Re: HAVE_FAST_UNALIGNED_ACCESS

["Mattias Engdegård" <mattias.engdegard@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 829 bytes --]

30 mars 2023 kl. 13.38 skrev Vibhav Pant <vibhavp@gmail.com>:

> I had recently made a few modifications to this on master. There,
> Fstring_lessp in uses the macro UNALIGNED_LOAD_SIZE instead, which is
> defined to __sanitizer_unaligned_loadXX in lisp.h if:
> 
> * We're building with AddressSaniziter,
> * <sanitizer/common_interface_defs.h> is available, and
> * USE_SANITIZER_UNALIGNED_LOAD is defined.

Thank you for making this improvement earlier -- I think using memcpy would subsume __sanitizer_unaligned_loadXX, so I'm going with that in my patch.
Another approach would be to disable this poor man's vectorisation when the sanitiser is enabled but that shouldn't be necessary.

The vectorisation is still only made on specific platforms, of course.

Eli, would this patch be acceptable for emacs-29?


[-- Attachment #2: string-lessp-uni-multi.diff --]
[-- Type: application/octet-stream, Size: 942 bytes --]

diff --git a/src/fns.c b/src/fns.c
index f5f222c39c..292d6fc405 100644
--- a/src/fns.c
+++ b/src/fns.c
@@ -514,7 +514,23 @@ DEFUN ("string-lessp", Fstring_lessp, Sstring_lessp, 2, 2, 0,
   else
     {
       /* string1 unibyte, string2 multibyte */
-      ptrdiff_t i1 = 0, i2 = 0, i2_byte = 0;
+
+      ptrdiff_t nb1 = n;
+      ptrdiff_t nb2 = SBYTES (string2);
+      ptrdiff_t nb = min (nb1, nb2);
+
+      /* Skip identical ASCII-only prefixes, a word at a time.  */
+      typedef size_t word_t;
+      int ws = sizeof (word_t);
+      word_t msbits = (word_t)0x8080808080808080;
+      const word_t *w1 = (const word_t *) SDATA (string1);
+      const word_t *w2 = (const word_t *) SDATA (string2);
+      ptrdiff_t b = 0;
+      while (b < nb - ws + 1 && w1[b / ws] == w2[b / ws]
+	     && !(w1[b / ws] & msbits))
+	b += ws;
+
+      ptrdiff_t i1 = b, i2 = b, i2_byte = b;
       while (i1 < n)
 	{
 	  int c1 = SREF (string1, i1++);