feat: add `Float32` primitive type with arithmetics and conversions to `Float` by ggreif · Pull Request #5906 · caffeinelabs/motoko

ggreif · 2026-03-12T11:42:39Z

Implement some Float32 functionality (basic arithmetics). This is experimental for now and somewhat starved in terms of supporting functions in the prelude (though basic arithmetic is there, everything else must pass via Float if missing, e.g. primitives like sin). But it should be efficient memory consumption-wise.

Can receive and send float32 by Candid messages. This was the main motivator for the feature. Stable variables of Float32 type work too.

TODO:

safeFloatToFloat32 : Float -> Epsilon -> ?Float32
verify that runtime tagging works too — see 0abf97e + follow-ups (OBSOLETE, supplanted by f32 scalars on 64-bit backend)
.most process working?
(re)move plans (opportunities inside, please consult the last commit(s) about the unboxed Float32 idea, and the unboxed Float – analogously on 64-bit backend!)

Documents all files to change for parsing, type-checking, IR lowering, value representation, Candid binding, and prim.mo additions. No codegen changes in scope. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- type.ml/mli: add Float32 prim variant (tag 19), string mapping, span, subtype - arrange_type.ml, typ_hash.ml: handle Float32 in exhaustive matches - numerics.ml/mli: Float32 module backed by Wasm.F64 - value.ml/mli: Float32 value variant + as_float32 accessor - show.ml: display Float32 values; coverage.ml: Float32 → Any - prim.ml: Float↔Float32 conversions via num_conv_trap_prim - mo_to_idl.ml: Float32 → Candid float32 - idl_to_mo.ml: Candid float32 → Motoko Float32 (was UnsupportedCandidFeature) - prelude.mo: type Float32 = prim "Float32" (global scope) - prim.mo: public type + floatToFloat32/float32ToFloat primitives - test/run/float32.mo: type-check + IR interpreter tests pass Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- prim.ml: floatToFloat32 now truncates through a real 32-bit float (Int32.float_of_bits ∘ Int32.bits_of_float) to purge excess f64 precision - error_codes.ml: retire M0161 as DEFUNCT (was 'Candid float32 cannot be imported'), consistent with other defunct codes - test/run/float32.mo: add precision assertions — verify 0.1 round-trip differs from original and that two f32 round-trips are idempotent Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…freedom "f32" violated the prefix-free invariant since "f" (Float) is a prefix of "f32". Use "h" (half the width of Float/f64) instead — no existing hash starts with "h". Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Prim module now exports Float32 type and floatToFloat32/float32ToFloat, so all snapshots that dump the full prim scope need updating. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add Float32->Text prim, @text_of_Float32 in internals, and ir_passes/show.ml case; test demonstrates '1.5f' output. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- to_idl_prim: Float32 -> Some 13l (Candid float32) - SR.of_type / stackrep_of_type: Float32 -> UnboxedFloat64 - ReadBuf.read_float32: load f32, promote to f64, advance 4 bytes - Serialization: demote f64->f32, store 4 bytes - Deserialization: read_float32 + Float.box - NumConvTrapPrim Float->Float32: demote+promote round-trip - NumConvTrapPrim Float32->Float: identity (already UnboxedFloat64) - @text_of_Float32: implemented via @text_of_Float + "f" suffix (avoids a Float32->Text OtherPrim in both codegen backends) - Remove //SKIP comp from float32.mo; accept wasm-run output Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

