Go 1.20

Unified IR becomes the default frontend with real dictionaries. PGO preview (hot-call inlining). Coverage redesign. Arenas experiment. 1-bit heap bitmap. Pacer smoothing via moving average. clear is type-checked here but gated to go1.21.

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)

• GC pacer redesign — Done: the PI controller was replaced with a 2-cycle moving average using the actual trigger; the PI controller moved to mgcscavenge.go, its only remaining user (41795528b0). gctrace prints lastStackScan instead of the max (56bdf7f7d9). Pacer fields were converted to atomic types (see Other, Go 1.20). Left open: right after a phase change the average lags, which can cause a transient rise in assists (tile38 WithinCircle, 41795528b0).
• Heap bitmap rewrite and allocation headers — Done: 1-bit bitmap rewrite (b589208c8c). It was reverted for wasm/386 failures (ad0287f496) and relanded unchanged (6a9c674a09) after making the GC never scan noscan objects (cb13022a24) and initializing pointer bits for noscan spans for cgo/heapdump callers (a6e6b11e3a). Word-at-a-time bitmap processing (“~30% faster in heapBitsSetType”) landed, was reverted and relanded the same way (c3833a5543, e1a8e0e05d, e49e876455). Left open: “we should make the gcdata bitmap an array of uintptr … Similarly for the processing of gc programs” (6a9c674a09).
• Scavenger redesign — Done: searchIdx used in scavengeOne to avoid N² behavior within a chunk (2caaad942e), and padding removed (f324355d1f). Left open: nothing noted.
• Finalizers, runtime.AddCleanup and the cleanup queue — Done: user-visible documentation that finalizers on tiny (<16 B) objects may never run (80d487111b); atomic fingStatus (−5% Finalizer benchmark, bd5595d7fa); test cleanup (eaf0e3d465). Left open: nothing noted.
• Memory arenas experiment — Done: user-visible runtime arenas (7866538d25) and the arena package with reflect.ArenaNew behind a GOEXPERIMENT (e0d01b8467); refactors (69fc74f3ee, 987f94fa03, 4c383951b9); a test skip (e72da1c15d). Left open: “A later CL can potentially export it” (7866538d25); it stayed experimental.
• Allocation publication, zeroing and preemptibility — Done: the scanner reads freeIndexForScan (febe7b8e2a); an accidental submit of the alternative design was reverted (bed2b7cf41, 2b59307ac2); Malloc benchmarks fixed (7717ac151a). Left open: nothing noted.
• Sweeper termination and pacing — Done: bgsweep heuristic (8451529e9a, “may be worth considering pushing this number higher”); unused sweepdata field removed (2df6c1abce). Left open: nothing noted.
• Address-space layout, tagged pointers and the mem.go layer — Done: arenaIdx accessors made nosplit (506e690a26). Left open: nothing noted.

Other changes

• GC controller and heap fields to typed atomics internal — about 20 conversions: heapLive, heapScan, stack/globals scan, bgScanCredit, mark times, dedicated workers, mSpanStateBox, mcache.flushGen, work.cycles, mspanset, pagesInUse (e76155614a, 3a9281ff61, 6e9925c4f7, e1b0da6144, 02fb9b8ca9, b8f4847d6f, ba89d59a3a, 9463638ca8, fe406c8b11, dcd10375bc, 014f0e8205, a25a34abe9, 28e388589b, d61ffe8b6c, 5a37965495, f1b7b2fc52). Manual alignment padding was removed as a result (a52f4952cb, e28cc362a8, 1c59199c91), and the cache-line padding formula fixed (f15761b50b).
• runtime/internal/sys.NotInHeap internal — replaces the //go:notinheap pragma with an embedded marker type (a719a78c1b).
• Write-barrier-enabled pointer atomics internal — UnsafePointer.Store/CompareAndSwap and Pointer[T] with barriers (9fedc481ea).
• GC bookkeeping fixes internal — userForced set under gcsema (6f7c9ea855); traceGCSTWDone only when tracing (8befe0e408); unused gcmarknewobject parameter removed (51c34e2f0b); _g_→gp renames (5e8d261918, 222799fde6); getStackMap as an stkframe method (b91e373729).

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 fields converted to atomic types: p.timer0When, p.timerModifiedEarliest, p.numTimers, p.deletedTimers, timer.status (e1fd51e076, bb5d656a3a, c82bbc0e8e, 6e74c4116a, f78efc0178); schedt.pollUntil likewise (0fc774a68f). Left open: the timer state machine itself is still CAS-based.
• Contended runtime-lock profiling (mutex profile for runtime-internal locks) — Done: user-visible /sync/mutex/wait/total:seconds runtime metric (63ceff95fa); wait reason for goroutine blocked on a mutex made specific (a2c396ce00, 686b38b5b2); mprof locks split (1.19, 89c0dd829f^1.19). Left open: runtime-internal locks invisible in the profile.
• Scheduler core: spinning Ms, work conservation, stealing, STW interaction — Done: sched.needspinning: if work submission (wakep, injectglist) finds no idle P it sets the flag, and a non-spinning M about to drop its P rechecks it and becomes spinning instead; fixes a worst-case deadlock when injectglist finds nothing as all Ps go idle (8cb350d69a). A long list of schedt fields converted to atomic types (see sync/atomic thread). goschedIfBusy yields if gp.preempt (50664c236f). bgsweep “low-priority” heuristic: don’t yield if idle Ps exist, sweep >=10 spans between yields, cutting trace events by 90% (8451529e9a). Left open: nothing noted.
• sync package API evolution — Done: user-visible Map.Swap, CompareAndSwap, CompareAndDelete (395323c4d0); Map and entry use atomic.Pointer (b0144b3843, 7366b199c5); WaitGroup state atomic.Uint64 (ee833ed72e); RWMutex counters atomic (e509452727); Once.done atomic conversion reverted (b53471a655). Left open: nothing noted.
• sync/atomic and runtime atomics — Done: the runtime-wide conversion (b464708b46, dd8cb66d0b, 5d7d50111f, 58a1dabfd1, 1eeef5d5b4, 88ef50e662, 3fbcf05d40, 951d2c657d, fa0e3bffb4, 3f65ddbfd3, 301ca7513f …); a revert for ncgocall (ef8414101f); user-facing hints toward typed API (0bd4710ca6). Left open: nothing noted.
• Runtime lock ranking — Done: DAG generation, unused ranks deleted, trace locks moved to the leaf (2b8a9a484f, ddfd639408, f42dc0de74, cc8bac8835, 44ff9bff0c, c5be4ed7df). Left open: names of ranks lost.
• Netpoller — Done: netpollWaiters, netpollWakeSig, netpollInited atomic (aab8d2b448, 901b9233e6 …). Left open: nothing noted.
• Goroutine lifecycle and identity — Done: goid uint64 (dd8cb66d0b). Left open: nothing noted.
• Fast random numbers for the scheduler — Done: loong64 wyrand (3a41094107). Left open: nothing noted.

