Go 1.21

Inference rewrite plus min/max/clear. Unified IR is the only frontend. abi.Type unification. Unwinder iterator replaces gentraceback. PGO devirtualization, and the compiler itself is built with PGO. PanicNilError. Pinner. wasip1. Tracer groundwork and FP unwinding.

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: cons/mark smoothing became the max over 5 cycles, biased against assists (b513bd808f, comment fix 600b982df9). Left open: after an application becomes less GC-active, the stale cons/mark persists for 5 cycles. The commit accepts this (b513bd808f).
• Soft memory limit and GC CPU limiter — Done: user-visible the limit is initialized in mallocinit, so early allocations do not scavenge needlessly (96e8e62821). Headroom became proportional (3%) to stop GOGC=off programs from scavenging on most allocations (e97bd776f9). Eager-scavenge details were added to scavtrace (a3e90dc377). The go119MemoryLimitSupport flag was removed (79edd1d19d), a lost idle-time flush was fixed (24617a068c), and a /gc/gomemlimit:bytes metric was added (2544b1051e). Left open: “This choice is still somewhat arbitrary … perhaps worth revisiting” (e97bd776f9).
• Scavenger redesign — Done: idle Ps’ page caches flushed at the end of each GC so the scavenger can see them (8d68b388d4); index initialization and minHeapIdx restored (8588a3fa08, 15c1276246); eager-scavenge statistics in scavtrace (a3e90dc377); density tracking (huge-page thread). Left open: nothing noted.
• Transparent huge page management — Done: explicit policy with 96% density and 1-GC-cycle hysteresis (8fa9e3beee), sysNoHugePage (84eaceaba7) with a page-alignment check (ada0eec827), and metadata huge pages only for large heaps (1f9d80e331). Left open: “The main risk here is churn” (8fa9e3beee).
• Finalizers, runtime.AddCleanup and the cleanup queue — Done: test portability only (fbf14ae166). Left open: nothing noted.
• Object pinning (runtime.Pinner) — Done: user-visible the Pinner API (251daf46fb), gcBits-based state (697644070c), ref preallocation and P cache (4bb38fe458, 0adcc5ace8, benchmarks 94c75232fb, 5d68121628), lock-rank edges (78df00dd6f, 807cce479f), and docs (c7b2f6494c). Alongside: heapObjectsCanMove() (always false) for go4.org/unsafe/assume-no-moving-gc, so that package stops breaking on each release (055c186f53, 789701e93a). Left open: nothing noted.
• Memory arenas experiment — Done: user-visible reflect.ArenaNew changed to take T and return *T, a breaking change (2b0ff4b629). Left open: nothing noted.
• Mark worker scheduling and GC work queues — Done: work.full/work.empty padded against false sharing (gRPC +8%, DeathStarBench +9.7%, bdc5533f39); duplicate work.full check removed (9411075748). Left open: nothing noted.
• Sweeper termination and pacing — Done: overflow in the sweep-assist delta could pin all allocators in sweeping until termination; fixed (a7b75972f2). Left open: nothing noted.
• Address-space layout, tagged pointers and the mem.go layer — Done: lfstack→taggedPointer (d959dd97b1); fallback to mmap when madvise is compiled out of Linux (5ffdc1f15c). Left open: nothing noted.
• Linear-memory heaps (Wasm and sbrk platforms) — Done: Wasm reuses blocks freed via sysFreeOS, sharing Plan 9’s sbrk strategy (dac514ea37). Left open: nothing noted.

Other changes

• Write-barrier buffer generalized for batching internal — the runtime no longer requires pointers in pairs and checks overflow first (d3daeb5267); the compiler batches barriered writes into gcWriteBarrier1..8 buffer slots (21d82e6ac8, compiler area). user-visible GODEBUG=cgocheck=2 became GOEXPERIMENT=cgocheck2 to enable these barrier optimizations (55044288ad).
• reflect typedmemmovepartial removed internal — unused (df6396fc22).
• Tracer state refactors at GC call sites internal — traceEnabled() and per-P trace state (8992bb19ad, b7e767b022), diag area.

Scheduler and concurrency

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

• Scheduler core: spinning Ms, work conservation, stealing, STW interaction — Done: stale now used by netpolling Ms fixed (6f13d0bfe4); checkdead double-lock/startm handling (f229787aff, ff059add10); extra-M API cleanup and exclusion from SetMaxThreads (77b1f23af7, 734b26d4b9). Left open: nothing noted.
• sync package API evolution — Done: user-visible OnceFunc, OnceValue, OnceValues (bb44c2b54e); Map memory model docs (f07910bd57). Left open: closure retention.
• Runtime lock ranking — Done: raceFiniLock, pinner and mspanSpecial/gcBitsArenas edges (f00e947cdf, 78df00dd6f, 807cce479f). Left open: nothing noted.
• Netpoller — Done: fd sequence tags (d5514013b6); wasip1 netpoll through poll_oneoff (c5c2184538); windows filter by completion key (7830180dcd). Left open: 32-bit tag.
• Goroutine lifecycle and identity — Done: user-visible parent goroutine ID in tracebacks (51225f6fc6). Left open: nothing noted.

