Repository navigation
Conversation
Tiny mode suppresses panic values and tracebacks to avoid leaking runtime information, but hiding the entire diagnostic makes a crashing binary look as if it exited without running. Keep a single panic: hidden line for unrecovered panics while leaving fatal runtime diagnostics suppressed. Replace printpanics after the generic runtime print stripping pass. Do not inspect or replace panic values, and keep preprintpanics empty so user-defined formatting methods cannot expose information. Recovery, nil-panic behavior, and GOTRACEBACK crash handling remain unchanged. Extend the existing tiny fixture without adding builds. Check nil panics with both panicnil settings, runtime panics, custom values, nil-panic recovery, and default-mode diagnostics. The pre-fix test failed with no match for ^panic: hidden$ in stderr; the fixed non-short fixture and runtime unit tests pass, as does go vet ./.... Fixes burrowers#604.
|
if we're showing a panic message, we might as well hash the original panic message, so that the user can reverse that to figure out what the original underlying panic was? |
|
e.g. |
|
@luames mvdan is right here |
|
I fully agree with not completely hiding panics btw. That is confusing. |
|
Agreed, a hash would be more useful than The missing piece is reversal. Should the initial implementation cover statically known string messages with reverse support, keeping the generic diagnostic for other panic values? That would make the reversibility guarantee explicit rather than printing a hash we cannot resolve. |
|
I think what needs to be able to be reconstructed is where the original panic came from, the message can then be read from source. We don't want stack traces, nor do we care about dynamic values for -tiny @luames please implement |
|
Yeah good idea. The location is more useful than the message, because the message could be repeated with other call sites. |
The generic diagnostic for burrowers#604 identifies a panic instead of a silent exit, but it cannot identify the failing source. Record an opaque source site before panic unwinding and print only that token for an unrecovered panic. Do not format panic values or print a stack trace. Reuse salted source-position hashes and teach reverse about tiny tokens. Retain PC-to-file and CU lookup metadata while stripping physical paths, line tables and function start lines. Public runtime source-location APIs remain stripped. Capture on the system stack so implicit runtime panics, inlining and deferred panics identify their originating site. Keep the generic diagnostic when no source token is available. Extend the existing tiny fixture with exact reversed-output assertions, both nil-panic settings, custom-value recovery and a literals build. The pre-fix test failed because stderr contained panic: hidden rather than a source token. Tiny, reverse, position and runtime tests, patcher tests, and go vet pass. Regenerated patch bytes and the freshly applied ordered patch-series tree match the readable toolchain commit. Retaining origin mappings has a size cost: on the unchanged tiny fixture with Go 1.27.0 and seed AAAAAAAAAAA, binaries grow from 2,211,964 to 2,683,004 bytes. Repeated builds of both versions are byte-identical. Fixes burrowers#604.
luames
changed the title
Tiny panic reversal intentionally retains salted source-site directives while stripping physical paths and public source locations. The debugdir fixture still asserted the previous absence of all line directives and failed on the full Linux, macOS and Windows CI runs. Assert that tiny output contains opaque panic-site directives and does not expose the physical main filename. The non-short debugdir fixture passes locally. Refs burrowers#604.
