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

[AArch64] Merge Fcmp32 and Fcmp64 (#4032)

Browse Source 
Copyright (c) 2022, Arm Limited.
This commit is contained in:
Sam Parker
2022-04-14 23:39:43 +01:00
committed by GitHub
parent a40b5c3985
commit cf533a8041
8 changed files with 202 additions and 220 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
cranelift/codegen/src/isa/aarch64/inst.isle
View File

@@ -327,13 +327,9 @@
(rm Reg) (rm Reg)
(ra Reg)) (ra Reg))
;; FPU comparison, single-precision (32 bit). ;; FPU comparison.
(FpuCmp32 (FpuCmp
(rn Reg) (size ScalarSize)
(rm Reg))
;; FPU comparison, double-precision (64 bit).
(FpuCmp64
(rn Reg) (rn Reg)
(rm Reg)) (rm Reg))

9
cranelift/codegen/src/isa/aarch64/inst/emit.rs
View File

@@ -2008,15 +2008,10 @@ impl MachInstEmit for Inst {
assert_eq!(machreg_to_vec(rn2), (machreg_to_vec(rn) + 1) % 32); assert_eq!(machreg_to_vec(rn2), (machreg_to_vec(rn) + 1) % 32);
sink.put4(enc_tbl(is_extension, 0b01, rd, rn, rm)); sink.put4(enc_tbl(is_extension, 0b01, rd, rn, rm));
} }
&Inst::FpuCmp32 { rn, rm } => { &Inst::FpuCmp { size, rn, rm } => {
let rn = allocs.next(rn); let rn = allocs.next(rn);
let rm = allocs.next(rm); let rm = allocs.next(rm);
sink.put4(enc_fcmp(ScalarSize::Size32, rn, rm)); sink.put4(enc_fcmp(size, rn, rm));
}
&Inst::FpuCmp64 { rn, rm } => {
let rn = allocs.next(rn);
let rm = allocs.next(rm);
sink.put4(enc_fcmp(ScalarSize::Size64, rn, rm));
} }
&Inst::FpuToInt { op, rd, rn } => { &Inst::FpuToInt { op, rd, rn } => {
let rd = allocs.next_writable(rd); let rd = allocs.next_writable(rd);

6
cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs
View File

@@ -5821,7 +5821,8 @@ fn test_aarch64_binemit() {
)); ));
insns.push(( insns.push((
Inst::FpuCmp32 { Inst::FpuCmp {
size: ScalarSize::Size32,
rn: vreg(23), rn: vreg(23),
rm: vreg(24), rm: vreg(24),
}, },
@@ -5830,7 +5831,8 @@ fn test_aarch64_binemit() {
)); ));
insns.push(( insns.push((
Inst::FpuCmp64 { Inst::FpuCmp {
size: ScalarSize::Size64,
rn: vreg(23), rn: vreg(23),
rm: vreg(24), rm: vreg(24),
}, },

13
cranelift/codegen/src/isa/aarch64/inst/mod.rs
View File

@@ -807,7 +807,7 @@ fn aarch64_get_operands<F: Fn(VReg) -> VReg>(inst: &Inst, collector: &mut Operan
collector.reg_use(rn); collector.reg_use(rn);
collector.reg_use(rm); collector.reg_use(rm);
} }
&Inst::FpuCmp32 { rn, rm } | &Inst::FpuCmp64 { rn, rm } => { &Inst::FpuCmp { rn, rm, .. } => {
collector.reg_use(rn); collector.reg_use(rn);
collector.reg_use(rm); collector.reg_use(rm);
} }
@@ -1765,14 +1765,9 @@ impl Inst {
let ra = pretty_print_vreg_scalar(ra, size, allocs); let ra = pretty_print_vreg_scalar(ra, size, allocs);
format!("{} {}, {}, {}, {}", op, rd, rn, rm, ra) format!("{} {}, {}, {}, {}", op, rd, rn, rm, ra)
} }
&Inst::FpuCmp32 { rn, rm } => { &Inst::FpuCmp { size, rn, rm } => {
let rn = pretty_print_vreg_scalar(rn, ScalarSize::Size32, allocs); let rn = pretty_print_vreg_scalar(rn, size, allocs);
let rm = pretty_print_vreg_scalar(rm, ScalarSize::Size32, allocs); let rm = pretty_print_vreg_scalar(rm, size, allocs);
format!("fcmp {}, {}", rn, rm)
}
&Inst::FpuCmp64 { rn, rm } => {
let rn = pretty_print_vreg_scalar(rn, ScalarSize::Size64, allocs);
let rm = pretty_print_vreg_scalar(rm, ScalarSize::Size64, allocs);
format!("fcmp {}, {}", rn, rm) format!("fcmp {}, {}", rn, rm)
} }
&Inst::FpuLoad32 { rd, ref mem, .. } => { &Inst::FpuLoad32 { rd, ref mem, .. } => {

