ZJIT: lobsters perf burndown

[k0kubun]

I'm focusing on lobsters again, so I filed a separate issue from #1030. Take 3 of the effort after #833 and #713.

I have a branch where ZJIT is faster than YJIT on lobsters. The goal of this issue is to upstream the most reasonable ideas out of it in cleaner patches, not necessarily porting everything.

yjit: ruby 4.1.0dev (2026-08-26T21:52:46Z master 0968fd4b3b) +YJIT +PRISM [x86_64-linux]
zjit-before: ruby 4.1.0dev (2026-08-26T21:52:46Z master 0968fd4b3b) +ZJIT +PRISM [x86_64-linux]
zjit-after: ruby 4.1.0dev (2026-08-26T23:15:25Z zjit/min-straight d2330a712c) +ZJIT +PRISM [x86_64-linux]

--------  ------------  ----------------  ---------------  -------------------  ------------------  ----------------  ---------------
bench        yjit (ms)  zjit-before (ms)  zjit-after (ms)  zjit-before 1st itr  zjit-after 1st itr  yjit/zjit-before  yjit/zjit-after
lobsters  421.7 ± 6.3%      461.9 ± 4.8%     399.5 ± 5.0%                0.533               0.459             0.913            1.056
--------  ------------  ----------------  ---------------  -------------------  ------------------  ----------------  ---------------

Branch

• Branch

List of patches

speedup is base ms / ms per itr from the same invocation. Δ is the change from the
previous row; + means this commit made lobsters faster.

