Go 1.26

Green Tea by default. SIMD (archsimd, amd64) lands behind an experiment. Old maps deleted. move2heap stack-allocated appends. panicBounds. Duff’s device removed. s390x register ABI. runtime/secret, heap-base randomization and freegc experiments. Goroutine-leak profile experiment. new(expr). Size-specialized malloc: on in master, off in the release.

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)

• Green Tea: span-based marking with vector scan kernels — Done: user-visible enabled by default except on loong64, where the compiler crashed on the Green Tea code (7fbf54bfeb); loong64 enabled later in the cycle (c53cb642de). The AVX-512 kernels were imported (889ab74169) and wired into scanSpan, giving “an extra 15-20%” on dense spans (5d040df092). scanobject was duplicated per mode to tell them apart in profiles (41b429881a, 1aa154621d). The global span queue was replaced by spanSPMC rings and work-stealing (d7a38adf4c); gcMarkWorkAvailable was split into a termination condition and a worker-spawn heuristic (fde10c4ce7). Follow-up fixes for spanSPMC: an unbounded loop when more than 64 live rings existed (f3312124c2), rings leaked on P destroy (3ee761739b, 7dd54e1fd7), list type (046dce0e54), lock rank (052c192a4c), and removal of the unused mspan.scanIdx (2baa1d1762). Kernel fixes and tuning: cache-size probing (adbf59525c), a size-class check (ff2ebf69c4), a register destination for VPCOMPRESSQ (−70% on Zen 4, 041f564b3e), an AVX-512 filterNil (+40% throughput, a9704f89ea), and VZEROUPPER (0d71bd57c9). A Go-simd-package kernel set (da92168ec8, −19% ns on the scan benchmark) was reverted for naming and merge conflicts with a plan to “do it again” (0c69e77343). Left open: freeSomeSpanSPMCs batching was dropped, “may want to revisit … for very high GOMAXPROCS” (f3312124c2). The simd-package kernels have not been relanded at HEAD, and scan_go.go still carries a TODO.
• Heap bitmap rewrite and allocation headers — Done: invariant work hoisted out of heapBitsSmallForAddrInline in Green Tea’s sparse scan loop (65c09eafdf). Left open: nothing noted.
• mallocgc decomposition and size-specialized malloc — Done: user-visible GOEXPERIMENT=sizespecializedmalloc added (60c1ee9183) and enabled by default (a7917eed70). Generator, templates and tables (a69395eab2, 411c250d64); the size-class generator moved into _mkmalloc (f37d75472d); direct compiler calls (19a30ea3f2); valgrind support (c54dc1418b); generator polish (21ebed0ac0, 2947cb0469, e2fef50def); an explicit nil check in heapSetTypeSmallHeader to avoid read-then-write page faults (tile38 at GOMAXPROCS=192: +9.6% ops/s, p99 −15%, 3032894e04). Left open: “I still need to investigate heapsampling.go” (411c250d64).
• Scavenger redesign — Done: on Linux, memory released with MADV_DONTNEED is known to be zero, so software zeroing is skipped on reuse (27937289dc); a duplicate scavengeIndex.sysGrow call removed (4fca79833f). Left open: nothing noted.
• Finalizers, runtime.AddCleanup and the cleanup queue — Done: user-visible panics when ptr == arg or when the cleanup closure captures ptr (06412288cf, eb63ef9d66); separate argument storage (−16.7% ns, −33% B/op, 1bb1f2bf0c); benchmark (9fd2e44439); /gc/cleanups/… and /gc/finalizers/… queue metrics (a4d99770c0); test log (43ebed88cc). Left open: nothing noted.
• Weak pointers and the unique package — Done: weak handles padded to 16 bytes so they bypass the tiny allocator and cannot leak neighbors (80c91eedbb); canon map fixes (ffa882059c, 47bf8f073e); test (a33bbf1988). Left open: “We could potentially do better by using the 8 byte noscan size class” (80c91eedbb).
• Object pinning (runtime.Pinner) — Done: documentation (856238615d, 76345533f7). Left open: nothing noted.
• Memory arenas experiment — Done: assist credit for arena chunks deducted only during GC, restoring behavior lost in the 1.24 malloc split (d79405a344); test skips (1a03d0db3f, 96e142ba2b). Left open: nothing noted.
• runtime.freegc: compiler-directed explicit freeing — Done: user-visible freegc with reuse lists and clobberfree (fecfcaa4f6), an assist-credit test (120f1874ef), size-specialized noscan reuse (50128a2154), the experiment renamed to runtimefreegc and off for 1.26 (5a347b775e), and compiler alias analysis with growsliceNoAlias (4879151d1d, compiler area). Left open: freegc “throws if noscan is false (support for which is implemented in a later CL)”. The specialized-malloc path disables pointerful freeing via a TODO, and clobberfree=2 (mprotect) is described as a later experimental CL (fecfcaa4f6, 50128a2154). None of these has landed at HEAD.
• Mark worker scheduling and GC work queues — Done: worker selection during start-the-world (829779f4fe, a18aff8057); atomic p.gcFractionalMarkTime (d3aeba1670). Left open: if fewer than 25% of Ps are idle, some user Ps still turn into dedicated workers, and stack-scan resumes still migrate goroutines (a18aff8057).
• Stack and register root scanning — Done: stack shrinking disabled whenever a goroutine waits for suspendG (tracer, mutex profiling), instead of trying to be shrink-safe (bd07fafb0a). Left open: nothing noted.
• Address-space layout, tagged pointers and the mem.go layer — Done: one more address bit for tagged pointers for amd64-on-arm64 emulation (2a7f1d47b0); riscv64 sv39/48/57 detection, cutting startup from ~85 mmaps/66 munmaps to 22/0 on sv39 (dae9e456ae); gccgo build tags (643f80a11f). Left open: nothing noted.
• Linear-memory heaps (Wasm and sbrk platforms) — Done: user-visible Wasm heap arena and page-chunk size reduced to 512 KiB; hello-world drops from 8 MB to <3 MB linear memory (53009b26dd). sbrk mem.go switches to the system stack to avoid memlock self-deadlock (11d5484190). Left open: nothing noted.
• Heap base address randomization — Done: user-visible randomization (6669aa3b14), on by default (72ba117bda), disabled with ASAN/MSAN (77aac7bb75), test fix (465b85eb76). Left open: nothing noted.

Other changes

• runtime/secret allocation size internal — the secret special records the object size so memclr does not overwrite neighbors (b8c4cc63e7).
• hexdumper internal — configurable hexdump (any word size, ASCII column, multi-annotation marks) that replaces hexdumpWords output in the runtime’s existing dumps, ahead of debuglog integration (e912618bd2).
• GC-based goroutine leak detection user-visible — goroutine leak profile computed by the garbage collector, behind GOEXPERIMENT=goroutineleakprofile (8c68a1c1ab, diag/sched area).

