T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

CL/aarch64: implement the wasm SIMD pseudo-max/min and FP-rounding instructions

Browse Source 
This patch implements, for aarch64, the following wasm SIMD extensions

  Floating-point rounding instructions
  https://github.com/WebAssembly/simd/pull/232

  Pseudo-Minimum and Pseudo-Maximum instructions
  https://github.com/WebAssembly/simd/pull/122

The changes are straightforward:

* `build.rs`: the relevant tests have been enabled

* `cranelift/codegen/meta/src/shared/instructions.rs`: new CLIF instructions
  `fmin_pseudo` and `fmax_pseudo`.  The wasm rounding instructions do not need
  any new CLIF instructions.

* `cranelift/wasm/src/code_translator.rs`: translation into CLIF; this is
  pretty much the same as any other unary or binary vector instruction (for
  the rounding and the pmin/max respectively)

* `cranelift/codegen/src/isa/aarch64/lower_inst.rs`:
  - `fmin_pseudo` and `fmax_pseudo` are converted into a two instruction
    sequence, `fcmpgt` followed by `bsl`
  - the CLIF rounding instructions are converted to a suitable vector
    `frint{n,z,p,m}` instruction.

* `cranelift/codegen/src/isa/aarch64/inst/mod.rs`: minor extension of `pub
  enum VecMisc2` to handle the rounding operations.  And corresponding `emit`
  cases.
This commit is contained in:
Julian Seward
2020-10-23 11:39:50 +02:00
committed by julian-seward1
parent fc1cedb2ff
commit c15d9bd61b
8 changed files with 265 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
cranelift/codegen/src/isa/aarch64/lower_inst.rs
View File

@@ -2373,6 +2373,43 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
}
}
Opcode::FminPseudo | Opcode::FmaxPseudo => {
let ty = ctx.input_ty(insn, 0);
if ty == F32X4 || ty == F64X2 {
// pmin(a,b) => bitsel(b, a, cmpgt(a, b))
// pmax(a,b) => bitsel(b, a, cmpgt(b, a))
let r_dst = get_output_reg(ctx, outputs[0]);
let r_a = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
let r_b = put_input_in_reg(ctx, inputs[1], NarrowValueMode::None);
// Since we're going to write the output register `r_dst` anyway, we might as
// well first use it to hold the comparison result. This has the slightly unusual
// effect that we modify the output register in the first instruction (`fcmgt`)
// but read both the inputs again in the second instruction (`bsl`), which means
// that the output register can't be either of the input registers. Regalloc
// should handle this correctly, nevertheless.
ctx.emit(Inst::VecRRR {
alu_op: VecALUOp::Fcmgt,
rd: r_dst,
rn: if op == Opcode::FminPseudo { r_a } else { r_b },
rm: if op == Opcode::FminPseudo { r_b } else { r_a },
size: if ty == F32X4 {
VectorSize::Size32x4
} else {
VectorSize::Size64x2
},
});
ctx.emit(Inst::VecRRR {
alu_op: VecALUOp::Bsl,
rd: r_dst,
rn: r_b,
rm: r_a,
size: VectorSize::Size8x16,
});
} else {
panic!("Opcode::FminPseudo | Opcode::FmaxPseudo: unhandled type");
}
}
Opcode::Sqrt | Opcode::Fneg | Opcode::Fabs | Opcode::Fpromote | Opcode::Fdemote => {
let ty = ty.unwrap();
let bits = ty_bits(ty);
@@ -2411,21 +2448,39 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
}
Opcode::Ceil | Opcode::Floor | Opcode::Trunc | Opcode::Nearest => {
let bits = ty_bits(ctx.output_ty(insn, 0));
let op = match (op, bits) {
(Opcode::Ceil, 32) => FpuRoundMode::Plus32,
(Opcode::Ceil, 64) => FpuRoundMode::Plus64,
(Opcode::Floor, 32) => FpuRoundMode::Minus32,
(Opcode::Floor, 64) => FpuRoundMode::Minus64,
(Opcode::Trunc, 32) => FpuRoundMode::Zero32,
(Opcode::Trunc, 64) => FpuRoundMode::Zero64,
(Opcode::Nearest, 32) => FpuRoundMode::Nearest32,
(Opcode::Nearest, 64) => FpuRoundMode::Nearest64,
_ => panic!("Unknown op/bits combination"),
};
let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
let rd = get_output_reg(ctx, outputs[0]);
ctx.emit(Inst::FpuRound { op, rd, rn });
let ty = ctx.output_ty(insn, 0);
if !ty.is_vector() {
let bits = ty_bits(ty);
let op = match (op, bits) {
(Opcode::Ceil, 32) => FpuRoundMode::Plus32,
(Opcode::Ceil, 64) => FpuRoundMode::Plus64,
(Opcode::Floor, 32) => FpuRoundMode::Minus32,
(Opcode::Floor, 64) => FpuRoundMode::Minus64,
(Opcode::Trunc, 32) => FpuRoundMode::Zero32,
(Opcode::Trunc, 64) => FpuRoundMode::Zero64,
(Opcode::Nearest, 32) => FpuRoundMode::Nearest32,
(Opcode::Nearest, 64) => FpuRoundMode::Nearest64,
_ => panic!("Unknown op/bits combination (scalar)"),
};
let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
let rd = get_output_reg(ctx, outputs[0]);
ctx.emit(Inst::FpuRound { op, rd, rn });
} else {
let (op, size) = match (op, ty) {
(Opcode::Ceil, F32X4) => (VecMisc2::Frintp, VectorSize::Size32x4),
(Opcode::Ceil, F64X2) => (VecMisc2::Frintp, VectorSize::Size64x2),
(Opcode::Floor, F32X4) => (VecMisc2::Frintm, VectorSize::Size32x4),
(Opcode::Floor, F64X2) => (VecMisc2::Frintm, VectorSize::Size64x2),
(Opcode::Trunc, F32X4) => (VecMisc2::Frintz, VectorSize::Size32x4),
(Opcode::Trunc, F64X2) => (VecMisc2::Frintz, VectorSize::Size64x2),
(Opcode::Nearest, F32X4) => (VecMisc2::Frintn, VectorSize::Size32x4),
(Opcode::Nearest, F64X2) => (VecMisc2::Frintn, VectorSize::Size64x2),
_ => panic!("Unknown op/ty combination (vector){:?}", ty),
};
let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
let rd = get_output_reg(ctx, outputs[0]);
ctx.emit(Inst::VecMisc { op, rd, rn, size });
}
}
Opcode::Fma => {