Assignee	#	sha	speedup	ms/itr	base ms	Δ	commit
		066814022d	1.000	441.0			master before ruby#18636 (base, --zjit)
@k0kubun	1	76b177533e	1.001	440.6	440.9		ZJIT: Profile invalidated ISEQs and function stubs (ruby#18636)
@k0kubun	2	91db83cf22	1.002	440.5	441.5	+0.16%	ZJIT: Wait a configured number of exits before invalidating an ISEQ
@k0kubun	3	b5c32853d6	1.011	435.5	440.3	+0.86%	ZJIT: Always inline every block handler case for getblockparamproxy
@tekknolagi	4	5cd3e16bbf	1.015	434.3	441.0	+0.43%	ZJIT: Fuse Test and fixnum compares into CondBranch
@tekknolagi	5	ddab20c8d3	1.025	429.8	440.6	+0.95%	ZJIT: Fuse type tests into the branch that consumes them
	6	4c733852ea	1.023	431.0	440.8	-0.23%	ZJIT: Compile exception handlers (ruby#18759)
	7	ca12f00c21	1.024	431.0	441.3	+0.12%	ZJIT: Read the throwing frame's ISEQ through CFP_ISEQ() in vm_throw_start
@k0kubun	8	35996b7ba5	1.036	426.1	441.4	+1.16%	ZJIT: Split cold code into an outlined code region
@tekknolagi	9	70203d64cc	1.037	425.2	441.1	+0.14%	ZJIT: Specialize polymorphic sends with blocks (ruby#18567)
@tekknolagi	10	0c0b0ccf4e	1.046	422.1	441.4	+0.79%	ZJIT: Fall back to a generic ivar access when a shape chain misses
@tekknolagi	11	421d05c0a2	1.047	421.4	441.2	+0.12%	ZJIT: Back ivar shape-chain misses with a per-name shape table
	12	9105450e49	1.051	419.9	441.2	+0.35%	ZJIT: Cache the ivar location of class and module shapes
	13	be2c0c3ac8	1.062	415.8	441.5	+1.05%	ZJIT: Pass a &blk argument through to a direct send
	14	f307e737da	1.057	417.2	440.8	-0.49%	ZJIT: Compile a yield to a Symbol block as the send it stands for
	15	75fd14733c	1.046	422.0	441.4	-1.01%	ZJIT: Let a frozen ivar dispatch earn back the shape it is missing
	16	3de96d3604	1.059	416.8	441.3	+1.21%	ZJIT: Weigh an ivar respecialization by the traffic it would take off the fallback
	17	d4f8303962	1.064	414.6	441.0	+0.46%	ZJIT: Profile expandarray and compile one shape per site
	18	91d3e3cddd	1.058	416.5	440.6	-0.55%	ZJIT: Let invalidated ISEQs recompile instead of exiting forever
	19	874a68f749	1.062	415.7	441.5	+0.40%	ZJIT: Inline def foo(...) callees and compile the bar(...) inside them
	20	7fbf7f0e09	1.061	415.7	441.2	-0.08%	ZJIT: Compile past polymorphic splatkw instead of side-exiting
	21	2d26d5bafa	1.063	415.1	441.2	+0.15%	ZJIT: Specialize bar(...) in standalone forwardable ISEQs
	22	c2f9ad7947	1.065	414.6	441.6	+0.21%	ZJIT: Expose bucket counts on DistributionSummary
	23	0549244c4e	1.069	413.3	441.9	+0.38%	ZJIT: Guard the profiled receiver types of megamorphic call sites
	24	f7f0e04a6a	1.061	415.7	441.0	-0.78%	ZJIT: Dispatch megamorphic-by-class call sites on one ancestor guard
	25	95e970ab5c	1.058	416.7	441.0	-0.25%	ZJIT: Dispatch to blocks that only return with a direct JIT-to-JIT call
	26	1a2305dc11	1.063	414.6	440.6	+0.41%	ZJIT: Look megamorphic send targets up in a class table
	27	234d901d50	1.059	416.0	440.7	-0.31%	ZJIT: Fix invokeblock block-handler profiling and gate its ISEQ chain
	28	e386eb9367	1.060	415.9	441.0	+0.09%	ZJIT: Inline a block whose non-local return escapes to the frame we compile
	29	2e004e33f3	1.061	416.0	441.5	+0.08%	ZJIT: Destructure a lone yielded Array for the direct block dispatch
	30	54f0503b88	1.062	414.9	440.6	+0.06%	ZJIT: Destructure a lone yielded Array at polymorphic yield sites too
	31	036864cea1	1.062	415.6	441.2	-0.03%	ZJIT: Specialize yield whose arity does not match the block
	32	cfa150ddf9	1.060	416.2	441.3	-0.12%	ZJIT: Lay the inlined block frame out on the reshaped yield arguments
	33	185d5eabd6	1.059	416.2	440.7	-0.13%	ZJIT: Narrow test reg, imm to the immediate's width
	34	7bbada3134	1.059	416.8	441.5	+0.04%	ZJIT: Call megamorphic ISEQ targets without leaving JIT code
	35	2aecabea46	1.060	416.1	441.0	+0.05%	ZJIT: Compile direct sends to callees with a **rest parameter
	36	b4e172f476	1.058	416.8	441.0	-0.16%	ZJIT: Let a yield-unlocking iterator past the caller's inline budget
	37	68a43cc289	1.072	411.4	440.9	+1.27%	ZJIT: Compile send/send with a profiled method name
	38	5e510e5c63	1.069	412.5	440.8	-0.29%	ZJIT: Size megamorphic send tables to the shapes that need them
	39	b1007b0b17	1.071	412.0	441.4	+0.26%	ZJIT: Guard profiled splatarray operands instead of calling rb_vm_splat_array
	40	93a9712b29	1.069	412.3	440.9	-0.19%	ZJIT: Read nil for an out-of-bounds Array#[] instead of side-exiting
	41	c928cc29a6	1.073	410.7	440.9	+0.38%	ZJIT: Dispatch a yield to a block that can break directly
	42	c96eac3688	1.076	410.2	441.2	+0.20%	ZJIT: Only branch on the block param proxy for an ISEQ callee
	43	ad7d7a19b7	1.072	411.2	440.9	-0.32%	ZJIT: Pass a &blk argument through to a direct C call
	44	05992a2874	1.077	409.4	441.1	+0.48%	ZJIT: Compile direct sends to callees with an anonymous **