Other changes

• Tracer refactors internal — per-g/m trace state types, trace clock, tracedSyscallEnter, block reasons (c213c905a2, 2f35a655e7, b60db8f7d9, 865179164e, 7c91e1e568, a2737b1aab, 62bf7a4809, c890d40d0d, e3ada56537); regular unwinding for cgo callbacks (3dd221a94d); frame-pointer unwinding off (608f204ac7) then on for amd64/arm64 (2d83b646d6).
• Unwinder rewrite internal — gentraceback deleted (3790ceca97), inlineUnwinder (166e5ee4f2), top+bottom 50 frames printed for deep stacks (9eba17ff90).
• Pinner caching internal — per-P inner pinner (0adcc5ace8).
• Misc internal — syscall fork avoids serialising on ForkLock (d6473a1263); wasm timeout calls (1cfacfbe8a); loong64 fcc in async preempt (48d61a46e2); extraMInUse (7911f7c21d); raceignore reset on new goroutine (48dbb6227a); linkname signature fix (8fb9565832).

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: equal/hash generation moved to walk and factored (ea3054aecc, 864b54e2c3, 4636c57fe1, 25f5d9d4a2; one revert/reland 95be547e0a/eee2697c38), AfterGlobalEscapeAnalysis removed (8ffcd3da93). Left open: nothing noted.
• internal/abi type layout unification and type metadata (rtype → abi.Type) — Done: above; bdc6ae579a, a2838ec5f2, 6cf7602b71, 3ee12d5702. Left open: “the new files in internal/abi are copied from the end of the stack; ultimately this will all end up being used” (bdc6ae579a), i.e. the rtype receiver change was reverted (6d2309b576).
• reflect implementation: performance, allocation and escape behaviour — Done: Value can be stack-allocated (0ac72f8b96), Type does not escape (be4fe08b57), Value.Clear (d42c08a2be), SliceHeader/StringHeader deprecated again (33d8cdeedc), unused typedmemmovepartial removed (df6396fc22). Left open: nothing noted.
• memmove / memclr / memequal / internal/bytealg assembly optimisation — Done: Power10 memclr/memmove/equal/index/compare (e6faa375b4, cdf77c7209, 8377f2014d, 56c4422770, 5f1a0320b9, db32aba508), riscv64 compare unaligned 8-byte loads (261fe25c83, -11..17%), a series of riscv64 alignment fixes (a4772a1a59, bfa25c3f6c, 9552a1122f, 03cd8a7b0e), PCALIGN modified to ensure alignment (e23322e2cc). Left open: nothing noted.
• Slice growth, string building and allocation-free bytes/strings helpers — Done: MakeNoZero (132fae93b7), unsafe.Slice usable from nowritebarrierrec (da384766a0). Left open: nothing noted.
• Classic map maintenance and maps package helpers (pre-swiss) — Done: e22bd2348c, 231f290e51, maps helpers. 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 STW events emitted for all pauses with kinds, not just GC (b1aadd034c; bumps trace version to 1.21 and adds test data); trace clock formalized (b60db8f7d9); a long series of preparatory refactors for the redesign: per-M trace state in a type (2f35a655e7), traceEnabled/traceShuttingDown helpers (3e9876cd3a, 62bf7a4809), tracedSyscallEnter replacing the subtle sysBlockTraced flag (865179164e), non-reentrant trace.lock (94a948d1b1), HeapGoal recorded at start (192814985d); regular unwinding for cgo callbacks via m.incgocallback (3dd221a94d). Left open: still one global buffer design; redesign proposal in #60773.
• Frame-pointer unwinding in tracing and profiling; profile stack depth — Done: new tracebackPCs API built on unwinder (3e360b035f, skip arg later removed c3db9af3a6); FP unwinding for tracing on amd64 (enabled, turned off, then re-enabled by default for amd64 and arm64 in 2d83b646d6, after crash fixes and arm64 frame-pointer fixup on stack copy); disabled on plan9 (bce7aec3cd). Left open: cgo-callback threads fall back to slow path.
• CPU profiling, goroutine profiling, pprof labels and the protobuf writer — Done: generic instantiated symbol names recorded in CPU profiles so PGO can match them (10ad6c91de); concurrent goroutine profile flag removed, making it the only path (469afbcc46); lazy regexps in internal/profile cut init 185µs → 6µs (f53137f95c). Left open: nothing noted.
• runtime/metrics and MemStats — Done: user-visible /gc/heap/live:bytes (ef2bb813c8), /gc/gogc:percent (1eae5c160d), /gc/gomemlimit:bytes (2544b1051e), /gc/scan/{stack,globals,heap,total}:bytes (8b51eb5199, a63afe4720, 131267f147, 5124371c1c; consistency fix 72bf8acee9), /godebug/non-default-behavior/* counters (213495a4a6); /sched/latencies:seconds marked cumulative (c2c787d73e); doc clarifications of OS stack metrics (b2215e49c7); ReadMemStats nil check before stack switch (e126572f8a). Left open: /gc/scan/* was always 0 (fixed in 1.22).
• Coverage redesign: runtime/coverage, internal/coverage, covdata — Done: user-visible atomic-only counter APIs (61e5ea4929); count-vs-emit discrepancy fix (39957b5d89); coverage.snapshot for testing.Coverage() (ef67022471); auxiliary meta-data files for packages without tests (c99d966c17); EmitPercent aggregation (380529d5c3); counter file ordering (0f91f92ee0). Left open: go command wiring for no-test packages (landed in 1.22).
• Race detector, ASan, MSan and Valgrind integration — Done: race tracebacks symbolize wrappers last (cb6b45a993); __tsan_fini marks entering non-Go code (30886390c2); 1-byte load for address checking in racecallatomic (1a7709d6af); ppc64le thunk cleanup (e78871ea07). Left open: nothing noted.

