SWT module clone - set original functions to be optimised and cloned as unoptimised.

[Pavel-Durov]

Interpreter starts always in optimised (non-tracing) mode, therefore the original module should be set as optimised.

[Pavel-Durov]

For some reason yk simple_peeling.c test is failing but I don't see how its related to these changes 🤔
That's the only test that is failing, with error:

Pattern (error at line 6):
   ...
   |...
   |header_start ...
   |...
>> |header_end [%{{0}}, %{{1}}, %{{2}}, %{{3}}, %{{4}}]
   |...
   |body_start [%{{19}}, %{{20}}, %{{21}}, %{{22}}, %{{23}}]
   |...
   ...

Text (error at line 17):
   ...
   |    %2: ptr = param Register(15, 8, [])
   |    %3: ptr = param Register(12, 8, [])
   |    header_start [%0, %1, %2, %3]
>> |    %5: ptr = lookup_global @stderr
   |    %6: ptr = load %5
   |    %7: i32 = load %3
   |    %8: ptr = lookup_global @.str
   ...

---- lang_tests::simple_peeling.c stdout ----

[ltratt]

That test is affected by a recent change to ykllvm -- but that change hasn't (yet) successfully merged into yk.

[ptersilie]

Does this change which functions we inline during tracing? I remember there was some discussion around this. For example, we know we need to trace the unopt functions, which then get inlined. But what happens when we outline a function. I think at the moment we just emit calls to the unopt functions from the trace. We might later want to rewrite these calls to actually call the opt functions. Though I think that's an optimisation. But it's worth thinking about.

[Pavel-Durov]

Does this change which functions we inline during tracing?

I don't think it changes anything in terms of tracing inlining. In this PR both versions will have the tracing calls so
essentially both versions will be unoptimised until the control point transition is implemented.

I remember there was some discussion around this. For example, we know we need to trace the unopt functions, which then get inlined. But what happens when we outline a function. I think at the moment we just emit calls to the unopt functions from the trace. We might later want to rewrite these calls to actually call the opt functions. Though I think that's an optimisation. But it's worth thinking about.

I don't remember what we said about funcion inclining tbh (and I can't find anything in my notes about it 😶 ).
But we can call opt functions instead of unopt when outlining, don't see an issue with that.

[ptersilie]

Why are we adding them to both at the moment?

[Pavel-Durov]

To remove tracing calls from the original functions (opt functions in this PR) we will need first to implement the control point transition between opt and unopt versions.
Right now the functions are cloned but they are not used at all, yk swt executes only the original functions.

[ptersilie]

I see. Okay, let's remove the commented code though. We can add it back later.

[Pavel-Durov]

removed 👉 24eceaa

[ptersilie]

Please squash,

[Pavel-Durov]

Squashed 👉 9504103