Go 1.25

Green Tea GC experiment. Container-aware GOMAXPROCS. Flight recorder. Spinbit everywhere. DWARF 5. Core types gone from the checker. Unlocker-blame runtime-lock profiling. Allocation-free interface conversions of constants and composite literals.

Each area lists the stage every thread reached in this release (with what was left open afterwards), then changes that are not part of a thread. Thread names link to the full thread on its topic page.

Memory management

From Runtime: memory allocator and garbage collector (Go 1.18–1.28)

• Soft memory limit and GC CPU limiter — Done: sched.idleTime was double-counted when the limiter forced an up-to-date flush; fixed (385000b004). Left open: nothing noted.
• Green Tea: span-based marking with vector scan kernels — Done: user-visible the greenteagc GOEXPERIMENT was added, off by default (f1ebad19bd). Size-class tables moved into internal/runtime/gc so generators could share them (528bafa049, e90ba1d208). The core landed: scan bits, per-P span queue, global overflow list, and per-size-class stats under GODEBUG=gctrace=2 (1b40dbce1a). Supporting work: a faster inner-loop bit clear (b30fa1bcc4); setting mspan.limit before publication so the conservative scanner cannot index past allocBits (4c7567290c); allocation-time publication made safe against conservative scanning (24ea1aa25c, see the publication thread); per-P overhead cut from ~2.6 KiB to 536 bytes when the experiment is off (d4bf716793); an explicit nil check in spanInlineMarkBits.init to avoid read-faulting fresh pages to the zero page, which made hugo slower (f8ccda2e05); and word-at-a-time mark merging, ~33% faster, with clearing folded in and skipped for fresh spans (397d2117ec, 0c6296ab12, 6d4a91c7a5, 75a19dbcd7). Left open: “A future extension … is to use SIMD-accelerated scanning kernels” (1b40dbce1a); experiment off by default.
• mallocgc decomposition and size-specialized malloc — Done: needzero removed from paths that must always zero (144d4e5d5f). Left open: nothing noted.
• Finalizers, runtime.AddCleanup and the cleanup queue — Done: user-visible checkfinalizer (3df078fc74, 913c069819, eec8dd0836, c58f58b9f8, 8b45a3f78b, 5b0b4c01ba), with queue lengths printed each cycle (0d42cebacd); finalizer blocks cleaned up (57cd75a351); the multi-goroutine cleanup queue (e46c8e0558, 1421b982dc); a missed-wakeup fix by re-checking the full list after marking asleep (b450b5409d); nil-P skip in procresize (8dd7d2111b); race-detector annotations (3ea94ae446). sync, reflect and unique moved from SetFinalizer to AddCleanup (954e2c0b06, b4e992b6e1, 0028532118). Left open: c58f58b9f8 flagged that a tiny-allocated cleanup arg “could end up co-located with the object”; fixed in 1.26 (1bb1f2bf0c^1.26).
• Weak pointers and the unique package — Done: user-visible weak.Make accepts linker-allocated objects (6d7cb594b3); Pointer conversions between types are blocked (8163ea1458); equality semantics clarified (5ec76ae5aa); weak pointers work in sbrk mode (53b9eae387). The unique canon map with AddCleanup landed (0028532118, e382bf5b32, 2f036e1475, ca3b474702, 644905891f). Compiler help for unique.Make: no copy for string(b) (27ff0f249c) and modeled data flow of non-string pointers for escape analysis (5e6a868b28). Left open: “We still need some kind of back-pressure mechanism” for cleanups (0028532118).
• Allocation publication, zeroing and preemptibility — Done: the fresh-from-OS zeroing skip restored plus atomic largeType (df9888ea4e); the freeIndexForScan snapshot (24ea1aa25c). Left open: nothing noted.
• Sweeper termination and pacing — Done: goschedIfBusy in bgsweep to avoid a livelock once lock is inlined (e666f1dabf); gcpacertrace end-of-sweep condition (ad7a6f8159). Left open: nothing noted.
• Stack and register root scanning — Done: GC STW now calls casGToWaitingForSuspendG, avoiding mutual deadlock with the tracer’s suspendG (c6ac736288). Left open: nothing noted.
• Address-space layout, tagged pointers and the mem.go layer — Done: user-visible anonymous mappings named via PR_SET_VMA_ANON_NAME (e.g. [anon: Go: page summary]), Linux 5.17+ (52eaed6633). 512-byte tag alignment (9d0320de25, aix fixes 3009566a46, 0c15d88974); freebsd/amd64 to 57 bits for LA57 (5bbac667c5). Left open: nothing noted.
• Linear-memory heaps (Wasm and sbrk platforms) — Done: missing unlock in sysReserveAlignedSbrk (456a90aa16). Left open: nothing noted.