When UnboxedFloat32 lands, @text_of_Float32 will need to handle an f32 on the stack directly; delegating via num_conv_Float32_Float would still work but is awkward. Keep Float32->Text as a proper OtherPrim in both backends (currently reads UnboxedFloat64, appends "f" suffix), ready for the TODO upgrade. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Consistent with Nat8 showing as bare '42', not '42n8'. The 'f' suffix belongs in AST/IR dumps, not debug_show. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add `Float32Lit of Numerics.Float32.t` to the AST (syntax.ml) and IR (ir.ml), analogous to `Nat8Lit`, so that `(3.14 : Float32)` is properly coerced (demote f64→f32 at compile time) rather than desugaring through floatToFloat32. Also fixes `buffer_size` (4 bytes) and `compile_eq` for Float32 in both backends, enabling `to_candid`/`from_candid` roundtrips. Adds drun test `idl-float32.mo` with echo, literal, and Float32→Float64 queries. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Add Float32 to sign_unop, num_binop, num_relop in operator.ml so that ==, !=, <, >, <=, >= and arithmetic work on Float32 in the interpreter; arithmetic results are demoted to f32 precision - Extract the f32-truncation into Numerics.Float32.demote (replaces the inline Int32.(float_of_bits (bits_of_float ...)) in prim.ml and typing.ml) - Fixes 4 remaining nix build .#tests.candid failures (Float32 == failing with M0060) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Tests Candid serialization of Float32 literals (3.14, 1.5) and verifies that a Float64 blob yields null when decoded as Float32. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…and add Changelog entry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…pported) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

github-actions · 2026-03-12T16:31:27Z

Comparing from 88aa95d to 0f0717c:
In terms of gas, 1 tests improved and the mean change is -0.0%.
In terms of size, 1 tests improved and the mean change is -0.0%.

crusso · 2026-03-12T17:06:24Z

Good use of AI to do something tedious!

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> test: switch float32.mo to -dl to show Float32Lit in IR dump Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> test: filter float32 IR dump to Float32Lit lines only Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> test: remove stale float32.tc.ok

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ggreif · 2026-03-13T11:27:48Z

Good use of AI to do something tedious!

If you have a moment available, can you verify that 0abf97e (+revs) is sane?

Replace per-pattern T.(...) local opens with a single let open T in at the top of check_lit. Save the local sub wrapper as is_sub before the open to avoid shadowing by T.sub (which has a different signature). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…just Add Const.Float32 to the Const.lit type (mirroring compile_classical.ml), emit it directly from Float32Lit, and handle it cleanly in build_constant_aux and StackRep.adjust — replacing the two ad-hoc cases (Vanilla extraction and Float64 demotion) with a single Const.Float32 f, UnboxedFloat32 case. Also fix check_lit in typing.ml: save shadowed `error` as `error'` before `let open T in` to avoid T.error (a typ value) taking over. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…T in Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Following the pattern of compile_comparison_f64, extract a compile_comparison_f32 helper for Float32 relops in both compile_enhanced.ml and compile_classical.ml. Also add compile_comparison_f64 helper to compile_classical.ml for symmetry. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add `let open Type in` after `let open Operator in` in compile_relop in both backends, removing repeated `Type.(Prim ...)` patterns. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

crusso · 2026-03-27T12:07:06Z

  | Const.Lit (Const.Float64 number) -> Float.constant env number
+  | Const.Lit (Const.Float32 f) ->
+    E.Vanilla Int64.(logor
+      (shift_left (logand (of_int32 (Wasm.F32.to_bits f)) 0xFFFFFFFFL) 32)


Isn't the logand redundant?

Absolutely! I even let Claude correct such an instance in the past :-)

Indeed redundant — shift_left _ 32 shifts the sign-extended upper 32 bits completely out. Removed.

crusso · 2026-03-27T12:08:59Z

+    (* Never-heap-boxed types (UnboxedWord64 / UnboxedFloat32) have bit 0 = 0
+       in their Vanilla encoding, so Opt.inject is a no-op — emit directly. *)


I guess you can delete the comment or move to Opt.inject

- Drop redundant `logand 0xFFFFFFFFL` in Float32 constant builder: `shift_left _ 32` already discards the sign-extended upper bits - Remove now-stale OptPrim comment (explained Opt.inject internals) - Delete .claude/plans/ — all implemented, outdated or deferred Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

crusso

LGTM apart from three comments

crusso

Fire away (but maybe improve the PR description and add a Changelog entry)

Context: caffeinelabs/motoko#5906.