15
cranelift/codegen/src/isa/aarch64/lower.rs
View File

@@ -1448,19 +1448,14 @@ pub(crate) fn lower_icmp<C: LowerCtx<I = Inst>>(
pub(crate) fn lower_fcmp_or_ffcmp_to_flags<C: LowerCtx<I = Inst>>(ctx: &mut C, insn: IRInst) { pub(crate) fn lower_fcmp_or_ffcmp_to_flags<C: LowerCtx<I = Inst>>(ctx: &mut C, insn: IRInst) {
let ty = ctx.input_ty(insn, 0); let ty = ctx.input_ty(insn, 0);
let bits = ty_bits(ty);
let inputs = [InsnInput { insn, input: 0 }, InsnInput { insn, input: 1 }]; let inputs = [InsnInput { insn, input: 0 }, InsnInput { insn, input: 1 }];
let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None); let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
let rm = put_input_in_reg(ctx, inputs[1], NarrowValueMode::None); let rm = put_input_in_reg(ctx, inputs[1], NarrowValueMode::None);
match bits { ctx.emit(Inst::FpuCmp {
32 => { size: ScalarSize::from_ty(ty),
ctx.emit(Inst::FpuCmp32 { rn, rm }); rn,
} rm,
64 => { });
ctx.emit(Inst::FpuCmp64 { rn, rm });
}
_ => panic!("Unknown float size"),
}
} }
/// Materialize a boolean value into a register from the flags /// Materialize a boolean value into a register from the flags

2
cranelift/codegen/src/isa/aarch64/lower/isle/generated_code.manifest
View File

@@ -1,4 +1,4 @@
src/clif.isle 443b34b797fc8ace src/clif.isle 443b34b797fc8ace
src/prelude.isle afd037c4d91c875c src/prelude.isle afd037c4d91c875c
src/isa/aarch64/inst.isle f7f03d5ea5411344 src/isa/aarch64/inst.isle 77984cc33a05be7
src/isa/aarch64/lower.isle 71c7e603b0e4bdef src/isa/aarch64/lower.isle 71c7e603b0e4bdef

309
cranelift/codegen/src/isa/aarch64/lower/isle/generated_code.rs generated
View File