Other changes

• Tracer/coverage/metrics internals internal — traceClockNow nosplit (e6fb1905cb), M passed to traceReleaseBuffer (07a25a9f6e), trace lock/extra-M refactors (a2737b1aab, c890d40d0d, e3ada56537), faketime trace.Stop deadlock (942c1c12d8), CPU stats accumulation refactor (3651d8e516), histogram metric doc (14f833f117), moduledata function-name dedup (dcb4c1c1aa), coverage encode/decode seek and func-literal fixes (7b89531860, 5322e66a12, 620399ef0d, 695a53620f), covdata percent format (5cdc3874b0).
• GODEBUG counters user-visible — non-default behaviour counters for settings in cmd/go cache and elsewhere (14ab998f95, 9b41418074, 6a063b01b0).

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: user-visible panic(nil) replaced by panic(new(runtime.PanicNilError)) so recover() can no longer return nil for a panicking goroutine (aa51c40b1c). Left open: nothing noted.
• Unwinder: gentraceback → unwinder iterator, inline unwinder, pcvalue cache — Done: traceback iterator (f758d648b0), waspanic eliminated (5acd2d658e), callee funcInfo resolved after the current frame (6f22d42c74), all callbacks migrated (2d99109cfc), PC back-up simplified (c9b2da3f28), gentraceback deleted (3790ceca97), inline unwinder API (f52bede354), tests for inlined frames in tracebacks (59d0de16e4), callee tracking fixed in printing (5f62ba621e), Callers/FPCallers benchmarks (2d93805171, 25f9af6661). Left open: Callers got slower; hot-path inlining of unwinder not solved.
• Symbol tables, runtime.Caller and CallersFrames — Done: name logic deduped into a moduledata method (dcb4c1c1aa); linker emits full symbol names for generic functions into the runtime table (1a612783c5); ID constants merged into internal/abi (9754521157, 7843ca83e7). Left open: nothing noted.
• Frame pointers: explicit NOFRAME, FP maintenance and FP unwinding — Done: explicit NOFRAME on windows/amd64 (083d94f69c), darwin (d939b3717f, after revert a9780d8dfd), linux (521d2614df), dragonfly, freebsd, netbsd (asm_netbsd_amd64.s deleted), openbsd, solaris, plan9; heuristic removed (d20e688fcf); unneeded NOFRAME removed (b35ee3b046, 43f911b0b6); FP adjusted on stack copy for ARM64 (a41a29ad19); FP unwinding in tracer on amd64 (ba71817390); frame pointer loop fix (b52f6d3721); systemstack FP adjustment fix and test (2380d17aea, 6dca1a29ab, 47e6fd05f7); tracer avoids FP unwinding in cgocallbackg (265d19ed52). Left open: FP unwinding only on amd64 at this point.
• Stack growth, stack bounds and the crash stack — Done: sigpanic is not injected when the fault is on a g0 stack (a0c9d153e0): output becomes a plain SIGSEGV ... signal arrived during cgo execution and C/debugger unwinders aren’t confused. Left open: nothing noted.
• C-to-Go callbacks: extra-M reuse, system-stack bounds, init gating — Done: M cached per C thread in a pthread key (ccad8a9f9c, 7b874619be, c426c87012, with reverts bfe3c678ab, 94850c6f79, 72c33a5ef0); needm better g0 stack bound (443eb9757c); extra-M API cleaned (77b1f23af7); tsan sync for cgo traceback function (618fb4ab06, 46bc9a6e1e); PPC64 vector registers preserved in crosscall (1e51b1fac3, 5c51e9f45b). Left open: stack-bound staleness (fixed in 1.21/1.22 follow-ups below).
• Library entry points and runtime initialization — Done: ppc64le/linux library startup cleaned (12e65dbdb1, ac571a388d), libfuzzer stack alignment (5d76600cc0), loong64 c-shared (80eeec6979). Left open: nothing noted.
• Security hardening: AT_SECURE, std fds, AT_RANDOM, gated debug features — Done: user-visible SUID/SGID protections (2496653d0a); GOTRACEBACK=wer on Windows (65ea4c5021); debugger hook for fatal signals (d8c9eece28). Left open: “in the future we may add additional protections”.
• Signals and async-preemption machinery — Done: debugger fatal-signal hook (d8c9eece28), sigreturn symbol named with “sigaction” for gdb (0ed364a463), EXCEPTION_IN_PAGE_ERROR (e03ee85ef4), sigtramp fixes on ppc64 (ac571a388d), no sigpanic off user G stack (a0c9d153e0). Left open: nothing noted.
• Windows runtime port — Done: TEB slots on 386 (28f8dbd7b9), LoadLibraryEx (a17d959deb), GOTRACEBACK=wer (65ea4c5021), procid (140000dd55), high-res timers on arm/arm64 (b6f29d2365), TEB fallback (715d53c090), windows sigtramp factored (cf9263dee1), NOFRAME cleanup (b35ee3b046). Left open: nothing noted.
• New, ported and removed platforms — Done: wasip1 (8890c3a05f, be50055eef), go:wasmimport limits (53c0158905), wasm memory reuse (dac514ea37), arm64 non-regabi removed (602e6aa979), loong64 c-shared (80eeec6979). Left open: nothing noted.
• CPU feature detection, GOAMD64 checks and cpu GODEBUG options — Done: auxv exposed (1a9893a969); Neoverse N2/V2 (e4b03f9425); ppc64 default osinit (ada6ffca83). Left open: nothing noted.

