Files
wasmtime/cranelift/codegen/src/isa/x64/inst/regs.rs
Chris Fallin 1dddba649a x64 regalloc register order: put caller-saves (volatiles) first.
The x64 backend currently builds the `RealRegUniverse` in a way that
is generating somewhat suboptimal code. In many blocks, we see uses of
callee-save (non-volatile) registers (r12, r13, r14, rbx) first, even in
very short leaf functions where there are plenty of volatiles to use.
This is leading to unnecessary spills/reloads.

On one (local) test program, a medium-sized C benchmark compiled to Wasm
and run on Wasmtime, I am seeing a ~10% performance improvement with
this change; it will be less pronounced in programs with high register
pressure (there we are likely to use all registers regardless, so the
prologue/epilogue will save/restore all callee-saves), or in programs
with fewer calls, but this is a clear win for small functions and in
many cases removes prologue/epilogue clobber-saves altogether.

Separately, I think the RA's coalescing is tripping up a bit in some
cases; see e.g. the filetest touched by this commit that loads a value
into %rsi then moves to %rax and returns immediately. This is an
orthogonal issue, though, and should be addressed (if worthwhile) in
regalloc.rs.
2020-12-06 22:37:43 -08:00

296 lines
8.9 KiB
Rust

//! Registers, the Universe thereof, and printing.
//!
//! These are ordered by sequence number, as required in the Universe.
//!
//! The caller-saved registers are placed first in order to prefer not to clobber (requiring
//! saves/restores in prologue/epilogue code) when possible. Note that there is no other heuristic
//! in the backend that will apply such pressure; the register allocator's cost heuristics are not
//! aware of the cost of clobber-save/restore code.
//!
//! One might worry that this pessimizes code with many callsites, where using caller-saves causes
//! us to have to save them (as we are the caller) frequently. However, the register allocator
//! *should be* aware of *this* cost, because it sees that the call instruction modifies all of the
//! caller-saved (i.e., callee-clobbered) registers.
//!
//! Hence, this ordering encodes pressure in one direction (prefer not to clobber registers that we
//! ourselves have to save) and this is balanaced against the RA's pressure in the other direction
//! at callsites.
use crate::settings;
use alloc::vec::Vec;
use regalloc::{
PrettyPrint, RealReg, RealRegUniverse, Reg, RegClass, RegClassInfo, NUM_REG_CLASSES,
};
use std::string::String;
// Hardware encodings for a few registers.
pub const ENC_RBX: u8 = 3;
pub const ENC_RSP: u8 = 4;
pub const ENC_RBP: u8 = 5;
pub const ENC_R12: u8 = 12;
pub const ENC_R13: u8 = 13;
pub const ENC_R14: u8 = 14;
pub const ENC_R15: u8 = 15;
fn gpr(enc: u8, index: u8) -> Reg {
Reg::new_real(RegClass::I64, enc, index)
}
pub(crate) fn rsi() -> Reg {
gpr(6, 16)
}
pub(crate) fn rdi() -> Reg {
gpr(7, 17)
}
pub(crate) fn rax() -> Reg {
gpr(0, 18)
}
pub(crate) fn rcx() -> Reg {
gpr(1, 19)
}
pub(crate) fn rdx() -> Reg {
gpr(2, 20)
}
pub(crate) fn r8() -> Reg {
gpr(8, 21)
}
pub(crate) fn r9() -> Reg {
gpr(9, 22)
}
pub(crate) fn r10() -> Reg {
gpr(10, 23)
}
pub(crate) fn r11() -> Reg {
gpr(11, 24)
}
pub(crate) fn r12() -> Reg {
gpr(ENC_R12, 25)
}
pub(crate) fn r13() -> Reg {
gpr(ENC_R13, 26)
}
pub(crate) fn r14() -> Reg {
gpr(ENC_R14, 27)
}
pub(crate) fn rbx() -> Reg {
gpr(ENC_RBX, 28)
}
pub(crate) fn r15() -> Reg {
// r15 is put aside since this is the pinned register.
gpr(ENC_R15, 29)
}
/// The pinned register on this architecture.