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: Timer.Reset(0) inside a synctest bubble no longer deadlocks on sendLock being taken twice (385dc33250). Left open: nothing noted.
• Scheduler core: spinning Ms, work conservation, stealing, STW interaction — Done: P-to-M affinity across STW: sched.midle made doubly linked so a specific M can be pulled out, Ps restart on their previous M and STW preemption via goschedImpl keeps the G on its P (2263d4aabd, 4ebf295b0b; TestTraceSTW error rate 99.8% -> 1.9%); GC mark workers chosen in procresize during start-the-world to avoid sequential wakeps (a18aff8057); steal path avoids the 3us usleep before taking runnext when the victim is in a syscall, now that such Ps stay _Prunning (d55ecea9e5); debug mode extending _Grunning-without-P windows (7a8d0b5d53). Left open: “best effort” affinity; spinning M can still steal from a slow-restarting P.
• Syscall path: eliminating _Psyscall, _Gdeadextra, and goroutine states — Done: the above commits, plus user-visible runtime/metrics for goroutine scheduling states, total goroutines created and thread counts (13df972f68, ab8121a407, 78a3968c2c). Left open: “_Grunning-without-P” debugging only via an env debug mode (7a8d0b5d53).
• GOMAXPROCS container awareness and dynamic updates — Done: cgroup v1 non-root mount points handled, path validation and fuzz tests (c2af9f14b4, cead111a77, 67851547d8). Left open: nothing noted.
• testing/synctest bubbles — Done: all bubble violations are fatal panics (so e.g. context cannot be left inconsistent by a recoverable panic) (a8564bd412); synctest GOEXPERIMENT removed (4c0fd3a2b4); Reset(0) hang fix (385dc33250); specials lock (5dcedd6550). Left open: nothing noted.
• Asynchronous preemption: extended register state — Done: scalar registers saved off-stack (af0c4fe2ca), AVX2/AVX-512 (55d961b202), arm64 Neon (955a5a0dc5), loong64 fcc/lasx/lsx (bd885401d5, 99cf4d671c), cherry-picked from dev.simd. Left open: GC scanning of that state.
• sync package API evolution — Done: WaitGroup.Go no longer calls Done when f panics (95a0e5adc1); race annotations kept when panicking (8f74f9daf4). Left open: nothing noted.
• Goroutine lifecycle and identity — Done: user-visible goroutine leak profile behind GOEXPERIMENT=goroutineleakprofile (8c68a1c1ab); DIT inheritance (af14f67911). Left open: experiment only.

Other changes

• GOTRACEBACK for user panics internal — 02d1f3a06b.
• Misc internal — recover detection (7b50024330); eventfd errno (d3be949ada); debug settings initialised earlier (300d9d2714); p.destroy assert (6469954203); castogscanstatus no panic (b5353fd90a); preemptPark dropg after trace (6455afbc6f); pMask dedupe (69e74b0aac); NetBSD sysmon workaround dropped (52b6d7f67a); internal/cpu inline (fe42628dae); loong64 CAS jump (d4b17f5869); secret goroutine behaviour restored to proposal (63fced531c); dev.simd merge commits (426cf36b4d, 9eeb1e7a9a, 574854fd86, 0710cce6eb, a8669c78f5); panic output priority (89dee70484); openbsd check (87e72769fa); internal/itoa removal (49c1da474d); fips WithoutEnforcement (86bbea0cfa).

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: user-visible GOEXPERIMENT=swissmap removed, ~6,000 lines of old map code deleted, map_swiss.go renamed to map.go (2ae059ccaf, 82a1921c3b); Clear only checks “any full slot” rather than computing the bitset (8098b99547) and only conditionally clears groups when sparse (de5d7eccb9); hoisted h2(hash) out of loops (b096ddb9ea); 32-bit hashes on wasm because 64x64->128 multiplies are not intrinsics there (a26f860fa4, -26% Hash geomean). Left open: nothing noted.
• Map hashing and hash functions (memhash, aeshash, fallback hash) — Done: 32-bit hash on wasm (a26f860fa4). Left open: nothing noted.
• Compiler-generated equality/hash functions and type descriptor emission (reflectdata) — Done: ptrmask no longer cleared in fillptrmask (e0807ba470), type-load folding (5b218461f9), TypeFor made inlineable (dc960d0bfe; also forces TypeLinksym in writeType for non-shape types so later lookups see the symbol). Left open: “TFlag … decent help … if statically known” (5b218461f9), implemented later (the “TFlag folding” CL 802360 cited in 8fc7e76411^1.28).
• internal/abi type layout unification and type metadata (rtype → abi.Type) — Done: the direct/indirect-interface bit moved from the high bit of Kind to TFlag, so kind access needs no mask; IfaceIndir → !IsDirectIface; gocore and delve had to be updated and the commit lists “TODO: any other tools need updating?” (f7d167fe71). Left open: debugger/tool updates explicitly open.
• reflect implementation: performance, allocation and escape behaviour — Done: user-visible iterator equivalents Type.Fields/Methods/Ins/Outs style (e8fdfeb72b); TypeAssert[I] allocates only on success (684e8d3363); TypeFor fully inlineable (dc960d0bfe); zero-sized fields of direct structs fix attempted then reverted (b3388569a1, caf9fc3ccd). Left open: the reverted fix.
• memmove / memclr / memequal / internal/bytealg assembly optimisation — Done: RISC-V vector indexbyte, equal, compare (3406a617d9, 75ea2d05c0, d83b16fcb8), mips64x/s390x memclr (b6cf1d94dc, a8edd99479), loong64 Index (dd3abf6bc5), arm64 DC ZVA mnemonic instead of WORD (e7351c03c8), benchmark for small sizes (e84ed38641). Left open: nothing noted.
• Runtime randomness: fastrand, chacha8rand and math/rand/v2 — Done: riscv64 vector implementation (6f04a92be3), loong64 instruction mnemonics (5defe8ebb3, 01c29d1f0b). Left open: nothing noted.
• Concurrent hash-trie map (internal/sync.HashTrieMap) — Done: warning about incorrect unsafe usage documented (1718828c81). 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: in-band end-of-generation event, giving unambiguous truncated trace detection and removing batch spilling (4a7fde922f, 77f911e31c); trace-writing fast path: seqlock replaced by write flag (d68aec8db1), m.profStack used instead of zeroed pcBuf (7fbd141de5, trace stacks now obey profstackdepth), no unique-string write for empty strings (0bc192368a); reentrancy generation-skew fix (dc9a3e2a65); P-clear vs. GOMAXPROCS race (1de9585be2); _Psyscall eliminated with tracer shielded from _Grunning-without-P (7244e9221f); Event.String improvements (de2d741667, b80ffb64d8); cmd/trace raw/text converters (bf58e7845e). Left open: alloc/free experiment tests skipped by default (4837bcc92c).
• Flight recorder and trace subscribers — Done: in-band EndOfGeneration for runtime.ReadTrace + parser (4a7fde922f); Log benchmark (8d9906cd34); test traces dumped on failure (ee4d2c312d, 89614ad264). Left open: nothing noted.
• CPU profiling, goroutine profiling, pprof labels and the protobuf writer — Done: fix double notewakeup when two writes hit a full profBuf (b604148c4e); remove fixed sleep in CPU profile reader (121bc3e464); GODEBUG=tracebacklabels=1 adds escaped pprof labels to goroutine headers in tracebacks, with a new runtime.quoted print type (6851795fb6); internal/profile.Parse allocs −36% (c5559344ac). Left open: escaping of non-ASCII runes and invalid UTF-8 “can be improved” (commit message); opt-in to experiment with format.
• Goroutine leak profile (GC-based leak detection) — Done: user-visible experiment added, default off (8c68a1c1ab); cleanup of profile writing, stress tests for concurrent leak and goroutine profile requests (8873e8bea2); noopt test fixes (625d8e9b9c). Left open: stress subtests flaky (8f45611e78).
• runtime/metrics and MemStats — Done: user-visible goroutine scheduling-state metrics (13df972f68), total goroutines created (ab8121a407), current thread count (78a3968c2c), cleanup/finalizer queue metrics (a4d99770c0). Left open: nothing noted.
• Race detector, ASan, MSan and Valgrind integration — Done: race on linux/riscv64 (8bcda6c79d); msan uses a dedicated copy routine (5f4b5f1a19); race darwin syso refreshed/restored (8282b72d62, 1b263fc604); ppc64x atomic And/Or wrappers (44cb82449e). Left open: nothing noted.
• Crash output, runtime/debug, BuildInfo and debuglog — Done: dlog.hexdump (8c41a482f9); tracebacklabels (6851795fb6, see profiling thread). Left open: nothing noted.