Other changes

• math/rand auto-seed user-visible — 90a67d052e; math/rand/v2 skeleton later (59f0ab4036^1.22, 1.22).
• Exit hooks internal — addExitHook for coverage (07bdf1dc54).
• godebug settings API internal — ea4631cc0c.
• _g_ removal and atomics sweep internal — e9d7f54a1a, 4358a53a97, 4509e951ec; further atomic conversions in the long tail (40fa2dabe0, 7666ec1c99, 449691b3ef, 09cc9bac72, b04e4637db, 3121c2bf64, 5b0ce94c07, b11b4b4de3, ce7aae5095, ec8ed5bd02, 9b988c9094, 04d8c2327d, dea67a9b34, 7e7ecf5cbb, bc805795bd, f6efcf1fbb, 2883af63c2, e0ec1d6770, 5634629f0b, 7a1ae605eb, 1110222bee (reverted), 5a6db7c48f).
• Lock rank annotations internal — d29a0282e9, 502b6057d2, 9923162f1c.
• Trace extra-M events internal — added, reverted, re-added (ea9c3fd42d, fb6c210dc3); schedtrace nil IDs (98e1648823).
• atomic pointer inlining internal — ef69718dd7, b14cf3d93a; misc cleanups (cc9584b54b, f00fa0b98d, 0063b9b036, 556c978400, ffe5f085cd, e406112011, 4fe4601d37, f63b268b9a, 185e1a7b27, 8232a09e3e).

Data structures and type metadata

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

• Compiler-generated equality/hash functions and type descriptor emission (reflectdata) — Done: Unified IR starts setting RType fields so reflectdata no longer re-derives runtime types from IR (7368647ac6, f70775ff22, 2cf632cd57), fewer basic types emitted in every compilation (cb5534c1c6), no content-addressable symbol for generic iface method names (8dc7710fae). Left open: “subsequent CLs will handle the remaining more complex cases” (7368647ac6).
• reflect implementation: performance, allocation and escape behaviour — Done: user-visible Value.Grow (61f0409c31), SetZero (bc1d0d8eb1), Comparable/Equal (acabf87127; Equal rewritten to avoid allocation a343f4017b, panics on two non-comparable values 7ffc1e47b4), slice→array conversion (805f174e5c), FuncOf supports >50 args (d7a3fa120d) by building func types through StructOf (63668a5155), rtype.MethodByName by binary search (0053ec452d), IsZero uses rtype.equal (1ab6b790be), unsafe.String for names/tags (26b2c9a620), SliceHeader/StringHeader deprecation added (d734203e39) then notices removed because no proposal discussion (9f2951b5d1). Race on funcTypes fixed (6d8ec89303). Left open: deprecation postponed to 1.22 per wiki rule.
• Register ABI completion for reflect, memmove/memclr and bytealg across architectures — Done: riscv64 always on and fallback deleted (7e72d384d6, 5d59fa143a). Left open: nothing noted.
• memmove / memclr / memequal / internal/bytealg assembly optimisation — Done: riscv64 memclr/memmove (44ad662a49, e18d07ddc5), alignment-check correctness (fd82718e06), ppc64x PCALIGN (92c7df116e). Left open: nothing noted.
• Slice growth, string building and allocation-free bytes/strings helpers — Done: panic when unsafe.Slice param is nil and length >0 (52d0667e6b). Left open: nothing noted.

Other changes

• Pointer bitmap one-word-at-a-time reverted internal — c3833a5543, e49e876455 reverted by e1a8e0e05d.
• Method wrappers only inlined without closures internal — f21514c7f8.

