Monday, November 5, 2012

JIT, Debugging, and Interrupts

I finally found some time to work on Octave last weekend. There has been some talk on the mailing list recently about releasing a new version of Octave, so I figured I should clean up a few loose ends in JIT.


Up until now JIT has been skipping breakpoints. While there are several issues with supporting breakpoints in JIT, the biggest one is that the Octave debugger allows for the execution of arbitrary statements. Arbitrary statement execution is a very powerful and useful feature of the Octave debugger, but it allows users to change the type of a variable in the middle of code execution.

JIT gets its performance improvement by making assumptions about variable types, so entering debug mode means we need to exit JIT. I took the simple way out here and do not start JIT if there are any breakpoints in the code (see hg).


In Octave if you hit control-c Octave will stop execution and return to the Octave prompt (unless debug on interrupt is set). The interpreter does this by calling octave_quit, which checks the interrupt state and throws an octave_interrupt_exception if an interrupt has occured.

Ideally, to support interrupts in JIT a call to octave_quit should be inserted once per loop iteration. Unfortunately, it is not that simple. After an interrupt occurs the current variable values need to be reported to the interpreter. For example,
i = 0;
while 1
  i += 1;
If the user interrupts the loop the interpreter needs to save the current value of i. This means JIT needs a way to catch and rethrow the octave_interrupt_exception. While LLVM does have a way of handling exceptions, the infrastructure in Octave does not yet exist to support the LLVM feature.

Instead, I inserted a check to octave_interrupt_state. If octave_interrupt_state is greater than 0, we need to exit to the Octave prompt. I reused the code for checking error_state to achieve this.

Now that JIT handles interrupts and breakpoints in a manner consistent with the interpreter, I can't think of ways in which JIT and the interpreter differ (besides speed). The amount of code which JIT can compile is still fairly limited. Hopefully, I will get some time over winter break to make it easier to extend JIT and improve what JIT can compile. In its current state JIT should be ready to include as an experimental feature in the next Octave release.

No comments:

Post a Comment