Multi-register value support: framework for Values wider than machine regs.

This will allow for support for `I128` values everywhere, and `I64`
values on 32-bit targets (e.g., ARM32 and x86-32). It does not alter the
machine backends to build such support; it just adds the framework for
the MachInst backends to *reason* about a `Value` residing in more than
one register.
This commit is contained in:
Chris Fallin
2020-12-12 20:48:56 -08:00
parent 6317290a1d
commit 6eea015d6c
18 changed files with 1024 additions and 561 deletions

View File

@@ -0,0 +1,185 @@
//! Data structure for tracking the (possibly multiple) registers that hold one
//! SSA `Value`.
use regalloc::{RealReg, Reg, VirtualReg, Writable};
use std::fmt::Debug;
#[cfg(feature = "arm32")]
const VALUE_REGS_PARTS: usize = 4;
#[cfg(not(feature = "arm32"))]
const VALUE_REGS_PARTS: usize = 2;
/// Location at which a `Value` is stored in register(s): the value is located
/// in one or more registers, depending on its width. A value may be stored in
/// more than one register if the machine has no registers wide enough
/// otherwise: for example, on a 32-bit architecture, we may store `I64` values
/// in two registers, and `I128` values in four.
///
/// By convention, the register parts are kept in machine-endian order here.
///
/// N.B.: we cap the capacity of this at four (when any 32-bit target is
/// enabled) or two (otherwise), and we use special in-band sentinal `Reg`
/// values (`Reg::invalid()`) to avoid the need to carry a separate length. This
/// allows the struct to be `Copy` (no heap or drop overhead) and be only 16 or
/// 8 bytes, which is important for compiler performance.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct ValueRegs<R: Clone + Copy + Debug + PartialEq + Eq + InvalidSentinel> {
parts: [R; VALUE_REGS_PARTS],
}
/// A type with an "invalid" sentinel value.
pub trait InvalidSentinel: Copy + Eq {
/// The invalid sentinel value.
fn invalid_sentinel() -> Self;
/// Is this the invalid sentinel?
fn is_invalid_sentinel(self) -> bool {
self == Self::invalid_sentinel()
}
}
impl InvalidSentinel for Reg {
fn invalid_sentinel() -> Self {
Reg::invalid()
}
}
impl InvalidSentinel for VirtualReg {
fn invalid_sentinel() -> Self {
VirtualReg::invalid()
}
}
impl InvalidSentinel for RealReg {
fn invalid_sentinel() -> Self {
RealReg::invalid()
}
}
impl InvalidSentinel for Writable<Reg> {
fn invalid_sentinel() -> Self {
Writable::from_reg(Reg::invalid_sentinel())
}
}
impl<R: Clone + Copy + Debug + PartialEq + Eq + InvalidSentinel> ValueRegs<R> {
/// Create an invalid Value-in-Reg.
pub fn invalid() -> Self {
ValueRegs {
parts: [R::invalid_sentinel(); VALUE_REGS_PARTS],
}
}
/// Is this Value-to-Reg mapping valid?
pub fn is_valid(self) -> bool {
!self.parts[0].is_invalid_sentinel()
}
/// Is this Value-to-Reg mapping invalid?
pub fn is_invalid(self) -> bool {
self.parts[0].is_invalid_sentinel()
}
/// Return the single register used for this value, if any.
pub fn only_reg(self) -> Option<R> {
if self.len() == 1 {
Some(self.parts[0])
} else {
None
}
}
/// Return an iterator over the registers storing this value.
pub fn regs(&self) -> &[R] {
&self.parts[0..self.len()]
}
}
#[cfg(feature = "arm32")]
impl<R: Clone + Copy + Debug + PartialEq + Eq + InvalidSentinel> ValueRegs<R> {
/// Create a Value-in-R location for a value stored in one register.
pub fn one(reg: R) -> Self {
ValueRegs {
parts: [
reg,
R::invalid_sentinel(),
R::invalid_sentinel(),
R::invalid_sentinel(),
],
}
}
/// Create a Value-in-R location for a value stored in two registers.
pub fn two(r1: R, r2: R) -> Self {
ValueRegs {
parts: [r1, r2, R::invalid_sentinel(), R::invalid_sentinel()],
}
}
/// Create a Value-in-R location for a value stored in four registers.
pub fn four(r1: R, r2: R, r3: R, r4: R) -> Self {
ValueRegs {
parts: [r1, r2, r3, r4],
}
}
/// Return the number of registers used.
pub fn len(self) -> usize {
// If rustc/LLVM is smart enough, this might even be vectorized...
(self.parts[0] != R::invalid_sentinel()) as usize
+ (self.parts[1] != R::invalid_sentinel()) as usize
+ (self.parts[2] != R::invalid_sentinel()) as usize
+ (self.parts[3] != R::invalid_sentinel()) as usize
}
/// Map individual registers via a map function.
pub fn map<NewR, F>(self, f: F) -> ValueRegs<NewR>
where
NewR: Clone + Copy + Debug + PartialEq + Eq + InvalidSentinel,
F: Fn(R) -> NewR,
{
ValueRegs {
parts: [
f(self.parts[0]),
f(self.parts[1]),
f(self.parts[2]),
f(self.parts[3]),
],
}
}
}
#[cfg(not(feature = "arm32"))]
impl<R: Clone + Copy + Debug + PartialEq + Eq + InvalidSentinel> ValueRegs<R> {
/// Create a Value-in-R location for a value stored in one register.
pub fn one(reg: R) -> Self {
ValueRegs {
parts: [reg, R::invalid_sentinel()],
}
}
/// Create a Value-in-R location for a value stored in two registers.
pub fn two(r1: R, r2: R) -> Self {
ValueRegs { parts: [r1, r2] }
}
/// Return the number of registers used.
pub fn len(self) -> usize {
// If rustc/LLVM is smart enough, this might even be vectorized...
(self.parts[0] != R::invalid_sentinel()) as usize
+ (self.parts[1] != R::invalid_sentinel()) as usize
}
/// Map individual registers via a map function.
pub fn map<NewR, F>(self, f: F) -> ValueRegs<NewR>
where
NewR: Clone + Copy + Debug + PartialEq + Eq + InvalidSentinel,
F: Fn(R) -> NewR,
{
ValueRegs {
parts: [f(self.parts[0]), f(self.parts[1])],
}
}
}
/// Create a writable ValueRegs.
pub(crate) fn writable_value_regs(regs: ValueRegs<Reg>) -> ValueRegs<Writable<Reg>> {
regs.map(|r| Writable::from_reg(r))
}
/// Strip a writable ValueRegs down to a readonly ValueRegs.
pub(crate) fn non_writable_value_regs(regs: ValueRegs<Writable<Reg>>) -> ValueRegs<Reg> {
regs.map(|r| r.to_reg())
}