github
/
cranelift
mirror of https://github.com/bytecodealliance/wasmtime.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

[machinst x64]: implement packed bitselect

pull/2284/head
Andrew Brown 4 years ago
parent
95f0e96e62
commit
1799b0947f
2 changed files with 80 additions and 27 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 84
      cranelift/codegen/src/isa/x64/inst/mod.rs
  2. 23
      cranelift/codegen/src/isa/x64/lower.rs

84
cranelift/codegen/src/isa/x64/inst/mod.rs
View File

@ -1105,33 +1105,6 @@ impl Inst {
}
}
/// Choose which instruction to use for comparing two values for equality.
pub(crate) fn equals(ty: Type, from: RegMem, to: Writable<Reg>) -> Inst {
match ty {
types::I8X16 => Inst::xmm_rm_r(SseOpcode::Pcmpeqb, from, to),
types::I16X8 => Inst::xmm_rm_r(SseOpcode::Pcmpeqw, from, to),
types::I32X4 => Inst::xmm_rm_r(SseOpcode::Pcmpeqd, from, to),
types::I64X2 => Inst::xmm_rm_r(SseOpcode::Pcmpeqq, from, to),
types::F32X4 => {
Inst::xmm_rm_r_imm(SseOpcode::Cmpps, from, to, FcmpImm::Equal.encode(), false)
}
types::F64X2 => {
Inst::xmm_rm_r_imm(SseOpcode::Cmppd, from, to, FcmpImm::Equal.encode(), false)
}
_ => unimplemented!("unimplemented type for Inst::equals: {}", ty),
}
}
/// Choose which instruction to use for computing a bitwise XOR on two values.
pub(crate) fn xor(ty: Type, from: RegMem, to: Writable<Reg>) -> Inst {
match ty {
types::F32X4 => Inst::xmm_rm_r(SseOpcode::Xorps, from, to),
types::F64X2 => Inst::xmm_rm_r(SseOpcode::Xorpd, from, to),
_ if ty.is_vector() => Inst::xmm_rm_r(SseOpcode::Pxor, from, to),
_ => unimplemented!("unimplemented type for Inst::xor: {}", ty),
}
}
/// Choose which instruction to use for loading a register value from memory. For loads smaller
/// than 64 bits, this method expects a way to extend the value (i.e. [ExtKind::SignExtend],
/// [ExtKind::ZeroExtend]); loads with no extension necessary will ignore this.
@ -1256,6 +1229,63 @@ impl Inst {
_ => false,
}
}
/// Choose which instruction to use for comparing two values for equality.
pub(crate) fn equals(ty: Type, from: RegMem, to: Writable<Reg>) -> Inst {
match ty {
types::I8X16 | types::B8X16 => Inst::xmm_rm_r(SseOpcode::Pcmpeqb, from, to),
types::I16X8 | types::B16X8 => Inst::xmm_rm_r(SseOpcode::Pcmpeqw, from, to),
types::I32X4 | types::B32X4 => Inst::xmm_rm_r(SseOpcode::Pcmpeqd, from, to),
types::I64X2 | types::B64X2 => Inst::xmm_rm_r(SseOpcode::Pcmpeqq, from, to),
types::F32X4 => {
Inst::xmm_rm_r_imm(SseOpcode::Cmpps, from, to, FcmpImm::Equal.encode(), false)
}
types::F64X2 => {
Inst::xmm_rm_r_imm(SseOpcode::Cmppd, from, to, FcmpImm::Equal.encode(), false)
}
_ => unimplemented!("unimplemented type for Inst::equals: {}", ty),
}
}
/// Choose which instruction to use for computing a bitwise AND on two values.
pub(crate) fn and(ty: Type, from: RegMem, to: Writable<Reg>) -> Inst {
match ty {
types::F32X4 => Inst::xmm_rm_r(SseOpcode::Andps, from, to),
types::F64X2 => Inst::xmm_rm_r(SseOpcode::Andpd, from, to),
_ if ty.is_vector() && ty.bits() == 128 => Inst::xmm_rm_r(SseOpcode::Pand, from, to),
_ => unimplemented!("unimplemented type for Inst::and: {}", ty),
}
}
/// Choose which instruction to use for computing a bitwise AND NOT on two values.
pub(crate) fn and_not(ty: Type, from: RegMem, to: Writable<Reg>) -> Inst {
match ty {
types::F32X4 => Inst::xmm_rm_r(SseOpcode::Andnps, from, to),
types::F64X2 => Inst::xmm_rm_r(SseOpcode::Andnpd, from, to),
_ if ty.is_vector() && ty.bits() == 128 => Inst::xmm_rm_r(SseOpcode::Pandn, from, to),
_ => unimplemented!("unimplemented type for Inst::and_not: {}", ty),
}
}
/// Choose which instruction to use for computing a bitwise OR on two values.
pub(crate) fn or(ty: Type, from: RegMem, to: Writable<Reg>) -> Inst {
match ty {
types::F32X4 => Inst::xmm_rm_r(SseOpcode::Orps, from, to),
types::F64X2 => Inst::xmm_rm_r(SseOpcode::Orpd, from, to),
_ if ty.is_vector() && ty.bits() == 128 => Inst::xmm_rm_r(SseOpcode::Por, from, to),
_ => unimplemented!("unimplemented type for Inst::or: {}", ty),
}
}
/// Choose which instruction to use for computing a bitwise XOR on two values.
pub(crate) fn xor(ty: Type, from: RegMem, to: Writable<Reg>) -> Inst {
match ty {
types::F32X4 => Inst::xmm_rm_r(SseOpcode::Xorps, from, to),
types::F64X2 => Inst::xmm_rm_r(SseOpcode::Xorpd, from, to),
_ if ty.is_vector() && ty.bits() == 128 => Inst::xmm_rm_r(SseOpcode::Pxor, from, to),
_ => unimplemented!("unimplemented type for Inst::xor: {}", ty),
}
}
}
//=============================================================================

23
cranelift/codegen/src/isa/x64/lower.rs
View File

@ -908,6 +908,29 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
}
}
Opcode::Bitselect => {
let ty = ty.unwrap();
let condition = put_input_in_reg(ctx, inputs[0]);
let if_true = put_input_in_reg(ctx, inputs[1]);
let if_false = input_to_reg_mem(ctx, inputs[2]);
let dst = get_output_reg(ctx, outputs[0]);
if ty.is_vector() {
let tmp1 = ctx.alloc_tmp(RegClass::V128, ty);
ctx.emit(Inst::gen_move(tmp1, if_true, ty));
ctx.emit(Inst::and(ty, RegMem::reg(condition.clone()), tmp1));
let tmp2 = ctx.alloc_tmp(RegClass::V128, ty);
ctx.emit(Inst::gen_move(tmp2, condition, ty));
ctx.emit(Inst::and_not(ty, if_false, tmp2));
ctx.emit(Inst::gen_move(dst, tmp2.to_reg(), ty));
ctx.emit(Inst::or(ty, RegMem::from(tmp1), dst));
} else {
unimplemented!("scalar bitselect")
}
}
Opcode::Ishl | Opcode::Ushr | Opcode::Sshr | Opcode::Rotl | Opcode::Rotr => {
let dst_ty = ctx.output_ty(insn, 0);
debug_assert_eq!(ctx.input_ty(insn, 0), dst_ty);

Write Preview
Loading…
Cancel
Save