core::num::f16b Rust's 16bit Brain Float - #160859
Open
Jamesbarford wants to merge 17 commits into
Open
Conversation
Collaborator
|
Some changes occurred in compiler/rustc_attr_ir cc @jdonszelmann, @JonathanBrouwer This PR changes rustc_public cc @oli-obk, @celinval, @ouz-a, @makai410
|
rustbot
added
A-attributes
rustbot
added
T-compiler
Collaborator
|
r? @jieyouxu rustbot has assigned @jieyouxu. Use Why was this reviewer chosen?The reviewer was selected based on:
|
This comment has been minimized.
This comment has been minimized.
Member
|
@rustbot reroll |
antoyo
reviewed
Aug 10, 2026
| } | ||
|
|
||
| fn type_f16b(&self) -> Type<'gcc> { | ||
| bug!("f16b is not supported by the GCC codegen backend") |
Contributor
There was a problem hiding this comment.
I believe GCC actually supports this type: https://github.com/rust-lang/gccjit.rs/blob/master/src/context.rs#L1482
Contributor
Author
There was a problem hiding this comment.
Thanks 😄, I will aim to add it in a follow up PR 👍
Jamesbarford
force-pushed
the
feat/fb16-pt1
branch
from
August 10, 2026 14:48
bf10f8a to
01c5c1b
Compare
Collaborator
|
cc @bjorn3 |
bjorn3
reviewed
Aug 10, 2026
bjorn3
reviewed
Aug 10, 2026
Member
There was a problem hiding this comment.
What is the calling convention of other targets?
Contributor
There was a problem hiding this comment.
I believe only these files were changed because they have an exhaustive match on Float.
However, my version of abi-cafe found two interesting failures: GCC and Clang are inconsistent on aarch64 and armv7
// callee, compiled with GCC 12
#include <inttypes.h>
#include <string.h>
#include <stdio.h>
#include <stdbool.h>
typedef struct Many1 {
__bf16 f0;
} Many1;
void struct_in_1(Many1 arg0) {
printf("%d", arg0.f0);
}// caller, compiled with clang 23
#include <inttypes.h>
#include <string.h>
#include <stdio.h>
#include <stdbool.h>
typedef struct Many1 {
__bf16 f0;
} Many1;
void struct_in_1(Many1 arg0);
void do_test(void) {
{
Many1 arg0 = { .f0 = (((union { uint16_t bits; __bf16 value; }){ .bits = 49600 }).value) };
printf("%d", arg0.f0);
struct_in_1(arg0);
}
}hits
func struct_in_1's values differed
values (native-endian hex bytes):
expect: C0 C1
caller: C0 C1
callee: 04 00
the value was arg0.f0: rustarithmeticty(f16b)
whose arg was arg0: Many1
The current Rust implementation matches clang, and is hence incompatible with GCC
armv7 with hardware floats also runs into incompatibilities
[target.armv7-unknown-linux-gnueabihf."f16::conv_c::repr_c::clang-nightly_calls_distro-gcc"]
busted = "check"
[target.armv7-unknown-linux-gnueabihf."f16::conv_c::repr_c::distro-gcc_calls_clang-nightly"]
busted = "check"
[target.armv7-unknown-linux-gnueabihf."f16b::conv_c::repr_c::clang-nightly_calls_distro-gcc"]
busted = "check"
[target.armv7-unknown-linux-gnueabihf."f16b::conv_c::repr_c::distro-gcc_calls_clang-nightly"]
busted = "check"
Finally, you can let this ICE on many targets, e.g. mips, powerpc, s390x, sparc
Contributor
There was a problem hiding this comment.
You can also (e.g. on loongarch64 https://godbolt.org/z/Y3hdhG4e6) emit a __truncsfbf2 libcall that is not provided (probably needs to be added to compiler-builtins).
I think the ICEs are probably a blocker? That needs a mechanism similar to has_reliable_f128.
Contributor
Author
There was a problem hiding this comment.
Yes I added it because of the exhaustive match statement in mips64.rs and sparc64.rs I've removed it; 40d3f6e and put in a panic!(...).
