Extend the shared type model so SPIR-V's OpTypeBool and OpTypeArray (whose length
is a constant <id>, not a literal) lower cleanly:
- ir.Type_Kind gains BOOL (a distinct boolean; LLVM i1 will use it too).
- ir.Type gains len_ref: Id -- an ARRAY length carried as a constant <id>
(alongside the existing literal count for dialects with literal lengths).
- type_bool / type_array constructors.
SPIR-V codec: OpTypeBool <-> Type{.BOOL}; OpTypeArray <-> Type{.ARRAY, elem,
len_ref}. Test bool_and_array round-trips byte-exact -> 6 passed.
NOTE: a spec-valid module orders an array's length constant before the array
type; the codec round-trips the shape byte-exact regardless, but emitting the
types/constants section in dependency order is a follow-up.
The decoder already detects endianness from the magic word and byte-swaps each
word; strings are read from word values, so both endiannesses decode. Lock it in:
the round-trip helper now also byte-swaps every word to big-endian, decodes, and
asserts the re-encoding reproduces the original little-endian bytes -- across all
5 cases.
build.lua was ISA-centric; add an IRS catalog (ir/<name>) and route it through
the gen / check / test tasks. IR packages have a single-stage generator (gen.odin
emits Odin directly), a plain odin check, and a tests/ suite -- no mnemonic
builders or external-assembler verify, so only those three tasks apply. Purely
additive: the ISA catalog, task functions, and verify-tool probing are untouched.
luajit build.lua --check --test now reports spirv alongside the ISAs
(check ok, test 5 passed).
fmt_enum prints the common ValueEnum operands by name (Capability Shader,
Addressing/Memory model Logical GLSL450, ExecutionModel GLCompute, StorageClass,
Decoration, ...); BitEnums stay numeric (their bit_set %v form is too verbose to
read inline). The header comment now reads version/generator/bound from the
encoded words, so Bound reflects the computed value, not the pre-encode 0.
- OpFunctionParameter: the entry block's Block.params round-trip as
OpFunctionParameter (emitted between OpFunction and the entry OpLabel, decoded
back onto the entry block).
- encode computes bound (max <id> + 1) when the caller leaves it 0; a non-zero
bound (e.g. from decode) is preserved, so re-encode is stable.
- Trailing operands an enum value/bit pulls in (MemoryAccess Aligned's alignment,
...) are captured on decode as literals, so enum-parameter instructions
re-encode byte-exact.
Tests: param_function + load_aligned added -> 5 passed.
print(m, sb, options): disassemble a Module into a spirv-dis-style listing.
Encodes to a scratch buffer (doubling on overflow) and walks the word stream
generically through the operand-layout table, so every opcode disassembles with
no per-op code -- Id operands as %id, LiteralString quoted, the result-id column
right-aligned before OpName:
; SPIR-V
; Version: 1.5
; Bound: 9
%2 = OpTypeInt 32 1
%4 = OpConstant %2 10
%6 = OpIAdd %2 %4 %5
OpReturn
Enums print numerically for now (symbolic names a refinement). Completes the
encode / decode / print verb triad.
A tests/ package (the ISA test convention) that, per module shape, does
encode -> decode -> re-encode and asserts byte-identical output -- exercising the
encoder, decoder, <id> side tables, and the generic operation codec together:
void_main header + preamble + a void function (116 B)
int_constants i32 type + two scalar OpConstants incl. 0xDEADBEEF (88 B)
iadd_function %r = OpIAdd %i32 %a %b ; OpReturn -- generic operand codec
with a result + value operands (164 B)
3 passed. Run: odin run core/rexcode/ir/spirv/tests. (build.lua is ISA-centric;
wiring the ir/ packages into the gen/check/test pipeline is a follow-up.)
The inverse of the encoder: read the header (detecting endianness from the
magic word), walk the instruction stream, and lower each instruction by opcode
back into the structured Module -- the ir core (types/constants/globals/
functions, the generic table-driven operation decode), the SPIR-V sections
(preamble / debug / annotations), and the flat <id> space into the side id
tables (with an id->Type_Ref map so type-naming operands recover TYPE operands).
