RFR: 8265218: trace_method_handle_stub fails to find calling frame on x86

[Vladimir Ivanov]

On Mon, 19 Apr 2021 20:02:30 GMT, Daniel D. Daugherty <dcubed at openjdk.org> wrote:

>> `trace_method_handle_stub` on x86 tries to dump the frame which initiated the tracing, but fails to find the proper frame when it is the immediate caller.  
>> 
>> The problem is `os::current_frame()` returns the caller (instead of the current frame as the name suggests) and then `os::get_sender_for_C_frame()` steps over one more frame. So, if there are no intermediate frames present, it steps over the desired frame.
>> 
>> Proposed fix is to start the search from the frame reported by `os::current_frame()`.
>> 
>> Testing:
>> - [x] failing test
>> - [x] hs-precheckin-comp,hs-tier1,hs-tier2 w/ `-Xlog:methodhandles=trace`
>
> src/hotspot/cpu/x86/methodHandles_x86.cpp line 565:
> 
>> 563:         // Assumes saved_regs comes from a pusha in the trace_calling_frame.
>> 564:         assert(cur_frame.sp() < saved_regs, "registers not saved on stack ?");
>> 565:         frame trace_calling_frame = cur_frame;
> 
> Since I'm expecting `cur_frame` to be `trace_method_handle_stub`,
> the original code makes sense to me since I want the sender for
> `trace_method_handle_stub`.
> 
> If `cur_frame` is not `trace_method_handle_stub`, then I definitely
> don't understand the code at L558 so I definitely don't understand
> the comment that you added either.

No, unfortunately, it's not `trace_method_handle_stub`. It has `trace_method_handle_stub` call in tail position, so even when inlining doesn't happen, the call may be turned into a tail-call. And that's what actually happens after 8264358. So, there's no stack frame for `trace_method_handle_stub` left when `os::current_frame()` is performed. 

The next frame under top frame corresponds to the caller of `trace_method_handle_stub` in generated code (which contains registers and has to be printed). And it's the one which is returned by `os::current_frame()`.

-------------

PR: https://git.openjdk.java.net/jdk/pull/3554