…#5947) (Cherrypicked reconstruction of #5939 after base was squash merged) Builds on the Opt.inject optimization in #5906, but extends it to more payload types. Implements #5938. * less code * more perf same correctness? Saves 10%(!) of instructions on bench/alloc (creating a large linked list, so not surprising) Background on our option representation: https://www.joachim-breitner.de/blog/787-A_mostly_allocation-free_optional_type - [x] Consider promoting, not just normalizing, to exploit bounds that are better than Any. - [x] Tests... Many iterators should get faster too. - [ ] Should we load_forwarding_pointer in fast path or not. If so, need to special case (or slow-path) bools coz of true literal looks like a pointer but can't be forwarded. For some reason, the existing Opt.simple_inject (a no-op) always forwards (but is only every used on reference types). - [x] Optimize Opt.project too. --------- Co-authored-by: Cycle and memory benchmark updater <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Gabor Greif <gabor.greif@caffeine.ai>

## Summary - Adds `e(Float32) = float32` and `i(float32) = Float32` to the type mapping tables. - Removes the note that said importing `float32` was unsupported/failing. ## Why `Float32` support was shipped in caffeinelabs#5906, which updated both `idl_to_mo.ml` and `mo_to_idl.ml` to handle the `float32 ↔ Float32` mapping. The spec doc (`design/IDL-Motoko.md`) was not updated at the time, leaving it contradicting the implementation. This is a companion to a fix in [dfinity/candid](https://github.com/dfinity/candid/pull/new/fix/motoko-binding-float32) where `didc bind --target mo` panics on any `.did` file containing `float32`, because it (correctly) follows the now-outdated spec. Made with [Cursor](https://cursor.com) Co-authored-by: Cursor <cursoragent@cursor.com>

…ding (#728) ## Summary - Replace `panic!("float32 not supported in Motoko")` with `str("Float32")` in the Motoko binding generator. - Add a `float.did` test fixture covering `float32`/`float64` round-trip methods. ## Why Motoko added `Float32` in version 1.4.0 ([caffeinelabs/motoko#5906](caffeinelabs/motoko#5906)), updating both the import and export IDL mappings. The spec (`design/IDL-Motoko.md`) is being updated in a companion PR. This brings `didc bind --target mo` in line with what `moc` already does when importing `.did` files directly. Concretely, any `.did` file using `float32` (e.g. the IC management canister interface) would previously cause `didc bind --target mo` to panic at runtime. --------- Co-authored-by: Cursor <cursoragent@cursor.com>

ggreif and others added 5 commits March 12, 2026 12:03

plan: add Float32 primitive type (frontend only)

9355990

Documents all files to change for parsing, type-checking, IR lowering, value representation, Candid binding, and prim.mo additions. No codegen changes in scope. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: typo 'construtors' → 'constructors' in typ_hash.ml

9804cbd

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ggreif mentioned this pull request Mar 12, 2026

chore: update ic.did from portal repo dfinity/sdk#4494

Merged

ggreif and others added 17 commits March 12, 2026 12:52

test: update fail-test snapshots for Float32 additions

d561be6

Prim module now exports Float32 type and floatToFloat32/float32ToFloat, so all snapshots that dump the full prim scope need updating. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

style: use Int32.(...) local open in floatToFloat32 conversion

e843b1d

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

chore: note next free tag number (20) after Float32=19

65463d6

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat: show Float32 values with 'f' suffix to distinguish from Float

23c07b5

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat: implement debug_show for Float32 with 'f' suffix

b28b1ff

Add Float32->Text prim, @text_of_Float32 in internals, and ir_passes/show.ml case; test demonstrates '1.5f' output. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

plan: Float32 backend (Candid serialization, NumConvTrapPrim codegen)

cf75548

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

chore: TODO note to switch Float32 to UnboxedFloat32 for compactness

e3eee5f

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: remove 'f' suffix from Float32 show/debug_show output

c47a30d

Consistent with Nat8 showing as bare '42', not '42n8'. The 'f' suffix belongs in AST/IR dumps, not debug_show. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

test: Float32 to_candid/from_candid round-trip in test/run

abcb8cc

Tests Candid serialization of Float32 literals (3.14, 1.5) and verifies that a Float64 blob yields null when decoded as Float32. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

test: fix float32-candid (import debugPrint, accept wasm-run output) …

0281bba

…and add Changelog entry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

chore: remove defunct M0161 from known_untested_codes (Float32 now su…

07b0e2b

…pported) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

test: accept blob-bounds with new @text_of_Float32 prelude entry

5253b1f

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: add Float32 to prim_js in astjs.ml (moc_js build)

