Skip to content

Make min/max and copysign intrinsics generic #162414

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
N1ark wants to merge 7 commits into
base: main
Choose a base branch
from
Open

Make min/max and copysign intrinsics generic #162414

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
0dcb734
cranelift: use fallbacks for abs f16, f128
N1ark Sep 6, 2026
7b86261
Make `copysign` generic
N1ark Sep 6, 2026
78c2b96
Make minimum/maximum generic
N1ark Sep 6, 2026
a707a46
Add `rustc_allow_const_fn_unstable` to `fabs`
N1ark Sep 7, 2026
c63c387
Replace float bound impls with a macro
N1ark Sep 7, 2026
a4edfff
De-constify the `FloatPrimitive` sealed trait
N1ark Sep 7, 2026
101eb51
Add test
N1ark Sep 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 0 additions & 14 deletions compiler/rustc_codegen_cranelift/src/codegen_f16_f128.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -160,20 +160,6 @@ pub(crate) fn neg_f128(fx: &mut FunctionCx<'_, '_, '_>, value: Value) -> Value {
fx.bcx.ins().bitcast(types::F128, MemFlagsData::new(), bits)
}

pub(crate) fn abs_f16(fx: &mut FunctionCx<'_, '_, '_>, value: Value) -> Value {
let bits = fx.bcx.ins().bitcast(types::I16, MemFlagsData::new(), value);
let bits = fx.bcx.ins().band_imm_u(bits, 0x7fff);
fx.bcx.ins().bitcast(types::F16, MemFlagsData::new(), bits)
}

pub(crate) fn abs_f128(fx: &mut FunctionCx<'_, '_, '_>, value: Value) -> Value {
let bits = fx.bcx.ins().bitcast(types::I128, MemFlagsData::new(), value);
let (low, high) = fx.bcx.ins().isplit(bits);
let high = fx.bcx.ins().band_imm_u(high, 0x7fff_ffff_ffff_ffff_u64 as i64);
let bits = fx.bcx.ins().iconcat(low, high);
fx.bcx.ins().bitcast(types::F128, MemFlagsData::new(), bits)
}

pub(crate) fn codegen_cast(
fx: &mut FunctionCx<'_, '_, '_>,
from: Value,
Expand Down
245 changes: 73 additions & 172 deletions compiler/rustc_codegen_cranelift/src/intrinsics/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -349,11 +349,6 @@ fn codegen_float_intrinsic_call<'tcx>(
sym::fmuladdf64 => ("fma", 3, fx.tcx.types.f64, types::F64),
sym::fmuladdf128 => return false, // has a fallback

sym::copysignf16 => return false, // has a fallback
sym::copysignf32 => ("copysignf", 2, fx.tcx.types.f32, types::F32),
sym::copysignf64 => ("copysign", 2, fx.tcx.types.f64, types::F64),
sym::copysignf128 => return false, // has a fallback

sym::floorf16 => return false, // has a fallback via f32
sym::floorf32 => ("floorf", 1, fx.tcx.types.f32, types::F32),
sym::floorf64 => ("floor", 1, fx.tcx.types.f64, types::F64),
Expand Down Expand Up @@ -417,7 +412,6 @@ fn codegen_float_intrinsic_call<'tcx>(
sym::fmaf32 | sym::fmaf64 | sym::fmuladdf32 | sym::fmuladdf64 => {
fx.bcx.ins().fma(args[0], args[1], args[2])
}
sym::copysignf32 | sym::copysignf64 => fx.bcx.ins().fcopysign(args[0], args[1]),
sym::floorf32 | sym::floorf64 => fx.bcx.ins().floor(args[0]),
sym::ceilf32 | sym::ceilf64 => fx.bcx.ins().ceil(args[0]),
sym::truncf32 | sym::truncf64 => fx.bcx.ins().trunc(args[0]),
Expand Down Expand Up @@ -455,6 +449,13 @@ fn codegen_float_intrinsic_call<'tcx>(
true
}

/// Used to distinguish fallbacks of float intrinsics. For some we have a codegen fallback,
/// while for others we fallback to an external libcall.
enum IntrinsicFallback {
Fallback(&'static str),

@RalfJung RalfJung Sep 7, 2026

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So this does not refer to a Rust fallback body? Please pick a different name then, this is confusing. Maybe Libcall or so?