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: v1 trace lock hygiene as preparation for scalability: trace.lock only on the system stack (6c2e327e35, 42794f3871), reader parking no longer under trace.lock (9923162f1c), stack table dumped directly into the trace buffer (d4a3466519, 0293a1f308, b648591b70); tracing.enabled made atomic (5e20f2e4df); extra-M (cgo callback) goroutines get synthetic GoCreate/GoInSyscall events at StartTrace so the parser stops reporting “no consistent ordering” (fb6c210dc3, after a revert 8adc63b3eb). Left open: “No consistent ordering” and the global lock remain structural problems of v1.
• Mutex and block profiles, including runtime-internal lock contention — Done: user-visible mutex profile samples scaled at record time, so changing the sample rate (even to 0) does not corrupt old samples, matching block profile (30b1af00ff); /sync/mutex/wait/total:seconds metric piggybacking on G runnable tracking and wait reasons (63ceff95fa). Left open: runtime-internal locks not in the profile.
• CPU profiling, goroutine profiling, pprof labels and the protobuf writer — Done: user-visible 4-column memprofile text form forced so pprof doesn’t collapse alloc==inuse into 2 columns (c2ede92a0d); Windows memory mappings (28afa5b176); Function.start_line (a401468b00); Windows MaxRSS (2cc6983a21); prof.hz, profbuf fields to atomic types (54cf1b107d, c3c10f5d67 etc.); MemProfileRate no longer set in init (f63b268b9a); empty-stack docs (0bbd67e52f). Left open: nothing noted.
• runtime/metrics and MemStats — Done: user-visible estimated CPU time breakdown metrics (/cpu/classes/...) with total CPU (b7c28f484d); /sync/mutex/wait/total:seconds (63ceff95fa); timeHistogram atomics and size cut (b04e4637db, 87eda2a782). Left open: CPU numbers are estimates, not on-CPU counters.
• Coverage redesign: runtime/coverage, internal/coverage, covdata — Done: user-visible full pipeline above: encoder/decoder (f951f697c4, 84f95aa811), exit hook (07bdf1dc54), runtime emit (87db4ffada), covdata tool (7a74829858), counter I/O (efa3f1749f), user APIs (ecd112c0d1), package registration hook (cf83a490e4); experiment GOEXPERIMENT=coverageredesign added (a3434b71a3) and enabled by default (9910f8a403, 2022-09-29); atomic counter reads (cddf792428); -coverprofile local-package path capture restored (2730170f63); package main handling fix (7973b0e508). Left open: covdata mode clash policy (relaxed in 1.21 a371fa5e70^1.21).
• Race detector, ASan, MSan and Valgrind integration — Done: ASan on ppc64le (9e810997c0); race syso subarch versioning and GOAMD64=v3 image (e665de2861, c75b10be0b, 18baca6765); no cgo needed for race on macOS (5bf9aeba1c). Left open: nothing noted.
• Crash output, runtime/debug, BuildInfo and debuglog — Done: docs and debuglog unsafe.Slice avoided (dadd80ae20, c45ebef05e). Left open: nothing noted.

Other changes

• Cleanups in profiling/coverage/trace code internal — _g_ removal in trace.go (74cee276fe, 4400238ec8), unsafe.String/Slice adoption (4fe4601d37, 7ffbcd1987, be3184c405), atomic types for profbuf (04d3d6bf48, 5b1658d691), Getrusage error check (2e06019dcd), coverage function renames (47a538842b), pprof mapping/label cleanups (aa4299735b, 0857633f8b).
• riscv64 SV57 internal (1e3c19f3fe); atomic-type conversions of profiler/trace fields (3121c2bf64, 2af9ee0727, 52d9e7f543, bcd1ac7120).
• lfstack vs checkptr internal (d6171c9be2).

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: panicking, runningPanicDefers converted to atomic types (bd302502d3, 7666ec1c99), canpanic and dopanic_m cleaned (7e952962df, 0ad2ec6596). Left open: nothing noted.
• Unwinder: gentraceback → unwinder iterator, inline unwinder, pcvalue cache — Done: switch gp when jumping stacks during traceback (511cd9b8af), only if M doesn’t change (60ef4b2c1e), not at entry of systemstack (500bc6b805); stkframe consolidation and function-context removal (8be94b82ab, 35026f3732); morestack_noctxt marked SPWRITE on LR arches (e4be2ac79f). Left open: the monolithic function and callbacks remain.
• Symbol tables, runtime.Caller and CallersFrames — Done: name/entry offset renames (196003f9fd, 7ebaa6587b). Left open: nothing noted.
• Stack growth, stack bounds and the crash stack — Done: -race stack guard increased (893964b972); cmd/dist guard multiplier fixed (b12366bbb5); _Stack* constants tidied (ef8e3b7fa4^1.21 then partly reverted df777cfa15^1.21); 104-byte guard removed (57236fe9f7^1.21, listed in Go 1.20 by the key file; log places it in 1.20 cycle). Left open: nothing noted.
• C-to-Go callbacks: extra-M reuse, system-stack bounds, init gating — Done: runtime/cgo.Incomplete replaces //go:notinheap in cgo (846c378b8c); -fno-stack-protector (365ca69472); darwin pthread stack size follows rlimit (7d3a5a501c); extram/extraMWaiters atomic (5634629f0b). Left open: nothing noted.
• Security hardening: AT_SECURE, std fds, AT_RANDOM, gated debug features — Done: page tracer behind GOEXPERIMENT (e4435cb844); -fno-stack-protector for cgo (365ca69472). Left open: nothing noted.
• vDSO, vgetrandom and fast clocks — Done: fast FreeBSD clock_gettime (e51b3ae0ee, 535fe2b226), freebsd/riscv64 vDSO (0e56c31b85), ARM VDSO call helper (876d477b0e). Left open: nothing noted.
• Signals and async-preemption machinery — Done: SI_TKILL like SI_USER (d11c58eedb, after revert), sig.state atomic (d01200e772), TestCgoSigfwd on all Unix (03272d14ed). Left open: nothing noted.
• Syscall layer: AllThreadsSyscall, internal/runtime/syscall, libc and Windows SyscallN — Done: epoll syscalls moved then reverted/relanded (c7cc2b94c6, 4a49af5755), OpenBSD libc (0d6a7f9d2e), MADV_DONTNEED on BSDs and Solaris (19a4804854), FreeBSD non-COMPAT kevent (e0e1ce9888), mips r2 zero (2c7c98c3ad). Left open: openbsd/mips64 stays on syscall(2).
• Windows runtime port — Done: TEB TLS slots (da564d0006), CFG (5c0d314adc), ignore exceptions from non-Go threads on windows arm/arm64 (6fe3ccd74c). Left open: nothing noted.
• New, ported and removed platforms — Done: freebsd/riscv64 (9aa7107cb5, d268504fd9, 0e56c31b85). Left open: nothing noted.
• CPU feature detection, GOAMD64 checks and cpu GODEBUG options — Done: SHA-NI, arm64 SHA512, openbsd/arm64 detection, M1 cache line (f1b1557cf3, 228f891c90, cd54ef1f61, 95a5e20d8f, 59865f1420, 1163acf3ea). Left open: nothing noted.