Other changes

• Trace parser/viewer fixes internal — GoUndetermined in GoroutineSummary (e38c38f0e5), Event.Range scope for EvGCSweepActive (2fe5d51d04), StateTransition stack docs (d2b3c1a504), CPU profile view not filtered by stack presence (00b8474e47), thread+proc annotations (704f841eab), recorder header-only write (d1e7f49e3d), s390x race ABIInternal (e92d2964fa).
• Trace test tooling internal — internal/trace testgen, event constructors for tests (54b82e944e), KernelVersionGE (68c6a73380).

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: recover decided by unwinding over logical frames (7b50024330, 08376e1a9c); panics printed before fatal-ing on defer (5cd1b73772); no “recovered” printed for a double panic with the same value (c93766007d); attempt to remove _defer.pc was reverted (361d51a6b5, 48c7fa13c6); gopanic linkname removed (62741480b8). Left open: wrapper-tracking in the old incremental scheme was only “ripped out in a subsequent CL”.
• Frame pointers: explicit NOFRAME, FP maintenance and FP unwinding — Done: FP cleared at thread entry points (eda2e8c683); getfp marked nosplit (d4972f6295). Left open: nothing noted.
• C-to-Go callbacks: extra-M reuse, system-stack bounds, init gating — Done: cgo call and callback benchmarks (70a2ff7648, 17b57078ab, 710abf74da); pointer-panic error messages (91ca80f970); cgo any params/returns (a593ca9d65); asmcgocall robust and 386 sigaction via libc when cgo (c5737dc21b). Left open: nothing noted.
• Library entry points and runtime initialization — Done: library entry unification (fdea7da3e6, 45d6bc76af, c9257151e5, eb1c7f6e69, 36863d6194) and early debug-setting init (300d9d2714). Left open: nothing noted.
• runtime/secret and data-independent timing (DIT) — Done: user-visible runtime/secret package (a3fb92a710); DIT inheritance, no thread locking (af14f67911), speculation barrier (2244bd7eeb); experiment guard (8d31244562), documented allocation caveats (516699848b), proposal-conforming goroutine behaviour (63fced531c). Left open: only linux/arm64+amd64; allocation guarantees explicitly loosened.
• vDSO, vgetrandom and fast clocks — Done: vDSO symbol hash test (a2baae6851^1.19), TestUsingVDSO check (98a031193b). Left open: nothing noted.
• Signals and async-preemption machinery — Done: AVX2/AVX-512 preempt state (55d961b202, cef381ba60), loong64 fcc save (bd885401d5), libc sigaction on 386 (c5737dc21b), loong64 (180795ceac^1.25, 1.25). Left open: nothing noted.
• Syscall layer: AllThreadsSyscall, internal/runtime/syscall, libc and Windows SyscallN — Done: Windows/darwin generic syscalln and moves (98f301cf68, ef40549786, e666972a67, 4a0115c886), 64-bit futex (3cf1aaf8b9, 4620db72d2), syscall clearenv fast path user-visible (836fa74518, −98% at 100 vars). Left open: nothing noted.
• Windows runtime port — Done: windows/arm removed (dc54d7b607, 9b0a507735); runtime types moved to internal/runtime/syscall/windows (889e71c2ac); CreateThread retry on access denied (e93f439ac4). Left open: nothing noted.
• New, ported and removed platforms — Done: windows/arm and openbsd/mips64 removed (dc54d7b607, d5dc36af45), riscv64 compressed instructions (b9ef0633f6), s390x regabi (41af98eb83, 58a48a3e3b). Left open: nothing noted.
• CPU feature detection, GOAMD64 checks and cpu GODEBUG options — Done: AVX-512 CD/DQ and “basic AVX-512” (e73afaae69), GFNI/VNNI fixes (1ee72a15a3, d00e96d3ae), AVX under macOS 15 Rosetta 2 (9125351583), DebugOptions inlined (fe42628dae). Left open: nothing noted.

