string: add and test an optimized memchr (DONE)

This version works on both SH3 and SH4.

CMakeLists.txt

@@ -167,7 +167,8 @@ endif()
 if(sh-generic IN_LIST TARGET_FOLDERS)
   list(APPEND SOURCES
     src/libc/setjmp/target/sh-generic/setjmp.S
-    src/libc/setjmp/target/sh-generic/longjmp.S)
+    src/libc/setjmp/target/sh-generic/longjmp.S
+    src/libc/string/target/sh-generic/memchr.S)
 endif()
 
 if(casiowin-fx IN_LIST TARGET_FOLDERS)

STATUS

@@ -14,6 +14,14 @@ taken from the C99 standard (ISO/IEC 9899:1999), section 7 ("Library").
   address of the function can be taken; don't rely on the macro being defined,
   as the user can remove it except in some special cases
 
+String functions (mainly in <string.h>) can use 4-byte accesses, and in doing
+so read up to 3 bytes after the end of the string if it is not padded (which
+malloc'd strings and literal strings both are, leaving only stack-allocated and
+statically-allocated ones). This allows important speed optimizations. The
+extra access cannot trigger memory protection because there is no valid memory
+less than 4 bytes before the end of any protection region. The extra access
+might trigger the UBC in very specific scenarios, but we don't really care.
+
 # Status
 
 In this file, every definition is classified in one of several implementation
@@ -102,7 +110,7 @@ DONE:       Function/symbol/macro is defined, builds, links, and is tested
 
 7.21 <string.h>
      7.21.2.1  memcpy: DONE
-     7.21.2.2  memmove: DONE
+     7.21.2.2  memmove: DONE    (Unoptimized: byte-by-byte)
 !    7.21.2.3  strcpy: TODO
 !    7.21.2.4  strncpy: TODO
 !    7.21.3.1  strcat: TODO
@@ -112,7 +120,7 @@ DONE:       Function/symbol/macro is defined, builds, links, and is tested
 !    7.21.4.3  strcoll: TODO
 !    7.21.4.4  strncmp: TODO
 !    7.21.4.5  strxfrm: TODO
-!    7.21.5.1  memchr: TODO
+     7.21.5.1  memchr: DONE
 !    7.21.5.2  strchr: TODO
 !    7.21.5.3  strcspn: TODO
 !    7.21.5.4  strpbrk: TODO
@@ -123,6 +131,13 @@ DONE:       Function/symbol/macro is defined, builds, links, and is tested
      7.21.6.1  memset: DONE
 !    7.21.6.2  strerror: TODO
 !    7.21.6.3  strlen: TODO
+     Extensions:
+     - strnlen: TODO
+     - strchrnul: TODO
+     - strcasecmp: TODO
+     - strncasecmp: TODO
+     - strdup: TODO
+     - strndup: TODO
 
 7.22 <tgmath.h> => GCC
 

src/libc/string/target/sh-generic/memchr.S

@@ -0,0 +1,68 @@
+.global _memchr
+.type _memchr, @function
+
+_memchr:
+	mov	r4, r0
+	exts.b	r5, r5
+
+	/* For small inputs, simply check bytes individually */
+	mov	#64, r2
+	cmp/hi	r6, r2
+	bt	.last
+
+.large:	/* Make a 4-byte version of r5 for cmp/str */
+	extu.b	r5, r3
+	swap.b	r3, r2
+	or	r3, r2
+	swap.w	r2, r3
+	or	r3, r2
+
+	/* First check 3 bytes to ensure we don't skip bytes when aligning */
+	mov.b	@r0+, r1
+	cmp/eq	r1, r5
+	bt	.end
+	mov.b	@r0+, r1
+	cmp/eq	r1, r5
+	bt	.end
+	mov.b	@r0+, r1
+	cmp/eq	r1, r5
+	bt	.end
+
+	/* Align to a 4-byte boundary */
+	shlr2	r0
+	shll2	r0
+	add	r4, r6
+	sub	r0, r6
+
+	mov	r6, r7
+	shlr2	r7
+	mov	#3, r3
+	and	r3, r6
+
+	/* Read longwords */
+1:	mov.l	@r0+, r1
+	cmp/str	r1, r2
+	bt	.found
+	dt	r7
+	bf	1b
+
+.last:	/* Don't read if there are no bytes left */
+	tst	r6, r6
+	bt	.none
+
+2:	mov.b	@r0+, r1
+	cmp/eq	r1, r5
+	bt	.end
+	dt	r6
+	bf	2b
+
+.none:	rts
+	mov	#0, r0
+
+.found:	/* Go back to find out which of the last 4 bytes is r5 */
+	add	#-4, r0
+	bra	2b
+	mov	#4, r6
+
+.end:	rts
+	add	#-1, r0