Other changes

• FreeBSD/OpenBSD/ppc64 odds internal — msan on freebsd/amd64 (5fde02e312), symbolic FreeBSD syscall constants (848febdcec, c04977fa76), SI_TKILL handling (7c87012f04), ppc64 start-up on newer kernels (7d1cab12c7), usleep fix (65f863529c), tfork change for openbsd/mips64 (e84ce0802d), consolidated low-level error reporting (79d9b395ad), atomic conversions of runtime fields (40fa2dabe0, bc805795bd, e1114fdf88, d3b35a4242, 296c40dbbe).
• _g_ cleanup internal — getg() locals replaced across os/signal/traceback/GC files (8666d89ca8, 4240ff764b, 54a0ab3f7b, b486518964, 5999a28de8, c647264619, 29b9a328d2, 0e18cf6d09, 399f50c9d7).
• Apple libc atfork hangs internal — osinit_hack calls notify_is_valid_token(0) and xpc_atfork_child() so fork/exec doesn’t deadlock (1f4394a0c9, retried 76d39ae349, revert 6c9b661867).
• Thread creation retry on EAGAIN internal — (14018c8bec).
• darwin DNS via libc user-visible — net uses libc not cgo (a3559f3301).
• internal/godebug internal — lightweight setting API: godebug.New → *Setting.Get (two atomic loads when unchanged) (ea4631cc0c, 8dc08394f0).
• Keep mp alive until mexit done internal — (e252dcf9d3).

Compiler frontend · types2

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

• Type sets, union terms, interfaces and constraint satisfaction — Done: user-visible new comparable semantics. 5a3900b23c added an alternative semantics under Config.AltComparableSemantics (experiment, off by default). 15e705ea96 made it the default, 2ca32a5b99 reverted it, and 8fd2875c3e reland it, with the flag renamed OldComparableSemantics and inverted. The commit also adds the exported predicate Satisfies; the compiler flag -oldcomparable was announced as temporary (“will be removed for Go 1.21”). 642fd5f7ce uses strict comparability for type-set intersection. Other: f60c77026b (empty type sets in operand descriptions), 0789ca4951 (signature instantiation when all type arguments are given). Left open: the flag removal was announced in the message; no removal commit appears in this bucket.
• Instantiation, Named types, Context deduplication and lazy expansion — Done: user-visible ddd8bc1b07 makes the instance lookup in LookupFieldOrMethod use a small inline array before a map: −60% time and −99% allocated bytes on the microbenchmark, ~5% faster type checking of std packages with ~10% fewer allocations. 0789ca4951 instantiates signatures if all type arguments were provided. 70f585f018 marks invalid generic types invalid.
• Type inference and unification — Done: user-visible 0b323a3c16 clearer CTI failure message; e738a2f19b always rename type parameters.
• Language features as seen by the type checker (by release) — Done: user-visible ba5deb408f unsafe.String/StringData/SliceData (+ tests 8c8429fe41); cd8aa40149 slice-to-array conversions; 5c834a2eb0 clear builtin; f80b12667c integer constant literal overflow check; 5a3900b23c→8fd2875c3e comparable.
• Error reporting: codes, messages, and sub-errors — Done: user-visible unification of messages with go/types (above), b52783c1e9, 4a954fafa3, 79d0d330a9 (field with the name of a missing method), 4b993bffb8, de0f4d190f (no unneeded error strings in implements).
• Sharing code between go/types and types2 — Done: test-data consolidation (above) and 157aae6eed for checker-specific local tests.
• Checker performance, memory and API ergonomics — Done: ddd8bc1b07: LookupFieldOrMethod −60% time, −98.8% bytes on a microbenchmark; c641d48afd^1.21 (1.21 by log).
• Robustness: invariants, incomplete types and importers — Done: 398d46d538 removes package height.

Other changes

• Range/shift/const fixes internal — 51af904456 constant shift value, 7ec0ec3645 skip comparison for invalid operands, 245e95dfab no field/method lookup when expecting a type, 5ba98b9756 report type mismatch when conversion impossible.
• Test helpers internal — bfcc3a755c removes global goVersion flag, 27f094b17c, 8190876240 (mustTypecheck), 54ad7f33e4.
• Small fixes, unifier/assignment refactors, message and tracing tweaks internal — 18 commits: cb357835d9, 5befb24bb5, 0fda8b1932, 2392b7061c, c7a0b15659, dad2966a83, 0a1118b8a1, 4c5a6f9555, 17f83e3473, 4360fd8d6f, b16501c08b, b2137e7dad, e22af33b48, 97d38f3002, 665992b515, 19095e109d, a8f9d3f0af, dfd13ce59d