|
Implemented. The tests cover explicit and implicit panics, inlined and deferred panics, recovery, both nil-panic settings, and reversal with The tradeoff is retained opaque origin metadata: the existing fixture binary grows by 21.3% compared with the generic-diagnostic version. That measurement is documented in the PR. Physical filenames and line tables remain stripped, as do the public runtime source-location APIs. |
|
@luames as tiny should be tiny can we get the increase in file size down |
Reversible panic locations for burrowers#604 retained the original CU lookup tables, including dead-file holes, and repeated the same token prefix and suffix thousands of times. Compact that metadata without dropping source coverage or changing the diagnostic format. Append a linker patch which assigns dense global file indices and rewrites PC-to-file data per function without mutating shared object symbols. Omit tables without opaque sites, coalesce hidden ranges, and deduplicate rewritten tables. Store token hashes alone and restore the common p_ and .go text only when printing an unrecovered panic. Extend the existing tiny build with binary-table assertions. Before the fix, the table check failed with 146,984 bytes for 13,276 files; after compaction, a second check caught the repeated token suffixes. Both checks now pass, as do non-short tiny, reverse, position and debugdir scripts, runtime/reverse unit tests, patcher tests and go vet. On the unchanged fixture with Go 1.27.0 and seed AAAAAAAAAAA, file size drops from 2,683,004 to 2,494,588 bytes. The generic-diagnostic baseline is 2,211,964 bytes, so the increase falls from 21.3% to 12.8%, removing 40% of the added size. Repeated builds are byte-identical and all three binaries execute with the expected diagnostics. The new patch matches its readable toolchain commit, and fresh ordered application reproduces the complete toolchain tree. Refs burrowers#604.
|
Yes. The fixture binary now drops from 2,683,004 to 2,494,588 bytes, removing 40% of the added size. Compared with the 2,211,964-byte generic-diagnostic baseline, the overhead falls from 21.3% to 12.8%. These are matching Go 1.27.0 builds with the same source and fixed seed; repeated builds are byte-identical. I replaced sparse per-CU tables with a dense global table, removed unused file mappings, deduplicated rewritten PC tables, and stopped storing the token prefix and suffix thousands of times. Panic-site coverage and the printed/reversed format are unchanged. The remaining overhead is the source-site metadata itself. Pushed the compaction and updated the measurements in the PR. The focused non-short tests and all six CI jobs pass. |
3 participants
Fixes #604. Tiny mode should distinguish an unrecovered panic from a silent exit and let the source owner identify the failing site, without printing the panic value or a stack trace.
Print one line such as
panic: p_abc123.go:1.garble -tiny reversemaps the token back to its original filename and line using the same source, build flags, and seed. The token hashes the source site, not the message, so dynamic values and user-defined formatting methods are never needed.Capture the origin before deferred calls run. This covers explicit panics, nil dereferences, bounds panics, inlined calls, and deferred panics. Panic values and recovery are unchanged, including both
panic(nil)settings. Fatal runtime errors remain silent, andGOTRACEBACK=crashstill crashes the process. If no source token is available, printpanic: hidden.The linker retains opaque PC-to-file mappings but removes physical filenames, line tables, and function start lines. Public runtime source-location APIs remain stripped. A dense global file table replaces sparse per-CU lookup tables. PC-to-file data is rewritten per function without mutating shared object symbols, omitted when no opaque sites are present, and deduplicated after rewriting. Hidden ranges coalesce. Each hash is stored without the repeated
p_prefix or.gosuffix; the panic printer restores them.On the unchanged existing tiny fixture, with Go 1.27.0 and fixed seed
AAAAAAAAAAA, the generic-diagnostic baseline is 2,211,964 bytes. The initial reversible version was 2,683,004 bytes; the compact version is 2,494,588 bytes. This removes 188,416 bytes, or 40% of the added size, reducing the overhead from 21.3% to 12.8%. Repeated builds of all three versions are byte-identical, and each binary executes with its expected diagnostic. The remaining overhead is retained source-site metadata.Extend the existing tiny fixture with exact reversed-output checks, recovery checks, a
-literalsbuild, and binary-table assertions for compact metadata. No separate script fixture or extra fixture build is added. The README documents reversal and the metadata tradeoff.Tests
go test -run '^TestScript$/^(tiny|reverse|position|debugdir)$|^TestRuntime|^TestReverse' -count=1passed without short mode.1851c8d: Test run.go test ./internal/patcher -count=1passed.go vet ./...passed.format-patchoutput. Applying all eight ordered patches to fresh Go sources reproduces the same tree.Implemented by Hermes Agent using gpt-6.1-sol, acting on behalf of @luantak.