[LV-k25-commit] kernel-2.5/include/asm-vax bitops.h,1.13,1.14

[Jan-Benedict G. <jb...@us...>]

Update of /cvsroot/linux-vax/kernel-2.5/include/asm-vax
In directory sc8-pr-cvs1.sourceforge.net:/tmp/cvs-serv18641

Modified Files:
	bitops.h 
Log Message:
- I've now taken Kenn's suggestion to force ffs's base operand into
  memory. mm/bootmem.c (as of 2.5.9) calls ffs() with a constant argument,
  which (with "g" constraint) may show up as an immediate value -- which
  the VAX CPUs as well as gas refuse to use.
- Since many (at least mostly all "important) architectures (including
  i386) use a C-only implementation for ffs(), I guess this isn't all that
  bad wrt. performance.


Index: bitops.h
===================================================================
RCS file: /cvsroot/linux-vax/kernel-2.5/include/asm-vax/bitops.h,v
retrieving revision 1.13
retrieving revision 1.14
diff -u -d -r1.13 -r1.14
--- bitops.h	2 Sep 2004 22:41:49 -0000	1.13
+++ bitops.h	19 Nov 2004 13:08:56 -0000	1.14
@@ -393,12 +393,18 @@
 {
 	int r;
 
+	/*
+	 * FIXME: We're forcing the input to be in memory, but it would
+	 * be okay if it was in a register, too. However, if it's "g",
+	 * gcc may decide to put in an immediate value, which is not accepted
+	 * by CPU (and gas).
+	 */
 	__asm__("ffs $0, $32, %1, %0\n"
 		"bnequ 1f\n"
 		"movl $-1,%0\n"
 		"1:" 
 		: "=ir" (r) 
-		: "g" (x));
+		: "m" (x));
 	return r+1;
 }
 #endif