Other changes

• GODEBUG=gccheckmark=1 repaired internal — had bit-rotted (root jobs not reset, noscan in workbufs, arenas, preemption); a dist smoke test was added (b25b5f3ff4).
• gQueue/gList carry their size internal — (e0dba45c62).
• plan9 heap sampling internal — the floating-point-free nextSampleNoFP was deleted now that Plan 9 can use floating point (40c953cd46).

Scheduler and concurrency

From Runtime: scheduler, timers, channels, locks and synchronization (Go 1.18–1.28)

• Timer rewrite: per-P timer heap, timer.lock, and GC-able synchronous timer channels — Done: bubbled timers behave like unbubbled ones: channel timers enter the bubble heap only when blocked on, and due timers run immediately rather than being queued, random order for same-instant timers (3432c68467, 3bd0eab96f); fake clock excludes monotonic time (5916bc5b57). Left open: nothing noted.
• Runtime mutex: tri-state futex/sema locks → “spinbit” mutex — Done: committed to spinbit everywhere including emulated-Xchg8 platforms, tri-state implementations deleted; wasm excepted (single thread) (7a177114df). Left open: nothing noted.
• Contended runtime-lock profiling (mutex profile for runtime-internal locks) — Done: unlocker blamed (4d0f955a5e); fix for cputicks overflow causing zero-duration events (ac645eaa0e). Left open: nothing noted.
• Scheduler core: spinning Ms, work conservation, stealing, STW interaction — Done: only one P at a time polls the network in findRunnable (352dd2d932), reverted for a CockroachDB benchmark regression (489917fc40), reintroduced (1e436ba668); explicit exclusion of a theoretical globrunqget->runqput->runqputslow recursion by refactoring runq/globrunq APIs (a42e337ede, +size field for gQueue/gList e0dba45c62); semrelease no direct G handoff on systemstack (a189516d3a); allpSnapshot stashed on M (740857f529). Left open: nothing noted.
• GOMAXPROCS container awareness and dynamic updates — Done: user-visible cgroup discovery and default-on behaviour (f12c66fbed, e6dacf91ff); rename to updateMaxProcsG (6160fa59b6); updatemaxprocs metric only when disabled (3b77085b40); guarantee of no GOMAXPROCS-update syscalls after a GOMAXPROCS call by synchronising with sysmon (for sandboxing) (e481a08e0e); custom GOMAXPROCS noted even when the value is unchanged (f069a82998). Left open: parent limits, cgroup migration (“we can consider changing this in the future”).
• testing/synctest bubbles — Done: fake clock stops advancing once the main goroutine exits (d6c8bedc7b); no monotonic reading in bubbled time.Time (5916bc5b57); race-detector happens-before for AfterFunc (c0392e7e49); “bubble” terminology (ef7724c432); closing bubbled chans from outside panics (d596bc0e81); hchan.bubble pointer (0afcf9192c); WaitGroup association (b170c7e94c); cond-misuse fatal (21b7e60c6b); distinct deadlock message once the main goroutine is done (049a5e6036); lock rank for weak pointers vs synctest (b78e38065e); clearer bubbled stack traces (848a768ba7); timers aligned with unbubbled (3432c68467, 3bd0eab96f). Left open: some violations still recoverable panics.
• Asynchronous preemption: extended register state — Done: arm64 resumes via return (77343fa646); internal/runtime explicitly excluded from async preempt (92a63bdfee). Left open: nothing noted.
• sync package API evolution — Done: user-visible WaitGroup.Go (822031dffc), f must not panic (9921537a96); fewer allocations in OnceFunc family via single captured struct (c5ff26a7a0); funcs not kept alive after panic (ee0d03fab6); Once.done as atomic.Bool (95611c0eb4). Left open: panic handling in WaitGroup.Go.

Other changes

• Panic printing internal — 478ad013f9.
• Misc internal — schedtrace output (ecda82012d, 5295722238); m padding for 64-bit (b199d9766a, 3968a5bca4); isExtraInC (577bb3d0ce); linknames kept for sonic (279da96532); frame-pointer preamble (742fda9524); atomic types not to be copied (42d3cdc909); Map.Delete doc (583d586ef9); synctest linker-alloc WG (4f86f22671); wording/cleanup (d365f2266d, c3a0859720, 4ed9943b26).