	45	f7af91226d	1.072	411.6	441.4	-0.47%	ZJIT: Inline callees that a &blk call site passes a block handler to
	46	b93d6d06f6	1.068	412.9	440.8	-0.45%	ZJIT: Stop side-exiting forever at unstable super sites
	47	cbfb8584cb	1.073	410.5	440.5	+0.51%	ZJIT: Dispatch polymorphic super on the frame's method entry
	48	480e3fdcf8	1.075	410.4	441.3	+0.20%	ZJIT: Compile protected calls behind a check on the caller's self
	49	ad5054a1aa	1.076	409.5	440.8	+0.10%	ZJIT: Drop the now-unreachable invokeblock may-throw fallback reason
	50	8d83692390	1.080	408.8	441.6	+0.35%	ZJIT: Pin the forwardable callee's fallback on a &blk call site
	51	67022c671b	1.076	409.5	440.6	-0.40%	ZJIT: Test the ivar reprofile budget inline
	52	05449ab73e	1.075	410.0	440.9	-0.05%	ZJIT: Hand the SP register back when returning past inlined frames
	53	b0151512b6	1.074	411.0	441.4	-0.13%	ZJIT: Let a getivar site's profile forget boot-time shapes
	54	91d2b5139d	1.076	410.3	441.3	+0.15%	ZJIT: Test profile eviction and cached class ivars
	55	805b7d7321	1.074	410.8	441.1	-0.17%	ZJIT: Size ivar shape tables to the names that need them
	56	098a399a61	1.073	411.3	441.2	-0.10%	ZJIT: Dispatch yield to a run-time block ISEQ without leaving JIT code
	57	7ad26cc3c8	1.072	410.8	440.3	-0.08%	ZJIT: Dispatch a Symbol block's send through the megamorphic class table
	58	fe932fdb94	1.075	410.6	441.3	+0.28%	ZJIT: Remove CheckInterrupts in empty inline frames
	59	4bb2ec5aec	1.073	410.9	440.8	-0.19%	ZJIT: Assert with_write_ptr patches the half it is on
	60	b2e41cbb27	1.074	410.6	440.9	+0.09%	ZJIT: Allocate callee-saved registers to call-crossing VRegs
	61	bdbee0ee60	1.084	407.5	441.7	+0.93%	ZJIT: cut the cost of the GC mark hooks
	62	70593e75e0	1.082	407.0	440.3	-0.19%	ZJIT: Coalesce block params with the arguments feeding them
	63	d1e33bdad2	1.080	408.6	441.3	-0.17%	ZJIT: Make the block-param coalescing analysis linear
	64	26f3f4f5c6	1.079	408.1	440.4	-0.08%	ZJIT: Remove the trivial inliner
	65	c4129b2a6b	1.076	410.1	441.2	-0.31%	ZJIT: Remove NoTracePoint patch points in empty inline frames
	66	e3eca3a258	1.077	409.4	440.9	+0.10%	ZJIT: Elide inline frames around leaf builtin calls
	67	71e7b305d2	1.083	407.4	441.4	+0.61%	ZJIT: Retry inlining annotated builtins after type inference
	68	311953991a	1.077	409.2	440.8	-0.58%	ZJIT: Grow the array instead of side-exiting on out-of-range Array#[]=
	69	1e4e643ca0	1.080	408.1	440.8	+0.27%	ZJIT: Compute the coderange instead of side-exiting when it is unknown
	70	c99b78293b	1.079	408.7	440.9	-0.12%	ZJIT: Shrink side-exit stubs with a shared metadata table
	71	18e29fe82e	1.082	407.8	441.4	+0.33%	ZJIT: Skip the bounds check on the side-exit metadata lookup
	72	f052b50585	1.082	407.8	441.2	-0.05%	ZJIT: Stop each half of the code region at its own end
	73	b2d270d7e9	1.080	407.5	440.2	-0.16%	ZJIT: Break down zjit_alloc_bytes by subsystem
	74	fef8304c3c	1.086	406.1	440.9	+0.51%	ZJIT: Size operand type profiles exactly
	75	058aa08487	1.081	408.0	441.2	-0.40%	ZJIT: Box the polymorphic tail of a type distribution
	76	0916e4784a	1.081	407.5	440.5	-0.04%	ZJIT: Skip redundant cfp->sp stores
	77	01ed8a350a	1.077	408.8	440.4	-0.34%	ZJIT: Release an ISEQ's profile and GC offsets when the ISEQ is freed
	78	1a6a886620	1.088	405.1	440.8	+0.99%	ZJIT: Free the payload of a GC-freed ISEQ and account dead-version bytes
	79	4d2f98f6cf	1.092	403.5	440.6	+0.35%	ZJIT: Box the rarely-used profile side tables
	80	488b48d0db	1.092	404.0	441.0	-0.03%	ZJIT: Don't break spilled parallel-copy cycles through SCRATCH_REG
	81	e194a2d75a	1.096	402.8	441.4	+0.38%	ZJIT: Test that spill elision keeps frame locals observable
	82	4d692e2f60	1.089	404.6	440.8	-0.58%	ZJIT: Inline a fast path for Array#push and Array#<<
	83	1f0f64ec01	1.088	405.0	440.8	-0.10%	ZJIT: Initialize the block param local when a function stub exits
	84	a7c79202c0	1.090	404.4	440.9	+0.17%	ZJIT: Stop over-reserving in long-lived metadata vectors
	85	1ee1ff0661	1.085	406.1	440.6	-0.49%	ZJIT: Keep the profiled shape reachable through guards and dispatch arms
	86	82ac8165f1	1.084	406.5	440.8	-0.06%	ZJIT: Branch on the ivar shape for attr_accessor sends in polymorphic arms
	87	28e750cff8	1.091	404.5	441.2	+0.58%	ZJIT: Materialize frames before pushing a caught break value
	88	f5eb199f23	1.086	405.8	440.8	-0.41%	ZJIT: Skip redundant local spill stores
	89	cd38d87122	1.089	405.1	441.2	+0.27%	ZJIT: Pin the profiling metadata layout the memdiet merge depends on
	90	3124ff46b7	1.089	405.1	441.3	+0.03%	ZJIT: Drop the now-unreachable invokeblock arity-mismatch fallback reason
	91	34ab69dd2d	1.090	404.8	441.1	+0.03%	ZJIT: Drop the splat-expansion path superseded by caller-splat dispatch
		066814022d	1.071	411.7	441.0		YJIT at the base commit (--yjit)

