Libitm on POWER hardware looks for the "htm" bit in AT_HWCAP2 to determine
whether it can use HTM when executing code within __transaction_atomic
code blocks. However, on newer hardware and kernels, the "htm" bit is no
longer set and instead the "htm-no-suspend" bit is set, so we currently
don't use HTM on new hw and kernels. The following patch adds support
for htm-no-suspend to libitm. I have also added code to use the
__builtin_cpu_supports() builtin if it is available, since that is
much faster than using the getauxval libc call.
This passed bootstrap and regtesting with no errors and someone within
IBM how had a POWER9 box with a newish kernel how ran into the problem
confirmed it works for his test case.
Ok for mainline? Should be backport this?
Peter
* config/powerpc/target.h (PPC_FEATURE2_HTM_NO_SUSPEND): Conditionally
define.
(htm_available): Add support for PPC_FEATURE2_HTM_NO_SUSPEND.
Use __builtin_cpu_supports if available.
Index: libitm/config/powerpc/target.h
===================================================================
--- libitm/config/powerpc/target.h (revision 267062)
+++ libitm/config/powerpc/target.h (working copy)
@@ -26,6 +26,11 @@
#include <sys/auxv.h>
#endif
+/* This is a fairly new feature bit, so handle it not being defined. */
+#ifndef PPC_FEATURE2_HTM_NO_SUSPEND
+# define PPC_FEATURE2_HTM_NO_SUSPEND 0
+#endif
+
namespace GTM HIDDEN {
typedef int v128 __attribute__((vector_size(16), may_alias, aligned(16)));
@@ -81,7 +86,16 @@ cpu_relax (void)
static inline bool
htm_available (void)
{
- return (getauxval (AT_HWCAP2) & PPC_FEATURE2_HAS_HTM) ? true : false;
+#ifdef __BUILTIN_CPU_SUPPORTS__
+ if (__builtin_cpu_supports ("htm-no-suspend")
+ || __builtin_cpu_supports ("htm"))
+ return true;
+#else
+ if (getauxval (AT_HWCAP2)
+ & (PPC_FEATURE2_HAS_HTM | PPC_FEATURE2_HTM_NO_SUSPEND))
+ return true;
+#endif
+ return false;
}
static inline uint32_t