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: Unified IR switches to real dictionaries: c9f2150cfb generates F[T]-dict globals, a F[T]-shaped function with .dict *[N]uintptr, and F[T] as wrapper; 38edd9bd8d is the shape-based stenciling commit (initially shape = underlying type, one dictionary per type, shared by all methods, deliberately so that a future “itab in closure register” scheme could find the dictionary; noted drawback: dictionary entries for pruned methods stay alive, possible fix by linker R_DICTTYPE/R_DICTUSE relocations). Pointer shaping b23d469e85, 6985ab27df, 07cf24bdfe, itab-section pruning 9fc09d2ffb, OCONVNOP for identical iface conversions 34f0029a85, shape rtypes avoided in ssagen (5e0258c700). ca634fa2c5 rejects notinheap type arguments. Left open: 5e0258c700 says shape rtypes are still used for runtime.newobject of heap-escaping variables and “longer term I would like to get rid of shape rtypes altogether”; no pruning of dictionary entries for dead methods.
• Unified IR: one serialization for noding, stenciling, inlining and export — Done: package height dropped from the importer and symbol ordering (3a1f1e1575, 4b78ece3d7, df7cb59de4); dictionary logic pruned (29b07037b1); 0a6e1fa986/ddc93a536f/aa6a7fa775 fixes for shapifying recursive instantiations and reflect naming of local generic types; default-on via 833367e98a (author date 2022-08-09, internal/buildcfg: enable unified IR by default, GOEXPERIMENT=nounified becomes the escape hatch); runtime dictionaries and shapes (see generics thread); pkgbits EnableSync becomes a dynamic header flag (9371a65584, bumping the version; “drchase’s benchmarks show it isn’t as expensive as feared”); body pruning heuristic f995946094 (include all inline bodies for… with a simple heuristic rather than crawler-based pruning); e8afb29111 “temporary” readers to reuse storage; caa8e1f977 type-specialized ReadUvarint; 6b80b62fd1 rawReloc perf. 8ef8b60e18 stops handling type expressions as expressions. Left open: f995946094 is explicitly a short-term heuristic vs. the old crawler.go.
• Old typechecker and IR node-model cleanup — Done: OFORUNTIL removed (a5370d038e), OVARKILL/OVARDEF/OVARLIVE removed from the frontend (e21c1f7896, a726c9f662), typecheck results stored in the syntax tree rather than a map (59bc93535b, ~3% compile time: SSA -4.0%, Unicode -9.0%), types.Rnd -> RoundUp, eagerly create LSym for closures (33738ddd0a), typecheck.EvalConst removal prepared.
• RType/ITab fields and implicit conversions in IR (RTTI plumbing) — Done: RType/ITab fields added (5960f4ec10), reflectdata.{TypePtr,ITabAddr} refactored to prefer them (f70775ff22), ssagen avoids reflectType (5e0258c700), set for OMAPLIT (3d432b6c4b), switch/case clauses (61ae2b734c), range assignments (d667be8831), selects (878439cfe5), writer-side implicit conversions for sends, index expr, complits, call args, assignments, returns, binary exprs and multi-valued expressions spilled to temps (5f5422a2dd, a3e474f867, 711dacd8cf, a3fea7796a, 421e9e9db2, c846fd8e13, 3635b07d16), RTTI written to export data (9b70178d58, 92798176e7, e376746e54), obsolete wiring removed (f48fa643f1), explicit nil / type-switch handling (318027044a, a4c5198a3c), function instantiations / selectors / receivers handled explicitly (2e6ffd6c5d, 88635b3862, de95dca32f), f2851c67fd allows inlining to fail gracefully. Left open: f773bef9ab^1.21 (1.21) notes the RType fields are mknode:"-" hidden fields “to avoid throwing off inlining cost heuristics”; “longer term we should unhide them”.
• Static initialization and package init — Done: inlined-call static init (b1678e508b, bed970b3ff, 249e51e5d9, c85848a4a6), initially unstable: 152119990f adds -d=inlstaticinit and temporarily disables it (“Google-internal testing is hitting additional failure cases”), f89b39c0af re-enables; coverage vars special-cased in init order (4fb35d6cee); Left open: fixes for hidden RType fields (f773bef9ab^1.21) still pending.
• Compile-time performance, concurrency and the self-PGO’d compiler — Done: one manager goroutine hands out work to at most nWorkers (2ff5fbfbd4) instead of hundreds of blocked goroutines; default concurrency = GOMAXPROCS, compiler downgrades silently if debug flags are incompatible (b725622ceb, f3d656be50, tagged flags via c81c027982 “concurrentOk”); if GOGC is unset, boost the GC to get rapid starting heap growth: ~14-17% less user build time and 3.5-7.8% wall (44cabb802a); typecheck results kept in the syntax tree (59bc93535b, ~3%); eager LSym creation for closures.
• Compiler debugging infrastructure: bisect/hash debugging, flags, dumps — Done: GOSSAHASH refactored to be usable outside ssa (522f0fc425, 03f6d81fc7), separator ; -> / (1bcfdc777a), file:line hashing incl. inlining (d33043d37d), concurrentOk flag property (c81c027982), coverage fixup mode (e6d9057e2f).
• Frontend support for new language features and spec changes — Done: user-visible unsafe.String/StringData/SliceData (c708532936), slice-to-array conversions (ceffdc8545), alternative comparable semantics (b95ea201d5; old removed in 1.21 354c8fb436^1.21), go:notinheap restricted (ec2ea40b31) and removed in favour of runtime/internal/sys.NotInHeap (0ee0bb1003, 8bbb362f4c), reject not-in-heap type arguments (ca634fa2c5), updated error messages (“satisfies”, “missing method m”, …) and internal/types/errors package (b440c15db7, 49372957ab).
• The syntax package (parser, scanner, tree) — Done: tolerant go/defer parse (BadExpr instead of a fake call c801e4b10f, no spurious “not used” errors 489f508ccf), missing-index handling (9485d4c1bd), stricter type instances (1be8fcdcdc), ErrorMap -> CommentMap (1.21 1e00516583^1.21), import decl tolerance (a330ca5c54).
• Method wrappers and promoted-method sharing — Done: a8780f94c3 missing method-value wrapper; 1519729c6a etc.

