Skip to content

Commit

Permalink
Merge pull request #2959 from afonso360/aarch64-i128-bit-ops
Browse files Browse the repository at this point in the history 
aarch64 add basic i128 bit ops
  • Loading branch information
@cfallin
cfallin committed Jun 9, 2021
2 parents 59ebe4f + b1475f3 commit caa85c2
Show file tree
Hide file tree
Showing 8 changed files with 709 additions and 8 deletions.
6 changes: 6 additions & 0 deletions cranelift/codegen/src/isa/aarch64/inst/emit.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -607,6 +607,8 @@ impl MachInstEmit for Inst {
ALUOp::Orr64 => 0b10101010_000,
ALUOp::And32 => 0b00001010_000,
ALUOp::And64 => 0b10001010_000,
ALUOp::AndS32 => 0b01101010_000,
ALUOp::AndS64 => 0b11101010_000,
ALUOp::Eor32 => 0b01001010_000,
ALUOp::Eor64 => 0b11001010_000,
ALUOp::OrrNot32 => 0b00101010_001,
Expand Down Expand Up @@ -694,6 +696,8 @@ impl MachInstEmit for Inst {
ALUOp::Orr64 => (0b101_100100, false),
ALUOp::And32 => (0b000_100100, false),
ALUOp::And64 => (0b100_100100, false),
ALUOp::AndS32 => (0b011_100100, false),
ALUOp::AndS64 => (0b111_100100, false),
ALUOp::Eor32 => (0b010_100100, false),
ALUOp::Eor64 => (0b110_100100, false),
ALUOp::OrrNot32 => (0b001_100100, true),
Expand Down Expand Up @@ -763,6 +767,8 @@ impl MachInstEmit for Inst {
ALUOp::Orr64 => 0b101_01010000,
ALUOp::And32 => 0b000_01010000,
ALUOp::And64 => 0b100_01010000,
ALUOp::AndS32 => 0b011_01010000,
ALUOp::AndS64 => 0b111_01010000,
ALUOp::Eor32 => 0b010_01010000,
ALUOp::Eor64 => 0b110_01010000,
ALUOp::OrrNot32 => 0b001_01010001,
Expand Down
68 changes: 68 additions & 0 deletions cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -151,6 +151,26 @@ fn test_aarch64_binemit() {
"A400068A",
"and x4, x5, x6",
));
insns.push((
Inst::AluRRR {
alu_op: ALUOp::AndS32,
rd: writable_xreg(1),
rn: xreg(2),
rm: xreg(3),
},
"4100036A",
"ands w1, w2, w3",
));
insns.push((
Inst::AluRRR {
alu_op: ALUOp::AndS64,
rd: writable_xreg(4),
rn: xreg(5),
rm: xreg(6),
},
"A40006EA",
"ands x4, x5, x6",
));
insns.push((
Inst::AluRRR {
alu_op: ALUOp::SubS32,
Expand Down Expand Up @@ -648,6 +668,34 @@ fn test_aarch64_binemit() {
"6A5D0C8A",
"and x10, x11, x12, LSL 23",
));
insns.push((
Inst::AluRRRShift {
alu_op: ALUOp::AndS32,
rd: writable_xreg(10),
rn: xreg(11),
rm: xreg(12),
shiftop: ShiftOpAndAmt::new(
ShiftOp::LSL,
ShiftOpShiftImm::maybe_from_shift(23).unwrap(),
),
},
"6A5D0C6A",
"ands w10, w11, w12, LSL 23",
));
insns.push((
Inst::AluRRRShift {
alu_op: ALUOp::AndS64,
rd: writable_xreg(10),
rn: xreg(11),
rm: xreg(12),
shiftop: ShiftOpAndAmt::new(
ShiftOp::LSL,
ShiftOpShiftImm::maybe_from_shift(23).unwrap(),
),
},
"6A5D0CEA",
"ands x10, x11, x12, LSL 23",
));
insns.push((
Inst::AluRRRShift {
alu_op: ALUOp::Eor32,
Expand Down Expand Up @@ -1015,6 +1063,26 @@ fn test_aarch64_binemit() {
"C7381592",
"and x7, x6, #288221580125796352",
));
insns.push((
Inst::AluRRImmLogic {
alu_op: ALUOp::AndS32,
rd: writable_xreg(21),
rn: xreg(27),
imml: ImmLogic::maybe_from_u64(0x80003fff, I32).unwrap(),
},
"753B0172",
"ands w21, w27, #2147500031",
));
insns.push((
Inst::AluRRImmLogic {
alu_op: ALUOp::AndS64,
rd: writable_xreg(7),
rn: xreg(6),
imml: ImmLogic::maybe_from_u64(0x3fff80003fff800, I64).unwrap(),
},
"C73815F2",
"ands x7, x6, #288221580125796352",
));
insns.push((
Inst::AluRRImmLogic {
alu_op: ALUOp::Orr32,
Expand Down
4 changes: 4 additions & 0 deletions cranelift/codegen/src/isa/aarch64/inst/mod.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,8 @@ pub enum ALUOp {
OrrNot64,
And32,
And64,
AndS32,
AndS64,
AndNot32,
AndNot64,
/// XOR (AArch64 calls this "EOR")
Expand Down Expand Up @@ -3186,6 +3188,8 @@ impl Inst {
ALUOp::Orr64 => ("orr", OperandSize::Size64),
ALUOp::And32 => ("and", OperandSize::Size32),
ALUOp::And64 => ("and", OperandSize::Size64),
ALUOp::AndS32 => ("ands", OperandSize::Size32),
ALUOp::AndS64 => ("ands", OperandSize::Size64),
ALUOp::Eor32 => ("eor", OperandSize::Size32),
ALUOp::Eor64 => ("eor", OperandSize::Size64),
ALUOp::AddS32 => ("adds", OperandSize::Size32),
Expand Down
205 changes: 205 additions & 0 deletions cranelift/codegen/src/isa/aarch64/lower.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1266,6 +1266,211 @@ pub(crate) fn lower_load<C: LowerCtx<I = Inst>, F: FnMut(&mut C, Writable<Reg>,
f(ctx, rd, elem_ty, mem);
}

