punkshell

1494 Commits

1 Branch

1 Tag

224 MiB

Commit Graph

Author	SHA1	Message	Date
Julian Noble	946a32da19	G-172 drafted (proposed) + the four recipe pins: distributed binaries get a baseline CPU floor The published tclsh9.0.5 runtime family is AVX-512-only. A flagless bin/punk-runtime.cmd fetch hands a fresh checkout the curated win32-x86_64 default (tclsh9.0.5-punk-r2.exe), which on a CPU without AVX-512 raises #UD and is killed with STATUS_ILLEGAL_INSTRUCTION (0xC000001D, errorlevel -1073741795) before writing a byte - no message, no prompt, no error. First reported as a Windows-version difference (25H2 works, 23H2 does not); it is not one. Cause: the recipes called b.standardTargetOptions(.{}) and no build path passed -Dtarget/-Dcpu, so zig resolved cpu_model .determined_by_arch_os by NATIVE detection and ReleaseFast auto-vectorised to the Zen 5 build host. Measured: all six published tclsh9.0.5-* artifacts (punk + plain, r1 + r2, both -bi- variants) carry 2756 zmm-operand instructions each - real EVEX encodings, first site at the front of .text, and Tcl has no runtime CPU dispatch - as do bin/punk905.exe, punk9-dev.exe, punk9_beta.exe, punk9bi_beta.exe, punkdeclare.exe and punkfiledemo.exe. bin/punkzip.exe carries 1724 and bin/punkres.exe 1423, so a bake on such a machine fails too. An older generation (punk901, punk9.exe, xcritcl.exe, tclsh90s) is AVX2-only from an earlier host: same defect, milder floor. Confirmed across three machines - Zen 2 dies, Zen 4 and Zen 5 run. Pins applied to all four in-repo recipes (suite_tcl90, suite_tcl86, punkzip, punkres): standardTargetOptions gains .default_target = .{ .cpu_model = .baseline }, so a flagless build is portable while -Dcpu=native remains the developer opt-in and -Dtarget=<triple> is unaffected. Floor chosen as baseline (x86-64 v1) - for a Tcl interpreter the codegen delta against v2/v3 is negligible set against the default runtime failing to start. This has to be a recipe DEFAULT rather than a build-command convention: "remember to pass -Dcpu= when publishing" is precisely what failed silently for six artifacts. Verified by zig ast-check on all four recipes, and by reproducing the mechanism with the same toolchain and optimize level on an equivalent recipe (flagless -> x86_64 baseline, zmm=0; -Dcpu=native -> znver5, zmm=805). No suite build was run: all four recipes gate on required_zig = "0.16.0" with a .lt SemVer compare, which rejects the bundled prerelease bin/tools/zig-0.16.0-dev.254 - a separate pre-existing defect, recorded in the goal's Notes and deliberately not fixed here. Nothing is rebuilt or republished, and no shipped surface changes, so no project version bump. Recording the floor in the artifact metadata record, an audit tool, punk-runtime surfacing it on use/run (fetch stays ungated - fetching for another platform is a normal workflow), and the rebuild/republish with the defaults.txt repoint all remain open as G-172. Assisted-by: harness=claude; primary-model=claude-opus-5[1m]; api-location=anthropic.com	4 days ago

Author

SHA1

Message

Date

Julian Noble

946a32da19

G-172 drafted (proposed) + the four recipe pins: distributed binaries get a baseline CPU floor

The published tclsh9.0.5 runtime family is AVX-512-only. A flagless
bin/punk-runtime.cmd fetch hands a fresh checkout the curated win32-x86_64
default (tclsh9.0.5-punk-r2.exe), which on a CPU without AVX-512 raises #UD and
is killed with STATUS_ILLEGAL_INSTRUCTION (0xC000001D, errorlevel -1073741795)
before writing a byte - no message, no prompt, no error. First reported as a
Windows-version difference (25H2 works, 23H2 does not); it is not one.

Cause: the recipes called b.standardTargetOptions(.{}) and no build path passed
-Dtarget/-Dcpu, so zig resolved cpu_model .determined_by_arch_os by NATIVE
detection and ReleaseFast auto-vectorised to the Zen 5 build host. Measured: all
six published tclsh9.0.5-* artifacts (punk + plain, r1 + r2, both -bi- variants)
carry 2756 zmm-operand instructions each - real EVEX encodings, first site at the
front of .text, and Tcl has no runtime CPU dispatch - as do bin/punk905.exe,
punk9-dev.exe, punk9_beta.exe, punk9bi_beta.exe, punkdeclare.exe and
punkfiledemo.exe. bin/punkzip.exe carries 1724 and bin/punkres.exe 1423, so a
bake on such a machine fails too. An older generation (punk901*, punk9.exe,
xcritcl.exe, tclsh90s*) is AVX2-only from an earlier host: same defect, milder
floor. Confirmed across three machines - Zen 2 dies, Zen 4 and Zen 5 run.

Pins applied to all four in-repo recipes (suite_tcl90, suite_tcl86, punkzip,
punkres): standardTargetOptions gains .default_target = .{ .cpu_model =
.baseline }, so a flagless build is portable while -Dcpu=native remains the
developer opt-in and -Dtarget=<triple> is unaffected. Floor chosen as baseline
(x86-64 v1) - for a Tcl interpreter the codegen delta against v2/v3 is
negligible set against the default runtime failing to start. This has to be a
recipe DEFAULT rather than a build-command convention: "remember to pass -Dcpu=
when publishing" is precisely what failed silently for six artifacts.

Verified by zig ast-check on all four recipes, and by reproducing the mechanism
with the same toolchain and optimize level on an equivalent recipe (flagless ->
x86_64 baseline, zmm=0; -Dcpu=native -> znver5, zmm=805). No suite build was
run: all four recipes gate on required_zig = "0.16.0" with a .lt SemVer compare,
which rejects the bundled prerelease bin/tools/zig-0.16.0-dev.254 - a separate
pre-existing defect, recorded in the goal's Notes and deliberately not fixed
here.

Nothing is rebuilt or republished, and no shipped surface changes, so no project
version bump. Recording the floor in the artifact metadata record, an audit tool,
punk-runtime surfacing it on use/run (fetch stays ungated - fetching for another
platform is a normal workflow), and the rebuild/republish with the defaults.txt
repoint all remain open as G-172.

Assisted-by: harness=claude; primary-model=claude-opus-5[1m]; api-location=anthropic.com

4 days ago

1 Commits (946a32da192c3b4e7c57b955eeb2fae4b69f5028)