[IA64] fix fls()

The ia64-version of fls() never worked as intended (the bitnumbering
was off by 1 and fls(0) was undefined).  This patch fixes the problem
by using a popcnt-based fls(), which on McKinley-derived cores is
slightly faster than both ia64_fls() and generic_fls().  The resulting
code, however, is bigger (7-8 bundles instead of about 3 bundles).
Also switch ia64_popcnt() to __builtin_popcountl() for GCC v3.4 or
newer since the compiler can predicate that and schedule it better.

Thanks to Simon Derr and Matt Mackall for tracking down this bug.

Signed-off-by: David Mosberger-Tang <davidm@hpl.hp.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>

1 files changed, 7 insertions, 3 deletions

diff --git a/include/asm-ia64/gcc_intrin.h b/include/asm-ia64/gcc_intrin.h
index 7c357dfbae5..4fb4e439b05 100644
--- a/include/asm-ia64/gcc_intrin.h
+++ b/include/asm-ia64/gcc_intrin.h
@@ -133,13 +133,17 @@ register unsigned long ia64_r13 asm ("r13") __attribute_used__;
         ia64_intri_res;                                                         \
 })
 
-#define ia64_popcnt(x)                                          \
-({                                                              \
+#if __GNUC__ >= 4 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 4)
+# define ia64_popcnt(x)         __builtin_popcountl(x)
+#else
+# define ia64_popcnt(x)                                         \
+  ({                                                            \
         __u64 ia64_intri_res;                                   \
         asm ("popcnt %0=%1" : "=r" (ia64_intri_res) : "r" (x)); \
                                                                 \
         ia64_intri_res;                                         \
-})
+  })
+#endif
 
 #define ia64_getf_exp(x)                                        \
 ({                                                              \