89caf3a

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ggreif commented Mar 13, 2026

View reviewed changes

Comment thread src/codegen/compile_enhanced.ml Outdated

ggreif and others added 2 commits March 13, 2026 12:23

style: use M.(local open) syntax in idl_to_mo.ml and T.(…) in typing.ml

847d0af

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs: plan for Float32 RTTI tag (Bits64 H, tag 47)

0abf97e

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ggreif and others added 5 commits March 27, 2026 11:29

style: combine sub', error' rebindings into one line before let open …

5fdff41

…T in Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

refactor: open Type in compile_relop to remove Type.( repetitions

af686cd

Add `let open Type in` after `let open Operator in` in compile_relop in both backends, removing repeated `Type.(Prim ...)` patterns. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ggreif added feature New feature or request opportunity More optimisation opportunities inside labels Mar 27, 2026

crusso reviewed Mar 27, 2026

View reviewed changes

Comment thread src/codegen/compile_enhanced.ml

crusso reviewed Mar 27, 2026

View reviewed changes

Comment thread src/prelude/prelude.mo

Comment thread test/bench/ok/palindrome.drun-run.ok

Comment thread test/bench/ok/palindrome.drun-run.ok

crusso approved these changes Mar 27, 2026

View reviewed changes

ggreif enabled auto-merge March 27, 2026 12:29

ggreif added this pull request to the merge queue Mar 27, 2026

Merged via the queue into master with commit c096bb4 Mar 27, 2026
30 checks passed

ggreif deleted the claude/float32 branch March 27, 2026 12:57

rvanasa added a commit to caffeinelabs/vscode-motoko that referenced this pull request Mar 27, 2026

Add Float32 type keyword (#438)

a2d55d3

Context: caffeinelabs/motoko#5906.

rvanasa added a commit to caffeinelabs/node-motoko that referenced this pull request Mar 27, 2026

Add Float32 type keyword (#163)

22a7809

Context: caffeinelabs/motoko#5906.

crusso restored the claude/float32 branch March 27, 2026 17:17

crusso mentioned this pull request Mar 27, 2026

perf: statically optimize no-op option injections and projections FTW #5947

Merged

4 tasks

This was referenced Mar 27, 2026

Float32: ModOp (%) not implemented in codegen backend #5948

Closed

Float32: ModOp (%) not implemented in codegen backend #5949

Closed

This was referenced May 6, 2026

spec: align IDL-Motoko.md with Float32 support #6093

Closed

spec: align IDL-Motoko.md with Float32 support #6094

Merged

fix(candid_parser): emit Float32 for Candid float32 in Motoko binding dfinity/candid#728

Merged

		(* Never-heap-boxed types (UnboxedWord64 / UnboxedFloat32) have bit 0 = 0
		in their Vanilla encoding, so Opt.inject is a no-op — emit directly. *)

Conversation

ggreif commented Mar 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Mar 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

crusso commented Mar 12, 2026

Uh oh!

Uh oh!

ggreif commented Mar 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

crusso Mar 27, 2026

Choose a reason for hiding this comment

Uh oh!

ggreif Mar 27, 2026

Choose a reason for hiding this comment

Uh oh!

ggreif Mar 27, 2026

Choose a reason for hiding this comment

Uh oh!

crusso Mar 27, 2026

Choose a reason for hiding this comment

Uh oh!

ggreif Mar 27, 2026

Choose a reason for hiding this comment

Uh oh!

crusso left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

crusso left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ggreif commented Mar 12, 2026 •

edited

Loading

github-actions Bot commented Mar 12, 2026 •

edited

Loading

ggreif commented Mar 13, 2026 •

edited

Loading

crusso left a comment •

edited

Loading