Repository navigation
Kilat 0.2.3 for Windows (x64). One program, no dependencies.
kilat-windows-x64.zip:kilat.exe,install.cmd, the language guide and examples. Unpack it and runinstall.cmd(orinstall.cmd /dry-runto see what it would do).kilat.exe: the program on its own.
The files are not code-signed, so Windows may show a SmartScreen prompt the first time.
New in 0.2.3, and it breaks some programs: a value with a unit is never shown without a named unit. print(d), "{d}" and str(d) are refused before the program runs, and the message suggests the unit the program itself used (write `{d:mm}` ). Before, 2.4 mm printed as 0.0024 m. 12 of the 61 Kilat program files in the repository had to be changed. The regression test tests/unit_hints_test.py now has 75 cases; 0.2.2 passes 40 of them.
Also new: a failed assert_eq on two values that print the same now says so and points to assert(approx(a, b, tolerance)). get_or on maps is faster: with that change alone the benchmark map_count ran 1.57 and 1.60 times faster on GitHub's Windows machine (best and median of 9 rounds) and 1.79 and 1.76 on its Linux machine, one sample per system.
Measured against CPython 3.13.2 on one Windows PC (best of 5 runs, CPU time, while the PC was busy with other work): 3.06 times faster for the whole process, 2.49 with start-up taken out, faster on 10 of 10 programs. An earlier attempt the same day logged 3.09 and 2.59 before it stopped in a later check. 0.2.2 measured 3.02 and 2.54, faster on 9 of 10. On GitHub's machines (best of 2 runs, each machine's own build, CPython 3.13.17 for Windows and 3.13.18 for Linux) four runs per system gave 3.09, 2.84, 3.09 and 2.92 for Windows (3.02, 2.76, 3.03 and 2.85 with start-up taken out) and 2.96, 3.13, 2.98 and 3.12 for Linux (2.91, 3.06, 2.92 and 3.06). The Windows build there uses Microsoft's toolchain and is not the file in this release.
Before and after on the same machine (GitHub's, 9 rounds, one sample per system): the ten benchmarks together ran at 1.06 times their 0.2.2 speed for Windows and 1.06 for Linux by best times, 1.06 and 1.04 by median times. More than half of that gain comes from one program, map_count. For Linux nbody ran slower with the final build in 4 of 4 comparisons (0.82 to 0.94 times its earlier speed) and the cause was not found. The goal of 5 times CPython was not reached in any run. 3 times with start-up taken out was not reached on this PC and was reached in 4 of the 8 GitHub runs. Details in RESULTS.md.
kilat.exe was built on GitHub's Windows machine with Rust's GNU toolchain from this tag's engine source. It needs only DLLs that are part of Windows.
SHA-256 of kilat.exe: db646696dc987dff95ebc9b64df5de275e0311cd8170479582aba07d0a0fe9cc
SHA-256 of kilat-windows-x64.zip: 6ca3a728f165acc73ef0e3b69d6650f6dd6f62a1636026ad3adbc018a8475466