Single allocator pass (context.allocator), one reused per-instruction word
scratch buffer. Validated: encode -> decode -> re-encode is byte-exact on the
void compute main module (116 bytes; caps/types/functions/blocks/ops/ids/bound
all recovered). Same gaps as the encoder (OpFunctionParameter, ARRAY/bool,
enum-parameter operands); big-endian sources a later refinement.
The second half of encode(): the <id> side tables (Module.type_ids /
global_ids / function_ids -- SPIR-V's flat id space, which ir.Type/Global/
Function don't carry) plus the lowering:
emit_types ir.Type -> OpTypeXxx (void/int/float/vector/pointer/struct/
function; INT signedness + POINTER storage class ride in aux)
emit_constants OpConstant / OpConstantComposite / true/false/null
emit_globals OpVariable (storage class from the pointer type)
emit_operation generic table-driven op emit: INSTRUCTION_INDEX gives the
result-type/result-id prefix, the rest stream from op.operands
emit_functions OpFunction / OpLabel / body / OpFunctionEnd
Validated: a complete void compute main module encodes to byte-exact-correct
SPIR-V (29 words, all checked). Known gaps: OpFunctionParameter, ARRAY/bool
types, explicit enum-parameter operands, computed bound. Decoder next.
The Module -> word-stream encoder's first half: a fast single-pass word writer
(one host-endian store per word; instruction headers written as a placeholder
and backpatched, so variable-length instructions need no measure pass) plus the
header and the preamble / debug / annotation sections in spec layout order
(capabilities, extensions, ext-inst imports, memory model, entry points,
execution modes, OpString/OpSource/OpName, decorations).
Validated: a minimal module encodes to byte-exact-correct SPIR-V (magic, v1.5,
OpCapability Shader, OpMemoryModel Logical GLSL450, OpEntryPoint GLCompute %main,
NUL-terminated 'main' string, correct backpatched word counts).
Types/constants/globals/function bodies (the <id> + type lowering) are next.
Extend the tablegen to emit encoding_table.odin from the grammar -- the
per-opcode operand layout that drives the codec:
INSTRUCTION_SPECS flat []Operand_Spec {Spec_Kind, Quantifier}
INSTRUCTION_INDEX [max_opcode+1]Spec_Run -- O(1) opcode -> its operand run
ENUM_PARAMS enumerant -> trailing-parameter operands, so the codec
PARAM_SPECS reads MemoryAccess.Aligned / OpExecutionMode LocalSize /
Decoration.SpecId etc. correctly
873 opcodes, 3693 operand specs. Package compiles.
Start the SPIR-V intermediate representation under core/rexcode/ir/spirv,
following the ir API (docs/ir_design.md) and the ISA package conventions. SPIR-V
is table-driven on encoding and SSA on dataflow, so it re-exports the shared ir
vocabulary and is laid out like an arch package.
spirv.odin package + ir re-exports + physical format (MAGIC, version,
Header, wordCount<<16|opcode framing)
module.odin Module :: struct { using base: ir.Module, ...sections... } --
capabilities, ext-imports, entry points, exec modes, constant
pool, decorations, debug; SPIR-V's module metadata has no
ir.Module slot, so it is carried alongside the core
reloc.odin Relocation (linkage import/export)
opcodes.odin GENERATED: the Opcode enum (873 opcodes)
operand_kinds.odin GENERATED: ValueEnum/BitEnum operand kinds + Spec_Kind
tablegen/gen.odin the generator (core:encoding/json)
tablegen/spirv.core.grammar.json vendored authoritative grammar (Khronos
SPIRV-Headers unified1) -- single source of truth
Generating from the authoritative grammar avoids hand-transcription errors
(hand-authoring had Quads/Aliased/... wrong). Package compiles. Codec
(encode/decode + flat<->structured lowering), printer, and tests are next.
Move ENCODE_RECIPES from a startup-built (@init) static table in the library to
a generated, #loaded table like every other one. The tablegen runs
form_to_recipe over every form, serializes to x86.encode_recipes.bin, and the
loader #loads it -- no @init, no allocation; the forms remain the single source
of truth and the recipe is derived from them.
