runtime: LLGo-owned frame-pointer unwinder (Stage 5)#2019
Open
cpunion wants to merge 57 commits into
Open
Conversation
Bring over the cross-branch runtime funcinfo benchmark (hot, deep, multipkg, cold, stdlib scenarios) so xgo-dev#2012 can reproduce its own performance numbers. cold.FirstCallersFrames now walks to the first fully symbolized frame, because synthetic runtime frames (LLGo's runtime.Callers placeholder) carry no file/line and the metric was silently skipped on LLGo. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
macOS previously had no entry/stub/pcline site sections, so first-use funcinfo initialization fell back to one dlsym per function and per stub (13ms cold on a small binary, 27ms with LTO), and statement-level pc-line records did not exist at all. Emit the same site records on Mach-O: - __DATA,__llgo_fie / __llgo_stub / __llgo_pcl sections with the live_support attribute: under ld64/lld -dead_strip a live_support atom survives only if the atom it references (the anchor label inside the function body) is live, which matches the records-follow-function semantics ELF gets from SHF_LINK_ORDER with --gc-sections. - One lowercase-l linker-private symbol per record so each record is its own atom and dead functions drop exactly their own records. - Assembler-local (L-prefixed) pc-site labels: Mach-O subsections-via-symbols treats visible labels as atom boundaries, and a visible label in the middle of a function let the linker split and reorder function bodies. - Boundary symbols via ld64's section$start$/section$end$, emitted with the \x01 verbatim-name prefix so LLVM does not prepend the Mach-O underscore. - A no_dead_strip zero record per section in the main module keeps the sections (and their boundary symbols) present even when no package contributed records. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
First-use initialization: - Skip the per-stub dlsym loop when the stub-site section provided the frames; each dlsym is a dynamic-loader query and the loop dominated cold latency. - Materialize per-function strings and entry PCs once per function and packed file strings once per file ID during pcline table construction instead of once per site. Cold FuncForPC fast path: before the frame table exists, resolve exact function-value PCs with a bounded linear scan of the raw entry-site and stub-site sections (compile-time data, no loader query), then one dladdr as fallback; both require an entry match within the warm path's slack so stripped-local misattribution is impossible. The path is budgeted: after a handful of cold lookups the sorted table amortizes better, so it is built as usual. cold.FirstFuncForPC drops from 13ms to ~35us on macOS. Find index: subbucket deltas are now uint16 and the whole-index abandonment on delta overflow is gone. Go stores uint8 deltas because its linker guarantees a 16-byte minimum function size; LLGo indexes call-site records that sit a few bytes apart, and a dense 4KiB bucket silently degraded every lookup in the process to a full binary search. A delta counts deduplicated PCs inside one bucket, so it is bounded by the bucket size and uint16 cannot overflow. Observability: LLGO_FUNCINFO_DEBUG=1 prints one line per lazily built table (frame/bucket counts, index built or fallback, sites vs dlsym sources) so benchmarks can tell which path they measured. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every Caller/Callers capture used to intern the frame into the synthetic table: a hash probe plus a full frame comparison per stack slot per call. Memoize the interned PC base in the shadow-stack slot and invalidate it when the recorded line changes (for one entry the instrumented name/file operands are constants, so the line is the only thing that varies between call sites). The three static frames emitted around every Callers walk get per-store memo slots, and the emit loop is unrolled so nothing escapes and skipped frames are never captured. macOS: hot.CallersOnly 182ns -> 125ns (Go 1.26: 118ns); with LTO 96ns. hot.CallersFramesFirst 528ns -> 471ns, 354ns with LTO (Go: 401ns). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…py limit Frames.Next allocated a fresh *Func per symbolized frame; route it through the FuncForPC 4-way cache so repeated CallersFrames walks over the same PCs stop allocating. hot.CallersFramesFirst: macOS 471->456ns (338ns with LTO, Go 1.26: 406ns); Linux LTO reaches parity at 433ns. Also document a pre-existing limitation at the entry-site emitter: the body-embedded inline-asm record is duplicated by LTO inlining into every inline site (~4x section growth on multipkg) and registers host-function PCs under the inlinee's symbol ID. Runtime only consults the table when native symbolization fails, which bounds the impact; the fix (data globals with !associated metadata) needs LLVMGlobalSetMetadata in the llvm binding and lands with the link-phase ftab work. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Record the experiment results at the emitter: !associated only guides linker GC and IR-level GlobalDCE deletes the records; llvm.compiler.used pins dead functions through the records' address initializers; and noduplicate blocks inlining. Section dedup is link-phase work. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Post-link table generation plan: parse the linked binary's metadata sections, dedup LTO inline copies against the symbol table, sort with a sentinel, build Go-layout findfunctab via internal/pclntab, and write back into a reserved section with ASLR-safe anchor offsets. Runtime adopts the prebuilt table when the header validates and keeps first-use construction as fallback. Includes the list of platform facts established in xgo-dev#2012 so implementation does not re-derive them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Only macOS ran tests with -coverprofile, so lines behind OS-specific branches (ELF emission, per-OS runtime shims) always showed as missed in codecov/patch even though the ubuntu job executed them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Covers the ELF and Mach-O directive branches, 32-bit pointer directives, quote-escaped symbol names and empty-table emission from one table-driven test, so single-platform coverage runs stop reporting the other platform's branches as dead. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
First stage of doc/design/pclntab-linkphase.md: parse a linked binary's funcinfo entry/stub sections (Mach-O and ELF), deduplicate LTO inline copies against the symbol table's text ranges, sort with a Go-style sentinel, and build findfunctab through internal/pclntab — the faithful port that has been waiting for exactly this caller. Read-only: prints what the P2 build integration would write back. Measured on the 576-target multipkg binaries: - non-LTO: 9319 records -> ftab 3161 + 207 buckets; lookup self-check 3160/3160; site sections 149KB -> 29KB (5.1x) - LTO: 15371 entry records -> 13857 inline copies dropped, 4144 kept; self-check 3045/3045; 299KB -> 28.5KB (10.5x) Findings for P2: on-disk Mach-O pointer slots hold dyld chained-fixup encodings (low 36 bits are the target; decoded here; the write-back design stores anchor-relative offsets and avoids pointers entirely), and some non-LTO stub symbols are absent from the symbol table (records conservatively dropped; needs tightening). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…adoption
pclnpost -write rewrites the entry-site section in place with the
prebuilt table (header + ftab {entryOff,funcIndex} + runtime-layout
findfunctab buckets), resolving funcinfo indexes through the binary's
symbol-index section, and voids the stub section (its records are
merged into the table). ASLR is handled by anchoring on the section's
own link-time address; entries are normalized to true symbol starts,
which retires the entry-PC slack on this path. macOS re-signs with an
ad-hoc codesign after rewriting.
The runtime adopts the table zero-copy when the magic header validates:
lookups binary-search the on-disk ftab directly through the shared
bucket index, nothing is materialized on first use (the funcIndex ->
entry map is built lazily and only for the pcline initializer), and the
cold scan/dladdr path is skipped since adoption is cheap. First-use
construction remains the fallback whenever the header is absent.
Linux end-to-end: entries=prebuilt, FuncForPC/FileLine correct,
first-FuncForPC 110µs (materializing) -> 6-8µs (zero-copy); 13ms on the
original macOS baseline. Known gap: on macOS the on-disk rewrite is
corrupted at load time because dyld still walks the stale chained-fixup
chain over the section; fix (unlinking the section's nodes from the
page chains in LC_DYLD_CHAINED_FIXUPS) is identified and next.