Other changes

• Misc internal — NetBSD sysmon workaround dropped (52b6d7f67a), netbsd/arm64 entry jump (6d51f93257), dead architectures cleaned from constraints (9a5a1202f4), timediv deleted (d32b1f02c3), bounds-check constants and PanicBounds in internal/abi/funcdata.h (741a19ab41, 46e5e2b09a), X15 zeroing guarded by GOEXPERIMENT=simd then low-128-bit only (de39282332, f38e968aba, 4c311aa38f), stack object record search tightened (ee0275d15b), osArchInit removed (b7c20413c5), loong64 R31 save (e8126bce9e).
• Linkname docs internal — 8d0bef7ffe; runtime/cgo tweaks (dec2b4c83d FreeBSD binuptime).
• Stack traces for durable blocking user-visible — “(durable)” tags corrected (7e0251bf58); Stack can traceback goroutines in syscall _Grunning window (c93cc603cd).
• Intrusive list internal — (5f11275457).
• procyield safety internal — (f838faffe2, 79ae97fe9b).

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: 6b32c613ca makes typeset return an iterator; 0c14000f87 removes under(Type) in favour of Type.Underlying().
• Type sets, union terms, interfaces and constraint satisfaction — Done: typeset returns an iterator (6b32c613ca); 1abc6b0204 permits type cycles through type-parameter lists (the old restriction cc14fcac2b^1.18 from 1.18 existed only because of a type-inference deadlock, #49439, and type inference for generic type arguments had since been removed).
• Instantiation, Named types, Context deduplication and lazy expansion — Done: a refactor of the type-declaration lifecycle by Mark Freeman. 1099436f1b defines Named.fromRHS as the RHS type expression and sets Named.underlying only once the underlying type is computed (before, underlying could temporarily point at another Named). Followups: 4bdb55b5b8 (Named.under becomes resolveUnderlying), cf826bfcb4 (set exactly once), 39fd61ddb0 (guarded by Named.mu), b8aa1ee442/d09a8c8ef4 (fewer locks), 8401512a9b/9fcdc814b2/03bd43e8bb rename namedState values and resolve to unpack, b2af92270f verifies state-mask transitions in debug mode, d8a32f3d4b (wrap fromRHS into Named.rhs). The documented lifecycle also removes the package-level under(Type) function in favour of Type.Underlying() (0c14000f87). Also: 7b53d8d06e adds a loaded state between loader calls and constraint expansion, to fix a deadlock when SetConstraint is called from a lazy loader that recurses back to the loaded type.
• Generic type aliases and Alias type nodes — Done: alias handling folded into the type-declaration lifecycle (1099436f1b). Also: d28b27cd8e uses nil rather than Invalid for incomplete explicit aliases, which also lets Invalid aliases be memoized.
• Type declaration cycles, valid-type detection and the object path — Done: the lifecycle refactor (1099436f1b, 57362e9814 direct cycles checked as a separate phase, 1f4d14e493 package-level object sort as its own phase, 4fef9f8b55 object path for grouped declarations, 3531ac23d4, ff2fd6327e, c048a9a11f, 1274d58dac).
• Per-file language versions and version gating — Done: 06e57e60a7 only report version errors if new(expr) is otherwise valid.
• Language features as seen by the type checker (by release) — Done: user-visible a27261c42f new(expr) (#45624; version gate in version.go); 06e57e60a7.
• Error reporting: codes, messages, and sub-errors — Done: cae45167b7, 532e320349 (go/types: diagnostics for type shadowing).

Other changes

• Simd tooling excludes internal — 257c1356ec; 0f6397384b relaxes NewSignatureType for append(slice, str...); cdf64106f6.
• Small fixes, unifier/assignment refactors, message and tracing tweaks internal — 6 commits: 5a42af7f6c, d8a32f3d4b, 3be9a0e014, e704b0993b, efe9ad501d, 2622d2955b

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: never shapify an already-shaped pointer (a5866ebe40).
• Unified IR: one serialization for noding, stenciling, inlining and export — Done: aliastypeparams GOEXPERIMENT deleted (00a7bdcb55); dedup in mknode (ba840c1bf9).
• Static initialization and package init — Done: global new(expr) fix (a8291eb614), dead code removal (317be4cfeb).
• Compile-time performance, concurrency and the self-PGO’d compiler — Done: starting heap boost made independent of -c and raised to 128 MiB nominal, plus -d=gcstart=N MiB knob (ab2829ec06).
• Frontend support for new language features and spec changes — Done: user-visible new(expr) (7bc1935db5, 0f31d742cd untyped fix), spec: cycle restriction for type parameters removed (4684a26c26) and alias RHS rule (790384c6c2), testing.B.Loop keep-alive semantic (22f24f90b5, 3fd9cb1895, 509ddf3868, 2ac1f9cbc3).
• SIMD frontend: the “midway” source-to-source translator — Done: dev.simd branch infrastructure: k-register support (11d2b28bff), bodyless intrinsics (88c013d6ff), devel compiler not treated as release (4150372a5d, ebb72bef44), early prototype removed (b4d1e018a8).

Other changes

• Stack-allocated append backing stores internal — 32f5aadd2f; growslice free 4879151d1d.
• Saturating float-to-int internal — conversion made consistent between amd64 and arm64 (0e466a8d1d), new saturating disabled (7056c71d32).
• Devirtualize analysis internal — de9da0de30.
• Heap start internal — ab2829ec06.
• Old typechecker and IR node-model cleanup (1.26) 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 (26338a7f69, 16c0f7e152)
• Housekeeping (1.26) internal — dead-code removal, ioutil/seek/slices/atomic-type replacements, README edits, debug-flag removal, assorted small bug fixes and doc comments (4ee0df8c46, 5024d0d884, 6346349f56, 9035f7aea5, aea881230d)

