Hi Yasumasa,
On 12/03/2015 9:58 PM, Yasumasa Suenaga wrote:
Hi all,
I tried to use NMT with details option on OpenJDK7 on RHEL6.6, but I got
address at AllocateHeap() as malloc() caller.
I checked symbol in libjvm.so <http://libjvm.so/> in OracleJDK8u40 Linux
x64, it has AllocateHeap()
symbol.
AllocateHeap() is defined as inline function, and it gives CURRENT_PC to
os::malloc(). I guess that implementation expects AllocateHeap() will be
inlined.
It seems so.
It may occur with GCC (g++) optimization only, however I want to fix it to
analyze native memory with NMT on Linux.
According to the docs [1]:
"GCC does not inline any functions when not optimizing unless you
specify the ‘always_inline’ attribute for the function"
I applied patch as below. This patch makes AllocateHeap() as inline
function.
--------------
diff -r af3b0db91659 src/share/vm/memory/allocation.inline.hpp
--- a/src/share/vm/memory/allocation.inline.hpp Mon Mar 09 09:30:16 2015
-0700
+++ b/src/share/vm/memory/allocation.inline.hpp Thu Mar 12 20:45:57 2015
+0900
@@ -62,11 +62,18 @@
}
return p;
}
+
+#ifdef __GNUC__
+__attribute__((always_inline))
+#endif
I dislike seeing the gcc specific directives in common code. I'm
wondering whether we should perhaps only use CURRENT_PC in product (and
optimized?) builds and use CALLER_PC otherwise. That would be imperfect
of course It also makes me wonder whether the inlining is occurring as
expected on other platforms.
I'd like to get other people's views on this.
Thanks,
David
[1] https://gcc.gnu.org/onlinedocs/gcc/Inline.html
inline char* AllocateHeap(size_t size, MEMFLAGS flags,
AllocFailType alloc_failmode = AllocFailStrategy::EXIT_OOM) {
return AllocateHeap(size, flags, CURRENT_PC, alloc_failmode);
}
+#ifdef __GNUC__
+__attribute__((always_inline))
+#endif
inline char* ReallocateHeap(char *old, size_t size, MEMFLAGS flag,
AllocFailType alloc_failmode = AllocFailStrategy::EXIT_OOM) {
char* p = (char*) os::realloc(old, size, flag, CURRENT_PC);
--------------
If this patch is accepted, I will file it to JBS and will upload webrev.
Thanks,
Yasumasa
