JDK-8150778 switched to StackWalker API for populating stacktrace elements in Throwable. We may optimize the VM parts even further:
 http://mail.openjdk.java.net/pipermail/core-libs-dev/2016-March/039280.html

Profiling shows we spent lots of time StringTable::intern-ing to convert Symbols to Strings:

                6.600 (60%) org.openjdk.StackTraceBench.doDepth(int)
                +- 3.920 (36%) java.lang.Throwable.getStackTrace()
                  +- 3.920 (36%) java.lang.Throwable.getStackTraceElements(java.lang.StackTraceElement[])
                    +- 3.920 (36%) JVM_GetStackTraceElements
                      +- 3.910 (36%) java_lang_Throwable::get_stack_trace_elements(Handle,objArrayHandle,Thread*)
                      | +- 3.910 (36%) java_lang_StackTraceElement::fill_in(Handle,InstanceKlass*,const methodHandle&,int,int,int,Thread*)
                      | +- 1.890 (17%) StringTable::intern(const char*,Thread*)
                      | | +- 0.480 (4%) StringTable::lookup_shared(unsigned short*,int)
                      | | +- 0.350 (3%) UTF8::unicode_length(const char*,bool&,bool&)
                      | | +- 0.320 (3%) java_lang_String::equals(oopDesc*,unsigned short*,int)
                      | | +- 0.300 (3%) StringTable::hash_string<unsigned short>(const unsigned short*,int)
                      | | +- 0.260 (2%) UTF8::convert_to_unicode<unsigned short>(const char*,unsigned short*,int)
                      | | +- 0.140 (1%) resource_allocate_bytes(unsigned long,AllocFailStrategy::AllocFailEnum)
                      | | +- 0.040 (0%) G1SATBCardTableModRefBS::enqueue(oopDesc*)
                      | +- 1.260 (12%) StringTable::intern(Symbol*,Thread*)
                      | | +- 0.490 (4%) java_lang_String::equals(oopDesc*,unsigned short*,int)
                      | | +- 0.360 (3%) StringTable::lookup_shared(unsigned short*,int)
                      | | +- 0.300 (3%) StringTable::hash_string<unsigned short>(const unsigned short*,int)
                      | | +- 0.080 (1%) Symbol::as_unicode(int&)const
                      | | +- 0.030 (0%) G1SATBCardTableModRefBS::enqueue(oopDesc*)
                      | +- 0.500 (5%) Klass::external_name()const
                      | | +- 0.380 (3%) Symbol::as_klass_external_name()const
                      | | +- 0.120 (1%) strlen
                      | +- 0.250 (2%) Backtrace::get_line_number(const methodHandle&,int)
                      | | +- 0.250 (2%) Method::line_number_from_bci(int)const
                      | +- 0.010 (0%) oopDesc::obj_field_put(int,oopDesc*)
                      +- 0.010 (0%) HandleMark::pop_and_restore()

At least class name may be cached in Class.name field in Java, e.g.:
 http://cr.openjdk.java.net/~shade/8151751/poc-cache-Class.name.patch

Method name and source file name might piggyback on ConstantPool constant string resolution?

The performance improvements are very promising, up to 3-4x performance improvement across stack trace depths, see the benchmark and results here:
  http://cr.openjdk.java.net/~shade/8151751/StackTraceBench.java
  (runnable JAR: http://cr.openjdk.java.net/~shade/8151751/benchmarks.jar)