Other changes

• Error messages and error codes user-visible — messages rephrased to match the old compiler (“missing method m” 336ce966e4, “method T.m already declared” 7398c3c0c6, “init… cycle” c58bfeeb04, “cannot use … value in …” 7997e5f254, “satisfies” 46e3d9d12a, 6ad27161f8); shared internal/types/errors package with exported codes (b440c15db7, 49372957ab, 470d146276, 578523e4a0) used by go/types and types2; structs package-annotated when ambiguous (ea2c27fe82).
• Std modernisation internal — mechanical ioutil -> os/io, strings.Builder in syntax/ir/pkginit (86f8b8d3f2, 1429a97520, 431446d0a0), SEEK_* constants, unsafeheader.Slice/String (41089704dd), mmap on solaris/aix (7c3284401f, 349d398ea3).
• Struct-compare inline heuristic internal — cost-based rather than field count (6605686e3b); non-CALL comparisons first in array-of-struct equality (c5ba9d2232^1.21, 1.21).
• PGO user-visible — profile-guided inlining first enabled (99862cd57d, CDF-based threshold b07e845e76, parameters 1b03568ae1; opt area).
• Ineffectual linkname user-visible — allowed again in -lang<=go1.17 (aa6240a445).
• Reproducible toolchain internal — cmd/dist builds cmd with CGO_ENABLED=0 (8a27154bcd^1.21, 1.21 log).
• Unified IR: noder/writer/reader refinements and pkgbits (1.20) 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 (55fc07e164, b39ac80871, 46ddf0873e, 9e5c968021, e7ef58542c, d6df08693c, c50c6bbc03, 7368647ac6, c70e93ff3d, 9cb784ac69, 46b01ec667, 76a82f09d6, 64cd6faa13, 131f981df0, 831fdf1dff, ac0844ec27, 2699065483, 7d70779db5, 9f8685f49e, db84f5315f, 2f103873c5, e4b624eae5, dfdf55158d, 5d213a3dc7, 157e018fc3)
• Generics: dictionary/shape fixes in the stenciler (1.20) internal — fixes in dictionary passes for type assertions, comparisons, switches, nested instantiation packages, and recursive generic interface instantiation (62654dfd4f, 9944ba757b)
• Old typechecker and IR node-model cleanup (1.20) 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 (13bd6ddbe7, b911771f12, bd901af30b, 68fa3696b6, 271f139f17)
• Syntax package: error messages and parser tolerance (1.20) internal — better syntax and go/parser error messages, accepted type-parameter lists, parser simplifications (aa5ff29dab, 1280ae7856, 8c29881dd1, c591d82ea9)
• Housekeeping (1.20) internal — dead-code removal, ioutil/seek/slices/atomic-type replacements, README edits, debug-flag removal, assorted small bug fixes and doc comments (fb1bfd4d37, 0c4db1e347, 3ea3d0e8a7, 2493072db6, 0a86cd6857, e99f285d52, 1e7f535475, 274528eca1, 00bee6d9a4, f8c659b62f, fac5338a6c, ae3abf16e4, 8bd803fd4e, 4bcf94b023, da6042e82e, 7ea0520234, 4725c71b73, b726b0cadb, d3726f3469, 96711e4d8b, 6a70292d1c)

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: user-visible Raj Barik’s initial implementation: pgo/graph.go and irgraph.go build the profile graph, and the inliner uses it (99862cd57d, #55022). Within the same cycle the heuristics were reworked:
  • Inlinability is decided by edge weight rather than node weight (fb4f7fdb26).
  • The hot threshold is a CDF over edge weights, not a fixed percentage (b07e845e76).
  • Calls are matched by line offset from start_line (bdd1e283a9, with a fix in 0409314db3).
  • Only the last two frames form an edge (6e0e492e12).
  • Defaults became 99% CDF and a 2000 budget, and the debug flags were renamed pgo* (1b03568ae1).
  • Supporting work: global state, label handling and per-location maps were removed for speed (ec0b540293, 204be97d24, d73885588a, 1a4cc091a9, a72e133422, 7e88b866f9, f187c6b08e, d9cc39b25c, 3f1bcc58b3); cross-package inlining was fixed (ada9385a5f); empty profiles are accepted (efe541d4e5); debug output and the missing-start_line hint were improved (e5d2814576, 64519baf38).
  Left open: the PGO package was still labelled “WORK IN PROGRESS”; the label was removed in 1.21.
• Inliner: budget, closures, and the interleaved inline/devirtualize driver — Done:
  • With unified IR on, inlining fails gracefully (f2851c67fd) and more inlining is allowed, lifting the shape restrictions (52016be3f4).
  • Transitive inlining is deferred until the AST is edited (f983a9340d).
  • Cost discounts for encoding/binary appends and zero-offset field addresses (629c0b3a6e, 638c9aad5f).
  • A latent CalleeEffects issue was fixed (25d74f324d).
  • Coverage-counter updates do not count against the budget (4a459cbbad, 317f2a7df6).
  Left open: nothing noted.
• Escape analysis: call modelling, closures, func-variable callees, and solver speed — Done: OITAB and OCHECKNIL are supported (0c2f87f23d). Left open: nothing noted.
• Prove pass and bounds-check elimination: from poset relations to a dense limit table, topo-sorted walk, and division-aware proofs — Done:
  • The relation between x+delta and x (d2e0587f77), with 8/16-bit deltas (2addbf3f2a).
  • Bounds for strided-loop induction variables were tightened (af7f067e0d).
  • Signed relations for constant integer phis: −1.4% total size (d96eb826cb).
  • Bitwise OR (3680b5e9c4), reverted for compile time (a6ddb15f8f^1.21) and relanded in 1.21.
  • An index out-of-bounds check in walkIndex was removed (98f5152368).
  Left open: the OR support was pending a compile-time fix.
• Memory-op optimization: memcombine, dead stores, store-to-load forwarding, alias analysis, mem2reg — Done: more constant stores combined on amd64 (09932f95f5), plus store combining of values (e7307034cc). These were per-arch rules, the predecessor of the memcombine pass. Left open: each architecture needed its own rules, which the 1.21 pass replaced.
• Lowering of builtins to memory ops: memmove/memclr/memequal inlining and generated equality — Done: the memmove optimization no longer trusts the destination pointer’s base type (e283473ebb). Struct-compare inlining uses a cost heuristic (6605686e3b). Left open: nothing noted.
• Switch lowering: jump tables, type-switch caches, and switch lookup tables — Done: a better split for string binary search (69aed4712d). Jump tables are off when spectre retpolines are on (1eb0465fa5). Left open: nothing noted.
• Static devirtualization and shape-aware type-assertion pruning — Done: fixed a unified IR devirtualization bug (d6294e00f0). defer/go calls are not devirtualized (c82304b712). OCALLMETH is desugared to OCALLFUNC within devirtualization (4f8bc6224b). Left open: nothing noted.
• Range-loop and range-over-func lowering in walk — Done:
  • OFORUNTIL is no longer used (661146bc0b).
  • Range desugaring was refactored for unified IR, with RType fields for range assignments (0a503cf43a, d667be8831).
  • Careful pointer incrementing (8477562ce5).
  Left open: nothing noted.
• Branch-free lowering: CondSelect-into-math and ARM64 CCMP — Done: CMOV disallowed when an argument is pointer arithmetic, which could be unsafe to speculate (c8057d8569).

