Skip to content
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

x86-32 float return for 'Rust' ABI: treat all float types consistently #131871

Merged
merged 2 commits into from
Oct 23, 2024
Merged

x86-32 float return for 'Rust' ABI: treat all float types consistently #131871

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
57d5f86
x86-32 float return for 'Rust' ABI: treat all float types consistently
RalfJung Oct 18, 2024
0906819
adjust asm test
RalfJung Oct 22, 2024
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
37 changes: 15 additions & 22 deletions compiler/rustc_ty_utils/src/abi.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,5 @@
use std::iter;

use rustc_abi::Float::*;
use rustc_abi::Primitive::{Float, Pointer};
use rustc_abi::{Abi, AddressSpace, PointerKind, Scalar, Size};
use rustc_hir as hir;
Expand Down Expand Up @@ -695,37 +694,31 @@ fn fn_abi_adjust_for_abi<'tcx>(
}

// Avoid returning floats in x87 registers on x86 as loading and storing from x87
// registers will quiet signalling NaNs.
// registers will quiet signalling NaNs. Also avoid using SSE registers since they
// are not always available (depending on target features).
if tcx.sess.target.arch == "x86"
&& arg_idx.is_none()
// Intrinsics themselves are not actual "real" functions, so theres no need to
// change their ABIs.
&& abi != SpecAbi::RustIntrinsic
{
match arg.layout.abi {
// Handle similar to the way arguments with an `Abi::Aggregate` abi are handled
// below, by returning arguments up to the size of a pointer (32 bits on x86)
// cast to an appropriately sized integer.
Abi::Scalar(s) if s.primitive() == Float(F32) => {
// Same size as a pointer, return in a register.
arg.cast_to(Reg::i32());
return;
let has_float = match arg.layout.abi {
Abi::Scalar(s) => matches!(s.primitive(), Float(_)),
Abi::ScalarPair(s1, s2) => {
matches!(s1.primitive(), Float(_)) || matches!(s2.primitive(), Float(_))
Comment on lines +706 to +708
Copy link
Contributor

Choose a reason for hiding this comment

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

Suggested change
Abi::Scalar(s) => matches!(s.primitive(), Float(_)),
Abi::ScalarPair(s1, s2) => {
matches!(s1.primitive(), Float(_)) || matches!(s2.primitive(), Float(_))
Abi::Scalar(s) => matches!(s.primitive(), Float(F16 | F32 | F64)),
Abi::ScalarPair(s1, s2) => {
matches!(s1.primitive(), Float(F16 | F32 | F64)) || matches!(s2.primitive(), Float(F16 | F32 | F64))

nit: There's no need to include f128 here as the 32-bit x86 ABI already guarantees it gets passed in memory.

Copy link
Member Author

@RalfJung RalfJung Oct 18, 2024 •
edited
Loading

Choose a reason for hiding this comment

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

IMO we should treat all float types uniformly here, that's more future-proof. The result is the same both ways but my code makes it more clear what happens, relying on fewer external assumptions.

We don't care about being consistent with the C ABI here, after all.

Copy link
Member

Choose a reason for hiding this comment

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

Yeah, I kinda agree morally with beetrees but not actually: this seems fine as-is.

I don't think it's worth spending too much time on this. While I do need to be able to read this code to refactor it, I do intend to bulldoze it.

}
Abi::Scalar(s) if s.primitive() == Float(F64) => {
// Larger than a pointer, return indirectly.
arg.make_indirect();
return;
}
Abi::ScalarPair(s1, s2)
if matches!(s1.primitive(), Float(F32 | F64))
|| matches!(s2.primitive(), Float(F32 | F64)) =>
{
_ => false, // anyway not passed via registers on x86
};
if has_float {
if arg.layout.size <= Pointer(AddressSpace::DATA).size(cx) {
// Same size or smaller than pointer, return in a register.
arg.cast_to(Reg { kind: RegKind::Integer, size: arg.layout.size });
} else {
// Larger than a pointer, return indirectly.
arg.make_indirect();
return;
}
_ => {}
};
return;
}
}

if arg_idx.is_none() && arg.layout.size > Pointer(AddressSpace::DATA).size(cx) * 2 {
Expand Down
6 changes: 4 additions & 2 deletions tests/codegen/float/f128.rs
View file
Open in desktop
Copy link
Member

Choose a reason for hiding this comment

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

this test feels like it is getting a bit silly, but otherwise seems fine.

Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
// 32-bit x86 returns `f32` and `f64` differently to avoid the x87 stack.
// 32-bit x86 returns float types differently to avoid the x87 stack.
// 32-bit systems will return 128bit values using a return area pointer.
//@ revisions: x86 bit32 bit64
//@[x86] only-x86
Expand Down Expand Up @@ -152,7 +152,9 @@ pub fn f128_rem_assign(a: &mut f128, b: f128) {

/* float to float conversions */

// CHECK-LABEL: half @f128_as_f16(
// x86-LABEL: i16 @f128_as_f16(
// bits32-LABEL: half @f128_as_f16(
// bits64-LABEL: half @f128_as_f16(
#[no_mangle]
pub fn f128_as_f16(a: f128) -> f16 {
// CHECK: fptrunc fp128 %{{.+}} to half
Expand Down
49 changes: 27 additions & 22 deletions tests/codegen/float/f16.rs
View file
Open in desktop
Copy link
Member

