Core: Correct the pool-duplication TODO: the dead allocas do cost stack

The previous note claimed LLVM deletes the unreferenced duplicate temps.
Measured: removing only the dead alloca from the IR moves the BPF frame
from r10-32 to r10-24, so llc keeps it through frame layout and each
duplicated i64 temp costs 8 bytes per nesting level. opt -O2 would
remove it; the pipeline is llc only.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Author: r41k0u

pythonbpf/functions/functions_pass.py

@@ -149,11 +149,12 @@ def update_max_temps_for_stmt(stmt):
 
     # TODO: allocate_mem is re-entered for every if body (handle_if_allocation),
     # and each level allocates a pool of its own even though the top-level count
-    # already covers nested statements. The duplicates are unreferenced and LLVM
-    # deletes them (verified: the BPF frame is identical with and without the
-    # nesting), so this costs nothing at runtime; it is IR noise and a trap,
-    # since only the last-allocated pool stays in local_sym_tab. Fix: count
-    # recursively, allocate once at the top level.
+    # already covers nested statements. Only the last-allocated pool stays in
+    # local_sym_tab; the others are unreferenced, but llc does not delete
+    # unreferenced allocas before frame layout (opt -O2 would, and the pipeline
+    # is llc only), so each duplicated temp slot costs its size in BPF stack:
+    # measured 8 bytes per i64 temp per nesting level. Fix: count recursively,
+    # allocate once at the top level.
     allocate_temp_pool(builder, max_temps_needed, local_sym_tab)
 
     return local_sym_tab