Details

Summary

• Stack so far: 441.0 → 404.8 ms/itr, 1.090× over the base (through row 91).
• ZJIT first drops below the YJIT reference (411.7 ms) at row 37 (68a43cc289).

Largest single-commit gains

• +1.27%, row 37, 68a43cc289, ZJIT: Compile send/send with a profiled method name
• +1.21%, row 16, 3de96d3604, ZJIT: Weigh an ivar respecialization by the traffic it would take off the fallback
• +1.16%, row 8, 35996b7ba5, ZJIT: Split cold code into an outlined code region
• +1.05%, row 13, be2c0c3ac8, ZJIT: Pass a &blk argument through to a direct send
• +0.99%, row 78, 1a6a886620, ZJIT: Free the payload of a GC-freed ISEQ and account dead-version bytes
• +0.95%, row 5, ddab20c8d3, ZJIT: Fuse type tests into the branch that consumes them
• +0.93%, row 61, bdbee0ee60, ZJIT: cut the cost of the GC mark hooks
• +0.86%, row 3, b5c32853d6, ZJIT: Always inline every block handler case for getblockparamproxy
• +0.79%, row 10, 0c0b0ccf4e, ZJIT: Fall back to a generic ivar access when a shape chain misses
• +0.61%, row 67, 71e7b305d2, ZJIT: Retry inlining annotated builtins after type inference

Largest single-commit regressions

• -1.01%, row 15, 75fd14733c, ZJIT: Let a frozen ivar dispatch earn back the shape it is missing
• -0.78%, row 24, f7f0e04a6a, ZJIT: Dispatch megamorphic-by-class call sites on one ancestor guard
• -0.58%, row 68, 311953991a, ZJIT: Grow the array instead of side-exiting on out-of-range Array#[]=
• -0.58%, row 82, 4d692e2f60, ZJIT: Inline a fast path for Array#push and Array#<<
• -0.55%, row 18, 91d3e3cddd, ZJIT: Let invalidated ISEQs recompile instead of exiting forever
• -0.49%, row 14, f307e737da, ZJIT: Compile a yield to a Symbol block as the send it stands for
• -0.49%, row 85, 1ee1ff0661, ZJIT: Keep the profiled shape reachable through guards and dispatch arms
• -0.47%, row 45, f7af91226d, ZJIT: Inline callees that a &blk call site passes a block handler to

Caveats

• Base spread 1.7 ms (0.39%) across 91 re-measurements; treat single-row Δ under ~±0.5% as noise.
• The stack was reordered (upstreaming candidates first); a commit's Δ is its effect at its new position.