/// It must be the same as Spidermonkey's HeapReg, as found in this file.
/// https://searchfox.org/mozilla-central/source/js/src/jit/x64/Assembler-x64.h#99
pub(crate) fn pinned_reg() -> Reg {
r15()
}
fn fpr(enc: u8, index: u8) -> Reg {
Reg::new_real(RegClass::V128, enc, index)
}
pub(crate) fn xmm0() -> Reg {
fpr(0, 0)
}
pub(crate) fn xmm1() -> Reg {
fpr(1, 1)
}
pub(crate) fn xmm2() -> Reg {
fpr(2, 2)
}
pub(crate) fn xmm3() -> Reg {
fpr(3, 3)
}
pub(crate) fn xmm4() -> Reg {
fpr(4, 4)
}
pub(crate) fn xmm5() -> Reg {
fpr(5, 5)
}
pub(crate) fn xmm6() -> Reg {
fpr(6, 6)
}
pub(crate) fn xmm7() -> Reg {
fpr(7, 7)
}
pub(crate) fn xmm8() -> Reg {
fpr(8, 8)
}
pub(crate) fn xmm9() -> Reg {
fpr(9, 9)
}
pub(crate) fn xmm10() -> Reg {
fpr(10, 10)
}
pub(crate) fn xmm11() -> Reg {
fpr(11, 11)
}
pub(crate) fn xmm12() -> Reg {
fpr(12, 12)
}
pub(crate) fn xmm13() -> Reg {
fpr(13, 13)
}
pub(crate) fn xmm14() -> Reg {
fpr(14, 14)
}
pub(crate) fn xmm15() -> Reg {
fpr(15, 15)
}
pub(crate) fn rsp() -> Reg {
gpr(ENC_RSP, 30)
}
pub(crate) fn rbp() -> Reg {
gpr(ENC_RBP, 31)
}
/// Create the register universe for X64.
///
/// The ordering of registers matters, as commented in the file doc comment: assumes the
/// calling-convention is SystemV, at the moment.
pub(crate) fn create_reg_universe_systemv(flags: &settings::Flags) -> RealRegUniverse {
let mut regs = Vec::<(RealReg, String)>::new();
let mut allocable_by_class = [None; NUM_REG_CLASSES];
let use_pinned_reg = flags.enable_pinned_reg();
// XMM registers
let first_fpr = regs.len();
regs.push((xmm0().to_real_reg(), "%xmm0".into()));
regs.push((xmm1().to_real_reg(), "%xmm1".into()));
regs.push((xmm2().to_real_reg(), "%xmm2".into()));
regs.push((xmm3().to_real_reg(), "%xmm3".into()));
regs.push((xmm4().to_real_reg(), "%xmm4".into()));
regs.push((xmm5().to_real_reg(), "%xmm5".into()));
regs.push((xmm6().to_real_reg(), "%xmm6".into()));
regs.push((xmm7().to_real_reg(), "%xmm7".into()));
regs.push((xmm8().to_real_reg(), "%xmm8".into()));
regs.push((xmm9().to_real_reg(), "%xmm9".into()));
regs.push((xmm10().to_real_reg(), "%xmm10".into()));
regs.push((xmm11().to_real_reg(), "%xmm11".into()));
regs.push((xmm12().to_real_reg(), "%xmm12".into()));
regs.push((xmm13().to_real_reg(), "%xmm13".into()));
regs.push((xmm14().to_real_reg(), "%xmm14".into()));
regs.push((xmm15().to_real_reg(), "%xmm15".into()));
let last_fpr = regs.len() - 1;
// Integer regs.
let first_gpr = regs.len();
// Caller-saved, in the SystemV x86_64 ABI.
regs.push((rsi().to_real_reg(), "%rsi".into()));
regs.push((rdi().to_real_reg(), "%rdi".into()));
regs.push((rax().to_real_reg(), "%rax".into()));
regs.push((rcx().to_real_reg(), "%rcx".into()));
regs.push((rdx().to_real_reg(), "%rdx".into()));
regs.push((r8().to_real_reg(), "%r8".into()));
regs.push((r9().to_real_reg(), "%r9".into()));
regs.push((r10().to_real_reg(), "%r10".into()));
regs.push((r11().to_real_reg(), "%r11".into()));
// Callee-saved, in the SystemV x86_64 ABI.