- tablegen/gen.odin: ENCODE_RECIPES added to the BLOBS manifest (which drives
both the Stage B serializer and the loader's #load); emit the generated-package
global plus an @init that fills it from ENCODE_FORMS via lib.form_to_recipe.
- encoder_recipe.odin: drop the @init, the static storage, and the ENCODE_RECIPES
declaration. form_to_recipe stays -- now solely the tablegen-time derivation.
- tables.odin, generated/{encode_tables,writer}.odin: regenerated.
- tables/x86.encode_recipes.bin: the generated blob (28260 bytes = 2355 x 12).
No behavior change: the recipes are identical to the @init ones. gen / builders /
check / test / idempotent green (2282 cases, 22 byte-stable artifacts); encode
perf unchanged.
The matcher path (generic builders, hand-built, decode->re-encode) resolves an
instruction to its encoding form by linearly scanning the forms for its mnemonic
and operand-matching each -- the dominant cost on that path. Memoize it: pack
(mnemonic, per-operand shape) into a key (immediates folded to the smallest size
class they fit, matching imm_matches_inline) and cache key -> form so a repeated
instruction shape skips the scan.
Direct-mapped, fixed 8192-slot table (64 KB, no allocation). Each slot packs the
full 48-bit key and form index into one u64, read/written with relaxed atomics,
so concurrent encode() stays safe -- a reader sees a matching key or rescans,
never a torn entry. The scan stays the source of truth (a miss runs it and
records the result), so the cache is exact.
Lookup + scan live in a non-inlined find_form() so they don't bloat encode()'s
hot loop and slow the hint path that shares it. (Routing the matcher path through
the recipe emit was tried and dropped: it costs the hint path ~1.2-1.5 ns however
isolated -- the hot loop is too codegen-sensitive -- while the cache alone is
free for the hint path.)
Realistic generic-builder mix: matcher ~52 -> ~35 ns/inst (~1.49x); hint path
unchanged. Byte-exact across 2282 + idempotent.
Extend emit_recipe to the full ModR/M + SIB + displacement addressing (register
direct, RIP-relative, absolute [disp32], and base/index/scale/disp), mirroring
the interpreter byte-for-byte, and drop the caller's reg-direct guard so memory
operands take the fast path too. Only a label/relative immediate (a relocation)
still falls back.
Realistic immediate-heavy mix: ~20.1 -> ~12.9 ns/inst vs the pre-recipe base
(~1.55x, 50 -> 77 M/s). Byte-exact across 2282 + idempotent.
Precompute each encoding form into a flat Form_Recipe -- prefix byte, escape+
opcode blob, role->operand-index slots, ext, imm size, flags -- so the encoder
replays common forms straight-line instead of re-interpreting enc.ops/enc.enc
on every instruction (the resolve scan, escape ladder, prefix/REX selection).
encode() takes the fast path when the form is hinted, eligible, has a register
r/m and a literal immediate; everything else falls through to the existing
interpreter, which stays the byte-exact source of truth. First cut:
- reg-direct ModR/M only (memory r/m falls back)
- hint path only (matcher / generic builders fall back)
- ~33% of forms eligible (VEX/EVEX, 16-bit operand-size, x87 fixed-ModR/M,
moffs/far/rel/implicit operands are marked ineligible)
Recipes are built at startup into static storage (no heap); this moves into the
table generator (#loaded like every other table) once the shape settles.
Realistic immediate-heavy mix: ~19.0 -> ~16.3 ns/inst (52.7 -> 61.3 M/s).
Byte-exact across 2282 cases + idempotent.
Next: memory r/m addressing in the fast path, then the matcher path, then the
gen-time port.
Roll the encode/decode buffer-sizing helpers (added for x86 in 49787b7de) out
to every other ISA, and document them in the cross-arch naming contract.
Per arch (arm32, arm64, mips, riscv, ppc, ppc_vle, rsp, mos6502, mos65816):
- encode_max_code_size / encode_max_relocation_count now key off the
[]Instruction slice (were int counts); bodies unchanged (* MAX_INST_SIZE).
- encode_reserve(code, relocs, instructions): grows the caller's code []u8 by
length and reserves relocs by capacity; allocates no new buffers.