With regard to has_reliable_f128, are you envisaging a has_reliable_f16b entry on TargetConfig?
Contributor
There was a problem hiding this comment.
With regard to
has_reliable_f128, are you envisaging ahas_reliable_f16bentry onTargetConfig?
Exactly
Contributor
There was a problem hiding this comment.
What's your take on those ABI mismatches? We should track that somewhere.
Contributor
Author
There was a problem hiding this comment.
I'm not particularly certain what mips64 should do.
It doesn't, at least to my knowledge, have hardware support. Given we aren't implementing scalar arithmetic and an f16b can only be created through a bit pattern or vendor intrinsics, I can't immediately see a practical application? Hence a panic! seems like a pragmatic choice for the time being.
This comment has been minimized.
This comment has been minimized.
folkertdev
reviewed
Aug 11, 2026
Contributor
There was a problem hiding this comment.
I believe only these files were changed because they have an exhaustive match on Float.
However, my version of abi-cafe found two interesting failures: GCC and Clang are inconsistent on aarch64 and armv7
// callee, compiled with GCC 12
#include <inttypes.h>
#include <string.h>
#include <stdio.h>
#include <stdbool.h>
typedef struct Many1 {
__bf16 f0;
} Many1;
void struct_in_1(Many1 arg0) {
printf("%d", arg0.f0);
}// caller, compiled with clang 23
#include <inttypes.h>
#include <string.h>
#include <stdio.h>
#include <stdbool.h>
typedef struct Many1 {
__bf16 f0;
} Many1;
void struct_in_1(Many1 arg0);
void do_test(void) {
{
Many1 arg0 = { .f0 = (((union { uint16_t bits; __bf16 value; }){ .bits = 49600 }).value) };
printf("%d", arg0.f0);
struct_in_1(arg0);
}
}hits
func struct_in_1's values differed
values (native-endian hex bytes):
expect: C0 C1
caller: C0 C1
callee: 04 00
the value was arg0.f0: rustarithmeticty(f16b)
whose arg was arg0: Many1
The current Rust implementation matches clang, and is hence incompatible with GCC
armv7 with hardware floats also runs into incompatibilities
[target.armv7-unknown-linux-gnueabihf."f16::conv_c::repr_c::clang-nightly_calls_distro-gcc"]
busted = "check"
[target.armv7-unknown-linux-gnueabihf."f16::conv_c::repr_c::distro-gcc_calls_clang-nightly"]
busted = "check"
[target.armv7-unknown-linux-gnueabihf."f16b::conv_c::repr_c::clang-nightly_calls_distro-gcc"]
busted = "check"
[target.armv7-unknown-linux-gnueabihf."f16b::conv_c::repr_c::distro-gcc_calls_clang-nightly"]
busted = "check"
Finally, you can let this ICE on many targets, e.g. mips, powerpc, s390x, sparc
| }, | ||
| Primitive::Float(float) => match float { | ||
| Float::F16 | Float::F32 => "f32", | ||
| Float::F16 | Float::F16B | Float::F32 => "f32", |
Contributor
There was a problem hiding this comment.
is that right? LLVM just crashes on bf16 right now, so it's probably at least untested?
Contributor
Author
There was a problem hiding this comment.
I could be mistaken, however I don't think WASM supports bf16? I've made it panic!(...) for now; 40d3f6e
tgross35
reviewed
Aug 11, 2026
Jamesbarford
force-pushed
the
feat/fb16-pt1
branch
from
August 12, 2026 08:19
53d3660 to
40d3f6e
Compare
This comment has been minimized.
This comment has been minimized.
rustbot
added
the
A-test-infra-minicore
Jamesbarford
force-pushed
the
feat/fb16-pt1
branch
from
August 12, 2026 10:08
da620d0 to
0ed984f
Compare
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
Collaborator
|
Some changes occurred in cfg and check-cfg configuration cc @Urgau
cc @rust-lang/miri |
This comment has been minimized.
This comment has been minimized.
