The four runtimes
One section each: how it reaches C, what it will not carry, and what it costs. Both games are in the tables on their own pages; reading-the-numbers.md says how to quote them.
The short comparison
| reaches C via | structs by value | primitive arrays | |
|---|---|---|---|
| babashka | babashka.ffi, built into bb | yes, as maps | aset-double over plain aset |
| Clojure/JVM | raylib-clj, coffi on Panama | yes, as maps | ^"[D" / ^"[I" on the defs |
| jank | cpp/ interop, the real header | yes, transparently | none at all |
| jolt | jolt.ffi via net.b12n/raylib | yes, :by-value plus a pointer | ^double/1 / ^int/1 |
Every one of the four carries structs by value. That is worth stating plainly, because the sibling repo's documentation claimed otherwise about two of them until September 2026.
babashka
babashka.ffi ships inside bb, so a port is a source file and nothing else. No build step, no dependencies, fastest of the four to get on screen.
It is not scalar-only. It passes structs by value in both directions, including a 120-byte Mesh returned from GenMeshCube. Where a port here packs a Color into a uint anyway, that is a choice rather than a workaround: a four-byte all-integer struct rides in a single register on both ABIs, so the packed int is the same memory with no map built per call.
The cost is that the game logic runs on an interpreter. At helitorus's default resolution that is invisible, since all four runtimes hit the frame cap. At 900 rings it is a third of the frame rate, and in mesh-instancing's spin mode it is the slowest of the four by a wide margin.
Needs bb 1.13.220; babashka.ffi is absent in 1.12.212.
Clojure on the JVM
raylib-clj binds raylib with coffi on Panama, so structs arrive as ordinary Clojure maps and nothing is packed. The comfortable end of the four.
It is also the fastest, and by more in mesh-instancing than in helitorus, because mem/write-float compiles to a direct store and that game is mostly writes.
Two things to know. The array type hints are load-bearing: without ^"[D" and ^"[I" every aget resolves reflectively, which measured thirty times slower. And where raylib-clj does not bind a call, coffi's defcfn takes the C symbol and the types and nothing else, so extending the binding inside your own file is a few lines. The mesh-instancing port declares seven that way.
Needs a JDK 22 or newer, and three JVM flags that are all mandatory.
jank
jank includes the real header, so Mesh and Material cross by value with no ceremony at all. Of the four FFIs it is the one that never has to think about struct layout.
What it does not have is primitive arrays. double-array, int-array and aget all exist in clojure.core.jank and all throw TODO: port ..., verified against upstream at 2026-09-20. Buffers are therefore malloc'd natives, cast to (:* double) or (:* float), and carried as cpp/box values, because a native pointer cannot be a jank fn's parameter: arguments arrive boxed and cpp/aget will not subscript them. That is the same constraint as the one a jank fn meets from the other side, where it may not return a native value.
Measure the -O3 binary. The :base profile builds at -O0 and the difference is eleven times. bb release then bb run-release.
There is no JVM, so no Math/*, no parse-long and no format. C maths comes from math.h, the clock from GetTime, integer parsing from TextToInteger.
jolt
jolt.ffi through net.b12n/raylib, which is a real wrapper rather than a raw binding, so most of the struct work is already done. It passes structs by value with [:by-value [:struct ...]] on the parameter and a pointer to an allocated layout buffer as the argument.
It sits second in helitorus and third in mesh-instancing, which is the one place the two games disagree about ordering between neighbours. The wrapper writes matrices field by field at word offsets, and that is a lot of individual writes.
Needs jolt 0.8.0 or newer; the bindings are pinned by sha.
One layout note that applies to all four
raylib declares Matrix as m0 m4 m8 m12 then m1 m5 m9 m13, so in declaration order the translation column sits at slots 3, 7 and 11. Every port here has to get that right, and getting it wrong stacks every instance at the origin, which is indistinguishable from several other bugs.