Choose a reason for hiding this comment

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

seems fine

Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
// 32-bit x86 returns `f32` and `f64` differently to avoid the x87 stack.
// 32-bit x86 returns float types differently to avoid the x87 stack.
// 32-bit systems will return 128bit values using a return area pointer.
//@ revisions: x86 bit32 bit64
//@[x86] only-x86
Expand Down Expand Up @@ -58,42 +58,44 @@ pub fn f16_le(a: f16, b: f16) -> bool {
a <= b
}

// CHECK-LABEL: half @f16_neg(
// This is where we check the argument and return ABI for f16.
// other-LABEL: half @f16_neg(half
// x86-LABEL: i16 @f16_neg(half
#[no_mangle]
pub fn f16_neg(a: f16) -> f16 {
// CHECK: fneg half %{{.+}}
-a
}

// CHECK-LABEL: half @f16_add(
// CHECK-LABEL: @f16_add
#[no_mangle]
pub fn f16_add(a: f16, b: f16) -> f16 {
// CHECK: fadd half %{{.+}}, %{{.+}}
a + b
}

// CHECK-LABEL: half @f16_sub(
// CHECK-LABEL: @f16_sub
#[no_mangle]
pub fn f16_sub(a: f16, b: f16) -> f16 {
// CHECK: fsub half %{{.+}}, %{{.+}}
a - b
}

// CHECK-LABEL: half @f16_mul(
// CHECK-LABEL: @f16_mul
#[no_mangle]
pub fn f16_mul(a: f16, b: f16) -> f16 {
// CHECK: fmul half %{{.+}}, %{{.+}}
a * b
}

// CHECK-LABEL: half @f16_div(
// CHECK-LABEL: @f16_div
#[no_mangle]
pub fn f16_div(a: f16, b: f16) -> f16 {
// CHECK: fdiv half %{{.+}}, %{{.+}}
a / b
}

// CHECK-LABEL: half @f16_rem(
// CHECK-LABEL: @f16_rem
#[no_mangle]
pub fn f16_rem(a: f16, b: f16) -> f16 {
// CHECK: frem half %{{.+}}, %{{.+}}
Expand Down Expand Up @@ -142,10 +144,13 @@ pub fn f16_rem_assign(a: &mut f16, b: f16) {

/* float to float conversions */

// CHECK-LABEL: half @f16_as_self(
// other-LABEL: half @f16_as_self(
// x86-LABEL: i16 @f16_as_self(
#[no_mangle]
pub fn f16_as_self(a: f16) -> f16 {
// CHECK: ret half %{{.+}}
// other-CHECK: ret half %{{.+}}
// x86-CHECK: bitcast half
// x86-CHECK: ret i16
a as f16
}

Expand Down Expand Up @@ -176,21 +181,21 @@ pub fn f16_as_f128(a: f16) -> f128 {
a as f128
}

// CHECK-LABEL: half @f32_as_f16(
// CHECK-LABEL: @f32_as_f16
#[no_mangle]
pub fn f32_as_f16(a: f32) -> f16 {
// CHECK: fptrunc float %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @f64_as_f16(
// CHECK-LABEL: @f64_as_f16
#[no_mangle]
pub fn f64_as_f16(a: f64) -> f16 {
// CHECK: fptrunc double %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @f128_as_f16(
// CHECK-LABEL: @f128_as_f16
#[no_mangle]
pub fn f128_as_f16(a: f128) -> f16 {
// CHECK: fptrunc fp128 %{{.+}} to half
Expand Down Expand Up @@ -273,70 +278,70 @@ pub fn f16_as_i128(a: f16) -> i128 {

/* int to float conversions */

// CHECK-LABEL: half @u8_as_f16(
// CHECK-LABEL: @u8_as_f16
#[no_mangle]
pub fn u8_as_f16(a: u8) -> f16 {
// CHECK: uitofp i8 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @u16_as_f16(
// CHECK-LABEL: @u16_as_f16
#[no_mangle]
pub fn u16_as_f16(a: u16) -> f16 {
// CHECK: uitofp i16 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @u32_as_f16(
// CHECK-LABEL: @u32_as_f16
#[no_mangle]
pub fn u32_as_f16(a: u32) -> f16 {
// CHECK: uitofp i32 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @u64_as_f16(
// CHECK-LABEL: @u64_as_f16
#[no_mangle]
pub fn u64_as_f16(a: u64) -> f16 {
// CHECK: uitofp i64 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @u128_as_f16(
// CHECK-LABEL: @u128_as_f16
#[no_mangle]
pub fn u128_as_f16(a: u128) -> f16 {
// CHECK: uitofp i128 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @i8_as_f16(
// CHECK-LABEL: @i8_as_f16
#[no_mangle]
pub fn i8_as_f16(a: i8) -> f16 {
// CHECK: sitofp i8 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @i16_as_f16(
// CHECK-LABEL: @i16_as_f16
#[no_mangle]
pub fn i16_as_f16(a: i16) -> f16 {
// CHECK: sitofp i16 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @i32_as_f16(
// CHECK-LABEL: @i32_as_f16
#[no_mangle]
pub fn i32_as_f16(a: i32) -> f16 {
// CHECK: sitofp i32 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @i64_as_f16(
// CHECK-LABEL: @i64_as_f16
#[no_mangle]
pub fn i64_as_f16(a: i64) -> f16 {
// CHECK: sitofp i64 %{{.+}} to half
a as f16
}

// CHECK-LABEL: half @i128_as_f16(
// CHECK-LABEL: @i128_as_f16
#[no_mangle]
pub fn i128_as_f16(a: i128) -> f16 {
// CHECK: sitofp i128 %{{.+}} to half
Expand Down
Loading