,13 +185,52 @@ string_trim (gfc_charlen_type *len, CHAR
gfc_charlen_type
string_len_trim (gfc_charlen_type len, const CHARTYPE *s)
{
+ const gfc_charlen_type long_len = (gfc_charlen_type) sizeof (unsigned long);
gfc_charlen_type i;
- for (i = len - 1; i >= 0; i--)
- {
- if (s[i] != ' ')
- break;
+ i = len - 1;
+
+ /* If we've got the standard (KIND=1) character type, we scan the string in
+ long word chunks to speed it up (until a long word is hit that does not
+ consist of ' 's). */
+ if (sizeof (CHARTYPE) == 1 && i >= long_len)
+ {
+ int starting;
+ unsigned long blank_longword;
+
+ /* Handle the first characters until we're aligned on a long word
+ boundary. Actually, s + i + 1 must be properly aligned, because
+ s + i will be the last byte of a long word read. */
+ starting = ((unsigned long) (s + i + 1)) % long_len;
+ i -= starting;
+ for (; starting > 0; --starting)
+ if (s[i + starting] != ' ')
+ return i + starting + 1;
+
+ /* Handle the others in a batch until first non-blank long word is
+ found. Here again, s + i is the last byte of the current chunk,
+ to it starts at s + i - sizeof (long) + 1. */
+
+ /* XXX: Do this reliably, i.e. working for 64-bit systems, too. */
+ blank_longword = 0x20202020L;
+
+ while (i >= long_len)
+ {
+ i -= long_len;
+ if (*((unsigned long*) (s + i + 1)) != blank_longword)
+ {
+ i += long_len;
+ break;
+ }
+ }
+
+ /* Now continue for the last characters with naive approach below. */
+ assert (i >= 0);
}
+
+ /* Simply look for the first non-blank character. */
+ while (i >= 0 && s[i] == ' ')
+ --i;
return i + 1;
}
Index: libgfortran/intrinsics/string_intrinsics.c
===================================================================
--- libgfortran/intrinsics/string_intrinsics.c (revision 144104)
+++ libgfortran/intrinsics/string_intrinsics.c (working copy)
@@ -39,6 +39,7 @@ Boston, MA 02110-1301, USA. */
#include <stdlib.h>
#include <string.h>
+#include <assert.h>
/* Helper function to set parts of wide strings to a constant (usually
Index: gcc/testsuite/gfortran.dg/trim_1.f90
===================================================================
--- gcc/testsuite/gfortran.dg/trim_1.f90 (revision 0)
+++ gcc/testsuite/gfortran.dg/trim_1.f90 (revision 0)
@@ -0,0 +1,41 @@
+! { dg-do run }
+
+! Torture-test TRIM and LEN_TRIM for correctness.
+
+
+! Given a total string length and a trimmed length, construct an
+! appropriate string and check gfortran gets it right.
+
+SUBROUTINE check_trim (full_len, trimmed_len)
+ IMPLICIT NONE
+ INTEGER, INTENT(IN) :: full_len, trimmed_len
+ CHARACTER(LEN=full_len) :: string
+
+ string = ""
+ IF (trimmed_len > 0) THEN
+ string(trimmed_len:trimmed_len) = "x"
+ END IF
+
+ IF (LEN (string) /= full_len &
+ .OR. LEN_TRIM (string) /= trimmed_len &
+ .OR. LEN (TRIM (string)) /= trimmed_len &
+ .OR. TRIM (string) /= string (1:trimmed_len)) THEN
+ PRINT *, full_len, trimmed_len
+ PRINT *, LEN (string), LEN_TRIM (string)
+ CALL abort ()
+ END IF
+END SUBROUTINE check_trim
+
+
+! The main program, check with various combinations.
+
+PROGRAM main
+ IMPLICIT NONE
+ INTEGER :: i, j
+
+ DO i = 0, 20
+ DO j = 0, i
+ CALL check_trim (i, j)
+ END DO
+ END DO
+END PROGRAM main
2009-02-12 Daniel Kraft <d@domob.eu>
* intrinsics/string_intrinsics.c: #include <assert.h>
* intrinsics/string_intrinsics_inc.c (string_trim): Use string_len_trim
instead of calculating the length directly.
(string_len_trim): For KIND=1, speed search up.
2009-02-12 Daniel Kraft <d@domob.eu>
* gfortran.dg/trim_1.f90: New test.