Middle end and SSA optimizer

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

• Inliner: budget, closures, and the interleaved inline/devirtualize driver — Done:
  • Superfluous re-slicing s[:] is ignored in cost (a67977da5e).
  • Functions using runtime.deferrangefunc are not inlined (cd668d744f).
  Left open: nothing noted.
• Allocation avoidance: stack-allocated slice backing stores and allocation-free literals — Done:
  • append backing stores are stack-allocated even when the result escapes, through move2heap (32f5aadd2f). move2heap is no longer placed inside loops deeper than the declaration (3c6bf6fbf3).
  • len([]T{}) is propagated into make (84db201ae1).
  • Alias analysis plus runtime.growsliceNoAlias (4879151d1d) sit behind GOEXPERIMENT=runtimefreegc, which was set off by default for 1.26 (5a347b775e).
  Left open: 4879151d1d notes that the final slice memory is never freed and is left to other work; runtimefreegc stays off.
• Prove pass and bounds-check elimination: from poset relations to a dense limit table, topo-sorted walk, and division-aware proofs — Done:
  • Transitive len/cap properties (a3295df873) were reverted (d0a64f7969, #74974). Transitive proofs for safe unsigned and signed adds and subs landed instead (e5f202bb60, 1a72920f09, d574856482, f32cf8e4b0).
  • Loop-return domination is checked in both controls (8cd85e602a).
  • StringLen rules (f703dc5bef, 62cd044a79).
  • Maximum len/cap from element size (9fcb87c352).
  • Bounds checks eliminated for slices of the same length (519ae514ab).
  • Double-offset IsInBounds (1caa95acfa), slice len/cap arithmetic (d91148c7a8, fixed in f84f8d86be), chunked indexing with reslicing (20c9377e47).
  • Slicemask elided for len==c>0 (ec70d19023, 1bca4c1673).
  • Topo-sort walk (53be78630a), with the #68857 workarounds removed (a50de4bda7, 2ce5aab79e).
  • The divisible/divmod passes and “middle opt” (9bbda7c99d). Unsigned rounding-up divide (77dc138030). Multiplies proved to be by 0 or 1 become CondSelect (24af441437).
  • Subtract idioms (c12c337099). Rsh becomes RshU for positive arguments, with unsigned relations propagated (ec92bc6d63, 3820f94c1d). Right shift modelled precisely (235b4e729d).
  • Unsigned-division control-flow fix (dc42565a20). Overflow fix (f87aaec53d).
  • Unsigned power-of-two detection (f3582fc80e, bd94ae8903, 7b67b68a0d, 26da1199eb, 72147ffa75, 84b070bfb1).
  • Cleanups (a0f33b2887, b8468d8c4e, bc15963813, e64023dcbf). Brute-force limit tests improved limit addition (03fcb33c0e, 37ce4adcd4).
  Left open: the 9bbda7c99d message debates the naming of the opt passes (“Perhaps instead they should be generic, opt, and late opt?”).
• Memory-op optimization: memcombine, dead stores, store-to-load forwarding, alias analysis, mem2reg — Done: DSE tracks multiple shadowed ranges (d474f1fd21). Dead auto elimination became more aggressive (c7ccbddf22). Stores to unread parameters are removed (38b26f29f1, also on dev.simd as 2b50ffe172). Left open: nothing noted.
• Lowering of builtins to memory ops: memmove/memclr/memequal inlining and generated equality — Done: the MemEq intrinsic, initially on arm64 (dda7c8253d). memequal is expanded for lengths 3, 5, 6, 7 (a5fa5ea51c). Left open: dda7c8253d says LoweredMemEq “may be lowered differently for more constant size cases in future (for other targets …)”.
• Switch lowering: jump tables, type-switch caches, and switch lookup tables — Done: propagation of an out-of-range jump-table index was fixed (b6b8b2fe6e). loong64 jump tables (af6999e60d). Left open: nothing noted.
• Static devirtualization and shape-aware type-assertion pruning — Done: user-visible more calls devirtualized.
  • The concrete-type analysis (de9da0de30, #69521, #64824).
  • Fixes for OCONVNOP (5a9ef44bc0), non-PAUTO assignments (4dbf1a5a4c), positions (0ddb5ed465) and typos (b9d86baae3).
  Left open: nothing noted.
• testing.B.Loop support in the compiler (bloop) — Done: user-visible the bloop KeepAlive pass (22f24f90b5). Only addressable nodes are kept alive (509ddf3868), the name lookup was fixed (3fd9cb1895), and unnecessary interface conversions are avoided (2ac1f9cbc3). Left open: the bloop.go TODO about OpVarLive.
• Sparse conditional constant propagation (sccp) — Done: sccp debug messages gained position info (9a77aa4f08).
• Branch-free lowering: CondSelect-into-math and ARM64 CCMP — Done: CondSelect→math for +1 and doublings/halvings (fcd28070fe, ce05ad448f); cmov(x,x,cond)→x (94645d2413); branchelim supported arches moved into Config (2d33a456c6); CCMP aux type and generation (10ac80de77); fused NaN checks with other comparisons (97fd6bdecc); bool→int→bool roundtrip removal (73d7635fae).
• Constant-folding loads from type metadata and dictionaries — Done: abi.PtrType.Elem loads (df29038486, symbol marked used d767064170); fixed-load logic consolidated (b915e14490); abi.Type.{Size_,PtrBytes,Kind_} loads, with counts in std/cmd such as 75 for Kind_ (5b218461f9); nil checks on .dict removed (97da068774).
• SSA values for zero-sized and multi-field direct-interface types — Done: multi-field structs stored directly in interfaces (cd55f86b8d, 03ed43988f); StructSelect of interface data fields and more StructMake args allowed (bcd25c79aa, 72e8237cc1).
• ARM64 load/store pairing (LDP/STP) — Done: stp merging for non-sequential cases (4c63d798cb).

Other changes

• Misc internal — dedupe instructions when rewriting func results (cc571dab91), containsUnavoidableCall fix (5045fdd8ff), no optimization of panicking interface compare (bd4dc413cd), phi line numbers (61d1ff61ad), bisect for float-to-unsigned (48bb7a6114), dev.simd generator/intrinsic plumbing (~10 commits, e.g. 88c013d6ff, ce0e803ab9, 31b664d40b).
• dev.simd compiler support internal — adapters, intrinsic generation for bodyless functions, CPU-feature scope tracking, TERNLOG rewrite for logical trees (f6b4711095), inliner tweaks favoring SIMD functions and closures-of-SIMD (48756abd3a, e452f4ac7d), prove enhancements for double-offset IsInBounds, slice len/cap and chunked indexing (832c1f76dc, 2ca96d218d, 58fa1d023e; “Prove pass” extension), slicemask removal (8e60feeb41, 7c84e984e6), d2270bccbd, 04b1030ae4.
• Bounds-check failure with less code internal — idx and len remembered without fixed registers before calling panicIndex (3024785b92).
• Size-specialized malloc calls internal — direct calls to generated size-specialized malloc functions when size is known at compile time (19a30ea3f2).
• Constant-time intrinsics user-visible — crypto/subtle ConstantTimeSelect/Eq intrinsics, exposed to FIPS packages (2c91c33e88, 00ee1860ce).
• Misc rules/fixes internal — Add64carry with unused carries → Add64 (5453b788fd), single-bit-difference comparisons (0a569528ea), negation into multiplication (2cdcc4150b, e7d47ac33d), CSE prefers statement-marked value (b1f3e38e41), duplicate nil check to both branches of write barrier (f9e61a9a32), checkptr for unsafe.Add (5ab9f23977), InlMark speculatable (084c0f8494), sparse map generics (6505fcbd0a).