Other changes

• loong64 cleanups internal — entry point, sigtramp callee-saved registers, RSB removed, hardware NOPs, cputicks from stable counter (574c620901, 5fa65d8c93, 6f2ef5ec98, 7789966286, f92a69f907, c8402cd330, b5c8ae9816, 93897960cb); PPC64 funcdesc consolidation (6827f0d772, 1064335047, landed after reverts of 8c2900bb9f), -fstack-protector linking (066b780502, 48f4728211); write-barrier buffer no longer requires pointer pairs (d3daeb5267, d49719b1f7); gentraceback stack tracking dropped (4f132b7ef8); stackguard1-style cgo fixes (fba8566cda reverted 99276aeb5e).
• SUID/exec details internal — syscall restores NOFILE rlimit in child (f5eef58e43); fcntl consolidated (547aed18e0, 10532fdb14, 58315a771a); os sigpipe removed (719e2b6f09).
• Wasm/wasip1 net user-visible — net.FileListener/FileConn on wasip1 (a17de43ef1), Fcntl (7756f82590).
• auxv exposed internal — for x/sys/cpu (1a9893a969); arm64 SHA512 hwcaps (b90971b292).
• Darwin core dumps user-visible — permitted on darwin-amd64 (d7d235c92f); AIX/Solaris _exit (6bbbc5dc70); Illumos pthread_attr_get_np (ad87a124be, reverted 0e9b2bc39a).

Compiler frontend · types2

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

• Type sets, union terms, interfaces and constraint satisfaction — Done: b44f2222b5 (methods considered when unifying a type parameter against its constraint) and the missing-method rework (09852e75ac, c10ba768fe, 9f2fe2d721, 2f9e38a1be, 4ad72feb92, c6cdfdabef ambiguity message).
• Instantiation, Named types, Context deduplication and lazy expansion — Done: ac9137f8d3 (NewChecker does not mutate its arguments); 1dd24d8216/2ce84c3a6e interface and missing-method fixes.
• Type inference and unification — Done: user-visible the big inference release (proposals #58650, #59338, #60353, #58671). 683b2c4600 (infer2), 969c3ba839 (exclusive use; two unifier tweaks: use underlying type when the other side is a type literal or basic type, and always continue with the core type for unbound type parameters). Reverse inference of generic function arguments in assignments (cc048b32f3, behind Config.EnableReverseTypeInference; flag removed by fc106b016c when proposal #59338 was accepted), corrected to model the assignment as a single function-typed argument g'(func(t1…)) so parameter types must be identical (3367475e83, #60688). Generic functions as arguments (099f5a985f permits partially instantiated functions, ea9097c9f7 reverse inference for function arguments, 8337ca094b rename generic function arguments, 13201d5728). Untyped constants: the minimum default type is inferred (3c3b1d39be, from proposal #58671) and 2e4fe8bbf0 removes the version check. Interface inference (#60353): c4afec232c and 948f11b8a8 unify against interfaces by shared methods, 527d0e8a91 reverted one attempt, 6b7aab7750 enabled it, 1d3daebc5f/d49017a7c3/7183e77cc0 fixes, d2db8267f4 keeps inferring from constraints for -lang < go1.21. Smaller: a5350558e6 exclude untyped nil early, d9dce0cd26 check for nil targs, 4cba0bd74a. Left open: c4afec232c added EnableInterfaceInference as a config flag, later made internal; in 1.22 9070f27039^1.22 disables interface inference for versions before 1.21 after 60dd831103^1.22 reverted the constant removal.
• Per-file language versions and version gating — Done: 804d786a30, 4f9a966926 (fix version downgrade bug without Config.GoVersion), ade3f3f5ef (check builtins even if there is a version error), b490bdc27d record Config.GoVersion for Package.GoVersion.
• Language features as seen by the type checker (by release) — Done: user-visible 8c37d486c8 refuses pointer-to-array for clear (the clear builtin itself, 5c834a2eb0^1.20, is in the 1.20 bullet); b5515eef56 min/max (b3ca8d2b3c records the final type for arguments; 8abc6e2597^1.22 in 1.22 rejects untyped value arguments); reverse inference (inference thread).
• Error reporting: codes, messages, and sub-errors — Done: ad3ea0134f, 1079a5c08a better inference failure message; dbdb3359b5, c6cdfdabef; test-framework: 5cd805c6df distinguishes substring vs regexp error patterns, 3af3810a3e, d4639ecdfc.
• Sharing code between go/types and types2 — Done: first generated files and util.go.

Other changes

• Assignment/expression refactors internal — c013ed5b52 refactor of multiExpr/exprList, 70d836debb and c02e1bfbdb/8c5e8a38df refactors of assignVar, assignVars, initVars, 91a40f43b6 no mismatch errors when others exist, d93f02010c only mark variables used if they are, b1f1fb2950 spurious “declared and not used”, 969ab34e46 no panic on invalid comma-ok assignments, 05293d6b49 incorrect logic in assignability check.
• Cgo type names internal — 40bdc56180 resolve cgo base type names.
• Concurrent stdlib test internal — 0aedfc196b TestStdlib type-checks concurrently.
• Small fixes, unifier/assignment refactors, message and tracing tweaks internal — 55 commits: cf5dbd4459, 97f843b51f, 21b4e0146a, 6b8b7823c7, bd749504b8, cda461bb79, 07dca0fe14, 756a8ac91a, d871f63bcf, 3313880bdc, 331f0c6976, da47cd6192, a4d5fbc3a4, 5e6ccd1223, e7201cfa9f, e2f2123e25, ebc2aaae24, c80b9912e6, 8e19d2bbaf, acce3abb7e, bd7b19356f, bf9d9b7dba, abf9b112fd, a933d06271, 1742691270, f46ea60f2b, 9c75c4b6d7, f05c1941dd, 35ec948de7, 57678d5e03, 6a5f35047f, 522eace4ca, 73a4684caa, fb4a306e3a, c486f74eeb, 26899ef3ee, 05ca41d3ef, 0a9875c5c8, f3247f6110, 2532352ffb, f30cd52051, 945a2b17f3, 803a7dfb80, cecf8b9328, 8e3dfe783a, 44b8f3907c, 72f448cb48, 1dad7ef572, a965318ac0, f78483eaac, 0cd9064cd7, ea51e92c8a, f35d2dae74, ee361ce66c, d983be9cb5

