arm64: Don't always materialise a 64-bit constant

This improves the mov/movk/movn sequnce when the high half of the
64-bit value is all zero.

Copyright (c) 2020, Arm Limited.
This commit is contained in:
Joey Gouly
2020-09-01 13:29:01 +01:00
parent c9ec60dcce
commit 650d48cd84
6 changed files with 130 additions and 29 deletions

View File

@@ -587,18 +587,21 @@ pub enum Inst {
MovZ {
rd: Writable<Reg>,
imm: MoveWideConst,
size: OperandSize,
},
/// A MOVN with a 16-bit immediate.
MovN {
rd: Writable<Reg>,
imm: MoveWideConst,
size: OperandSize,
},
/// A MOVK with a 16-bit immediate.
MovK {
rd: Writable<Reg>,
imm: MoveWideConst,
size: OperandSize,
},
/// A sign- or zero-extend operation.
@@ -1122,9 +1125,9 @@ pub enum Inst {
},
}
fn count_zero_half_words(mut value: u64) -> usize {
fn count_zero_half_words(mut value: u64, num_half_words: u8) -> usize {
let mut count = 0;
for _ in 0..4 {
for _ in 0..num_half_words {
if value & 0xffff == 0 {
count += 1;
}
@@ -1176,10 +1179,18 @@ impl Inst {
pub fn load_constant(rd: Writable<Reg>, value: u64) -> SmallVec<[Inst; 4]> {
if let Some(imm) = MoveWideConst::maybe_from_u64(value) {
// 16-bit immediate (shifted by 0, 16, 32 or 48 bits) in MOVZ
smallvec![Inst::MovZ { rd, imm }]
smallvec![Inst::MovZ {
rd,
imm,
size: OperandSize::Size64
}]
} else if let Some(imm) = MoveWideConst::maybe_from_u64(!value) {
// 16-bit immediate (shifted by 0, 16, 32 or 48 bits) in MOVN
smallvec![Inst::MovN { rd, imm }]
smallvec![Inst::MovN {
rd,
imm,
size: OperandSize::Size64
}]
} else if let Some(imml) = ImmLogic::maybe_from_u64(value, I64) {
// Weird logical-instruction immediate in ORI using zero register
smallvec![Inst::AluRRImmLogic {
@@ -1191,15 +1202,22 @@ impl Inst {
} else {
let mut insts = smallvec![];
// If the top 32 bits are zero, use 32-bit `mov` operations.
let (num_half_words, size, negated) = if value >> 32 == 0 {
(2, OperandSize::Size32, (!value << 32) >> 32)
} else {
(4, OperandSize::Size64, !value)
};
// If the number of 0xffff half words is greater than the number of 0x0000 half words
// it is more efficient to use `movn` for the first instruction.
let first_is_inverted = count_zero_half_words(!value) > count_zero_half_words(value);
let first_is_inverted = count_zero_half_words(negated, num_half_words)
> count_zero_half_words(value, num_half_words);
// Either 0xffff or 0x0000 half words can be skipped, depending on the first
// instruction used.
let ignored_halfword = if first_is_inverted { 0xffff } else { 0 };
let mut first_mov_emitted = false;
for i in 0..4 {
for i in 0..num_half_words {
let imm16 = (value >> (16 * i)) & 0xffff;
if imm16 != ignored_halfword {
if !first_mov_emitted {
@@ -1208,15 +1226,15 @@ impl Inst {
let imm =
MoveWideConst::maybe_with_shift(((!imm16) & 0xffff) as u16, i * 16)
.unwrap();
insts.push(Inst::MovN { rd, imm });
insts.push(Inst::MovN { rd, imm, size });
} else {
let imm =
MoveWideConst::maybe_with_shift(imm16 as u16, i * 16).unwrap();
insts.push(Inst::MovZ { rd, imm });
insts.push(Inst::MovZ { rd, imm, size });
}
} else {
let imm = MoveWideConst::maybe_with_shift(imm16 as u16, i * 16).unwrap();
insts.push(Inst::MovK { rd, imm });
insts.push(Inst::MovK { rd, imm, size });
}
}
}
@@ -2870,18 +2888,18 @@ impl Inst {
let rm = show_ireg_sized(rm, mb_rru, OperandSize::Size32);
format!("mov {}, {}", rd, rm)
}
&Inst::MovZ { rd, ref imm } => {
let rd = rd.to_reg().show_rru(mb_rru);
&Inst::MovZ { rd, ref imm, size } => {
let rd = show_ireg_sized(rd.to_reg(), mb_rru, size);
let imm = imm.show_rru(mb_rru);
format!("movz {}, {}", rd, imm)
}
&Inst::MovN { rd, ref imm } => {
let rd = rd.to_reg().show_rru(mb_rru);
&Inst::MovN { rd, ref imm, size } => {
let rd = show_ireg_sized(rd.to_reg(), mb_rru, size);
let imm = imm.show_rru(mb_rru);
format!("movn {}, {}", rd, imm)
}
&Inst::MovK { rd, ref imm } => {
let rd = rd.to_reg().show_rru(mb_rru);
&Inst::MovK { rd, ref imm, size } => {
let rd = show_ireg_sized(rd.to_reg(), mb_rru, size);
let imm = imm.show_rru(mb_rru);
format!("movk {}, {}", rd, imm)
}