- decode_max_instruction_count / decode_estimate_instruction_count: exact
ceiling and typical estimate, keyed off the min/avg instruction size per
arch (fixed-4: arm64/mips/ppc/rsp; min-2: arm32/riscv/ppc_vle; min-1: mos).
- decode_reserve(instructions, inst_info, label_defs, data, exact=false).
docs/cross_arch_design.md: helpers added to the naming contract.
No behavior change to the existing size helpers (signature only). All 10 ISAs
check + test green (x86 2282, arm32 600, arm64 461, mips 281, riscv 154, ppc 31,
ppc_vle 281, rsp 70, mos6502 148, mos65816 53).
Give callers a clean way to pre-size their own buffers so the encode/decode
hot paths never allocate or resize, instead of decode() silently reserving the
caller's arrays itself (removed). The library allocates nothing -- these only
grow the caller's own dynamic arrays, and only when not already big enough
(Odin's reserve no-ops when capacity already suffices).
Size-only helpers (caller manages its own memory), keyed off the input slice:
encode_max_code_size(instructions) - exact code bytes
encode_max_relocation_count(instructions) - exact reloc upper bound
decode_max_instruction_count(data) - exact ceiling (1 byte/inst)
decode_estimate_instruction_count(data) - typical estimate (~3 B/inst)
Reserve helpers (pre-size the caller's dynamic arrays; nil to skip an array):
encode_reserve(code, relocs, instructions)
code is a [dynamic]u8 grown by LENGTH (so code[:] is a valid emit
target); relocs reserved by capacity on top of existing elements.
decode_reserve(instructions, inst_info, label_defs, data, exact=false)
reserves capacity on top of existing; exact=true for the ceiling.
Error arrays grow only on the failure path, so they are intentionally not
covered. check/test green; 2282 cases; exercised end-to-end (the [dynamic]u8
code pattern, factor-in-existing, nil args, exact ceiling, reserve no-op).
Data-oriented pass on the encode hot path. Profiling showed bounds checks
already elided by -o:speed; the cost was per-instruction loop/scan machinery
and immediates falling off the hint path.
- Gather the immediate slot in the single resolve pass and emit it straight-
line (no scan over enc.enc); likewise drive the legacy REX prefix from the
precomputed reg/mr/opr slots instead of a per-form scan.
- Fold the separate needs_66 (GPR16) and SPL/BPL/SIL/DIL operand loops into
the resolve pass, so user operands are visited exactly once. This was the
big one: mov r,r 27 -> 21 ns.
- Gate the whole legacy-prefix block on a single flags!=0 test (a legacy
prefix is almost always absent) instead of four branches per instruction.
- Make immediate forms hintable. A typed immediate builder names its width
(inst_add_r32_imm32), the matcher already keys off the operand's declared
size, so baking the form is byte-identical AND drops immediates from the
full match scan: mov r32,imm32 55.7 -> 17.8 ns (3.1x).
Floor (no-op) 14.55 -> 10.3 ns; realistic immediate-heavy typed mix
30 -> 20.5 ns/inst (~49 M inst/s). gen/builders/check/test/idempotent green;
2282 cases (typed==generic byte-identical, incl. the new immediate cases).
Targeted branchless revert + the pre-matched form fast path, and a fix
for a pre-existing bug the latter surfaced.
(a) Revert the two speculative-write spots from the prior branchless pass
(legacy-prefix emission, widened displacement store, ENCODE_TAIL_SLACK)
back to predicted branches. In real streams a legacy prefix is almost
always absent and disp size is stable, so those branches are ~free and
the unconditional stores only added work. Every class got faster
(RET 19->17.5, MOV r,r 52->46.6, VADDPS 42.8->39.3 ns).
(b) Pre-matched form hint. Instruction.enc_hint (in the existing 11-byte
padding, idx+1 biased; 0 = matcher path) lets a typed builder that maps
to a single value-independent form bake the global form index, so
encode() skips the O(forms) match scan -- and, in a varied stream, its
unpredictable branches. Generated for non-immediate forms only (value-
dependent imm8/imm32 selection stays on the matcher). On a 100k mixed
typed-builder stream: 47.3 -> 30.2 ns/inst (-36%), byte-identical to the
matcher path -- ~2x the original baseline for codegen.