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: user-visible reverse type inference enabled (1f570787a8, 767fbe01ae handles Info.Instances in assignment contexts), more lenient inference of untyped arguments (956d31ecd5, 4088e97fc2); instantiated generic methods recompiled in -linkshared mode (bcd82125f8); reproducible builds of aliased generic types (f58c6cccc4). Old nounified frontend and TTYPEPARAM/TUNION types removed (see unified IR thread).
• Unified IR: one serialization for noding, stenciling, inlining and export — Done: 4f467f1082 removes the nounified knob, de9bffb5f1 turns unreachable code into Fatalf, da9761303c removes the nounified frontend, a7de684e1b+f5eb0a7a5a stop creating/remove TUNION/TTYPEPARAM, 13c70b12d7 removes Markdcl/Pushdcl/Popdcl, 178080740c unused noding code; 7240d7e9e4 suppresses a union-type consistency check.
• Old typechecker and IR node-model cleanup — Done: constant arithmetic overflow checks removed from typecheck (fd208c8850), typecheck.EvalConst removed (25bc96dbd2), division-by-zero no longer reported there (60e6afb689), HasBreak removed (4c20c0a826), typecheck.Assignop stops calling base.Errorf (374bd6a3ac), Implements exported (13e9696437).
• Static initialization and package init — Done: hidden-field fix f773bef9ab; wrap global map init enabled (b46e44a399); inlstaticinit disabled again (b7736cbceb) then re-enabled (20c349e534) with fixes (abd55d8483 substituted arguments, 89138ce740 unhides closure funcs moved to staticinit, 3161081c12); 5b603f79fb drops special treatment of cover vars in initorder; 3494a72600 no DWARF for static vars.
• Compile-time performance, concurrency and the self-PGO’d compiler — Done: the compiler is built with PGO (ad6dcf9da4; reverted 298ff30958, reapplied 6111ac826e, reverted 6ffd135256, reapplied 45b74d20a0 after x/tools and ppc64 failures): Template -1.4%, Unicode -5.0%, Compiler -5.6%; SCC-reordering for inlining (f5c7416511).
• Compiler debugging infrastructure: bisect/hash debugging, flags, dumps — Done: HashDebug moves onto the new internal/bisect package (da5a3146ec), 134c9b2b17 (drop FS hash form), 1099644f52 (extra info in bisect output), 9dd0c7fc78, ef9c211de8 (match exclusion), 42866e566a, 81a31f5dd5 (godebug bisect). user-visible error URLs: 7a21f799a5 -url, e0db8ab6e0, enabled by default e7a9ca0a53 and reverted 42f387aae1; 2cbab4e98c passes types2 error codes through; 626d478d82 generates Markdown docs for each error code.
• Frontend support for new language features and spec changes — Done: user-visible clear (b89a840d65), min/max (b0f15b4ac0), //go:wasmimport (02411bcd7c, aff2f4a724), reverse type inference (1f570787a8), lenient untyped-arg inference (956d31ecd5), experimental loop-variable semantics GOEXPERIMENT=loopvar + -d=loopvar (c20d959163, 5a7793b7b8 flag on FOR/RANGE that preserves loop semantics across inlining), wasip1 (7f5af094f8).
• The syntax package (parser, scanner, tree) — Done: Crawl removed (3c357409d1), better = vs == message (db3f952b1f).

Other changes

• Runtime hash for known types internal — memhash and runtime hash functions (b1fd277318, fcff7226f3).
• Type eq/hash non-inlineable internal — 7fb075ddc0, later f123f5c768^1.22 (1.22 //go:noinline on generated eq/hash).
• Inlined-location preservation internal — full inlined position for locals (6dd20f4f12); ffc4cc05f5 outer-to-inner position lists.
• PGO call devirtualization user-visible — 8c445b7c9f (opt area).
• Race internal — no inlining from norace packages in race mode (87e69c1812).
• Unified IR: noder/writer/reader refinements and pkgbits (1.21) 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 (fa9efd9171)
• Old typechecker and IR node-model cleanup (1.21) 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 (6d4101ea68, 532e34dd38, f0b1563535, b94dc384ca, 617cf132ae, 171850f169, 89567a35c1, 47e2d7495f, 9fab498ab1)
• Syntax package: error messages and parser tolerance (1.21) internal — better syntax and go/parser error messages, accepted type-parameter lists, parser simplifications (afbcf2138c, 63f5596788)
• Housekeeping (1.21) internal — dead-code removal, ioutil/seek/slices/atomic-type replacements, README edits, debug-flag removal, assorted small bug fixes and doc comments (8ffcd3da93, 37a2004b43, 231806c83d, fc99c4b3f3, ed9744dff5, e539461e34, 1e5955aabd, 4bd77619b7, 9a0c506a4e, 70451004bd, 750e91152b, 761e813829, 6b2ad9ef50, 7c2cd0bbe2)