Backends and assemblers · architectures, ABI

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

• Register-based calling convention (ABIInternal) ports beyond amd64 — Done: s390x register ABI: info 73b6aa0fec, initial spill 81242d034c, CALL ops 24697419c5, prologue 85e6080089, enabled 89552911b3. [dev.simd] morestack spills the correct SIMD register c0f031fcc3. obj: argp adjustment for wrapper frames removed 6f6c6c5782. Left open: nothing noted for 386/arm/mips (they never got commits).
• Register allocator, scheduler and block-level backend passes — Done: ops may specify clobbering of an input register (copy made before clobber, like resultInArg0) 89a0af86b8, used by ensure we use allowed registers fix 5a56d8848b; faster regalloc liveness (fill loop headers instead of iterating to a fixpoint; passes over CFG from loop-depth to 2; liveness time 225→150 ms on cmd/compile/internal/ssa) 16705b962e; rematerializable values’ output reginfo respected 10e7968849 and removed from the live set across calls (arm64 .text −0.34% on preprofile) 284379cdfc, 9bbea0f21a fixedreg values always available; 4d0658bb08 prefer fixed registers (zero register) for values; containsUnavoidableCall fixed 5045fdd8ff; loopRotate handles nested loops 592c2db868; irreducible loops no longer trusted f15cd63ec4; stackalloc liveness switched to sparse path exploration (−74 ms CPU on a cmd/compile build) 4e761b9a18; reduce stack object liveness etc.
• amd64 code generation and GOAMD64 microarchitecture levels — Done: duff devices replaced by generated loops (see Move/Zero thread); zerorange simplified b10eb1d042; do not Zext bools before CMOV 1662d55247; bits.Mul64 on 32-bit machines 1e5bb416d8; VEX/k-register/AVX-512 plumbing for SIMD (see SIMD hooks thread).
• arm64 code generation and assembler — Done: generated loops replace duff (below); new bounds-check strategy 394d0bee8d; NEON in LoweredMemmove/MemmoveLoop (Raspberry Pi 5: −22.5% geomean) 5c9a26c7f8; CCMP generation introduces ARM64ConditionalParams aux (condition, NZCV, constant) 10ac80de77; 32×32→64 multiplies e1a12c781f; PAC instructions 11d3d2f77d; shorter constant integer loads 3b3d6b9e5d; fold boolean NOT into branches ba634ca5c7; conditional-instruction type correction d0d0a72980; ARM64RegisterExtension moved into obj 45eee553e2. [dev.simd] arm64 helpers (FMOVQ vector load/store, VMOVI16B) 9076ed8750^1.27 (1.27).
• riscv64 compiler and assembler (RVA profiles, vector, compressed) — Done: compressed (RVC) instructions: encodings 924fe98902, enabled for runtime/sys b9ef0633f6, assembler support with better bit-pattern encoding a15d036ce2,0f9c8fb29d, use in generated code 9859b43643; FCLASS for infinity/subnormal tests 320df537cc,34aef89366; float↔︎int register moves via FMV fcc036f03b; Zicond assembler ef05b66d61; generated loops replace duff (below); segment/fault-only-first vector loads 371c1d2fcb,6b8d507508; store-to-load forwarding rules 4373754bc9; assembler docs b24aec598b; generic CSR ops dd8276657f.
• loong64 backend and assembler — Done: ALSL.{W,WU,V} + ADDshiftLLV multiplication rules 4c7362e41c,bd71b94659,90b7d7aaa2; jump table af6999e60d; generated loops replace duff 7bba745820,91e76a513b,4f7bbc62c7; LSX-based LoweredZeroLoop c4bb9653ba; multiplications e071617222, a552737418; 32×32→64 feae743bdb; new bounds-check strategy 5216fd570e; MOVVP prologue optimization reverted eb7c67fdc9→dd1d597c3a; many instruction additions (LDPTR/STPTR 882335e2cb, ADDU16I.D ddce0522be, FSEL ee7bb8969a, VPERMI/VSHUF/VEXTRINS etc.).
• Duff’s device replaced by generated loops; Zero/Move lowering — Done: amd64: DUFFZERO and DUFFCOPY replaced by generated loops (eb7f515c4d,ec9e1176c3; the DUFFCOPY change was reverted 4e182db5fc and relanded ca66f907dd); arm64 (28aa529c99,15d6dbc05c) with runtime duff removed e4291e484c; riscv64 and loong64 (879ff736d3,4dac9e093f,7bba745820,91e76a513b) with runtime duff removed (b9a4a09b0f,4f7bbc62c7); zerorange simplified on amd64/arm64/riscv64/loong64/mips/mips64 (b10eb1d042,5e94d72158,77643dc63f,e0870a0a12,06e791c0cd,bb48272e24; mips fix 0b26678db2); zero constant register reuse on mips64x bdd51e7855.
• Bounds-check calls: new “panicBounds” strategy — Done: per-arch migration: arm64 394d0bee8d, amd64/386 076eae436e, arm32 7dccd6395c, loong64 5216fd570e, mips64 9f9d7b50e8, mips32 8330fb48a6, riscv64 95693816a5, s390x 7a1679d7ae, ppc64 802d056c78; old-style calls removed 21ab0128b6; bounds-check slice SUB/SUBconst compare optimization dcb479c2f9, shift-through-AND folding for slices bd80f74bc1; bounds report fix af03343f93; ppc64 panicBounds frame made smaller e0a1ea431c. Left open: nothing noted.
• Late-lower pass, memory-op combining and arch-independent rule migration — Done: ARM64 and others’ rules for fold negation into multiplication 2cdcc4150b; branchelim supported arches moved to Config 2d33a456c6; SIMD rule file plumbing ce0e803ab9.
• Jump tables for switch statements on more architectures — Done: loong64 af6999e60d.
• Multi-word arithmetic, intrinsics and carry-chain support — Done: bits.Mul64 (Mul64uhilo, Hmul64, Hmul64u, Avg64u) on all 32-bit systems, making constant 64-bit division emit directly (i386 i/999: 7 mul→4 mul) 1e5bb416d8; wasm Avg64u/Hmul64 6e165b4d17.
• WebAssembly backend — Done: signext/satconv always enabled 742f92063e; unaligned memory ops 53ad68de4b; Avg64u/Hmul64(u) 6e165b4d17; FP→int32/64 match change reverted 3bc9d9fa83.
• s390x, MIPS, 386 and 32-bit ARM backends — Done: user-visible all remaining windows/arm code deleted 98238fd495; 386/wasm FP→int conversion change reverted 6d5b13793f; s390x float min/max cedf63616a; 386 8-bit semantics 78d75b3799-class reverted; s390x MVCLE d71d8aeafd.
• Backend hooks for GOEXPERIMENT=simd (amd64 AVX-512, arm64 SVE, wasm) — Done: k registers 11d2b28bff, register shapes (1161228bf1,6bc3505773,884f646966,1b87d52549), simdGen prog writer c81cb05e3e, AVX-512 upper regs/masked loads/stores (8db7f41674,1ac5f3533f), VPTEST f1e3651c33.

