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

[AArch64] Port min/max to ISLE (#4374)

Browse Source 
Copyright (c) 2022, Arm Limited.
This commit is contained in:
Sam Parker
2022-07-05 17:16:45 +01:00
committed by GitHub
parent e91f493ff5
commit d9e0e6a6a9
6 changed files with 511 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
cranelift/codegen/src/isa/aarch64/inst.isle
View File

@@ -925,6 +925,7 @@
(rule (vector_size (multi_lane 8 16)) (VectorSize.Size8x16))
(rule (vector_size (multi_lane 16 4)) (VectorSize.Size16x4))
(rule (vector_size (multi_lane 16 8)) (VectorSize.Size16x8))
(rule (vector_size (multi_lane 32 2)) (VectorSize.Size32x2))
(rule (vector_size (multi_lane 32 4)) (VectorSize.Size32x4))
(rule (vector_size (multi_lane 64 2)) (VectorSize.Size64x2))

144
cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs
View File

@@ -3576,6 +3576,18 @@ fn test_aarch64_binemit() {
"sshl v8.2d, v22.2d, v2.2d",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umin,
rd: writable_vreg(0),
rn: vreg(11),
rm: vreg(2),
size: VectorSize::Size8x8,
},
"606D222E",
"umin v0.8b, v11.8b, v2.8b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umin,
@@ -3588,6 +3600,18 @@ fn test_aarch64_binemit() {
"umin v1.16b, v12.16b, v3.16b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umin,
rd: writable_vreg(29),
rn: vreg(19),
rm: vreg(9),
size: VectorSize::Size16x4,
},
"7D6E692E",
"umin v29.4h, v19.4h, v9.4h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umin,
@@ -3600,6 +3624,18 @@ fn test_aarch64_binemit() {
"umin v30.8h, v20.8h, v10.8h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umin,
rd: writable_vreg(7),
rn: vreg(21),
rm: vreg(20),
size: VectorSize::Size32x2,
},
"A76EB42E",
"umin v7.2s, v21.2s, v20.2s",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umin,
@@ -3612,6 +3648,18 @@ fn test_aarch64_binemit() {
"umin v8.4s, v22.4s, v21.4s",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smin,
rd: writable_vreg(2),
rn: vreg(13),
rm: vreg(4),
size: VectorSize::Size8x8,
},
"A26D240E",
"smin v2.8b, v13.8b, v4.8b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smin,
@@ -3624,6 +3672,18 @@ fn test_aarch64_binemit() {
"smin v1.16b, v12.16b, v3.16b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smin,
rd: writable_vreg(3),
rn: vreg(2),
rm: vreg(1),
size: VectorSize::Size16x4,
},
"436C610E",
"smin v3.4h, v2.4h, v1.4h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smin,
@@ -3636,6 +3696,18 @@ fn test_aarch64_binemit() {
"smin v30.8h, v20.8h, v10.8h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smin,
rd: writable_vreg(9),
rn: vreg(22),
rm: vreg(20),
size: VectorSize::Size32x2,
},
"C96EB40E",
"smin v9.2s, v22.2s, v20.2s",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smin,
@@ -3660,6 +3732,30 @@ fn test_aarch64_binemit() {
"umax v6.8b, v9.8b, v8.8b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umax,
rd: writable_vreg(5),
rn: vreg(15),
rm: vreg(8),
size: VectorSize::Size8x16,
},
"E565286E",
"umax v5.16b, v15.16b, v8.16b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umax,
rd: writable_vreg(12),
rn: vreg(14),
rm: vreg(3),
size: VectorSize::Size16x4,
},
"CC65632E",
"umax v12.4h, v14.4h, v3.4h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umax,
@@ -3672,6 +3768,18 @@ fn test_aarch64_binemit() {
"umax v11.8h, v13.8h, v2.8h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umax,
rd: writable_vreg(9),
rn: vreg(13),
rm: vreg(15),
size: VectorSize::Size32x2,
},
"A965AF2E",
"umax v9.2s, v13.2s, v15.2s",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Umax,
@@ -3684,6 +3792,18 @@ fn test_aarch64_binemit() {
"umax v8.4s, v12.4s, v14.4s",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smax,
rd: writable_vreg(7),
rn: vreg(8),
rm: vreg(9),
size: VectorSize::Size8x8,
},
"0765290E",
"smax v7.8b, v8.8b, v9.8b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smax,
@@ -3696,6 +3816,18 @@ fn test_aarch64_binemit() {
"smax v6.16b, v9.16b, v8.16b",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smax,
rd: writable_vreg(11),
rn: vreg(12),
rm: vreg(13),
size: VectorSize::Size16x4,
},
"8B656D0E",
"smax v11.4h, v12.4h, v13.4h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smax,
@@ -3708,6 +3840,18 @@ fn test_aarch64_binemit() {
"smax v11.8h, v13.8h, v2.8h",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smax,
rd: writable_vreg(14),
rn: vreg(16),
rm: vreg(18),
size: VectorSize::Size32x2,
},
"0E66B20E",
"smax v14.2s, v16.2s, v18.2s",
));
insns.push((
Inst::VecRRR {
alu_op: VecALUOp::Smax,

14
cranelift/codegen/src/isa/aarch64/lower.isle
View File

@@ -499,6 +499,20 @@
(result Reg (msub $I64 div y64 x64)))
result))
;;; Rules for integer min/max: umin, imin, umax, imax ;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type ty @ (not_i64x2) (imin x y)))
(vec_rrr (VecALUOp.Smin) x y (vector_size ty)))
(rule (lower (has_type ty @ (not_i64x2) (umin x y)))
(vec_rrr (VecALUOp.Umin) x y (vector_size ty)))
(rule (lower (has_type ty @ (not_i64x2) (imax x y)))
(vec_rrr (VecALUOp.Smax) x y (vector_size ty)))
(rule (lower (has_type ty @ (not_i64x2) (umax x y)))
(vec_rrr (VecALUOp.Umax) x y (vector_size ty)))
;;;; Rules for `uextend` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; General rule for extending input to an output which fits in a single

29
cranelift/codegen/src/isa/aarch64/lower_inst.rs
View File

@@ -1245,34 +1245,7 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
ctx.emit(Inst::gen_move(dst.regs()[1], src_hi, I64));
}
Opcode::Imax | Opcode::Umax | Opcode::Umin | Opcode::Imin => {
let ty = ty.unwrap();
if !ty.is_vector() || ty.lane_bits() == 64 {
return Err(CodegenError::Unsupported(format!(
"{}: Unsupported type: {:?}",
op, ty
)));
}
let alu_op = match op {
Opcode::Umin => VecALUOp::Umin,
Opcode::Imin => VecALUOp::Smin,
Opcode::Umax => VecALUOp::Umax,
Opcode::Imax => VecALUOp::Smax,
_ => unreachable!(),
};
let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
let rm = put_input_in_reg(ctx, inputs[1], NarrowValueMode::None);
ctx.emit(Inst::VecRRR {
alu_op,
rd,
rn,
rm,
size: VectorSize::from_ty(ty),
});
}
Opcode::Imax | Opcode::Umax | Opcode::Umin | Opcode::Imin => implemented_in_isle(ctx),
Opcode::IaddPairwise => implemented_in_isle(ctx),