Handle divide overflow on x86_64
Andrew Haley
aph@redhat.com
Mon Apr 10 15:51:00 GMT 2006
I never bothered to enable the divide overflow exception on x86_64
because I didn't think it was an important component of total
runtimes. To my surprise, I discovered that under heavy application
server load, the divide subroutines account for 4% of total CPU time.
samples % image name app name symbol name
176796 2.1502 libgcj.so.7.0.0 gij _Jv_remI
104605 1.2722 libgcj.so.7.0.0 gij _Jv_divI
57126 0.6948 libgcj.so.7.0.0 gij _Jv_remJ
392 0.0048 libgcj.so.7.0.0 gij _Jv_divJ
It seems that divide and remainder are used a great deal, particularly
for hash tables, so it's worth inlining them rather than using library
routines.
Andrew.
2006-04-10 Andrew Haley <aph@redhat.com>
* include/x86_64-signal.h (HANDLE_DIVIDE_OVERFLOW): New.
(SIGNAL_HANDLER): Mark arg as unused.
* configure.host (x86_64-* DIVIDESPEC): Use
fno-use-divide-subroutine.
Index: configure.host
===================================================================
--- configure.host (revision 112821)
+++ configure.host (working copy)
@@ -115,7 +115,7 @@
libgcj_flags="${libgcj_flags} -fomit-frame-pointer"
libgcj_cxxflags=
libgcj_cflags=
- DIVIDESPEC=-f%{m32:no-}use-divide-subroutine
+ DIVIDESPEC=-fno-use-divide-subroutine
enable_hash_synchronization_default=yes
slow_pthread_self=yes
libgcj_interpreter=yes
Index: include/x86_64-signal.h
===================================================================
--- include/x86_64-signal.h (revision 112821)
+++ include/x86_64-signal.h (working copy)
@@ -19,9 +19,88 @@
#include <sys/syscall.h>
#define HANDLE_SEGV 1
+#define HANDLE_FPE 1
-#define SIGNAL_HANDLER(_name) \
-static void _Jv_##_name (int, siginfo_t *, void *_p)
+#define SIGNAL_HANDLER(_name) \
+static void _Jv_##_name (int, siginfo_t *, \
+ void *_p __attribute__ ((__unused__)))
+
+#define HANDLE_DIVIDE_OVERFLOW \
+do \
+{ \
+ struct ucontext *_uc = (struct ucontext *)_p; \
+ volatile struct sigcontext *_sc = (struct sigcontext *) &_uc->uc_mcontext; \
+ \
+ register unsigned char *_rip = (unsigned char *)_sc->rip; \
+ \
+ /* According to the JVM spec, "if the dividend is the negative \
+ * integer of largest possible magnitude for the type and the \
+ * divisor is -1, then overflow occurs and the result is equal to \
+ * the dividend. Despite the overflow, no exception occurs". \
+ \
+ * We handle this by inspecting the instruction which generated the \
+ * signal and advancing ip to point to the following instruction. \
+ * As the instructions are variable length it is necessary to do a \
+ * little calculation to figure out where the following instruction \
+ * actually is. \
+ \
+ */ \
+ \
+ bool _is_64_bit = false; \
+ \
+ if ((_rip[0] & 0xf0) == 0x40) /* REX byte present. */ \
+ { \
+ unsigned char _rex = _rip[0] & 0x0f; \
+ _is_64_bit = (_rex & 0x08) != 0; \
+ _rip++; \
+ } \
+ \
+ /* Detect a signed division of Integer.MIN_VALUE or Long.MIN_VALUE. */ \
+ if (_rip[0] == 0xf7) \
+ { \
+ bool _min_value_dividend = false; \
+ unsigned char _modrm = _rip[1]; \
+ \
+ if (((_modrm >> 3) & 7) == 7) \
+ { \
+ if (_is_64_bit) \
+ _min_value_dividend = (_sc->rax == 0x8000000000000000L); \
+ else \
+ _min_value_dividend = ((_sc->rax & 0xffffffff) == 0x80000000); \
+ } \
+ \
+ if (_min_value_dividend) \
+ { \
+ unsigned char _rm = _modrm & 7; \
+ _sc->rdx = 0; /* the remainder is zero */ \
+ switch (_modrm >> 6) \
+ { \
+ case 0: /* register indirect */ \
+ if (_rm == 5) /* 32-bit displacement */ \
+ _rip += 4; \
+ if (_rm == 4) /* A SIB byte follows the ModR/M byte */ \
+ _rip += 1; \
+ break; \
+ case 1: /* register indirect + 8-bit displacement */ \
+ _rip += 1; \
+ if (_rm == 4) /* A SIB byte follows the ModR/M byte */ \
+ _rip += 1; \
+ break; \
+ case 2: /* register indirect + 32-bit displacement */ \
+ _rip += 4; \
+ if (_rm == 4) /* A SIB byte follows the ModR/M byte */ \
+ _rip += 1; \
+ break; \
+ case 3: \
+ break; \
+ } \
+ _rip += 2; \
+ _sc->rip = (unsigned long)_rip; \
+ return; \
+ } \
+ } \
+} \
+while (0)
extern "C"
{
@@ -65,12 +144,29 @@
} \
while (0)
-/* We use syscall(SYS_rt_sigaction) in INIT_SEGV instead of
- * sigaction() because on some systems the pthreads wrappers for
- * signal handlers are not compiled with unwind information, so it's
- * not possible to unwind through them. This is a problem that will
- * go away if all systems ever have pthreads libraries that are
- * compiled with unwind info. */
+#define INIT_FPE \
+do \
+ { \
+ struct kernel_sigaction act; \
+ act.k_sa_sigaction = _Jv_catch_fpe; \
+ sigemptyset (&act.k_sa_mask); \
+ act.k_sa_flags = SA_SIGINFO|0x4000000; \
+ act.k_sa_restorer = restore_rt; \
+ syscall (SYS_rt_sigaction, SIGFPE, &act, NULL, _NSIG / 8); \
+ } \
+while (0)
+
+/* You might wonder why we use syscall(SYS_sigaction) in INIT_FPE
+ * instead of the standard sigaction(). This is necessary because of
+ * the shenanigans above where we increment the PC saved in the
+ * context and then return. This trick will only work when we are
+ * called _directly_ by the kernel, because linuxthreads wraps signal
+ * handlers and its wrappers do not copy the sigcontext struct back
+ * when returning from a signal handler. If we return from our divide
+ * handler to a linuxthreads wrapper, we will lose the PC adjustment
+ * we made and return to the faulting instruction again. Using
+ * syscall(SYS_sigaction) causes our handler to be called directly
+ * by the kernel, bypassing any wrappers. */
#endif /* JAVA_SIGNAL_H */
More information about the Java
mailing list