Or, given that this is pre-existing, at least please add a comment.

@beetrees beetrees Sep 7, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think all cases which use IntrinsicFallback::Fallback should be able to be removed and just use the Rust fallback body? In which case adding a FIXME here to remove it in a future PR seems like a good idea.

Codegen(Value),
}

fn codegen_regular_intrinsic_call<'tcx>(
fx: &mut FunctionCx<'_, '_, 'tcx>,
instance: Instance<'tcx>,
Expand Down Expand Up @@ -1163,6 +1164,66 @@ fn codegen_regular_intrinsic_call<'tcx>(
ret.write_cvalue(fx, old);
}

sym::copysign
| sym::minimum
| sym::maximum
| sym::minimum_number_nsz
| sym::maximum_number_nsz => {
intrinsic_args!(fx, args => (arg1, arg2); intrinsic);
let layout = arg1.layout();
let ty::Float(float_ty) = layout.ty.kind() else {
span_bug!(
source_info.span,
"expected float type for fabs intrinsic: {:?}",
layout.ty
);
};
use FloatTy::*;
use IntrinsicFallback::*;
let x = arg1.load_scalar(fx);
let y = arg2.load_scalar(fx);
let res = match (intrinsic, float_ty) {
(sym::copysign, F32 | F64) => Codegen(fx.bcx.ins().fcopysign(x, y)),

(sym::minimum, F32 | F64) => Codegen(fx.bcx.ins().fmin(x, y)),
(sym::maximum, F32 | F64) => Codegen(fx.bcx.ins().fmax(x, y)),
// FIXME(bytecodealliance/wasmtime#8312): Use `fmin`/`fmax` directly for `f16` and
// `f128` once the lowerings have been implemented in Cranelift.
(sym::minimum, F128) => Codegen(codegen_f16_f128::fmin_f128(fx, x, y)),
(sym::maximum, F128) => Codegen(codegen_f16_f128::fmax_f128(fx, x, y)),
(sym::minimum, F16) => {
Codegen(codegen_f16_f128::maybe_with_f16_to_f32_pair(fx, x, y, |fx, x, y| {
fx.bcx.ins().fmin(x, y)
}))
}
(sym::maximum, F16) => {
Codegen(codegen_f16_f128::maybe_with_f16_to_f32_pair(fx, x, y, |fx, x, y| {
fx.bcx.ins().fmax(x, y)
}))
}

(sym::minimum_number_nsz, _) => Codegen(crate::num::codegen_float_min(fx, x, y)),
(sym::maximum_number_nsz, _) => Codegen(crate::num::codegen_float_max(fx, x, y)),

(sym::copysign, F128) | (_, F16) => {
// We use the intrinsic fallback bodies for the rest
return Err(Instance::new_raw(instance.def_id(), instance.args));
}

_ => unreachable!(),
};
let val = match res {
Codegen(val) => val,
Fallback(name) => {
let ty = fx.clif_type(layout.ty).unwrap();
let arg = AbiParam::new(ty);
fx.lib_call(name, vec![arg, arg], vec![arg], &[x, y])[0]
}
};

@bjorn3 bjorn3 Sep 9, 2026 •
edited by rustbot
Loading

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The fallback variant is never used here.

View changes since the review

let val = CValue::by_val(val, layout);
ret.write_cvalue(fx, val);
}

sym::fabs
| sym::exp
| sym::exp2
Expand All @@ -1176,24 +1237,16 @@ fn codegen_regular_intrinsic_call<'tcx>(
let ty::Float(float_ty) = layout.ty.kind() else {
span_bug!(
source_info.span,
"expected float type for fabs intrinsic: {:?}",
"expected float type for {:?} intrinsic: {:?}",
intrinsic,
layout.ty
);
};
enum IntrinsicFallback {
Fallback(&'static str),
Codegen(Value),
}
use FloatTy::*;
use IntrinsicFallback::*;
let x = arg.load_scalar(fx);
let res = match (intrinsic, float_ty) {
(sym::fabs, F32 | F64) => Codegen(fx.bcx.ins().fabs(x)),
// FIXME(bytecodealliance/wasmtime#8312): Use `fabsf16` once Cranelift
// backend lowerings are implemented.
(sym::fabs, F16) => Codegen(codegen_f16_f128::abs_f16(fx, x)),
(sym::fabs, F128) => Codegen(codegen_f16_f128::abs_f128(fx, x)),

