Implementing a debugger: How to call lua_singlestep and/or suspend lua execution?

[JDaance]

My idea was to have singlestep = false as default, and when suspending on a breakpoint, if the user of the debugger then chooses to step I will call lua_singlestep(true) before continuing execution. But this does not seem to come into effect immediately, its only after a bunch of instructions that the step hook will be called. If I run with singlestep = true all the time, the hook will be called for the line right after the breakpoint. I suspect the singlestep value is cached for a run of vm, and only refreshed at some point.

This might be due to how I suspend the lua execution -- the way I did it in the lua-5.1/luajit days I just infinite loop with a thread sleep to halt lua execution. And this works fine except for this single step issue. So right now upon breaking the suspend loop I set lua_singlestep(true) if the user wants to step. But maybe there is a better way? I tried lua_break and lua_yield to suspend execution without knowing what I'm doing, but these complained about crossing c-boundary.

Or maybe I simply need to run singlestep = true all the time?

[zeux]

You definitely shouldn't run with singlestep=true all the time if performance is important. The way our debugger does this is that a breakpoint suspension yields the thread (via lua_break), and sets it to single-step before resuming when the user decides to step. We disable single step when the user resumes execution (instead of using step in/out).

[JDaance]

Ok, good! I must have been on the main state when I tried lua_break. Maybe I can run singlestep = true on main state and set it to false on all threads as they are created.

[JDaance]

I ended up running with singlestep = true whenever there is an active breakpoint. When the last breakpoint is gone I remove it. Performance with singlestep on is still orders of magnitude better then my old lua_sethook implementation. After I changed it so that I ignore all step calls on the current line (too many instructions on some lines), it steps magically good, much more consistent and predictable then the lua-5.1/luajit hook.

[ClxS]

This is quite an old post so sorry for reviving, I tried to find a more recent thread but this is the only place I saw mentioned the same issue I have been fighting today.
But in case anyone hits similar issues with lua_break erroring about crossing the C boundary, check your Lua Binding library. I am using LuaBridge which uses lua_pcall to call into lua. This results in nCcalls being incremented without incrementing the baseCalls count. Replacing it with lua_resume to execute instead solves this and got things working for me.

[cppmadguy]

How to resume execution after hitting a breakpoint ? once I hit one I can't resume execution, the vm always return LUA_BREAK, and it's not possible to disable the breakpoint because pushing to the stack is prevented by an assert, so calling lua_breakpoint() is impossible.
Now, this could work if I set a VM dirty flag (using the userdata pointer) that tells my callback if I am already in 'paused execution' state. Is this the way to go ? We need more documentation on this api...
Ok, after a bit more work I managed to get it to work: their should be a worker/runner thread that runs your code separately from the main thread, then it is possible to remove the breakpoint by calling lua_breakpoint on the main thread.