pub(crate) fn emit_shl_i128<C: LowerCtx<I = Inst>>(
ctx: &mut C,
src: ValueRegs<Reg>,
dst: ValueRegs<Writable<Reg>>,
amt: Reg,
) {
let src_lo = src.regs()[0];
let src_hi = src.regs()[1];
let dst_lo = dst.regs()[0];
let dst_hi = dst.regs()[1];

// mvn inv_amt, amt
// lsr tmp1, src_lo, #1
// lsl tmp2, src_hi, amt
// lsr tmp1, tmp1, inv_amt
// lsl tmp3, src_lo, amt
// tst amt, #0x40
// orr tmp2, tmp2, tmp1
// csel dst_hi, tmp3, tmp2, ne
// csel dst_lo, xzr, tmp3, ne

let xzr = writable_zero_reg();
let inv_amt = ctx.alloc_tmp(I64).only_reg().unwrap();
let tmp1 = ctx.alloc_tmp(I64).only_reg().unwrap();
let tmp2 = ctx.alloc_tmp(I64).only_reg().unwrap();
let tmp3 = ctx.alloc_tmp(I64).only_reg().unwrap();

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::OrrNot32,
rd: inv_amt,
rn: xzr.to_reg(),
rm: amt,
});

ctx.emit(Inst::AluRRImmShift {
alu_op: ALUOp::Lsr64,
rd: tmp1,
rn: src_lo,
immshift: ImmShift::maybe_from_u64(1).unwrap(),
});

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::Lsl64,
rd: tmp2,
rn: src_hi,
rm: amt,
});

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::Lsr64,
rd: tmp1,
rn: tmp1.to_reg(),
rm: inv_amt.to_reg(),
});