(sym::exp, F32) => Fallback("expf"),
(sym::exp, F64) => Fallback("exp"),
(sym::exp, F128) => Fallback("expf128"),
Expand Down Expand Up @@ -1222,8 +1275,10 @@ fn codegen_regular_intrinsic_call<'tcx>(
(sym::cos, F64) => Fallback("cos"),
(sym::cos, F128) => Fallback("cosf128"),

(_, F16) => {
// We implement fallbacks for other f16 intrinsics via f32
(sym::fabs, F128) | (_, F16) => {
// FIXME(bytecodealliance/wasmtime#8312): Use the native operations once
// Cranelift backend lowerings for `f16` are implemented.
// We use the intrinsic fallback bodies for the rest
return Err(Instance::new_raw(instance.def_id(), instance.args));
}

Expand All @@ -1240,160 +1295,6 @@ fn codegen_regular_intrinsic_call<'tcx>(
ret.write_cvalue(fx, val);
}

sym::minimumf16 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = codegen_f16_f128::maybe_with_f16_to_f32_pair(fx, a, b, |fx, a, b| {
fx.bcx.ins().fmin(a, b)
});
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f16));
ret.write_cvalue(fx, val);
}
sym::minimumf32 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = fx.bcx.ins().fmin(a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
ret.write_cvalue(fx, val);
}
sym::minimumf64 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = fx.bcx.ins().fmin(a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
ret.write_cvalue(fx, val);
}
sym::minimumf128 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

// FIXME(bytecodealliance/wasmtime#8312): Use `fmin` once Cranelift
// backend lowerings are implemented.
let val = codegen_f16_f128::fmin_f128(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f128));
ret.write_cvalue(fx, val);
}
sym::maximumf16 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = codegen_f16_f128::maybe_with_f16_to_f32_pair(fx, a, b, |fx, a, b| {
fx.bcx.ins().fmax(a, b)
});
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f16));
ret.write_cvalue(fx, val);
}
sym::maximumf32 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = fx.bcx.ins().fmax(a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
ret.write_cvalue(fx, val);
}
sym::maximumf64 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = fx.bcx.ins().fmax(a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
ret.write_cvalue(fx, val);
}
sym::maximumf128 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

// FIXME(bytecodealliance/wasmtime#8312): Use `fmax` once Cranelift
// backend lowerings are implemented.
let val = codegen_f16_f128::fmax_f128(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f128));
ret.write_cvalue(fx, val);
}

sym::minimum_number_nsz_f16 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_min(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f16));
ret.write_cvalue(fx, val);
}
sym::minimum_number_nsz_f32 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_min(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
ret.write_cvalue(fx, val);
}
sym::minimum_number_nsz_f64 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_min(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
ret.write_cvalue(fx, val);
}
sym::minimum_number_nsz_f128 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_min(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f128));
ret.write_cvalue(fx, val);
}
sym::maximum_number_nsz_f16 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_max(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f16));
ret.write_cvalue(fx, val);
}
sym::maximum_number_nsz_f32 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_max(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
ret.write_cvalue(fx, val);
}
sym::maximum_number_nsz_f64 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_max(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
ret.write_cvalue(fx, val);
}
sym::maximum_number_nsz_f128 => {
intrinsic_args!(fx, args => (a, b); intrinsic);
let a = a.load_scalar(fx);
let b = b.load_scalar(fx);

let val = crate::num::codegen_float_max(fx, a, b);
let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f128));
ret.write_cvalue(fx, val);
}

sym::catch_unwind => {
let ret_block = fx.get_block(destination.unwrap());

Expand Down
Loading
Loading