Data structures and type metadata

From Runtime: maps, built-in data structures, hashing, randomness and type metadata (Go 1.18–1.28)

• Swiss-table maps (GOEXPERIMENT=swissmap → only implementation) — Done: clear also erases tombstones (4e63ae46e0), prune tombstones before growing (05ed8a00e0, fixes #70886, pathological insert/delete workloads), faster small-map lookups (bfc209518e), empty/zero map lookups inline better (8a85a2e70a, -12.75% geomean), race PC fix (14b15a2bea), faster maps.Clone (a645bc5eb9). Left open: the noswissmap fallback still existed (“noswissmap is going away soon”, 252c939445).
• reflect implementation: performance, allocation and escape behaviour — Done: user-visible reflect.TypeAssert[T] (2541a68a70; zero-alloc for concrete T, allocates for methods or interface T from Elem()), later fixed for nil interface (f1e6ae2f6f); faster TypeFor (679cd8e779), IsZero pointer check (a8e0641d5b), packEface leaks input to result rather than heap (ef20ccc10b), runtime.AddCleanup instead of finalizers (b4e992b6e1), Seq fixes (d93f6df0cc, 194696f1d1). Left open: nothing noted.
• memmove / memclr / memequal / internal/bytealg assembly optimisation — Done: arm64 Count (f2e9076764) and Index* (5b36f61356, ABIInternal), loong64 SIMD memmove/memclr/memequal/indexbyte/compare/Count (13b7c7d8d2, 1763ee199d, ecdd429a3b, 7bc714d601, a1c3e2f008, 17b9c9f2ad), s390x memclr (a177448765), mips64x Count (1776255742), riscv64 simplified (ad39fd9486) and synthesised MIN/MAX (d88f93f720). Left open: nothing noted.
• Runtime randomness: fastrand, chacha8rand and math/rand/v2 — Done: loong64 assembly block (b8ed752d6f: 40-45% faster). Left open: nothing noted.
• iter.Pull and range-over-func iterator types — Done: smaller Pull footprint (ba7b8ca336). Left open: nothing noted.

Diagnostics and knobs

From Runtime: tracing, profiling, metrics and diagnostics (Go 1.18–1.28)

• The new execution tracer (trace v2): generations, per-M buffers, new parser — Done: user-visible runtime/trace flight recorder (83df0afc4e, see thread below); tracev2 package split out with spec validation tests and an experiments guide (0f62125487, 34e8541d24, fadfe2fc80, b5f34aa4ab); trace v1 reader renamed (8028731648); sync batch with EvClockSnapshot (trace, mono and wall clock) at the start of each generation, format bumped to 1.25 (112c23612f, 2d216141a1); stricter per-version event checks (715754ba86); sync event emitted even when the next generation is broken (283296195b, 85f8e240fe); readTrace no longer spins on shutdown (816199e421). Left open: truncated-versus-broken traces cannot be told apart (fixed in 1.26).
• Frame-pointer unwinding in tracing and profiling; profile stack depth — Done: account for missing frame pointer in function preamble (742fda9524); traceStack testable plus benchmark (75b43f9a97). Left open: nothing noted.
• Mutex and block profiles, including runtime-internal lock contention — Done: verification test using cputicks bounds comparing fast vs. slow critical sections (591c4b50c0); unlocker blame (4d0f955a5e); GODEBUG=runtimecontentionstacks removed because the semantics now match (94c4cdc948); failing tests log the profile (35139d6e45). Left open: nothing noted.
• Flight recorder and trace subscribers — Done: user-visible flight recorder (83df0afc4e), TestSubscribers infrastructure (f2d96272cb^1.26 in 1.26 log). Left open: “Future optimizations are planned for” the multi-subscriber mechanism (commit message).
• CPU profiling, goroutine profiling, pprof labels and the protobuf writer — Done: write errors from profile output propagated (4aeb9ba0de); system goroutines join the goroutine profile state machine (281cfcfc1b); TestCgoCallbackPprof skipped where profiling is broken (7dd7d70c03). Left open: nothing noted.
• runtime/metrics and MemStats — Done: ptr/scalar bitmap metric removed (60d3bcdec3); HeapGoal reported 0 when GC disabled (1947c4233a). Left open: nothing noted.
• Race detector, ASan, MSan and Valgrind integration — Done: race on loong64 (bce8486ebe); gsignal checks in asancall/msancall/racecall (11c86ddcb8); user-visible ASan builds run __lsan_do_leak_check at exit (645ea53019) and unregister root regions on free (56e5476e10); msan memmove through cgo (969a0da362); precise Go data/bss bounds for race (c1fc209c41); user-visible Valgrind instrumentation (40b19b56a9); GDB minimum version raised to 10 (8d767ff38d). Left open: Valgrind --leak-check=yes reports errors because “likely need more annotations”; async preemption and GC very slow under Valgrind (commit message).
• Crash output, runtime/debug, BuildInfo and debuglog — Done: DefaultGODEBUG documented as a build setting (c4136a433c); SetCrashOutput example no longer passes parent env (3672a09a48). Left open: nothing noted.

