On Tue, 6 Jun 2023 09:51:09 GMT, JoKern65 <d...@openjdk.org> wrote:
> The sys_thread_3() function contains an empty while loop, which by the
> standard can be optimized away. Please refer to discussion in
> https://github.com/llvm/llvm-project/issues/60622
> The xlc17 compiler is doing so, and IBM claims that they are following the
> standard and will not fix this on compiler side.
> So we have (at least) 3 ways to circumvent this behavior.
>
> 1. we can introduce the call of a nop library function, which will hinder the
> optimizer to throw away the loop (This is our proposed solution, but instead
> of a heavy looping thread, the result is a more or less idle thread):
> `#include <unistd.h>`
> `static void`
> `sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p)`
> `{`
> ` while (1) {`
> ` sleep(1);`
> ` }`
> `}`
>
> 2. We can make use of a volatile variable in the loop body which also hinders
> the optimizer to throw away the loop:
> `static void`
> `sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p)`
> `{`
> ` volatile int i = 1;`
> ` while (i) {`
> ` i += 2;`
> ` }`
> `}`
>
> 3. we can use the __attribute__ ((optnone)) modifier in the function
> declaration to suppress the optimization at all:
> `static void`
> `sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p) __attribute__
> ((optnone))`
> `{`
> ` while (1) {`
> ` }`
> `}`
>
> To make the third approach platform independent, we can implement it in the
> following way:
> In globalDefinitions.hpp
> `#ifndef OPTNONE`
> `#define OPTNONE`
> `#endif`
>
> In globalDefinitions_xlc.hpp
> `// optnone support`
> `//`
> `// To use if a function should not be optimized`
> `// Usage:`
> `// void* func(size_t size) OPTNONE {...}`
> `#define OPTNONE __attribute__(( optnone))`
>
> With this we can change libagentthr001.cpp in a platform independent way to
> `static void`
> `sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p) OPTNONE`
> `{`
> ` while (1) {`
> ` }`
> `}`
The solution to use sleep looks fine. Please address the comments and also
update the copyright year.
test/hotspot/jtreg/vmTestbase/nsk/jvmti/RunAgentThread/agentthr001/agentthr001.cpp
line 141:
> 139: sys_thread_3(jvmtiEnv* jvmti, JNIEnv* jni, void *p) {
> 140: while (1) {
> 141: sleep(1);
Maybe you could add a comment here that the call to sleep is necessary to avoid
the compiler optimization to elide the loop.
-------------
Changes requested by clanger (Reviewer).
PR Review: https://git.openjdk.org/jdk/pull/14330#pullrequestreview-1466167108
PR Review Comment: https://git.openjdk.org/jdk/pull/14330#discussion_r1220417459
