Track PHI first state per control slot

[guosran]

Summary

Addresses #292 with a minimized PHI fix. This PR only updates PhiRTL.py and its focused test.

Changes

• Track PHI first-state independently for each control-memory slot.
• Clear the corresponding state only after a valid PHI transaction completes.
• Add concrete examples in the focused PHI tests.

No MemUnit changes, markdown files, new ports, or new interfaces are included.

[tancheng]

Thanks @guosran, this is the only PR that I understand what it is trying to solve. Can we provide a concrete example in the description about why we need this fix? What issue it would encounter if we don't fix it?

[tancheng]

Well, the Drop unrelated PHI PR noise actually adds MemUnit fix in this PR. Can we avoid that?

[guosran]

Thanks @guosran, this is the only PR that I understand what it is trying to solve. Can we provide a concrete example in the description about why we need this fix? What issue it would encounter if we don't fix it?

Concrete example: two different control-memory slots can both execute PHI operations. Before this fix, PhiRTL used one shared first-state flag, so executing PHI_START/PHI_CONST in one slot could incorrectly affect whether a PHI in another slot is treated as first iteration.

Without the fix, a later PHI slot may choose the loop-carried input when it should still choose the initial/const input, or vice versa. This can corrupt the value entering the loop body.

The fix tracks first-state per control slot, so each PHI instruction has independent first-iteration state.

[guosran]

Well, the Drop unrelated PHI PR noise actually adds MemUnit fix in this PR. Can we avoid that?

The unrelated MemUnit change has been removed from this PR.

[guosran]

Concrete example:
PHI needs to select the first-state value only for the first dynamic execution.

first iteration:
first_state = 1
PHI selects init input

later iterations:
first_state = 0
PHI selects loop-carried input

[yyan7223]

The proposed Track PHI first state per cotrol slot has already been implemented in the initial code, I don't understand what is this PR trying to resolve. Most of code changes are just modifying the signal name, e.g., renaming s.first[s.ctrl_addr_inport] to s.cur_first.

Besides, I agree with that OPT_PHI_START selects the value from port 0 for the first iteration, but I remember it should select the ports value with predicate = 1 in the following iterations just like OPT_PHI, rather than just select the value from port 1 regardless of its predicate. @tancheng WDYT?

[yyan7223]

What if the loop-carried input comes from s.recv_in[s.in0_idx]? Will that happen for OPT_PHI_START?
Sorry I forget the definition of PHI_START. What's the difference between PHI_START, PHI_CONST, and PHI?

[yyan7223]

I think directly using s.phi_executed_first_time[s.ctrl_addr_inport] is more straightforward.