mcount: Restore/reset rstack only when the return address is hijacked

[jyf111]

When specifying estimate_return, the return address is not hijacked. This patch ensures that mcount_rstack_reset_exception() is not called in this case, to avoid recording incorrect end times.

Before:
I got an inverted time bug in estimate_return mode (it only occurs when using dynamic tracing).
I found that mcount_rstack_reset_exception() was called unexpectedly, which sets the end time using mcount_gettime(), but the end time should be estimated by mcount_rstack_inject_return().
This PR should fix this small problem.

$ clang++ tests/s-exception3.cpp -o test/exp
$ uftrace -P. -e test/exp
# DURATION     TID     FUNCTION
            [117574] | main() {
            [117574] |   A::A() {
            [117574] |     foo1() {
            [117574] |       foo2() {
            [117574] |         foo4() {
   5.050 us [117574] |           C::C();
            [117574] |           foo5() {
 129.894 us [117574] |             __cxa_allocate_exception();
  81.546 us [117574] |           } /* foo5 */
  36.698 us [117574] |           C::~C();
 515.376 us [117574] |         } /* foo4 */
 547.074 us [117574] |       } /* foo2 */
                     |          /* inverted time: broken data? */
 289.136 us [117574] |     } /* foo1 */
  13.199 us [117574] |     B::B();
  17.199 us [117574] |     B::~B();
                     |        /* inverted time: broken data? */
 335.734 us [117574] |   } /* A::A */
   5.700 us [117574] |   A::~A();
            [117574] |   catch_exc() {
            [117574] |     bar() {
   3.450 us [117574] |       B::B();
            [117574] |       bar1() {
            [117574] |         bar2() {
   3.549 us [117574] |           C::C();
   7.199 us [117574] |           __cxa_allocate_exception();
  10.399 us [117574] |           C::~C();
  64.897 us [117574] |         } /* bar2 */
                     |            /* inverted time: broken data? */
  42.348 us [117574] |       } /* bar1 */
   3.749 us [117574] |       B::~B();
            [117574] |       catch_exc() {
  73.298 us [117574] |         baz();
  86.498 us [117574] |       } /* catch_exc */
 235.792 us [117574] |     } /* bar */
 261.492 us [117574] |   } /* catch_exc */
 988.659 us [117574] | } /* main */

[namhyung]

Thanks for the fix. I think it's better to move this check to mcount_rstack_reset_exception() to minimize the change and to be in sync with mcount_rstack_reset().

[jyf111]

Ok, I also found that if I don't do this, mtdp->in_exception will not be reset to false, which may cause other issues.
Similarly, I removed the check for mcount_estimate_return in __cxa_begin_catch().

[namhyung]

Actually I don't remember why I put it in __cxa_begin_catch() but it seems ok.

[jyf111]

This line is introduced by 27b0f44

[honggyukim]

Thanks for the fix. The change looks correct and I've confirmed that the inverted time: broken data? message got disappeared by this change.