Member
|
I'm not a good reviewer for this change. Can somebody here pick it up rather than blindly rerolling? |
Contributor
|
r? me |
This comment has been minimized.
This comment has been minimized.
folkertdev
reviewed
Aug 14, 2026
Contributor
There was a problem hiding this comment.
What's your take on those ABI mismatches? We should track that somewhere.
| // This is similar to <https://github.com/llvm/llvm-project/issues/94434>, however | ||
| // does not work until LLVM 23 on Windows. | ||
| (Arch::Arm64EC, _) => major >= 23, | ||
| (Arch::AArch64, _) | (Arch::X86_64, _) | (Arch::RiscV64, _) => major >= 21, |
Contributor
There was a problem hiding this comment.
based on https://github.com/loongson/la-abi-specs/blob/release/lapcs.adoc loongarch also has __bf16.
Contributor
Author
There was a problem hiding this comment.
62a7f72 Added loongarch64 along with a test, of which I'd hesitate to say the output is correct. I verified it by comparing the output in the ./build... folder to see if it bore a resemblance to what C produced for something similar; https://godbolt.org/z/o17e1e5bc
Comment on lines
+1
to
+2
| //@ edition: 2021 | ||
| //@ only-aarch64 |
Contributor
There was a problem hiding this comment.
this probably needs -O too
Contributor
There was a problem hiding this comment.
can you match tests/ui/feature-gates/feature-gate-f16.rs more closely here?
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
RalfJung
reviewed
Aug 16, 2026
| F16 => tcx.types.f16, | ||
| F16B => Ty::new_adt( | ||
| tcx, | ||
| tcx.adt_def(tcx.require_lang_item(LangItem::BFloat, DUMMY_SP)), |
Member
There was a problem hiding this comment.
This is... interesting.^^
Why does this need both a lang item and a new variant in abi::Float?
Is the key thing that we do not have a new variant in ty::FloatTy?
Contributor
There was a problem hiding this comment.
Because we don't (at present) want to make it a true builtin (like f16), hence the lang item, but bfloat is used in LLVM intrinsic signatures so we want it to make it all the way to the backend.
This is different from Complex which has custom ABI requirements but does not occur in any LLVM signatures.
RalfJung
reviewed
Aug 16, 2026
| #[cfg_attr(feature = "nightly", derive(StableHash))] | ||
| pub enum Float { | ||
| F16, | ||
| F16B, |
Member
There was a problem hiding this comment.
Suggested change
| F16B, | |
| /// `f16b`. This is not a builtin type in Rust (it is exposed as a lang item), | |
| /// but it is a builtin type in LLVM so needs to be explicitly represented | |
| /// in the backend. | |
| F16B, |
… and backend handling. GCC and Cranelift explicitly unsupported
…long with traits agreed on in the RFC
…6b` to be treated as a primitive scalar
- Add documentation aliases to `f16b` - Add Wikipedia link to `bfloat16` - Remove `@ only-x86_64` flag from test
- ICE on `sparc64`, `mips64` & WASM
Jamesbarford
force-pushed
the
feat/fb16-pt1
branch
from
August 17, 2026 13:59
4ff6c4b to
fc04ded
Compare
This comment has been minimized.
This comment has been minimized.
Jamesbarford
force-pushed
the
feat/fb16-pt1
branch
from
August 17, 2026 14:50
fc04ded to
6754cda
Compare
This comment has been minimized.
This comment has been minimized.
Jamesbarford
force-pushed
the
feat/fb16-pt1
branch
from
August 17, 2026 17:24
6754cda to
b5496fb
Compare
Collaborator
|
The job Click to see the possible cause of the failure (guessed by this bot) |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
10 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
View all comments
Implements the RFC: f16b type. Best reviewed commit by commit, happy to split into separate PRs if that is deemed easier to review. However the line count and surface area is, in my opinion, reasonably small.
Adds;
f16balong withbfloatlang item to work with LLVM, GCC is explicitlyunimplemented!(...)f16bfeature gate, page forf16bon libruscdoc and astruct bf16incore::numf16bas a scalar primitive for scalable vectorsIssues;