Other changes

• Optimize abi.Type.GCData loads f1fd13016a; MemEq intrinsic for runtime.memequal dda7c8253d.
• prove div/mod internal — new divisible/divmod passes run after prove (“middle opt” and “late opt”) so prove can reason about div/mod bounds 9bbda7c99d.
• Enforce SP-based locals eef5f8d930; uintptr→unsafe.Pointer unsafe-point (1.21) 0384235a15^1.21.
• 32 small obj/loong64 and loong64 asm commits internal — e.g. obj/loong64: add {V,XV}{BITCLR/BITSET/BITREV}[I].{B/H/W/D} i; obj/loong64: add [X]VLDREPL.{B/H/W/D} instructions support; cc: avoid extending when already sufficiently masked on loon. (b2960e3580,d44749b65b,b1e933d955,b266318cf7,a62f72f7a7,73ff6d1480,a7f05b38f7,9763ece873,9632ba8160,1843f1e9c0,fa706ea50f,6b837a64db,b8cc907425,731e546166,c552ad913f,f5b20689e9,6447ff409a,916e682d51,954fdcc51a,dadbac0c9e,ad5e941a45,2e2072561c,8111104a21,f01a1841fd,abb241a789,7f2ae21fb4,cdba82c7d6,a318843a2a,c0f02c11ff,c6d64f8556,a9093067ee,e0a4dffb0c)
• 2 small obj/x86 and amd64/386 rules commits internal — e.g. [dev.simd] cc: add simdFp1gp1fp1Imm8 helper to amd64 code ge; obj/x86: handle global reference in From3 in dynlink mode. (7c6ac35275,4b740af56a)
• 59 small ssa/regalloc/liveness/obj generic commits internal — e.g. [dev.simd] cc: glue codes for Shift and Rotate; [dev.simd] cc: support simd(imm,fp) returns gp; [dev.simd] cc: make regalloc simd aware on copy. (4fda27c0cc,0d8cb89f5c,d8fa853b37,0870ed04a3,5429328b0c,9ea33ed538,08cd62e9f5,ab7f839280,ccb43dcec7,3f789721d6,a0b87a7478,957f06c410,6f7a1164e7,08bec02907,6fbad4be75,4ee0df8c46,880ca333d7,fb8bbe46d5,2b622b05a9,82d056ddd7,bcd25c79aa,b226bcc4a9,d5dea86993,08ab8e24a3,9783f86bc8,e001300cf2,908e3e8166,7380213a4e,a5137ec92a,8d874834f1,4a3ea146ae,6890aa2e20,83714616aa,3492e4262b,356c48d8e9,fcb893fc4b,1e5631d4e0,3ec0b25ab7,dabe2bb4fb,0e590a505d,18cd4a1fc7,78ef487a6f,d9751166a6,e34ad6de42,ad3db2562e,0ab038af62,7ae0eb2e80,63a09d6d3d,578777bf7c,003b5ce1bc,f0e281e693,d4830c6130,ba72ee0f30,647c790143,a5a249d6a6,30c047d0d0,d8269ab0d5,e84983fa40,9b2e3b9a02)
• 1 small cmd/asm commits internal — e.g. cmd/asm: process forward jump to PCALIGN. (1ad30844d9)
• 10 small obj/riscv and riscv64 rules commits internal — e.g. cc: remove sign extension before MULW on riscv64; obj/riscv: remove ACFLWSP/ACFSWSP and ACFLW/ACFSW; obj/riscv: simplify validation and encoding of raw instructi. (27ce6e4e26,76d088eb74,e5688d0bdd,2469e92d8c,e2cfc1eb3a,d7a52f9369,a7d174ccaa,1e5e6663e9,4ca048cc32,a0e738c657)
• 1 small mips/s390x/arm/wasm commits internal — e.g. cc: make wasm match other platforms for FP->int32/64 convers. (8d810286b3)
• 1 small obj/arm64 and arm64 rules commits internal — e.g. obj/arm64, image/gif, runtime, sort: use math/bits to calcul. (a18294bb6a)

Backends and assemblers · SIMD

From SIMD: the simd packages and their compiler support (Go 1.18–1.28)

• Compiler SIMD type system, ABI and register-allocation support — Done: register shapes, vec/mask SSA types and the first AMD64 SIMD lowering (8ecbd59ebb generated 63k lines: simdssa.go, simdAMD64.rules, simdAMD64ops.go, simdgenericOps.go, simdintrinsics.go); helpers moved from generated code into the compiler (6c50c8b892); vectors made uncomparable (5a78e1a4a1); CPU-feature pass (d2270bccbd, ba72ee0f30); X15 low-128 zeroing (f38e968aba). Left open: spilling masks (“will depend on the architecture”), 256/512-bit zero values built on the fly, VZEROUPPER not evaluated (f38e968aba).
• simdgen: the instruction-table-driven generator — Done: first generated amd64 code (8ecbd59ebb); generator output ordering fixes (3a4d10bfca removed a map iteration, 35b8cf7fed, e62e377ed6); simdgen migrated into the repo (b7c8698549) with etetest.sh rewrite (8b90d48d8c); tests generated in repo (88568519b4, a24ffe3379); memory operand parsing and merge (f42c9261d3, 0b323350a5, 5a0446d449); SHA features parsed from XED (2b8eded4f4); feature implications (9ef1692c93). Left open: the generator depends on XED data path ($XEDDATA); the API source of truth is spread over YAML categories, hand-written files and docs.
• AMD64 archsimd API: AVX, AVX2 and AVX-512 operations — Done: user-visible the whole API above reached master through reverse merges starting d4f5650cc5 (2025-11-24), 2b62144069 (12-03), 9e09812308 (12-08); package renamed simd -> simd/archsimd with package doc (144cf17d2c, 1da0c29c2a); doc rewrites (05e41225f6..d64add4d60), Mask.AsIntMxN->ToIntMxN and ToMask only on integer vectors (f110ba540c, c0ba519764); Broadcast rename (8ac4477d83); tests skipped without AVX (1b291b70df). Left open: many names still to rename (see API shaping thread); DotProductQuadruple removed pending redesign (70c22e0ad7).
• Mask representation, masked operations and merge peepholes — Done: user-visible mask types and conversions (a034826e26, 6f7a1164e7), .Masked() peepholes (ede64cf0d8, cf31b15635), merging ops (86b4fe31d9), emulations of mask/merge methods for AVX2 (94d72355f6, 2080415aa2), mask load/store removal (2e71cf1a2a), ToMask integer-only (f110ba540c). Left open: IfElse naming, Merge emulation wrongly named at first (“under a wrong name; subsequent CLs will … rename”, 94d72355f6).
• Memory operands, constant-immediate forms and jump tables — Done: 38b76bf2a3, 5a0446d449, 48f366d826, 3ec0b25ab7, 4eb5c6e07b, 443b7aeddb, c4fbf3b4cf, 3f6eabdf09. Left open: bounds-checked jump table “in the next CL” (38b76bf2a3).
• dev.simd branch, GOEXPERIMENT=simd lifecycle and merges to master — Done: user-visible flag added (4d2c71ebf9), branch default-on (2ef7106881), reverse merges: d4f5650cc5 (2025-11-24), 2b62144069 (2025-12-03), 9e09812308 (2025-12-08); package moved to simd/archsimd (144cf17d2c). Left open: experiment default off; simd name was reserved for a portable package.
• API shaping and renames before stabilisation — Done: ead249a2e2, 2c25f3e846, c2d775d401, 416332dba2, 0978935a99, 3fdd183aef. Left open: more renames.
• Emulation of missing instructions and test infrastructure — Done: 94d72355f6, 2080415aa2, be57d94c4c (Not), 88568519b4. Left open: emulations named “wrong name” first.

