Jos Visser <[EMAIL PROTECTED]> wrote:
> My current code generator generates code to trap the LEX_NOT_FOUND
> exception. However, the "set_eh" instruction is *terribly* slow...
After having a closer look at this, I can describe what's happening:
- the exception handler is a continuation
- these have COWed copies of the interpreter's stacks
- the C<set_eh> pushes an exception handler onto the control stack
this unCOWes the interpreter's control stack
- when the exception is resumable, a return continuation is built
again setting the control stack as COWed
- so resuming after the exception puts a saved copy of the original
control_stack in place, which has the COW flag set.
- this causes a chunk_copy (effectively constructing a new stack) at
the next set_eh :-(
The COW logic for stacks is wrong.
Test program follows:
new_pad 0
new P0, .PerlInt
set P0, 42
store_lex -1, "yep", P0
set I0, 10000
time N0
lp1:
find_lex P1, "yep"
dec I0
if I0, lp1
time N1
sub N2, N1, N0
print N2
print "\n"
newsub P3, .Exception_Handler, _failed
set_eh P3
set I0, 10000
time N0
lp2:
find_lex P1, "no"
ret:
set_eh P3
dec I0
if I0, lp2
time N1
sub N2, N1, N0
print N2
print "\n"
end
_failed:
set P0, P5["_invoke_cc"]
invoke # resume at ret:
