force_array -> is_consecutive

The actual ABI implication here is that in some cases the values
are required to be "consecutive", i.e. must either all be passed
in registers or all on stack (without padding).

Adjust the code to either use Uniform::new() or Uniform::consecutive()
depending on which behavior is needed.

Then, when lowering this in LLVM, skip the [1 x i128] to i128
simplification if is_consecutive is set. i128 is the only case
I'm aware of where this is problematic right now. If we find
other cases, we can extend this (either based on target information
or possibly just by not simplifying for is_consecutive entirely).
This commit is contained in:
Nikita Popov 2024-03-21 16:10:23 +01:00
parent 009280c5e3
commit 1b7342b411
13 changed files with 53 additions and 57 deletions

View File

@ -150,7 +150,10 @@ impl LlvmType for CastTarget {
// Simplify to a single unit or an array if there's no prefix.
// This produces the same layout, but using a simpler type.
if self.prefix.iter().all(|x| x.is_none()) {
if rest_count == 1 && !self.rest.force_array {
// We can't do this if is_consecutive is set and the unit would get
// split on the target. Currently, this is only relevant for i128
// registers.
if rest_count == 1 && (!self.rest.is_consecutive || self.rest.unit != Reg::i128()) {
return rest_ll_unit;
}

View File

@ -31,7 +31,7 @@ where
RegKind::Vector => size.bits() == 64 || size.bits() == 128,
};
valid_unit.then_some(Uniform { unit, total: size, force_array: false })
valid_unit.then_some(Uniform::consecutive(unit, size))
})
}
@ -60,7 +60,7 @@ where
let size = ret.layout.size;
let bits = size.bits();
if bits <= 128 {
ret.cast_to(Uniform { unit: Reg::i64(), total: size, force_array: false });
ret.cast_to(Uniform::new(Reg::i64(), size));
return;
}
ret.make_indirect();
@ -100,9 +100,9 @@ where
};
if size.bits() <= 128 {
if align.bits() == 128 {
arg.cast_to(Uniform { unit: Reg::i128(), total: size, force_array: false });
arg.cast_to(Uniform::new(Reg::i128(), size));
} else {
arg.cast_to(Uniform { unit: Reg::i64(), total: size, force_array: false });
arg.cast_to(Uniform::new(Reg::i64(), size));
}
return;
}

View File

@ -21,7 +21,7 @@ where
RegKind::Vector => size.bits() == 64 || size.bits() == 128,
};
valid_unit.then_some(Uniform { unit, total: size, force_array: false })
valid_unit.then_some(Uniform::consecutive(unit, size))
})
}
@ -49,7 +49,7 @@ where
let size = ret.layout.size;
let bits = size.bits();
if bits <= 32 {
ret.cast_to(Uniform { unit: Reg::i32(), total: size, force_array: false });
ret.cast_to(Uniform::new(Reg::i32(), size));
return;
}
ret.make_indirect();
@ -78,11 +78,7 @@ where
let align = arg.layout.align.abi.bytes();
let total = arg.layout.size;
arg.cast_to(Uniform {
unit: if align <= 4 { Reg::i32() } else { Reg::i64() },
total,
force_array: false,
});
arg.cast_to(Uniform::consecutive(if align <= 4 { Reg::i32() } else { Reg::i64() }, total));
}
pub fn compute_abi_info<'a, Ty, C>(cx: &C, fn_abi: &mut FnAbi<'a, Ty>)

View File

@ -18,7 +18,7 @@ fn classify_ret<Ty>(arg: &mut ArgAbi<'_, Ty>) {
if total.bits() > 64 {
arg.make_indirect();
} else if total.bits() > 32 {
arg.cast_to(Uniform { unit: Reg::i32(), total, force_array: false });
arg.cast_to(Uniform::new(Reg::i32(), total));
} else {
arg.cast_to(Reg::i32());
}
@ -38,7 +38,7 @@ fn classify_arg<Ty>(arg: &mut ArgAbi<'_, Ty>) {
if arg.layout.is_aggregate() {
let total = arg.layout.size;
if total.bits() > 32 {
arg.cast_to(Uniform { unit: Reg::i32(), total, force_array: false });
arg.cast_to(Uniform::new(Reg::i32(), total));
} else {
arg.cast_to(Reg::i32());
}

View File

@ -195,11 +195,7 @@ where
if total.bits() <= xlen {
arg.cast_to(xlen_reg);
} else {
arg.cast_to(Uniform {
unit: xlen_reg,
total: Size::from_bits(xlen * 2),
force_array: false,
});
arg.cast_to(Uniform::new(xlen_reg, Size::from_bits(xlen * 2)));
}
return false;
}
@ -282,11 +278,10 @@ fn classify_arg<'a, Ty, C>(
if total.bits() > xlen {
let align_regs = align > xlen;
if is_loongarch_aggregate(arg) {
arg.cast_to(Uniform {
unit: if align_regs { double_xlen_reg } else { xlen_reg },
total: Size::from_bits(xlen * 2),
force_array: false,
});
arg.cast_to(Uniform::new(
if align_regs { double_xlen_reg } else { xlen_reg },
Size::from_bits(xlen * 2),
));
}
if align_regs && is_vararg {
*avail_gprs -= *avail_gprs % 2;

View File

@ -27,10 +27,7 @@ where
if arg.layout.is_aggregate() {
let pad_i32 = !offset.is_aligned(align);
arg.cast_to_and_pad_i32(
Uniform { unit: Reg::i32(), total: size, force_array: false },
pad_i32,
);
arg.cast_to_and_pad_i32(Uniform::new(Reg::i32(), size), pad_i32);
} else {
arg.extend_integer_width_to(32);
}

View File

@ -68,7 +68,7 @@ where
}
// Cast to a uniform int structure
ret.cast_to(Uniform { unit: Reg::i64(), total: size, force_array: false });
ret.cast_to(Uniform::new(Reg::i64(), size));
} else {
ret.make_indirect();
}
@ -139,7 +139,7 @@ where
let rest_size = size - Size::from_bytes(8) * prefix_index as u64;
arg.cast_to(CastTarget {
prefix,
rest: Uniform { unit: Reg::i64(), total: rest_size, force_array: false },
rest: Uniform::new(Reg::i64(), rest_size),
attrs: ArgAttributes {
regular: ArgAttribute::default(),
arg_ext: ArgExtension::None,

View File

@ -256,13 +256,15 @@ pub struct Uniform {
/// this size will be rounded up to the nearest multiple of `unit.size`.
pub total: Size,
/// Force the use of an array, even if there is only a single element.
pub force_array: bool,
/// Indicate that the argument is consecutive, in the sense that either all values need to be
/// passed in register, or all on the stack. If they are passed on the stack, there should be
/// no additional padding between elements.
pub is_consecutive: bool,
}
impl From<Reg> for Uniform {
fn from(unit: Reg) -> Uniform {
Uniform { unit, total: unit.size, force_array: false }
Uniform { unit, total: unit.size, is_consecutive: false }
}
}
@ -270,6 +272,18 @@ impl Uniform {
pub fn align<C: HasDataLayout>(&self, cx: &C) -> Align {
self.unit.align(cx)
}
/// Pass using one or more values of the given type, without requiring them to be consecutive.
/// That is, some values may be passed in register and some on the stack.
pub fn new(unit: Reg, total: Size) -> Self {
Uniform { unit, total, is_consecutive: false }
}
/// Pass using one or more consecutive values of the given type. Either all values will be
/// passed in registers, or all on the stack.
pub fn consecutive(unit: Reg, total: Size) -> Self {
Uniform { unit, total, is_consecutive: true }
}
}
/// Describes the type used for `PassMode::Cast`.

View File

@ -35,7 +35,7 @@ where
16 => Reg::i128(),
_ => unreachable!("Align is given as power of 2 no larger than 16 bytes"),
};
arg.cast_to(Uniform { unit, total: Size::from_bytes(2 * align_bytes), force_array: false });
arg.cast_to(Uniform::new(unit, Size::from_bytes(2 * align_bytes)));
} else {
// FIXME: find a better way to do this. See https://github.com/rust-lang/rust/issues/117271.
arg.make_direct_deprecated();

View File

@ -37,7 +37,7 @@ where
RegKind::Vector => arg.layout.size.bits() == 128,
};
valid_unit.then_some(Uniform { unit, total: arg.layout.size, force_array: false })
valid_unit.then_some(Uniform::consecutive(unit, arg.layout.size))
})
}
@ -81,7 +81,7 @@ where
Reg::i64()
};
ret.cast_to(Uniform { unit, total: size, force_array: false });
ret.cast_to(Uniform::new(unit, size));
return;
}
@ -117,11 +117,10 @@ where
// of i64s or i128s, depending on the aggregate alignment. Always use an array for
// this, even if there is only a single element.
let reg = if arg.layout.align.abi.bytes() > 8 { Reg::i128() } else { Reg::i64() };
arg.cast_to(Uniform {
unit: reg,
total: size.align_to(Align::from_bytes(reg.size.bytes()).unwrap()),
force_array: true,
})
arg.cast_to(Uniform::consecutive(
reg,
size.align_to(Align::from_bytes(reg.size.bytes()).unwrap()),
))
};
}

View File

@ -201,11 +201,7 @@ where
if total.bits() <= xlen {
arg.cast_to(xlen_reg);
} else {
arg.cast_to(Uniform {
unit: xlen_reg,
total: Size::from_bits(xlen * 2),
force_array: false,
});
arg.cast_to(Uniform::new(xlen_reg, Size::from_bits(xlen * 2)));
}
return false;
}
@ -288,11 +284,10 @@ fn classify_arg<'a, Ty, C>(
if total.bits() > xlen {
let align_regs = align > xlen;
if is_riscv_aggregate(arg) {
arg.cast_to(Uniform {
unit: if align_regs { double_xlen_reg } else { xlen_reg },
total: Size::from_bits(xlen * 2),
force_array: false,
});
arg.cast_to(Uniform::new(
if align_regs { double_xlen_reg } else { xlen_reg },
Size::from_bits(xlen * 2),
));
}
if align_regs && is_vararg {
*avail_gprs -= *avail_gprs % 2;

View File

@ -27,10 +27,7 @@ where
if arg.layout.is_aggregate() {
let pad_i32 = !offset.is_aligned(align);
arg.cast_to_and_pad_i32(
Uniform { unit: Reg::i32(), total: size, force_array: false },
pad_i32,
);
arg.cast_to_and_pad_i32(Uniform::new(Reg::i32(), size), pad_i32);
} else {
arg.extend_integer_width_to(32);
}

View File

@ -192,7 +192,7 @@ where
arg.cast_to(CastTarget {
prefix: data.prefix,
rest: Uniform { unit: Reg::i64(), total: rest_size, force_array: false },
rest: Uniform::new(Reg::i64(), rest_size),
attrs: ArgAttributes {
regular: data.arg_attribute,
arg_ext: ArgExtension::None,
@ -205,7 +205,7 @@ where
}
}
arg.cast_to(Uniform { unit: Reg::i64(), total, force_array: false });
arg.cast_to(Uniform::new(Reg::i64(), total));
}
pub fn compute_abi_info<'a, Ty, C>(cx: &C, fn_abi: &mut FnAbi<'a, Ty>)