Non-prebuilt paths verified regression-free: cl + test/go suites pass,
smoke behavior unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every llgo-linked executable (linux/darwin, sites enabled) now gets the prebuilt ftab/findfunctab automatically: internal/build runs internal/pclnpost.Rewrite after linkMainPkg, and any failure degrades silently to the first-use construction fallback. Moves the tool core into internal/pclnpost and hardens it: - Canonical-record detection by FNV: a record survives when its anchor's owning symbol hashes to the record's symbolID (or is the __llgo_stub. wrapper of it). The previous one-per-symbolID rule wrongly collapsed a function with its stub — they share the target's symbolID by design — which broke exact-entry lookups (caught by TestRuntimeLineInfoAndStack on Linux). LTO inline copies are now identified exactly: 8.4k/9.5k copies removed in the LTO probes. - Mach-O chained-fixups surgery: unlink the rewritten sections' pointer slots from the dyld page chains (repointing predecessors' next links and page_start entries) so dyld neither rebases slots inside the new table nor skips unrelated fixups after the zeroed stub section, then re-sign ad hoc. Without this the table was corrupted at load. - LTO-safe metadata location: the entry section carries a meta record whose relocations hold the addresses of the symbol-index pointer and count globals; LTO internalization strips those names from the symbol table but relocations always resolve. Runtime skips the meta rows (pc==0 / symbolID==0). - Idempotence guard (already-rewritten binaries are left alone). Runtime fixes that surfaced during validation: - materializePrebuiltEntries is now two-phase so concurrent losers wait for the winner's store instead of reading a nil entries slice. - pcLineFrameForPC rejects nearest-below sites whose entry is unresolved when the caller knows the function entry, instead of leaking a neighboring function's file/line. Validation: macOS cl (full) + test/go + LLDB 194/194; Linux test/go TestRuntime suite; probes on both platforms report entries=prebuilt with first-FuncForPC at 7-21µs (Linux) from 13ms on the original baseline, and LTO builds drop 8-9.5k inline copies. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…table On Mach-O, pointer slots that name exported functions — every __llgo_stub.* wrapper and any exported Go function — are emitted as chained-fixup BIND nodes, not rebases. The rewriter only decoded rebase nodes, so all stub records (and some entry records) were dropped as unowned and never reached the prebuilt ftab; FuncForPC on function values silently fell back to dladdr (~6µs per fresh pc on darwin). - Parse the LC_DYLD_CHAINED_FIXUPS imports table and resolve bind ordinals to their in-image definitions. - Match canonical owners against the record symbolID with underscore normalization (debug/macho's suffix-shared string table can surface one mangling underscore more or less than the source-level name). - Splice the prebuilt header's base slot back into the fixup chain as a live rebase node: dyld writes the slid text base at load, so the runtime reads a ready runtime PC with no slide arithmetic (non-PIE ELF link-time values already equal runtime addresses). - LLGO_PCLNPOST=0 escape hatch keeps first-use construction. Fresh-pc FuncForPC slow path: darwin 6-8µs -> 1.2-1.7µs, linux 6.8µs -> 0.5µs; first-in-process lookup: darwin ~32µs -> ~14µs, linux ~6.8µs -> ~4µs. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Pure-compute probes (recursive fib, JSON round-trip, sort.Ints, map churn) with no runtime introspection, so one harness run covers both the introspection extremes and what the funcinfo machinery costs code that never asks for it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Go's pclntab pages are touched by its own runtime (traceback, GC) long before user code queries it, so its first FuncForPC never pays page-in. Mirror that: when the prebuilt table is present, init adopts it (zero-copy, sub-µs), touches the pages the lookup path reads (blob, funcinfo records, string offsets, strings), runs one synthetic lookup to warm the code paths, and write-warms the FuncForPC cache pages. First-in-process FuncForPC: darwin ~17µs -> ~2.8µs, linux ~6.6µs -> ~1.0µs. Startup cost is page-count-bound (tens of µs on stdlib-sized tables, invisible next to ~3ms process startup; hello-world medians unchanged). Non-prebuilt binaries stay fully lazy: first-use construction allocates, which has no place in init, and programs that never introspect pay nothing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
-depths generates deep_<N> scenarios at configurable call depths; -bigsizes generates bigfunc scenarios (funcs x statements) whose large bodies stress statement-level pcline density, mid-function pc symbolization, and ordinary performance of big method bodies. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Blob overflow: function-value stubs can double the row count, and at ~9k functions the prebuilt blob no longer fit the entry section, so the rewrite silently fell back to first-use construction (cold.FirstFuncForPC 96x96 non-LTO: 2.4ms). On overflow, retry with function entries only — stub pcs degrade to dladdr, real entries keep the prebuilt table. - FuncForPC cache thrash: the set-associative pc cache holds 4k entries; batch workloads over 9k+ distinct functions evicted constantly and paid the string-materializing slow path on every call (multipkg.FuncForPCMany 96x96: 8-11ms vs Go 172µs). Add a per-ftab-row *Func cache for exact-entry lookups, so batch lookups are O(binary search) after the first pass at any scale. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…erflow