Other changes

• Scheduler/other internal — m0 padding (3968a5bca4), cleanup race hints (3ea94ae446).

OS layer, stacks and panics

From Runtime: OS and platform layer, stacks, panics and tracebacks (Go 1.18–1.28)

• Defer, panic and recover re-implementation — Done: panic transfers to the deferreturn PC (c2ae5c7443); re-raised panic values no longer duplicated in printpanics (478ad013f9); “raised” wording replaced by “panicked” (d365f2266d). Left open: nothing noted.
• Frame pointers: explicit NOFRAME, FP maintenance and FP unwinding — Done: FP cleared in mcall and morestack (9c1d19a183, 93fb2c9074). Left open: nothing noted.
• C-to-Go callbacks: extra-M reuse, system-stack bounds, init gating — Done: isExtraInC set only if no Go frames left (577bb3d0ce); Windows CreateThread instead of _beginthread (00e793cfe3); TSAN lock wrappers around cgo symbolizer/traceback (d7abfe4f0d^1.26), noting wide cleanup potential since most runtime/cgo functions carry manual TSAN instrumentation. Left open: “lots of room for future cleanup” of TSAN instrumentation.
• vDSO, vgetrandom and fast clocks — Done: vgetrandom M-exit cleanup and stress test (0b31e6d4cc, 8969771cc3); s390x vdsoSP fix (6c3b5a2798). Left open: nothing noted.
• Signals and async-preemption machinery — Done: async preempt resumes by return on arm64 (77343fa646), Windows context-call injection deduped (3108366214). Left open: nothing noted.
• New, ported and removed platforms — Done: 32-bit function index on Wasm (9782dcfd16), loong64 cgo traceback (0694718389). Left open: nothing noted.
• CPU feature detection, GOAMD64 checks and cpu GODEBUG options — Done: RISC-V runtime extension detection on Linux (bcfa00cbd2), ARM64 SHA3 (e4216d54f5), LSE check relanded (81c66e71d4), Neoverse N3/V3 (adf21a9b56). Left open: nothing noted.

Other changes

• Misc internal — linknames restored for sonic (279da96532), gobuf.ret removed (c594762ad5), no restore from g0.sched in arm64 systemstack (739fb752e3), loong64 SIMD memclr/Count and libFuzzer tracing (1763ee199d, 17b9c9f2ad, dc1e255104), loong64 exitThread barrier (11c847e536), usleep s390x (ca4649747a), cgo docs and warning fixes (a8e532b0f2, 76c7028253, 2e71ae33ca, 01ba8bfe86), HACKING.md update (8c6fec6d25).
• Mapping decoration, cpu detection internal — RISC-V extensions on Linux (bcfa00cbd2), Neoverse N3/V3 (adf21a9b56), HasSHA3 (e4216d54f5), rva23u64 (664aebab7d).
• Cleanup goroutine race internal — re-check work list after marking asleep (b450b5409d).
• Synctest tracebacks internal — stack traces for bubbled goroutines (86101b083a, b7382cc1f0, 848a768ba7); runfinq always shown (9c88db5f1e).
• arm64 backoff internal — ISB/backoff for lfstack.pop/spanSet.pop, procyield via counter-timer (Garbage −40%) (50586182ab^1.26).
• Plan 9 internal — monotonic time, randomness (4c32b1cc75, 9f55e7bc21, adb29670fc).

Compiler frontend · types2

From Compiler frontend: the types2 type checker (Go 1.18–1.28)

• Bringing up generics in types2 (type-parameter operands, core/structural types, tparam-as-interface) — Done: removal of the checker’s use of core types (see Design). Left open: nothing noted.
• Type sets, union terms, interfaces and constraint satisfaction — Done: commonUnder replaces coreType (see previous thread).
• Type inference and unification — Done: nothing substantive besides core-type removal (previous thread).
• Type declaration cycles, valid-type detection and the object path — Done: 5e7f0b947c infrastructure to record declaration positions for panics (off by default); 608e9fac90^1.26 (“flip on position tracing”, 1.26 by log) enables it after compilebench showed no significant change (geomean −0.77%).
• Language features as seen by the type checker (by release) — Done: user-visible core types removed from the language spec: checker side in “Bringing up generics”; Parser-detected invalid ... (ad7b46ee4a).
• Error reporting: codes, messages, and sub-errors — Done: the rewrite listed above; 18ad74dd36 dumps the position stack on non-bailout panics.
• Sharing code between go/types and types2 — Done: 8a7742e78c range.go.
• Checker performance, memory and API ergonomics — Done: 9189921e47 and fe9b292b11 (used maps allocated in initFiles); 4ac729283c/1cf6386b5e (interface methods not registered in Info.Types, then reverted).