Other changes

• Fixes and small rules (~55 commits) internal — booleans registerizable (de561dc766), arm64 RotateLeft8/16 and MulUintptr intrinsic (121344ac33, d654117075), loong64 rotate intrinsics (e8f0340fa4), combine rotates (33a7e5a4b4), stack offset alignment beyond PtrSize (1211a62bdc), static-init slice walk/typecheck fixes (fce449680a, 03a1810473), libfuzzer instrumentation tuning (6b113c0fec), sparsemaps with position separated (7ddc45263c).
• VarKill removal internal — with stack objects, VARKILL is unnecessary; usage stopped, SSA op, frontend OVARKILL, OVARDEF and OVARLIVE ripped out, VarDef only for pointer-ful types (908499adec, 6dc7b060cd, e21c1f7896, a726c9f662, a74e5f584e).
• Growslice calling convention redone internal — growslice(oldPtr, newLen, oldCap, inc, typ) so old length is dead across the call and registers align better (dd323fe205); append-with-constant-length recognition and post-decomposed growslice rule (0156b797e6, 9ce27feaeb); unsafe.Slice codegen (c2a9c55823, nil-with-length panic 52d0667e6b).
• Late lower pass internal — extra pass so last rules run (ddc7d2a80c); arch-independent rotate detection (60ad3c48f5); >>c<<c to mask (5b1fbfba1c); Floor/Ceil/Trunc/RoundToEven constant folding (ab8a2c5e44); sqrt assembly removed in favor of intrinsics (c8000a18d6).
• Compiler memory reuse internal — cache of sizeable objects via sync.Pool of power-of-two slices (68bd383368); fuse largest runs of plain blocks (9e0149da3c); quadratic IsNewObject fix (0202ad0b3a); Not-in-Phi block control elision (fd59c6cf8c, fixed 8893da7c72); overlap handling for assignments (332a5981d0, 6485e8f503).
• riscv64 optimization series (~15 commits) internal — redundant SEQZ/SNEZ, extension avoidance, immediate combining (9085ff5859, 77da976419, 7234c90352, 5760fde4df); riscv64 Add64/Sub64 intrinsics (969f48a3a2, a6219737e3); arm64 TST/shift/Muluhilo changes (efe5929dbd, 54c7bc9cff, 8bf9e01473).
• dev.unified rtype plumbing internal — extract rtype code from walk, avoid reflectType in ssagen (93833cd5d8, 5e0258c700, 3d432b6c4b).

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: temporary registers: ops can declare a temp so regalloc reserves one instead of a fixed register; amd64 converted 5f7abeca5a, with a fix when mask is empty 9088c691da; issueSpill removed e81263c791; fix score for Select{0,1} of flags 897ad2fe90; sparsemaps separated 7ddc45263c; carry chain scheduling priority readjusted d031e9e07a; OpVarKill dropped from the backend 6dc7b060cd.
• amd64 code generation and GOAMD64 microarchitecture levels — Done: combine more constant stores 09932f95f5; 3-operand LEA split in the new late-lower pass 1c783f7c68, {SHL,SHR,SAR}X rules moved to late lower 44563340cc; Windows TEB TLS slots on amd64/386 (da564d0006).
• arm64 code generation and assembler — Done: late-lower pass makes constants materialize once (movz+movk) instead of repeatedly folded into adds ddc7d2a80c; TST optimizations in-block efe5929dbd; shift by v&63 uses hardware masking 54c7bc9cff; CMP↔︎CMN rules cf53990b18; LDP/STP memory moving 0a52d80666; carry-chain scheduling enabled on arm64 (reverted once fec83c8a7d, relanded) 4c414c7673,669ec549b5; Duff’s device on darwin/arm64 4bcc138bc6; Muluhilo split 8bf9e01473.
• riscv64 compiler and assembler (RVA profiles, vector, compressed) — Done: a run of ~12 constant/extension peepholes: SLTI/SLTIU known outcomes a7bcc94719, SLT constants→immediate ba8c94b5f2, remove redundant SEQZ/SNEZ 77da976419, combining ops with immediates 7234c90352, mask+zero-ext 0ca355318f, avoid extension after word arithmetic 5760fde4df, FP compare 239115c3ef,646c3eee06, negation folding 4274ffd4b8, sub const 3e11e61f3c, Slicemask c011270fa5, bounded shift da6556968f; CAS32 arg sign-extension fix (LL/SC loads sign-extend; shared with loong64/mips64) 0b2ad1d815.
• loong64 backend and assembler — Done: brachelim enabled 268f4629df; tail-call flag 8a9e2d9d49; MASKEQZ/MASKNEZ 1dcef7b3bd; ROTR + intrinsics 7d574466a9,e8f0340fa4; Add64/Sub64 intrinsics 97760ed651,90a3527427; resultInArg0 register checks removed c1c6b0ca79.
• PPC64 / Power10 backend (GOPPC64, prefixed instructions, PC-relative code) — Done: ISA 3.1 and MMA registers 3a067b288e,85c0d26c6a,e6eaf39cfa; atomicCas without branch 62125c9b79; condition-code ops (…CC) reuse to drop compares d0b0b10b5c; lateLower enabled f841722853 with ISEL merge dc6b7c86df; Mul64uhilo rules 871a3a409a; big uint32 constants in registers 4e6f90fecd.
• Duff’s device replaced by generated loops; Zero/Move lowering — Done: Duff’s device enabled on darwin/arm64 4bcc138bc6; fine-tuned memory moving on arm64 0a52d80666.
• Late-lower pass, memory-op combining and arch-independent rule migration — Done: late lower on arm64 ddc7d2a80c, PPC64 f841722853,bd6cd7db07, amd64 1c783f7c68,44563340cc; riscv64 late lower type fix 2952fe4f16.
• Multi-word arithmetic, intrinsics and carry-chain support — Done: arm64 carry scheduling 669ec549b5; Add64/Sub64 on loong64 97760ed651,90a3527427; Muluhilo split on ARM64 8bf9e01473; 386 TrailingZeros (1.21) cedfcba3e8^1.21.
• WebAssembly backend — Done: bulk memory ops 50557edf10; docs on Go→wasm mapping 660d4815ea.