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::Lsl64,
rd: tmp3,
rn: src_lo,
rm: amt,
});

ctx.emit(Inst::AluRRImmLogic {
alu_op: ALUOp::AndS64,
rd: xzr,
rn: amt,
imml: ImmLogic::maybe_from_u64(64, I64).unwrap(),
});

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::Orr64,
rd: tmp2,
rn: tmp2.to_reg(),
rm: tmp1.to_reg(),
});

ctx.emit(Inst::CSel {
cond: Cond::Ne,
rd: dst_hi,
rn: tmp3.to_reg(),
rm: tmp2.to_reg(),
});

ctx.emit(Inst::CSel {
cond: Cond::Ne,
rd: dst_lo,
rn: xzr.to_reg(),
rm: tmp3.to_reg(),
});
}

pub(crate) fn emit_shr_i128<C: LowerCtx<I = Inst>>(
ctx: &mut C,
src: ValueRegs<Reg>,
dst: ValueRegs<Writable<Reg>>,
amt: Reg,
is_signed: bool,
) {
let src_lo = src.regs()[0];
let src_hi = src.regs()[1];
let dst_lo = dst.regs()[0];
let dst_hi = dst.regs()[1];

// mvn inv_amt, amt
// lsl tmp1, src_lo, #1
// lsr tmp2, src_hi, amt
// lsl tmp1, tmp1, inv_amt
// lsr/asr tmp3, src_lo, amt
// tst amt, #0x40
// orr tmp2, tmp2, tmp1
//
// if signed:
// asr tmp4, src_hi, #63
// csel dst_hi, tmp4, tmp3, ne
// else:
// csel dst_hi, xzr, tmp3, ne
//
// csel dst_lo, tmp3, tmp2, ne

let xzr = writable_zero_reg();
let inv_amt = ctx.alloc_tmp(I64).only_reg().unwrap();
let tmp1 = ctx.alloc_tmp(I64).only_reg().unwrap();
let tmp2 = ctx.alloc_tmp(I64).only_reg().unwrap();
let tmp3 = ctx.alloc_tmp(I64).only_reg().unwrap();
let tmp4 = ctx.alloc_tmp(I64).only_reg().unwrap();

let shift_op = if is_signed {
ALUOp::Asr64
} else {
ALUOp::Lsr64
};

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::OrrNot32,
rd: inv_amt,
rn: xzr.to_reg(),
rm: amt,
});

ctx.emit(Inst::AluRRImmShift {
alu_op: ALUOp::Lsl64,
rd: tmp1,
rn: src_hi,
immshift: ImmShift::maybe_from_u64(1).unwrap(),
});

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::Lsr64,
rd: tmp2,
rn: src_lo,
rm: amt,
});

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::Lsl64,
rd: tmp1,
rn: tmp1.to_reg(),
rm: inv_amt.to_reg(),
});

ctx.emit(Inst::AluRRR {
alu_op: shift_op,
rd: tmp3,
rn: src_hi,
rm: amt,
});

ctx.emit(Inst::AluRRImmLogic {
alu_op: ALUOp::AndS64,
rd: xzr,
rn: amt,
imml: ImmLogic::maybe_from_u64(64, I64).unwrap(),
});

if is_signed {
ctx.emit(Inst::AluRRImmShift {
alu_op: ALUOp::Asr64,
rd: tmp4,
rn: src_hi,
immshift: ImmShift::maybe_from_u64(63).unwrap(),
});
}

ctx.emit(Inst::AluRRR {
alu_op: ALUOp::Orr64,
rd: tmp2,
rn: tmp2.to_reg(),
rm: tmp1.to_reg(),
});

ctx.emit(Inst::CSel {
cond: Cond::Ne,
rd: dst_hi,
rn: if is_signed { tmp4 } else { xzr }.to_reg(),
rm: tmp3.to_reg(),
});

ctx.emit(Inst::CSel {
cond: Cond::Ne,
rd: dst_lo,
rn: tmp3.to_reg(),
rm: tmp2.to_reg(),
});
}

//=============================================================================
// Lowering-backend trait implementation.

Expand Down
Loading

0 comments on commit caa85c2

Please sign in to comment.