Function-value stubs can push the row count past what the entry section
holds (~9k functions with taken addresses). Instead of dropping stub
rows, write the full blob into the (larger) stub section and leave a
32-byte redirect header ("LLGOFTB2" + a live-relocation pointer) in the
entry section; the runtime follows it and adopts the same zero-copy
view. Function-value lookups keep the prebuilt table at any scale
instead of degrading to dladdr.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
funcForPCSlow treated any unaligned pc as a shadow-stack synthetic marker. arm64 function entries are always 4-aligned so this never fired, but amd64 function and stub entries need not be: an unaligned function-value pc skipped the prebuilt exact-entry path entirely and fell through to nearest-below symbolization, reporting the previous function's name (test/go TestRuntimeLineInfoAndStack on ubuntu CI, "bad function value func: main.renamedPC"). Hoist the prebuilt exact-entry + per-row-cache lookup ahead of the alignment heuristic; a genuine synthetic pc just misses the cheap search and proceeds as before. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The overflow fallback dropped stub rows to fit the entry section. That leaves pc ranges the table claims to cover but does not: a function value whose stub falls in a gap resolves nearest-below to the previous function and silently returns the wrong name — exactly what ubuntu CI caught (amd64 --icf=safe layouts overflow by a few hundred bytes, and non-PIE ELF dladdr cannot rescue). If the blob fits neither the entry section nor the (larger) stub section, skip the rewrite entirely: first-use construction is slower but covers every record. Reproduced and verified on linux/amd64 (qemu): the stub pc had no exact row and nearest-below returned the neighbouring function's name. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…rgery Fabricated fixtures make the IO paths testable in-process: a minimal ELF exercises load/Rewrite end to end (in-place, stub-section spill, and the overflow fallback that must leave the binary untouched), and a synthetic Mach-O image drives the chained-fixup chain surgery (remove+splice, empty-page insert, unconsumed-insert error). Package coverage 16% -> 69%. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
cpunion
force-pushed
the
codex/stage5-unwinder
branch
from
0ab4e62 to
f01f644
Compare
A fabricated Mach-O (segments, sections, symtab, chained-fixup imports and an empty page chain) drives load, bind-target resolution, record decoding and both Rewrite outcomes (in-place and stub-section spill) end to end. codesign now runs only when the input carries LC_CODE_SIGNATURE: real lld executables always do, unsigned inputs need no signature and codesign rejects them. Also cover asmQuoteELFSymbol, the empty-table initializers and the Rewrite error paths. Package coverage: pclnpost 69% -> 86%. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
cpunion
force-pushed
the
codex/stage5-unwinder
branch
3 times, most recently
from
d70dd76 to
db8a078
Compare
Every Go function on supported targets keeps the frame-pointer chain
("frame-pointer"="non-leaf", gated by Program.NeedsFramePointer to
linux/darwin — on embedded targets the unwinder does not exist and the
layout change perturbed the conservative GC on ESP32-C3). runtime.Caller,
Callers, CallersFrames, Stack and the unrecovered-panic dump walk
[fp]/[fp+w] directly and symbolize through the prebuilt ftab and pcline
tables:
- Return addresses resolve at pc-1 (Go's convention); statement labels
can land exactly on a return address, so raw-pc nearest-below reported
the following line. The convention holds with or without the prebuilt
table (text bounds fall back to the first-use frame table — link-phase
overflow layouts otherwise silently disabled it, the root cause of the
amd64 CI failures).
- The walk is bounded to the program's own text: libc frames without FP
discipline decode as wild pcs that nearest-below would attribute to
arbitrary functions.
- Methods and anonymous functions are now trackable (methods had no
pcline labels; closures lost their innermost frame to tail-call
optimization), and mid-function aligned pcs merge statement records
instead of returning declaration lines.
- frameSymbol results are memoized per pc (deep re-walks paid a dladdr
per frame: 32-frame walks 8µs -> 180ns) and the pcline table is built
during the startup pre-warm (lazily building it inside the first
Caller cost ~200µs at scale).
- Shadow-stack instrumentation is no longer emitted; LLGO_SHADOW_STACK=1
keeps the legacy emitters for one release. Tracked functions retain
noinline, no-tail-call and the data-only pcline records.
- libunwind is gone: the clite stacktrace fallback walks the FP chain
with dladdr names (same output format), and linux binaries no longer
link -lunwind.