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
  • 8c445b7c9f adds PGO devirtualization of interface calls, based on Uber’s CL 484838. Follow-ups:
    • 3637132233 handles methods from other packages when the current package references the concrete type.
    • fd03e6ebc9 checks the method name.
    • 38e2376f35 adjusts the diagnostics.
  • Hot calls are no longer inlined into “big” functions. The PGO budget had bypassed the callee cap of 20 that applies to big functions (c1ac63e973).
  • Symbol matching escapes package paths (0fd6ae548f), and profile open and parse errors are reported (afbcf2138c).
  • Graph fixes: 7180a09745, 5e4e2acc8e, a4b66b14e2.
  • The pprof graph moved to pgo/internal/graph and dead code was deleted (d740b365b7, 7484dd30fd). Node weights left IRNode (4b6a542048), the post-inlining graph dump was removed (cce67690b8), and the “WORK IN PROGRESS” label was dropped (446ad131f6).
  • -d=pgoinline became -d=pgodebug (7c2cd0bbe2).
  • The compiler itself is now built with PGO (45b74d20a0).
  • Functions from norace packages are not inlined in race mode, an issue PGO exposed (87e69c1812).
  Left open: the 8c445b7c9f message lists its limitations: only interface calls, only the single hottest callee, and no lookup of callees missing from export data.
• Inliner: budget, closures, and the interleaved inline/devirtualize driver — Done:
  • -m=2 flags “large” functions (733ba92187), and isBigFunc recognizes unified IR codegen (6f6276ce36).
  • Operations inside SCCs are reordered so callees are analyzed before callers: +47 inlinable functions in the runtime (f5c7416511, followed by 6b2ad9ef50 and 63edd418b6).
  • user-visible Functions that construct closures became inlinable, +57/−10 inlinable functions in std (445e520d49). This landed after several revert rounds caused by closure-name and position issues.
  • Hidden closures that become unreferenced after inlining are dead-coded (f1caf1aa1c, reworked in ea69de9b92).
  Left open: nothing noted.
• Escape analysis: call modelling, closures, func-variable callees, and solver speed — Done: fixed memory corruption from escape analysis of unsafe.SliceData/StringData (198074abd7) and generic go/defer calls (93f10b8829). -m=2 output was fixed for recursive functions with closures (1e12c63aac). Left open: nothing noted.
• Allocation avoidance: stack-allocated slice backing stores and allocation-free literals — Done: an empty makeslice is rewritten to a zerobase pointer (fc814056aa, 0d8d181bd5). 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:
  • More bounds information from logic operators (9be533a8ee).
  • loopbce handles 8-, 16- and 32-bit induction variables (bb5ff5342d).
  • Unsigned division, modulus and right shift (35755d772f). OR relanded (63a08e61bd).
  • Slice-to-array-pointer casts are known non-nil (0d9eb8bea2), while a slice’s pointer is not assumed non-nil (61bc17f04e).
  Left open: nothing noted.
• Memory-op optimization: memcombine, dead stores, store-to-load forwarding, alias analysis, mem2reg — Done: the separate memcombine pass (cedf5008a8). It was made robust to reassociated expressions (6042a062dc). Left open: nothing noted.
• Lowering of builtins to memory ops: memmove/memclr/memequal inlining and generated equality — Done:
  • Type equality functions are generated during walk (864b54e2c3). The unused eqFor was removed (89c3f7c119).
  • Known-size memclrNoHeapPointers becomes OpZero (1fc585dc2f), with PPC64 support (ebe49f98c8). Negative-size memclr is not inlined (07559ceb72).
  • clear on slices with zero-size elements was fixed (0b65b02ba5).
  • memequal was removed from more string compares (6b165577fe).
  Left open: nothing noted.
• Switch lowering: jump tables, type-switch caches, and switch lookup tables — Done: the go119UseJumpTables flag was removed (4e3abee245). Better code for switches folded to a constant early (74b52d9519). Left open: nothing noted.
• Loop variable per-iteration semantics (loopvar) — Done: [user-visible, experiment]
  • GOEXPERIMENT=loopvar, -d=loopvar and -d=loopvarhash were added (c20d959163).
  • A flag on FOR/RANGE preserves the semantics across inlining (5a7793b7b8).
  • The return-in-loop optimization was restored correctly (23e8f43cb1).
  • “loop-transformed” logopt (dace96b9a1), logging tweaks (59d19ba797, 1524bd1c78), and bisect report improvements (b3d1cce3eb, 48f8c5c037).
  Left open: 5a7793b7b8 says the change would later be “triggered … by the specified version for compilation”.
• Sparse conditional constant propagation (sccp) — Done: first landing of the pass (fa50248ce6, Yi Yang, #59399). Left open: it was withdrawn and had to be resubmitted.
• Constant-folding loads from type metadata and dictionaries — Done: fold loads from constant dictionaries and types (635839a17a, bd3f44e4ff, #59591); ir.Func exposed to SSA (598cf5e6ac).

Other changes

• Misc internal — 386 ReverseBytes (d7ac5d1480), phiElimValue from removePhiArg (d29dd2ecf7), wasmimport type restrictions (53c0158905), PGO sample position fix (cdc65b83fe), callsite line for arg marshaling (c643b29381).
• min/max builtins user-visible — codegen, order-of-evaluation and mayCall fixes (a031f4ef83, 4ad4128d3c, 79d4defa75, 8b5fe5980c, 96d16803c2).
• Inliner: more inlining of functions that construct closures internal — three-part series (2ff684a541, 39986d28e4, f8162a0e72), closure naming incorporates inlined names (7f1467ff4d), hidden-closure dead marking reworked (d240226fe5); “Inliner” thread extension.
• Write barrier internal — reorganized and completion detected differently (6b5b7b3240, 44d22e75dd).
• Compiler allocations shared by shape internal — one allocator for similar-shape slices (0e42632301).
• Misc rules internal — De Morgan rewrite (95c4f320d5), arithmetic rules (da4687923b), zero-constant comparisons (85d54a7667), len(string([]byte)) without conversion (35a71dc56d), first instruction not inlined for FuncForPC (103f37497f, cbb9cd03f8), huge-function -N tightening (40a6e2dae5).