Other changes

• AMD64 API fix-ups from the first months (grouped, ~38 commits) user-visible — directory bootstrap and docs (fdb067d946, 3df41c856e, bbb6dccd84), regenerated/ordering fixes (7392dfd43e, ee1d9f3f85, 21d6573154, 9ea33ed538, 08ab8e24a3, 6af8881adb, e4d9484220, fe040658b2, f7c6fa709e), signature/param-order fixes (1be5eb2686, ef5f6cc921, c61743e4f0, 17baae72db, b69622b83e, 8eb5f6020e, 4d26d66a49, d3a0321dba, baa0ae3aaa, 42cda7c1df, 76dddce293), new ops (f5f42753ab VDPPS, 58cfc2a5f6 VPSADBW, c39b2fdd1e VPLZCNT, 0d8cb89f5c simd(imm,fp)->gp returns, 9a934d5080 float GetElem, ec5c20ba5a conversions, fa1e78c9ad/baea0c700b AVX shuffles, 74ebdd28d1 Select128FromPair types, 896f293a25 DotProductQuadruple peepholes), AVX512/AES feature checks (dd63b7aa0e, 20b3339542), removals from the exported API (972732b245 move, 4004ff3523 FlattenedTranspose, bf77323efa unexported methods in own file), testing hooks (af6475df73), generator cleanup (19b4a30899, 0f620776d7).
• Fast transpose example internal — demonstrates data-movement instructions for transposes (1c961c2fb2).
• Part loads and subvector load/store user-visible — {Int,Uint}{8x{16,32},16x{8,16}} subvector loads/stores from slices (24f2b8ae2e); slice funcs moved to generated code (d375b95357).
• Checking loops replaced by slice-checker internal (f51ee08905).
• Package restructure internal — internal/simd deleted (62e1fccfb9), test tweaks in go build and go/doc (eba2430c16).

Linker and object files

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

• Relocation-free pclntab and funcdata (offsets instead of relocations) — Done: .gosymtab no longer generated (empty since Go 1.2) 7347b54727; textStart removed from pcHeader (the only pclntab relocation) 0e1bd8b5f1; pclntab moved out of relro ad3ccd92e4; funcdata symbols placed in .gopclntab b437d5bf36; findfunctab too 312b2034a4; pclntab magic moved to internal/abi 43cfd785e7; moduledata in a dedicated .go.module section 9f5cd43fe6; DWARF is generated for moduledata b5f6816cea; cutab length fix ae8eba071b. Left open: the pcHeader zero placeholder “can be removed at some point in coordination with Delve” (0e1bd8b5f1).
• Type, itab and typelink layout (.go.type section, relro rework) — Done: ELF section headers sorted by address and shdr/shstrtab built from actual sections (no more .go.fuzzcntrs string in every binary) 0ff323143d, 4bc3410b6c. Left open: nothing noted.
• //go:linkname restrictions and the blocklist — Done: runtime.addmoduledata linkname disallowed 9f6936b8da; more names blocked 34af879dde. Left open: nothing noted.
• DWARF quality for debuggers (types, variables, trampolines) — Done: types only referenced through interfaces are exported to DWARF by scanning function relocations, fixing unreadable runtime types in Delve/viewcore after 1.25 80038586ed. Left open: nothing noted.
• ELF/relro/hardening and external-link policy — Done: .def file for Windows DLL exports instead of __declspec(dllexport) eaf2345256; weak binding on Darwin 8cf7a0b4c9. Left open: nothing noted.
• Per-platform object formats and internal linking (Mach-O, PE, XCOFF, wasm, ppc64, riscv64, loong64) — Done: internal linking for windows/arm64 9e25c2f6de; MSVC clang c70713da82; Mach-O ARM64 POINTER_TO_GOT and SUBTRACTOR 6e95748335, 7c8166d02d; darwin/386 and windows/arm relocations removed 17a0fabc43, ca0e03560d; loong64 relocs 4b0e3cc1d6. Left open: nothing noted.

Other changes

• Windows and MSVC clang internal — -w made consistent and passed in external mode (22ac328856, 6dceff8bad), no -Wl,-S on AIX/Solaris (459f3a3adc, 047c2ab841), .def with MSVC clang and LIBRARY only for cshared (adce7f196e, ee5369b003), race with MSVC clang (e603e9834e), PE reloc mapping simplified (594deca981).
• ELF cleanups internal — shdr/phdr slices (6525f46707, b0c278be40), section name offset fix (32a9804c7b), DWARF compress alignment per section (8806d53c10), runtime.buildVersion with experiments valid (9daaab305c), -V=goexperiment special case removed (ccd389036a), get_pc_thunk duplicate (9ba0948172), plugin capability flag (d14b6427cf), posix_fallocate on netbsd/freebsd (ce3f3e2ae7, d52a56cce1), OpenBSD buildid workaround removed (3bea95b277), riscv64 relocs (ffc85ee1f1), Mach-O SUBTRACTOR fix (61bf26a9ee), goobj error output (5c8624a396), arm64 LR/FP save redo reverted (719dfcf8a8, c938051dd0).
• Data section selection in internal linker user-visible (cf5e993177).