* lib/rawmemchr.c (rawmemchr) [__CHERI__]: Use memchr instead of
one-byte reads. This sped up a simple benchmark (rawmemchr on
100,000 bytes) by 6x on a Research Morello SoC r0p0 on CheriBSD 14.
[!__CHERI__]: Use sizeof, not alignof, as better alignment
should help performance a bit on some platforms.
* modules/rawmemchr (Depends-on): Remove alignasof.
---
ChangeLog | 10 ++++++++++
lib/rawmemchr.c | 27 +++++++++++++++++----------
modules/rawmemchr | 1 -
3 files changed, 27 insertions(+), 11 deletions(-)
diff --git a/ChangeLog b/ChangeLog
index c7fa04a173..09f0577925 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,13 @@
+2023-11-12 Paul Eggert <egg...@cs.ucla.edu>
+
+ rawmemchr: speed up, particularly on CHERI
+ * lib/rawmemchr.c (rawmemchr) [__CHERI__]: Use memchr instead of
+ one-byte reads. This sped up a simple benchmark (rawmemchr on
+ 100,000 bytes) by 6x on a Research Morello SoC r0p0 on CheriBSD 14.
+ [!__CHERI__]: Use sizeof, not alignof, as better alignment
+ should help performance a bit on some platforms.
+ * modules/rawmemchr (Depends-on): Remove alignasof.
+
2023-11-12 Bruno Haible <br...@clisp.org>
jit/cache: New module.
diff --git a/lib/rawmemchr.c b/lib/rawmemchr.c
index 137d7282a6..9386ffc628 100644
--- a/lib/rawmemchr.c
+++ b/lib/rawmemchr.c
@@ -19,7 +19,7 @@
/* Specification. */
#include <string.h>
-/* A function definition is only needed if HAVE_RAWMEMCHR is not defined. */
+/* A function definition is needed only if HAVE_RAWMEMCHR is not defined. */
#if !HAVE_RAWMEMCHR
# include <limits.h>
@@ -30,24 +30,30 @@
void *
rawmemchr (const void *s, int c_in)
{
-#ifdef __CHERI__
- /* Most architectures let you read an aligned word, even if the unsigned char
- array at S ends in the middle of the word. However, CHERI does not. */
- typedef unsigned char longword;
-#else
- /* Change this typedef to experiment with performance. */
+# ifdef __CHERI__
+ /* Most architectures let you read an aligned word,
+ even if the unsigned char array at S ends in the middle of the word.
+ However CHERI does not, so call memchr
+ with the underlying object's remaining length.
+ This cannot return NULL if S points to a C_IN-terminated array.
+ Use builtins rather than including <cheri.h> which is less stable. */
+ return memchr (s, c_in, (__builtin_cheri_length_get (s)
+ - __builtin_cheri_offset_get (s)));
+# else
+
+ /* You can change this typedef to experiment with performance. */
typedef uintptr_t longword;
/* Verify that the longword type lacks padding bits. */
static_assert (UINTPTR_WIDTH == UCHAR_WIDTH * sizeof (uintptr_t));
-#endif
const unsigned char *char_ptr;
unsigned char c = c_in;
/* Handle the first few bytes by reading one byte at a time.
- Do this until CHAR_PTR is aligned on a longword boundary. */
+ Do this until CHAR_PTR is aligned on a natural longword boundary,
+ as using alignof (longword) might be slower. */
for (char_ptr = (const unsigned char *) s;
- (uintptr_t) char_ptr % alignof (longword) != 0;
+ (uintptr_t) char_ptr % sizeof (longword) != 0;
++char_ptr)
if (*char_ptr == c)
return (void *) char_ptr;
@@ -123,6 +129,7 @@ rawmemchr (const void *s, int c_in)
while (*char_ptr != c)
char_ptr++;
return (void *) char_ptr;
+# endif
}
#endif
diff --git a/modules/rawmemchr b/modules/rawmemchr
index 5f3eeaf2d0..aba5088651 100644
--- a/modules/rawmemchr
+++ b/modules/rawmemchr
@@ -7,7 +7,6 @@ lib/rawmemchr.valgrind
m4/rawmemchr.m4
Depends-on:
-alignasof
assert-h
extensions
stdint
--
2.40.1
