I made about 73 SpecApplication runs, over about 50 CPU hours.
The runs were 3 at a time on Krajy, a 4-way machine.
Eight runs raised the following assertion from gcLocker.cpp:
  assert(!h->gc_is_active(), "GC active during No_GC_Verifier")

(NOTE: There were also 2 failures with the message "Fatal: not a call imm32".
These were a more serious unrelated bug, 4681417.)

Here is a backtrace from one of the runs:

  [6] report_assertion_failure(code_str = ???, file_name = ???, line_no = ???, message = ???) (optimized), at 0xdebceae1 (line ~144) in "jrose/hopper/src/share/vm/utilities/debug.cpp"
  [7] No_GC_Verifier::No_GC_Verifier(this = ???, activate = ???) (optimized), at 0xdec2f38e (line ~27) in "jrose/hopper/src/share/vm/memory/gcLocker.cpp"
  [8] SharedRuntime::d2i(x = ???) (optimized), at 0xdf16b3f8 (line ~103) in "jrose/hopper/src/share/vm/runtime/sharedRuntime.cpp"
  [9] ConvD2INode::Value(this = ???, phase = ???) (optimized), at 0xdebbd89a (line ~554) in "jrose/hopper/src/share/vm/opto/connode.cpp"

The assertion check is noticing that the C2 optimizer is calling
the JRT_LEAF routine SharedRuntime::d2i, while a GC is in progress.
Although this is harmless since the compiler runs in the state
_thread_in_native, the assertion is overly conservative, since
JRT_LEAF routines were designed to be called from Java code,
not C code.

Inspection of the code suggests that the only place where non-Java
code is calling JRT_LEAF routines is in opto/connode.cpp, in four
places. However, the interface is publicly available to all C code,
and it is impossible to ensure that C routines will not call leaf
routines.