Other changes

• API additions user-visible — 932a4a4bef Var.Kind() / VarKind (go/types), b3aff930cf LookupSelection, 34b70684ba/34cf5f6205 interface-method field type tests, 34b70684ba append(bytes, y...) typing.
• Small fixes, unifier/assignment refactors, message and tracing tweaks internal — 4 commits: aa8d4df9d9, 011b7ce8d1, 3ada42ffed, bc5aa2f7d3

Compiler frontend · syntax, IR, generics

From Compiler frontend: syntax, IR, unified IR and the generics implementation (Go 1.18–1.28)

• Generics via GC-shape stenciling and runtime dictionaries — Done: shapify fixes: no shapify when reading a reshaping expression (bfbf736564), relaxed reshaping (da0e8c4517).
• Unified IR: one serialization for noding, stenciling, inlining and export — Done: formal documentation of UIR: grammar (cad1fc5207, 019d798726), sections renamed Reloc->Section (52624e533f), string/PosBase/Pkg/Name/Obj sections documented (a24f4db2a2, 2ab210bc74, d93bea0e59, 11660d537b, 25ca686a0b), reference-table rationale (f14f3aae1c).
• Compile-time performance, concurrency and the self-PGO’d compiler — Done: fece9871bb profile refresh.
• Compiler debugging infrastructure: bisect/hash debugging, flags, dumps — Done: 2c9689ab0e (variablemakehash), e46d586edd (literalallochash) so allocation optimizations are bisectable.
• Frontend support for new language features and spec changes — Done: label requirement after goto (go/parser), scalar-allocation aggregation for heap escapes (5fc596ebe7), type-based alias analysis (11f7ea8ce0), bdef177831 go/parser requires a label after goto.

Other changes

• Allocation optimizations internal — type-based alias analysis (11f7ea8ce0), 5fc596ebe7, 8203265d5e concatbytes, f27d86c631 shift tests.
• Inline tail calls internal — enabled 6c70f2b960, reverted 1aa3362093.
• Error message user-visible — importing package embed (5b748eed9c).
• Unified IR: noder/writer/reader refinements and pkgbits (1.25) internal — RType/implicit-conversion wiring, *At helpers, optional expression handling, UIR documentation and grammar, pkgbits.Index/SectionKind renames, synthetic closure positions, unsafe.Sizeof handling and related reader/writer fixes (68ca584ffa, 30b2b767d6, 6681ff9c9e, d73aa535de, 195e64232d, fd6afa352d, 96d2211c61, bfb8f13274, 26e05b95c2, b6e251b264, cc119ee391, 497cb7c0c3)
• Generics: dictionary/shape fixes in the stenciler (1.25) internal — fixes in dictionary passes for type assertions, comparisons, switches, nested instantiation packages, and recursive generic interface instantiation (b1f259b1b4)
• Old typechecker and IR node-model cleanup (1.25) internal — removal of unused typecheck/ir/types code (Ntype, CurFunc uses, vargen, Resolve dependencies), more typed IR node construction, simplified DeclFunc/NewMethodExpr and constant handling (67e0681aef)
• Syntax package: error messages and parser tolerance (1.25) internal — better syntax and go/parser error messages, accepted type-parameter lists, parser simplifications (c8664ced4e)
• Housekeeping (1.25) internal — dead-code removal, ioutil/seek/slices/atomic-type replacements, README edits, debug-flag removal, assorted small bug fixes and doc comments (586e205522, b38b0c0088, deb6790fcf, 116b82354c, 198c3cb785, 3e82316a43, 961818e013, bb917bb030)

Middle end and SSA optimizer

From Compiler: middle end and SSA optimization (Go 1.18–1.28)