File diff suppressed because it is too large Load Diff

58
cranelift/codegen/src/isa/aarch64/lower_inst.rs
View File

@@ -735,20 +735,11 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap(); let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
if !ty.is_vector() { if !ty.is_vector() {
match ty_bits(ty) { ctx.emit(Inst::FpuCmp {
32 => { size: ScalarSize::from_ty(ty),
ctx.emit(Inst::FpuCmp32 { rn, rm }); rn,
} rm,
64 => { });
ctx.emit(Inst::FpuCmp64 { rn, rm });
}
_ => {
return Err(CodegenError::Unsupported(format!(
"Fcmp: Unsupported type: {:?}",
ty
)))
}
}
materialize_bool_result(ctx, insn, rd, cond); materialize_bool_result(ctx, insn, rd, cond);
} else { } else {
lower_vector_compare(ctx, rd, rn, rm, ty, cond)?; lower_vector_compare(ctx, rd, rn, rm, ty, cond)?;
@@ -1076,7 +1067,8 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
rd: tmp, rd: tmp,
rn: tmp.to_reg(), rn: tmp.to_reg(),
}); });
ctx.emit(Inst::FpuCmp64 { ctx.emit(Inst::FpuCmp {
size: ScalarSize::Size64,
rn: tmp.to_reg(), rn: tmp.to_reg(),
rm: tmp.to_reg(), rm: tmp.to_reg(),
}); });
@@ -1672,8 +1664,12 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
size, size,
}); });
} else { } else {
ctx.emit(Inst::FpuCmp {
size: ScalarSize::from_ty(lane_type),
rn: ra,
rm: rb,
});
if lane_type == F32 { if lane_type == F32 {
ctx.emit(Inst::FpuCmp32 { rn: ra, rm: rb });
ctx.emit(Inst::FpuCSel32 { ctx.emit(Inst::FpuCSel32 {
rd, rd,
rn, rn,
@@ -1681,7 +1677,6 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
cond: Cond::Gt, cond: Cond::Gt,
}); });
} else { } else {
ctx.emit(Inst::FpuCmp64 { rn: ra, rm: rb });
ctx.emit(Inst::FpuCSel64 { ctx.emit(Inst::FpuCSel64 {
rd, rd,
rn, rn,
@@ -1897,11 +1892,11 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
// in-bounds conversion, per wasm semantics. // in-bounds conversion, per wasm semantics.
// Check that the input is not a NaN. // Check that the input is not a NaN.
if in_bits == 32 { ctx.emit(Inst::FpuCmp {
ctx.emit(Inst::FpuCmp32 { rn, rm: rn }); size: ScalarSize::from_ty(input_ty),
} else { rn,
ctx.emit(Inst::FpuCmp64 { rn, rm: rn }); rm: rn,
} });
let trap_code = TrapCode::BadConversionToInteger; let trap_code = TrapCode::BadConversionToInteger;
ctx.emit(Inst::TrapIf { ctx.emit(Inst::TrapIf {
trap_code, trap_code,
@@ -1950,7 +1945,8 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
// >= low_bound // >= low_bound
lower_constant_f32(ctx, tmp, low_bound); lower_constant_f32(ctx, tmp, low_bound);
ctx.emit(Inst::FpuCmp32 { ctx.emit(Inst::FpuCmp {
size: ScalarSize::Size32,
rn, rn,
rm: tmp.to_reg(), rm: tmp.to_reg(),
}); });
@@ -1962,7 +1958,8 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
// <= high_bound // <= high_bound
lower_constant_f32(ctx, tmp, high_bound); lower_constant_f32(ctx, tmp, high_bound);
ctx.emit(Inst::FpuCmp32 { ctx.emit(Inst::FpuCmp {
size: ScalarSize::Size32,
rn, rn,
rm: tmp.to_reg(), rm: tmp.to_reg(),
}); });
@@ -2003,7 +2000,8 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
// >= low_bound // >= low_bound
lower_constant_f64(ctx, tmp, low_bound); lower_constant_f64(ctx, tmp, low_bound);
ctx.emit(Inst::FpuCmp64 { ctx.emit(Inst::FpuCmp {
size: ScalarSize::Size64,
rn, rn,
rm: tmp.to_reg(), rm: tmp.to_reg(),
}); });
@@ -2015,7 +2013,8 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
// <= high_bound // <= high_bound
lower_constant_f64(ctx, tmp, high_bound); lower_constant_f64(ctx, tmp, high_bound);
ctx.emit(Inst::FpuCmp64 { ctx.emit(Inst::FpuCmp {
size: ScalarSize::Size64,
rn, rn,
rm: tmp.to_reg(), rm: tmp.to_reg(),
}); });
@@ -2180,8 +2179,12 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
lower_constant_f64(ctx, rtmp1, 0.0); lower_constant_f64(ctx, rtmp1, 0.0);
} }
} }
ctx.emit(Inst::FpuCmp {
size: ScalarSize::from_ty(in_ty),
rn,
rm: rn,
});
if in_bits == 32 { if in_bits == 32 {
ctx.emit(Inst::FpuCmp32 { rn, rm: rn });
ctx.emit(Inst::FpuCSel32 { ctx.emit(Inst::FpuCSel32 {
rd: rtmp2, rd: rtmp2,
rn: rtmp1.to_reg(), rn: rtmp1.to_reg(),
@@ -2189,7 +2192,6 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
cond: Cond::Ne, cond: Cond::Ne,
}); });
} else { } else {
ctx.emit(Inst::FpuCmp64 { rn, rm: rn });
ctx.emit(Inst::FpuCSel64 { ctx.emit(Inst::FpuCSel64 {
rd: rtmp2, rd: rtmp2,
rn: rtmp1.to_reg(), rn: rtmp1.to_reg(),