[ZweZeya]

Stacked on #6 (feat/spsc-runtime).

What this changes

--select-queue-impl, from 253a2bc (feat/select-queue-impl), ported to the pin: for a cir.global whose verdict is SPSC it retargets the storage, construction, push/pop and destructor registration to the module's spsc_queue<T>. Three commits:

1. The pass. Changes from 253a2bc, each found at the pin:
   • The destructor is retargeted through this global's __cxa_atexit registration. 253a2bc rewrote every reference to the original destructor, and two globals of one type share one, so substituting @q1 also registered spsc_queue's destructor for @q2.
   • Every use of the global is checked first; one the pass cannot retarget means no substitution, with a remark naming the use.
   • Alignment comes from cir.record_layouts (the record type's data layout ignores alignas; the substitute aligns to 64).
   • Pops with a result slot are substituted only when the substitute's try_pop takes one too.
   • Diagnostics are MLIR remarks.
2. The shim makes the substitute available. queue<T>'s constructor odr-uses a [[gnu::used]] anchor that instantiates spsc_queue<T>, so no program has to name it; the pass erases the anchor and its @llvm.used entry.
3. Benchmarks, with the result checked (count, sum, order) so a fast wrong queue fails rather than gets timed; run.sh is portable and reports the median.

Only globals are substituted.

How it was checked

29 of 29 pass, no warnings (SelectQueueImplementation/global-spsc.mlir covers the shared-destructor case; refusals.mlir the four ways a queue is left alone). On the prototype's example (q1 SPSC, q2 MPMC, same type) only q1 is substituted and q2's destructor is untouched; the result runs as a native binary, and for x86-64 lowers to LLVM IR whose substituted try_pop call carries sret and the new size.

Measured at the pin on an Apple M1 Pro, median of five, all runs checked:

variant	items/s	vs baseline
wtc::queue<int>	2.07M	1.0x
substituted	7.44M	3.6x
spsc_queue by hand	43.1M	20.9x

The remaining gap is one shared_ptr allocation per pop, kept to match try_pop's interface.

Checklist

• CI is green
• New behaviour has a test; a fixed bug has a test that failed before
• No file moves or renames mixed in with substantive changes
• The pinned LLVM revision is untouched, or this PR does nothing else

🤖 Generated with Claude Code

[apaznikov]

Closing as superseded by #10. The substitution there rewrites the topology tag at the queue's constructor rather than retargeting the global to another type, so it applies to locals as well as globals and needs no anchor or destructor retargeting; on the pipeline benchmark the compiler's build is byte for byte the hand-tagged one. Your M1 numbers are a useful independent replication: 3.6× substituted and 20.9× by hand line up with the 3.07× and the 10 to 41× of the inline ring measured on x86 for the gate, and the gap you name, one shared_ptr allocation per pop, is exactly what the borrow analysis and the inline queue in #10 remove.