Semantics are gc ground truth, verified against go: physical stacks show
every real frame; interface-chain Caller marks land at skip 3 and
closure chains at skip 4 (the old expectations encoded shadow-stack
frame loss). Perf (best-of, mac/linux): hot.Caller0 17/37ns (Go
155/241), deep.Direct512 2.8µs (Go 9.7µs; was 87-95µs), bigfunc.Work
18µs (Go 30µs; was 433µs), binary size unchanged or smaller.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
IR goldens gain the frame-pointer attribute (out.ll files carry no attribute groups and needed no regeneration); the legacy shadow-stack emitter assertions opt into LLGO_SHADOW_STACK; statement-line probes move to gc ground-truth skip counts; NeedsFramePointer target matrix and pclnpost symbolAddr/decodePtr edges covered. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
cpunion
force-pushed
the
codex/stage5-unwinder
branch
from
db8a078 to
4925b82
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stage 5 of #2004, based on #2016: replace the shadow-stack instrumentation and libunwind with an LLGo-owned frame-pointer unwinder.
What changes
"frame-pointer"="non-leaf", verifiedstp x29,x30 / mov x29,spprologues). Cost is the standard ~1-2% envelope — Go itself always keeps FP on arm64.runtime.Caller/Callers/CallersFrames/Stackwalk real stacks (fpCallers: [fp]/[fp+8] chain with stride/alignment guards), symbolized through runtime: link-phase ftab/findfunctab generation (Stage 2, P1–P3) #2016's prebuilt ftab + pcline labels. Return addresses resolve atpc-1(Go's convention); the machine scheduler can place the next statement's label exactly on a return address, so raw-pc lookups mis-attributed frames.LLGO_SHADOW_STACK=1keeps the legacy emitters for one release). Tracked functions retain onlynoinline, no-tail-call and the data-only pcline records.-lunwind.__llgo_fp_chainbyte next to the funcinfo table recordingssa.Program.NeedsFramePointer();fpUnwindAvailabletrusts that declaration plus table presence, so a target that keeps tables but not the FP attribute can never take the physical walk by accident.refinePCSymbolLinehelper (same-function statement record at pc, then pc-1) —FuncForPC,FileLineandCallersFramescannot disagree on line attribution.Semantics: now Go-conformant, verified against gc
Physical stacks see every real frame — the shadow stack only ever recorded instrumented ones. Skip counts in the statement-line probes moved to gc ground truth (each verified by running the same chain shapes under go): interface chain MARK at skip 3 (was 2), closure chain at skip 4 (was 3). Adjacent
runtime.Stackcalls report their own lines. Mid-body pcs,FuncForPC(pc-1)andFileLine(pc-1)agree withFrames.Performance
All numbers
best/trimmed-avg, 5 runs (scale rows 3), 24×24=576-target scenarios, ±LTO.s5= this PR.Hot paths — macOS / Linux (ns)
Deep stacks — the Stage 5 headline (macOS / Linux)
The ~160ns/frame shadow-stack tax is gone: 512-deep chains with a
Callerat the bottom are now 3.4× faster than Go, and instrumented-graph big bodies dropped 24×:Cold paths (first use per process, macOS / Linux)
Ordinary code and size
plain.*(fib/json/sort/map) is identical across 2012/2016/s5 within noise on both platforms — code that never asks for caller info pays nothing. Binary sizes are unchanged (±0.1MiB across every scenario); linux binaries no longer link-lunwind.Known follow-ups (single cells, documented)
bigfunc.FirstFileLineMid(32k-site probe): 18.5µs vs 2016's 1.0µs — first-use interaction between the startup pcline warm and site-dense binaries; P4's prebuilt pcline table subsumes it.stdlib.Workis unchanged (~23µs vs Go 6.2µs): the residual isnoinline/no-tail-call on Caller-reachable functions plus the LLGo baseline, no longer unwind bookkeeping. Lifting noinline needs the P4 inline tree.mac cold.FirstCallersFrames4.6µs vs 2016's 1.3µs: first-walk memoization allocations; steady-state walks are 5–8× faster than 2016.Validation
cl(462s),test/go,internal/build,ssa, LLDB suite 0 failed — all with shadow-stack emission off.test/gogreen with-lunwindremoved.attributes #0now carries the FP attribute; out.ll files carry no attribute groups and needed no regeneration).Known scope cuts (documented)
noinline. The P4 inline tree lifts this.Depends on #2016 (branched from it; rebases after it merges).
🤖 Generated with Claude Code