• Profile-guided optimization (PGO): hot-call inlining, devirtualization, profile preprocessing — Done: a zero-weight edge is never chosen as the hottest callee (13b1261175). Left open: nothing noted.
• Inliner: budget, closures, and the interleaved inline/devirtualize driver — Done:
  • Infinite recursion and out-of-memory when inlining closures were fixed (34073a736a, 4f45b2b7e0).
  • The recursion test is now based on inline positions (cad4dca518), which made older checks unnecessary (1cf6b50263, +0.5% cmd/compile text).
  • ir.ContainsClosure was added (a08984bc8f), and CallExpr.NoInline was removed (3033ef0016).
  Left open: nothing noted.
• Escape analysis: call modelling, closures, func-variable callees, and solver speed — Done:
  • The solver speedups 04a9b16f3d, 74304cda29 and f506ad2644.
  • Escape uses an ir.ReassignOracle (e89791983a).
  • unique.Make models the data flow of non-string pointers (5e6a868b28), and SSA eliminates the string copy in calls to unique.Make (27ff0f249c).
  • -m=2 logs became more accessible (8bf816ae68, c1aad08901).
  • The literal and constant allocation avoidance described in the next thread.
  Left open: nothing noted.
• Allocation avoidance: stack-allocated slice backing stores and allocation-free literals — Done: user-visible fewer allocations.
  • Constants are propagated into interface conversions (ed24bb4e60, 326e5e1b7a). Zero values use the global zeroVal (c3bb27bbc7). Composite literals convert to interfaces without allocating (f4de2ecffb).
  • Follow-ups preserve side effects when rewriting with literals (988a20c8c5), keep the optimization on under coverage (479b51ee1f), add literalallochash (e46d586edd), and fix line numbering (66536242fc).
  • Variable-sized makeslice is stack-allocated, pointer-free elements only (7a427143b6). Pointerful elements were then allowed (2d050e91a3), and static len/cap are resolved (ba3f568988).
  • append backing stores are stack-allocated, cmd/go <0.1% bigger (ce88e341b9).
  • variablemakehash and the threshold flag were added; they found the culprit in #73199 (2c9689ab0e). The optimization is off under -N (21acfdc4ef).
  • Stack buffers are over-aligned to match heap alignment (c8bf388bad).
  • In a related backend change, scalar heap allocations are aggregated (5fc596ebe7).
  Left open: the optimization caused #73199, a user code bug exposed by stack allocation.
• Prove pass and bounds-check elimination: from poset relations to a dense limit table, topo-sorted walk, and division-aware proofs — Done:
  • isGenericIntConst (bdc2d856a8).
  • Bool-to-uint8 limits (b7c9cdd53c). Duplicate unsat reports are suppressed (00635de759).
  • min/max inside flowLimit (d2842229fc). OnesCount limits (c00647b49b, 99411d7847).
  • Bit size derived from type size (644b984027).
  • Bounds on signed x % 2^k (3baf53aec6).
  Left open: nothing noted.
• Memory-op optimization: memcombine, dead stores, store-to-load forwarding, alias analysis, mem2reg — Done:
  • OpLocalAddr is not a load in DSE (0825475599).
  • Type-based alias analysis (11f7ea8ce0, #70488).
  • memcombine handles different-size stores (113b25774e), with an offset fix (d6c29c7156).
  • Store-to-load forwarding works with copy-compatible types (16a6b71f18).
  Left open: nothing noted.
• Loop variable per-iteration semantics (loopvar) — Done: an ICE in the transformation was fixed (a0dc7bf084). Left open: nothing noted.
• testing.B.Loop support in the compiler (bloop) — Done: a stopped timer in B.Loop is detected (5918101d67). The B.Loop documentation was clarified per #73137 (ab2926291b). Left open: ab2926291b hints that “the exact approach might change in the future”.
• Range-loop and range-over-func lowering in walk — Done: side effects of range over an array are not evaluated when there are none (8af32240c6). A cast was added in the final-value computation (3452d80da3). Left open: nothing noted.
• Constant-folding loads from type metadata and dictionaries — Done: load constants from itabs (43e6525986); avoid ifaceeq call when the interface is known direct (072eea9b3b).
• ARM64 load/store pairing (LDP/STP) — Done: pair loads and stores on arm64 (20d7c57422); write-barrier code shaped for pairing (187fd2698d).

Other changes

• Cleanups internal — linkedBlocks removed (9dc572eab5), Sym/Aux replace interface{} (eec3745bd7), hv2 <= 127 (ef7f09149b), intrinsic aliases for bits.Len/OnesCount (7e8ceadf85, 03cb8d408e), PPC64 AND-shift merge fix (ee7bfbdbcc).
• Escape/alloc-related lowering internal — []byte→string map-key lookup moved to SSA (89c2f282dc), concatbytes optimized (8203265d5e), phiopt cases (c9d0fad5cb), shortcircuit patterns (4d6722a8fd), load-op merging in more situations (d681270714).
• Constant folding internal — PopCount, bits.Add64, 128-bit multiplies (3e033b7553, b60b9cf21f, 7d0cb2a2ad), 8/16-bit multiplication distribution (f7dbbf2519), MakeTuple op (4ff70cf868).
• Fixes internal — 32-bit alignment of values with 64-bit fields (1647896aa2), side effects of len() arg (336626bac4), closure LSym for type conversions (fc641e7fae).