Repair the typed inst_/emit_ builders. They were non-functional: the
generator cast the hw-only typed enum straight to Register
(Register(GPR64.RAX) -> class 0), so every typed-builder operand was
rejected by the matcher (encode returned empty). Untested because the
suite builds via the generic constructors. Now they build through the
class-correct op_gpr64/op_xmm/... path (op_* already used by 3+ operand
builders), emit_ reuses inst_, and a new 30-case consistency suite
asserts typed == generic (llvm-verified) and hint == matcher.
gen/builders/check/test/idempotent all green; 2276 cases.
Three layers on the x86 encode/decode hot paths, all byte-exact (2246
LLVM-verified cases) and roundtrip-clean:
1. Branchless: legacy-prefix emission (speculative write + conditional
advance), REX/VEX/EVEX extension-bit accumulation (gate-and-mask),
ModRM mod/disp-size selection (cmov selects), displacement emission
(widened store + ENCODE_TAIL_SLACK); decoder REX/VEX/EVEX register
extensions (arithmetic instead of if/+=8).
2. Resolve-operands-once: the previous code re-derived each user operand
~5-10x per instruction (a fresh O(n) scan of enc.ops per emission
pass). Now resolved into a [4]^Operand map a single time.
3. Single-pass gather: fold the opcode-+rb and ModR/M slot-detection
scans into that one resolve pass (3 enc.enc passes -> 1).
Net on a 100k mixed-instruction benchmark: encode ~58 -> ~54 ns/inst
(best 52). Branchless alone was a ~7% encode regression (predicted
branches, nothing to recover); the algorithmic passes recovered it and
beat baseline.
A sibling to core:rexcode/isa for the intermediate representations (WASM,
SPIR-V, LLVM bitcode + the LLVM dialects AIR/DXIL). Holds the shared
vocabulary every IR package builds on, implements no specific IR.
Design stance (see docs/ir_design.md): keep the ISA layer's spirit, but
where IRs are structurally MORE uniform than ISAs (SSA + a type system
regularize the operand/module shape), the shared core is richer. ir/ owns:
status.odin Error/Error_Code (shape-identical to isa.Error)
refs.odin Id/Ref/Ref_Space/Symbol_Table (the label analog: structural
id references, not PC-relative byte offsets)
types.odin Type/Type_Ref/Type_Kind (the type table -- no ISA analog)
module.odin Module/Function/Block/Operation/Operand/Result/Dataflow
(the structured model; Operation = isa.Instruction + an
optional typed Result, opcode a u16 like Mnemonic)
print.odin token kinds + options + num-fmt (parallels isa.print)
Three honest concessions vs the ISA API, made explicit not inert: a
structured Module replaces the flat []Instruction; a first-class type
system; id-based entity refs replace labels. The encode/decode verbs take
a Module and drop label_defs/resolve/base_address. Dataflow hosts both the
WASM value stack and SSA; the codec is pluggable (table for WASM/SPIR-V,
bitstream for the LLVM family -- AIR/DXIL are LLVM dialects, not peers).
Package compiles; a hand-built SSA module round-trips through the types.
Move all ten ISA packages (x86, arm32, arm64, mips, riscv, ppc, ppc_vle,
rsp, mos6502, mos65816) from core/rexcode/<arch> to core/rexcode/isa/<arch>,
so the import pattern is now `import "core:rexcode/isa/x86"`. The shared
core stays at core:rexcode/isa.
Mechanical: relative `import "../isa"` / "../../isa" -> absolute
"core:rexcode/isa" (the only path that survives the move; the "../" and
"../.." self/generated imports move with their packages). build.lua now
builds paths as <root>/isa/<name>; stale `cd <arch>` hints in the verify
tools and the doc.odin paths updated.
WASM stays at core/rexcode/wasm for now -- it is an IR, not an ISA, and
will move under the forthcoming core:rexcode/ir once that layer lands.
All 10 arches gen/builders/check/test green; import core:rexcode/isa/x86
verified working; wasm still compiles.