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: scheduling pass improved (forward scheduling) 12befc3ce3; register args scheduled before stack args b08d5ee5e9; values with no in-block uses scheduled later 6224db9b4d; SP scheduled early 40ed359182, InitMem after Args 518889b35c; tighten pass enhanced for memory values ea8f037996 (reverted aa6e168480, relanded e0ceba8139); heuristics for resultInArg0 overwrite decisions 3c7884a21b (reverted 2bf0f54bbd); “avoid registers” priority lowered 4237dea5e3; drop overwritten block input requirements 2c6aaaea64; anchored SP (SPanchored) so LEAQ of a variable schedules after its VARDEF f959fb3872; explicit boolean for “clobbers flags” mov marker 8354f6b5bb.
• amd64 code generation and GOAMD64 microarchitecture levels — Done: X15 + XORL for zerorange in ggen 326df693d7; shorter movzx for LoweredHASCPUFeature 1ddab59229; simplified signed division 2a41dbf13c,ee522e2cda; broken LEA “optimization” removed 6a97a60b4b; RDPID in the assembler 05cd6cbb98; frame pointer: obj/x86 uses PUSH/POP to save/restore BP (Windows toolchain −0.7%) f66581ead6, then MOV instead of LEA for FP setup adc1db23ee, with the implicit NOFRAME heuristic for frameless nosplit functions removed after every amd64 OS marked asm explicitly (d939b3717f…ea75b9458f, d20e688fcf).
• arm64 code generation and assembler — Done: redundant sign/zero extension omitted 4d180f71dc; address folding for globals on ARM64 except -dynlink b675a75c3d; CMP→CMN under <,>= 42f99b203d; assembler code cleanup series (8aa0175368…e7c7f33263).
• riscv64 compiler and assembler (RVA profiles, vector, compressed) — Done: loading large immediates with trailing zeros 481a6beba5; invalid shift amount check 5454834f01; FMA negation commutativity fix 3fce111535; regenerate encoding table 8e8303e1d0.
• loong64 backend and assembler — Done: c-shared 1f908bd060; save LR after SP decrement, remove branch delay slots (3ed8a1e629,b561ebab46); all unconditional jumps realized with B/BL (inherited BEQ ZERO,ZERO from MIPS dropped) 09f1ddb158; BEQ/BNE with 0 → beqz/bnez 22f9317f20; auto-align loop heads to 16 bytes via synthetic PCALIGN $16 (≤3 NOOPs) f736a9ad01, PCALIGN directive 47b22b6548; MUL/DIV rewrites 14015be5bb,96428e160d; PCALAU12I b4ac4b4b42.
• PPC64 / Power10 backend (GOPPC64, prefixed instructions, PC-relative code) — Done: prefixed MOV*/ADD for power10 72301a9863; SETBC/SETBCR rules (2.5% on ISEL-heavy loops, fewer registers) a432d89137; byte-reverse intrinsics for power10 cd1fc87156; NOTOC 29d5272da8; PCrel in linker ccb8db88c5; unbounded-shift rework f98dd29910; SETB 99f811ecca; xxspltidp (1.23) ebf7747dbe^1.23.
• Late-lower pass, memory-op combining and arch-independent rule migration — Done: memcombine pass cedf5008a8; generic rules cleanup: store rules use store type 60140a86b3.
• Prologue, frame pointer, stack split and function alignment (cmd/internal/obj) — Done: amd64 BP save via PUSH/POP f66581ead6, explicit NOFRAME on every OS + removal of heuristic (d939b3717f…d20e688fcf).
• Atomics and memory-ordering code generation — Done: MIPS64x And/Or 75add1ce0e.
• s390x, MIPS, 386 and 32-bit ARM backends — Done: 386 TrailingZeros intrinsics cedfcba3e8; s390x breakpoint fix 3e77efc720; Windows 386 TEB TLS slots 28f8dbd7b9; arm HAUTO addpool fdb3dc471d.