Backends and assemblers · architectures, ABI

From Compiler backend: architectures, register allocation, ABI and assemblers (Go 1.18–1.28)

• Register allocator, scheduler and block-level backend passes — Done: handle desired registers of 2-output instructions a0029e95e5; tighten register-pressure heuristic relaxed 1e756dc5f7; nilcheck ops not moved by tighten a1ddbdd3ef; “don’t preload registers if destination already scheduled” 524946d247; induction-variable increments scheduled late 19f05770b0; temp register reuse fixed cc16fb52e6; GOSSAFUNC output now includes liveness 7472b4c324; residual GC register map code removed 3b25b3c195.
• amd64 code generation and GOAMD64 microarchitecture levels — Done: x < 128 → x <= 127 for imm8 (−3.15% archive/tar geomean) d524e1eccd; (x&1)==0 via AND/XOR e57769d5ad; shift-by-1 → ADD cd595be6d6; reuse flags rather than TEST x x 6729fbe93e; x*z+y lowers to FMA when enabled (v3) 43b7e67040; multiplication strength reduction by automatic linear-combination search (amd64, arm64) 12110c3f7e; PCMPEQB commutative 8cb6d3b826; Plan 9 SSE/FMA special-cases removed c18ff21cc8,35cb497d6e,c9b07e8871.
• arm64 code generation and assembler — Done: LDP/STP load/store combining variants 0b88a87879, sign-extension bug in paired 32-bit loads fixed beac2f7d3b; BTI instruction 9ce47e66e8; epilogue SP increment in one instruction f7cc61e7d7; register encoding moved into oprrr/opldrr helpers (f90f3cfcf7,4f5863c2ce,fd0805b660); more conservative use of zero register 16dbd2be39; constant-classification refactor (391dde29a3…06f82af183); OpMove used more than memmove calls 11fa0de475; fold negation into add/sub dfebef1c04.
• riscv64 compiler and assembler (RVA profiles, vector, compressed) — Done: vector load/store, integer arithmetic, fixed-point, reductions, mask, FP and permutation instruction support, vsetvl configuration (37e9c5eaba,535e0daefd,4c1b8ca98c,2e60916f6e,936ecc3e24,5a342266e6,1cc624fd62,c8545439b5); CSR map 5a7db813a6; synthesized MIN/MAX d88f93f720; GORISCV64-gated intrinsics TrailingZeros (CTZ), Len (CLZ), OnesCount (CPOP, with Zbb check for rva20u64: +5 ns without Zbb vs up to −11 ns with), Bswap (REV8) 6fb7bdc96d,b70244ff7a,4d10d4ad84,90e8b8cdae; constant construction d37624881f; plugin build mode cdc9560794; R_GOT_PCREL_ITYPE_RELOC 371ee1469c; generic simplifications 123141166b; fold zero register for *zero instructions e373771490.
• loong64 backend and assembler — Done: ~25 vector assembler commits (e.g. 53d689ce99,b0e2f185c5,b10c35945d,47fd73a51c,d65c209b4b); immediate loading optimized b9630c56db,f24d2e175e; shifts 2a772a2fe7,b143c98169; unalignedOK for memcombine 09d76e59d2; negation folding c31a5c571f; prefetch intrinsic da9c5b142c.
• Duff’s device replaced by generated loops; Zero/Move lowering — Done: nil check done before calling duff functions on arm64/amd64 dbaa2d3e65; noDuffDevice flag removed 26040b1dd7.
• Late-lower pass, memory-op combining and arch-independent rule migration — Done: automated multiplication strength reduction 12110c3f7e; loong64 equivalent e071617222^1.26 (1.26); constant folding flags af278bfb1f.
• Stack slot merging, liveness and stack maps — Done: residual register GC map code removed 3b25b3c195; stack object alignment 1.28 5d29d80b6c^1.28.
• WebAssembly backend — Done: indirect calls use 64-bit ops because >65536 functions overflowed the 16-bit shift of the PC c876bf9346; large return address as i64 d60a684c87.
• s390x, MIPS, 386 and 32-bit ARM backends — Done: negation folding for mips64x/mipsx/s390x (8097cf14d2,d52679006c,7589e96042); publicationBarrier on mips (94e3caeec1,3dbc775d60); s390x LCDBR 4ae6ab2bdf.