regs.push((r12().to_real_reg(), "%r12".into()));
regs.push((r13().to_real_reg(), "%r13".into()));
regs.push((r14().to_real_reg(), "%r14".into()));
regs.push((rbx().to_real_reg(), "%rbx".into()));
// Other regs, not available to the allocator.
debug_assert_eq!(r15(), pinned_reg());
let allocable = if use_pinned_reg {
// The pinned register is not allocatable in this case, so record the length before adding
// it.
let len = regs.len();
regs.push((r15().to_real_reg(), "%r15/pinned".into()));
len
} else {
regs.push((r15().to_real_reg(), "%r15".into()));
regs.len()
};
let last_gpr = allocable - 1;
regs.push((rsp().to_real_reg(), "%rsp".into()));
regs.push((rbp().to_real_reg(), "%rbp".into()));
allocable_by_class[RegClass::I64.rc_to_usize()] = Some(RegClassInfo {
first: first_gpr,
last: last_gpr,
suggested_scratch: Some(r12().get_index()),
});
allocable_by_class[RegClass::V128.rc_to_usize()] = Some(RegClassInfo {
first: first_fpr,
last: last_fpr,
suggested_scratch: Some(xmm15().get_index()),
});
// Sanity-check: the index passed to the Reg ctor must match the order in the register list.
for (i, reg) in regs.iter().enumerate() {
assert_eq!(i, reg.0.get_index());
}
RealRegUniverse {
regs,
allocable,
allocable_by_class,
}
}
/// If `ireg` denotes an I64-classed reg, make a best-effort attempt to show its name at some
/// smaller size (4, 2 or 1 bytes).
pub fn show_ireg_sized(reg: Reg, mb_rru: Option<&RealRegUniverse>, size: u8) -> String {
let mut s = reg.show_rru(mb_rru);
if reg.get_class() != RegClass::I64 || size == 8 {
// We can't do any better.
return s;
}
if reg.is_real() {
// Change (eg) "rax" into "eax", "ax" or "al" as appropriate. This is something one could
// describe diplomatically as "a kludge", but it's only debug code.
let remapper = match s.as_str() {
"%rax" => Some(["%eax", "%ax", "%al"]),
"%rbx" => Some(["%ebx", "%bx", "%bl"]),
"%rcx" => Some(["%ecx", "%cx", "%cl"]),
"%rdx" => Some(["%edx", "%dx", "%dl"]),
"%rsi" => Some(["%esi", "%si", "%sil"]),
"%rdi" => Some(["%edi", "%di", "%dil"]),
"%rbp" => Some(["%ebp", "%bp", "%bpl"]),
"%rsp" => Some(["%esp", "%sp", "%spl"]),
"%r8" => Some(["%r8d", "%r8w", "%r8b"]),
"%r9" => Some(["%r9d", "%r9w", "%r9b"]),
"%r10" => Some(["%r10d", "%r10w", "%r10b"]),
"%r11" => Some(["%r11d", "%r11w", "%r11b"]),
"%r12" => Some(["%r12d", "%r12w", "%r12b"]),
"%r13" => Some(["%r13d", "%r13w", "%r13b"]),
"%r14" => Some(["%r14d", "%r14w", "%r14b"]),
"%r15" => Some(["%r15d", "%r15w", "%r15b"]),
_ => None,
};
if let Some(smaller_names) = remapper {
match size {
4 => s = smaller_names[0].into(),
2 => s = smaller_names[1].into(),
1 => s = smaller_names[2].into(),
_ => panic!("show_ireg_sized: real"),
}
}
} else {
// Add a "l", "w" or "b" suffix to RegClass::I64 vregs used at narrower widths.
let suffix = match size {
4 => "l",
2 => "w",
1 => "b",
_ => panic!("show_ireg_sized: virtual"),
};
s = s + suffix;
}
s
}