Other changes

• Write-barrier batching internal — barrier call returns a buffer pointer into which generated code records pointers (42 files) 21d82e6ac8; detection of write-barrier completion changed 44d22e75dd; uninterruptible range simplified (1.22) c94f39a80e^1.22,bf2a6d1957^1.22,727ebce6ce^1.22.
• SEH aux symbols for windows/amd64 76ac54b50e; flag init functions in object file 55bd193575.
• internal/abi consts merge 2668a190ba (reverted pieces 466e6dae95).
• ssa go:generate 47a0d46716; GetCallerSP memory arg 45dc81d856.
• reproducible aix/ppc64 builds ccf75b36ff.
• 1 small mips/s390x/arm/wasm commits internal — e.g. cc: support -buildmode=c-shared on linux/mips64{,le}. (d5fea5078b)
• 6 small obj/loong64 and loong64 asm commits internal — e.g. cmd/asm: add RDTIME{L,H}.W, RDTIME.D support for loong64; obj: remove redundant cnames on loong64; cc: improve logical shift with MASKEQZ on loong64. (bf8d142b4e,e29dd78ffe,83907206de,7789966286,2ef70d9d0f,c054f223e7)
• 17 small ssa/regalloc/liveness/obj generic commits internal — e.g. cc,cmd/link: set DW_AT_decl_line for function declarations; obj: drop Filesym from DwarfAbstractFunc; cc: improve register overwrite decision for resultInArg0 ops. (9e45b1d53e,a34cec8e41,4a0e84a1be,d29dd2ecf7,21f434058c,a42bb79dd4,40a6e2dae5,3360be4a11,4481042c43,d8d22664c6,b3bc8620f8,a3f3868c7a,9a0c506a4e,f5f7755da8,03ad1f1a34,4e8c6af239,10141676d1)
• 9 small obj/x86 and amd64/386 rules commits internal — e.g. runtime: use explicit NOFRAME on darwin/amd64; runtime: use explicit NOFRAME on linux/amd64; runtime: use explicit NOFRAME on dragonfly/amd64. (cea70301e2,521d2614df,516753278d,ddb423a7da,f76fc28c6b,0b41b678c2,1a5cf03b1b,b958d4a597,f9cf2c4d04)
• 14 small obj/arm64 and arm64 rules commits internal — e.g. obj/arm64: use obj.REG_NONE where appropriate; obj/arm64: invert SP check; obj/arm64: remove spurious assignments with OR. (8f3a13de9a,ecff9c9335,71c02bed67,f07fafacef,8b67cf0bc6,d389f779d4,0a092d7f80,d15fcbc79f,91a2e921dd,e81cc9119f,e7af0e0cac,d75cc4b9c6,6ed8474317,9fc84363d1)
• 5 small cmd/asm commits internal — e.g. cmd/asm: remove the incorrect check of LDADDx-like instructi; cmd/asm: reject avx512 .Z instructions without a mask regist; cmd/asm: preserve 3rd arg when GOT-rewriting assembly instru. (41e8679611,ee5ce77c62,1e0c32cccb,455168d113,cf624a6127)
• 3 small obj/ppc64 and PPC64 rules commits internal — e.g. obj/ppc64: add VC[LT]ZLSBB instructions; obj/ppc64: fix incorrect base reg causing segv; cc: on PPC64, fix sign/zero extension when masking. (7f4a54c0df,2716dfd995,5e4000ad7f)

Linker and object files

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

• DWARF quality for debuggers (types, variables, trampolines) — Done: DW_AT_decl_line for functions 9e45b1d53e; inlined DW_AT_call_line honours //line f60a2a9c94; DWARF section sizes checked separately c2becd70c5. Left open: nothing noted.
• Nosplit stack checking and stack constants — Done: StackLimit/StackGuard rationalized into abi.StackNosplitBase (landed, reverted d11ff3f081, relanded) 921699fe5f, 0f099a4bc5; stack check forced to 1 in dist b12366bbb5^1.20 (1.20). Left open: nothing noted.
• Dead code elimination and reflection-aware pruning — Done: user-visible linker side of dead-map removal e2ca417ee7; ReflectMethod flag in -dumpdep 0bbb54a02d. Left open: nothing noted.
• Linker performance and memory — Done: symbol index Sym becomes uint32 (−17.7% peak memory linking cmd/compile externally) d33a5136e1; outer-symbol slice and special/generator bitmaps (Symtab 12.9 to 6.4 ms) 21c2fdd91c; data symbols sorted by index not name (Dodata −4.7%) ddd822e5ca; type name decode without allocation 630ef2edc2; per-binary slice cleanup 8e8f4be3f4 Left open: nothing noted.
• Reproducible build IDs, buildinfo and symbol tables — Done: go.buildinfo kept with --gc-sections 54bdb8ee70; plugin build id 6c8ea391cf; reproducible toolchain build 8a27154bcd. Left open: nothing noted.
• ELF/relro/hardening and external-link policy — Done: only needed symbols exported in ELF external links 1f29f39795; external linking default for cgo std packages 06b67591e7 reverted 20da76da91, replaced by preferlinkext token for odd cgo flags 035db07d7c; c-archive consolidation 43115ff094. Left open: nothing noted.
• Per-platform object formats and internal linking (Mach-O, PE, XCOFF, wasm, ppc64, riscv64, loong64) — Done: PE .xdata 39ca989b88 and .pdata 14cf82aa37 (start of SEH unwinding) with SEH aux symbols 76ac54b50e; power10 PCrel and NOTOC ccb8db88c5, 29d5272da8; loong64 psABI v2 a3dd959229; wasip1 99de78e175; ARM trampoline fixes 0b5affb193, f26bf203ac; macho label symbols for duff 7dbd6de7d4; c-shared on loong64 1f908bd060. Left open: nothing noted.

Other changes

• Internal-linking additions internal — mips64 SDYNIMPORT (4a1829b65a), PPC64 PCrel trampolines and power10 relocs (a3df6c0e81, de788efeac, 27500d88f0, e946f0e1ff, 4cf79e479b, f742ddc349, 3c46d8f511), Windows ADDR32NB and DOS header (f57f02fcd5, e587a7691b), COFF complex types (d9c29ec6a5, b37c0602cd, 6e57ecdb03), loong64 padding/alignment (71f82df19a, 0d3c23fe87), etext padding (a6c382eaa8), ELFArch refactor (2ab0e04681), PAX_FLAGS dropped (ffb07d0c66), -fstack-protector libs (066b780502, c3d3be1066), dsymutil temp file workaround (a9a01ea280), builtin list update (be91040ea7).
• Package init order in linker user-visible — dependable ordering computed by the linker, then reverted (ce2a609909, 642542cb3c).
• Zero-sized data at runtime.zerobase internal (19fd96512c).