Other changes

• Start line numbers in func metadata user-visible — runtime._func and inlinedCall record the func/TEXT line so pprof can use relative lines; +0.83% binary size vs master, −0.31% vs 1.19 f2656f20ea.
• debug-hash for FMA internal — -d=fmahash bisects FMA-sensitive floating-point bugs 667c53e159; hash-debug on file:line, including inlining d33043d37d.
• cmd/go support for long assembler/cgo commands 5bc8fa74f5; ssa _gen/_builtin renamed 164406ad93.
• ssa Config cache of sizeable objects internal 68bd383368.
• 10 small obj/arm64 and arm64 rules commits internal — e.g. obj/arm64: remove the transition from $0 to ZR; cmd/asm: add VTBX instruction on arm64; cc: Add shiftIsBounded check for logic shifts of arm64. (9f0f87c806,9e8020b2d4,454a058ffc,205f636e0a,0613418c98,eeb1ba7a89,ebc75ac1c6,a813be86df,9f78c3cb04,612bb34af5)
• 10 small obj/ppc64 and PPC64 rules commits internal — e.g. cmd/asm: add new classification for index memory operands on; obj/ppc64: eliminate dead code from asm9.go; cc: fix confusion with ANDCCconst in PPC64 rules. (67d85ad00f,f67c766069,c1bfefe9d1,403f91c244,59b15726d1,a50c434308,4350c4a131,6e1b769377,1985507257,888047c310)
• 11 small ssa/regalloc/liveness/obj generic commits internal — e.g. obj: remove redundant code in genFuncInfoSyms; abt: use strings.Builder; all: remove unnecessary allocations from w.WriteString(fmt.S. (f7e6986d6b,0cf996a9c1,4baa486983,00bee6d9a4,1a6af5f7a0,2b1c414fa8,b314eea4cd,85196fc982,e09bbaec69,5497300d9c,e8f78cb60c)
• 6 small obj/riscv and riscv64 rules commits internal — e.g. cc: deadcode for LoweredMuluhilo on riscv64; cc: avoid extending when already sufficiently masked on risc; cc: remove NEG when used with SEQZ/SNEZ on riscv64. (b60432df14,9085ff5859,971373f56a,b6a6847b2f,83d94daec2,af668c689c)
• 6 small cmd/asm commits internal — e.g. cmd/asm: use strings.Builder; cmd/asm: use opd.String() instead of fmt.Sprintf; cmd/asm: optimize if statement to return directly. (cf4edac16e,c4a3fae27e,58093e8e2b,09daf313b2,0b7aa9fa5b,9dde2de454)
• 1 small obj/x86 and amd64/386 rules commits internal — e.g. obj/x86: return comparison directly. (574b5decf2)
• 1 small obj/loong64 and loong64 asm commits internal — e.g. cc: remove output registers limit for MUL/DIV on loong64. (ffc4496306)

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: unused fields removed from inlinedCall c83e1b6d5b; linker _func names aligned with runtime 3e75859390; panic if an inlined function lacks FuncInfo 8a0cf719a6. Left open: nothing noted.
• DWARF quality for debuggers (types, variables, trampolines) — Done: DIEs for zero-sized variables e59d873ff9; trampolines no longer break line info 2340d37179. Left open: nothing noted.
• Linker performance and memory — Done: fewer allocations in object writing (−3–6% compiler allocs) 5497300d9c. Left open: nothing noted.
• ELF/relro/hardening and external-link policy — Done: glibc vs musl ldso detection 3315066f46; ":" used for compiler symbols 0f8dffd0aa; coverage counter hooks 072c7d4969. Left open: nothing noted.
• Per-platform object formats and internal linking (Mach-O, PE, XCOFF, wasm, ppc64, riscv64, loong64) — Done: PPC64 power10 prefixed relocs 69abfab979, -fno-plt support 7fda98a8d9; freebsd/riscv64 9faf6b7929. Left open: nothing noted.

Other changes

• Platform work internal — PPC64 inline PLT and ldso paths (c318f191e4, 15d395a196), musl interpreter (0aad4d3257), riscv64 loadelf relocs (bc73996fac), darwin libc msync and no pagezero (d36466fe2a, 5231ba2f05, 706d84fca2), Windows no timestamp (fe67a21625), PE DLL import symbols (cf93b25366), arm64 ADRP+ADD+LD/ST fusion with -1.06% text on HelloSize (537c4354cb), entry symbol/runtime.a missing errors (77420fa119, fffda6b3ad), libfuzzer counters section (1ba7341cb2), AIX coverage (bf5805eae4), basic types not emitted from every compilation (cb5534c1c6).
• Object header records GOARCH setting internal (819e3394c9).
• Drop Go 1.12 darwin compat internal (8df21a7802).
• -capturehostobjs debugging flag internal (bda0235466).