[libc] Add inline assembly implementation of flsl() using BSR instruction

Signed-off-by: Michael Brown <mcb30@ipxe.org>

3 files changed, 33 insertions, 15 deletions

diff --git a/src/arch/x86/include/bits/strings.h b/src/arch/x86/include/bits/strings.h
new file mode 100644
index 000000000..62e0fbfba
--- /dev/null
+++ b/src/arch/x86/include/bits/strings.h
@@ -0,0 +1,31 @@
+#ifndef _BITS_STRINGS_H
+#define _BITS_STRINGS_H
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+/**
+ * Find last (i.e. most significant) set bit
+ *
+ * @v value		Value
+ * @ret msb		Most significant bit set in value (LSB=1), or zero
+ */
+static inline __attribute__ (( always_inline )) int __flsl ( long value ) {
+	long msb_minus_one;
+
+	/* If the input value is zero, the BSR instruction returns
+	 * ZF=1 and leaves an undefined value in the output register.
+	 * Perform this check in C rather than asm so that it can be
+	 * omitted in cases where the compiler is able to prove that
+	 * the input is non-zero.
+	 */
+	if ( value ) {
+		__asm__ ( "bsr %1, %0"
+			  : "=r" ( msb_minus_one )
+			  : "rm" ( value ) );
+		return ( msb_minus_one + 1 );
+	} else {
+		return 0;
+	}
+}
+
+#endif /* _BITS_STRINGS_H */
diff --git a/src/core/bitops.c b/src/core/bitops.c
deleted file mode 100644
index 1bca9e47b..000000000
--- a/src/core/bitops.c
+++ /dev/null
@@ -1,13 +0,0 @@
-#include <strings.h>
-
-FILE_LICENCE ( GPL2_OR_LATER );
-
-int __flsl ( long x ) {
-	unsigned long value = x;
-	int ls = 0;
-
-	for ( ls = 0 ; value ; ls++ ) {
-		value >>= 1;
-	}
-	return ls;
-}
diff --git a/src/include/strings.h b/src/include/strings.h
index c7063d682..924a084fc 100644
--- a/src/include/strings.h
+++ b/src/include/strings.h
@@ -5,6 +5,7 @@ FILE_LICENCE ( GPL2_OR_LATER );
 
 #include <limits.h>
 #include <string.h>
+#include <bits/strings.h>
 
 static inline __attribute__ (( always_inline )) int
 __constant_flsl ( unsigned long x ) {
@@ -42,8 +43,7 @@ __constant_flsl ( unsigned long x ) {
 	return r;
 }
 
-/* We don't actually have these functions yet */
-extern int __flsl ( long x );
+int __flsl ( long x );
 
 #define flsl( x ) \
 	( __builtin_constant_p ( x ) ? __constant_flsl ( x ) : __flsl ( x ) )