Other changes

• panic transfer via deferreturn PC internal — recover from range-over-func defers uses the PC of deferreturn as target, removing the register-value handshake (go cmd −0.2%) c2ae5c7443,9ddeac30b5.
• Signed loads from read-only memory a7e331e671; preamble preemptible (reverted) 3f3782feed,fa2bb342d7.
• Generated-files up-to-date test internal d4c6effaa7.
• endbr64 and BTI (CET/BTI) user-visible 04de448c6d,9ce47e66e8.
• 14 small obj/riscv and riscv64 rules commits internal — e.g. cc: combine negation and word sign extension on riscv64; obj/riscv: support MOVD with floating point constants; ssa: optimise more branches with zero on riscv64. (21417518a9,d7c242a19a,e6c2e12c63,215de81513,7a2f757c52,24b395119b,1aee4f3464,c386ed107e,af133d86e4,10d070668c,d000963d04,5a1f47a7f7,0d7dc6842b,2363897932)
• 4 small obj/arm64 and arm64 rules commits internal — e.g. ssa: correct MOVDnop handling for arm64; obj/arm64: replace range checks with isaddcon2; obj/arm64: simplify conclass. (fba83cdfc6,dab10bca37,c524db9ca8,4464a7d94f)
• 1 small obj/ppc64 and PPC64 rules commits internal — e.g. cmd/internal/sys: allow unaligned loads on big-endian ppc64. (3981b446dd)
• 16 small obj/loong64 and loong64 asm commits internal — e.g. obj/loong64: add {V,XV}NEG{B/H/W/V} instructions support; obj/loong64: add {V,XV}{FSQRT/FRECIP/FRSQRT}.{S/D} instructi; obj/loong64: add {V,XV}{ADD/SUB}.{B,H,W,D,Q} and {V,XV}{ADD/. (f41fdd962d,80f068928f,39d7a49736,2ce1fb4220,e3ea8e68fb,ac8474c0a2,0095b5d098,005d7f29d1,fb2c88147d,3474c52e4b,b386b62852,d13da63929,f49ca5f525,c62099cfac,045b5c1bfb,d37a1bdd48)
• 10 small ssa/regalloc/liveness/obj generic commits internal — e.g. cc: remove reduntant Zero rule; cmd: use cmd/internal/hash.New32 and Sum32 only; cc: simplify intrinsification of TrailingZeros16 and Trailin. (46fd6b4e37,eab8e987c0,927fdb7843,1b1c6b838e,8b8bff7bb2,a20d583bb9,afe11db4a7,47ab9cbd82,8a8efafa88,4ce1c8e9e1)
• 1 small obj/x86 and amd64/386 rules commits internal — e.g. cc: remove carriage returns from AMD64latelower.rules. (a65078204c)
• 1 small mips/s390x/arm/wasm commits internal — e.g. obj/s390x: fix potential recursive String call. (9f3dd846e2)

Linker and object files

From Linker, object files and debug info (Go 1.18–1.28)

• //go:linkname restrictions and the blocklist — Done: user-visible more blocked 683810a368, 47b941f445; linkname reference to a TEXT symbol allowed regardless of size 70109eb326. Left open: nothing noted.
• DWARF 5 migration — Done: dead R_DWARFFILEREF removed 38825ef347; DW_LNCT/DW_UT constants 41a30dd192; R_DWTXTADDR_* relocations 31b800abe6; SDWARFADDR 282a14ec61; line tables 767e7680ec; .debug_addr and v5 .debug_info e7cd4979be; rnglists 4c0a47a8ff; loclists bef2bb80a9; fix of wrong inlined-subroutine pc range (always DW_AT_ranges) 955cf0873f, bdfa604b2e; DWARF 5 enabled by default ca19f987ca; loclist fix for heap return vars without optimizations 71c2bf5513. Left open: e7cd4979be says DWARF stays damaged until the later patches landed; not enabled on macOS.
• Linker performance and memory — Done: -e unlimited errors 76c18e2ed2; -funcalign=N 9302a57134; size fixups after references load 88cf0c5d55. Left open: nothing noted.
• Per-platform object formats and internal linking (Mach-O, PE, XCOFF, wasm, ppc64, riscv64, loong64) — Done: loong64 internal linking f212b6499a; Wasm method/func tables hold only function index so >65536 funcs fit 9c507e7942; Mach-O X86_64_RELOC_SUBTRACTOR for race syso 5fb9e5dc19; .got in __DATA_CONST 61641c1145; >4 GB base for 64-bit PE 176a2154aa. Left open: nothing noted.

Other changes

• Misc linker internal — __got section name (2298215f5b), loong64 PLT trampoline removed (6cb8c839f9), LookupPkgSpecial(pkg).Runtime shared by go and linker (f3c69c2e78), riscv64 gcc-15 and mapping symbols (12e5efd710, a2fbb50322), macOS 12.0 versions (33d3f603c1), devel detection (214b208e24), mmap check (b338f6bfa6), pe64 bool (ba01453bbe).
• Linker block of duplicate BSS internal — larger-size BSS wins (8f6c083d7b, 0f8ab2db17).