fluencelabs/wasmer
1
0
Fork 0
mirror of https://github.com/fluencelabs/wasmer synced 2025-06-15 09:51:21 +00:00
Code Issues Projects Releases Wiki Activity

Implement F64Min and F64Max.

Browse Source
This commit is contained in:
Nick Lewycky
2019-10-14 14:15:18 -07:00
parent 06ffb00deb
commit 26a4f073f0
3 changed files with 235 additions and 222 deletions
Download Patch File Download Diff File Expand all files Collapse all files

217
lib/singlepass-backend/src/codegen_x64.rs
View File

@ -3183,18 +3183,211 @@ impl FunctionCodeGenerator<CodegenError> for X64FunctionCode {
&mut self.value_stack,
Assembler::emit_vdivsd,
),
Operator::F64Max => Self::emit_fp_binop_avx(
a,
&mut self.machine,
&mut self.value_stack,
Assembler::emit_vmaxsd,
),
Operator::F64Min => Self::emit_fp_binop_avx(
a,
&mut self.machine,
&mut self.value_stack,
Assembler::emit_vminsd,
),
Operator::F64Max => {
let src2 =
get_location_released(a, &mut self.machine, self.value_stack.pop().unwrap());
let src1 =
get_location_released(a, &mut self.machine, self.value_stack.pop().unwrap());
let ret = self.machine.acquire_locations(
a,
&[(WpType::F64, MachineValue::WasmStack(self.value_stack.len()))],
false,
)[0];
self.value_stack.push(ret);
let tmp1 = self.machine.acquire_temp_xmm().unwrap();
let tmp2 = self.machine.acquire_temp_xmm().unwrap();
let tmpg1 = self.machine.acquire_temp_gpr().unwrap();
let tmpg2 = self.machine.acquire_temp_gpr().unwrap();
let src1 = match src1 {
Location::XMM(x) => x,
Location::GPR(_) | Location::Memory(_, _) => {
a.emit_mov(Size::S64, src1, Location::XMM(tmp1));
tmp1
}
Location::Imm32(_) => {
a.emit_mov(Size::S32, src1, Location::GPR(tmpg1));
a.emit_mov(Size::S32, Location::GPR(tmpg1), Location::XMM(tmp1));
tmp1
}
Location::Imm64(_) => {
a.emit_mov(Size::S64, src1, Location::GPR(tmpg1));
a.emit_mov(Size::S64, Location::GPR(tmpg1), Location::XMM(tmp1));
tmp1
}
_ => unreachable!(),
};
let src2 = match src2 {
Location::XMM(x) => x,
Location::GPR(_) | Location::Memory(_, _) => {
a.emit_mov(Size::S64, src2, Location::XMM(tmp2));
tmp2
}
Location::Imm32(_) => {
a.emit_mov(Size::S32, src2, Location::GPR(tmpg1));
a.emit_mov(Size::S32, Location::GPR(tmpg1), Location::XMM(tmp2));
tmp2
}
Location::Imm64(_) => {
a.emit_mov(Size::S64, src2, Location::GPR(tmpg1));
a.emit_mov(Size::S64, Location::GPR(tmpg1), Location::XMM(tmp2));
tmp2
}
_ => unreachable!(),
};
let tmp_xmm1 = XMM::XMM8;
let tmp_xmm2 = XMM::XMM9;
let tmp_xmm3 = XMM::XMM10;
static CANONICAL_NAN: u128 = 0x7FF8_0000_0000_0000;
a.emit_mov(Size::S64, Location::XMM(src1), Location::GPR(tmpg1));
a.emit_mov(Size::S64, Location::XMM(src2), Location::GPR(tmpg2));
a.emit_cmp(Size::S64, Location::GPR(tmpg2), Location::GPR(tmpg1));
a.emit_vmaxsd(src1, XMMOrMemory::XMM(src2), tmp_xmm1);
let label1 = a.get_label();
let label2 = a.get_label();
a.emit_jmp(Condition::NotEqual, label1);
a.emit_vmovapd(XMMOrMemory::XMM(tmp_xmm1), XMMOrMemory::XMM(tmp_xmm2));
a.emit_jmp(Condition::None, label2);
a.emit_label(label1);
a.emit_vxorpd(tmp_xmm2, XMMOrMemory::XMM(tmp_xmm2), tmp_xmm2);
a.emit_label(label2);
a.emit_vcmpeqsd(src1, XMMOrMemory::XMM(src2), tmp_xmm3);
a.emit_vblendvpd(tmp_xmm3, XMMOrMemory::XMM(tmp_xmm2), tmp_xmm1, tmp_xmm1);
a.emit_vcmpunordsd(src1, XMMOrMemory::XMM(src2), src1);
// load float canonical nan
a.emit_mov(
Size::S64,
Location::Imm64((&CANONICAL_NAN as *const u128) as u64),
Location::GPR(tmpg1),
);
a.emit_mov(Size::S64, Location::Memory(tmpg1, 0), Location::XMM(src2));
a.emit_vblendvpd(src1, XMMOrMemory::XMM(src2), tmp_xmm1, src1);
match ret {
Location::XMM(x) => {
a.emit_vmovapd(XMMOrMemory::XMM(src1), XMMOrMemory::XMM(x));
}
Location::Memory(_, _) | Location::GPR(_) => {
a.emit_mov(Size::S64, Location::XMM(src1), ret);
}
_ => unreachable!(),
}
self.machine.release_temp_gpr(tmpg2);
self.machine.release_temp_gpr(tmpg1);
self.machine.release_temp_xmm(tmp2);
self.machine.release_temp_xmm(tmp1);
}
Operator::F64Min => {
let src2 =
get_location_released(a, &mut self.machine, self.value_stack.pop().unwrap());
let src1 =
get_location_released(a, &mut self.machine, self.value_stack.pop().unwrap());
let ret = self.machine.acquire_locations(
a,
&[(WpType::F64, MachineValue::WasmStack(self.value_stack.len()))],
false,
)[0];
self.value_stack.push(ret);
let tmp1 = self.machine.acquire_temp_xmm().unwrap();
let tmp2 = self.machine.acquire_temp_xmm().unwrap();
let tmpg1 = self.machine.acquire_temp_gpr().unwrap();
let tmpg2 = self.machine.acquire_temp_gpr().unwrap();
let src1 = match src1 {
Location::XMM(x) => x,
Location::GPR(_) | Location::Memory(_, _) => {
a.emit_mov(Size::S64, src1, Location::XMM(tmp1));
tmp1
}
Location::Imm32(_) => {
a.emit_mov(Size::S32, src1, Location::GPR(tmpg1));
a.emit_mov(Size::S32, Location::GPR(tmpg1), Location::XMM(tmp1));
tmp1
}
Location::Imm64(_) => {
a.emit_mov(Size::S64, src1, Location::GPR(tmpg1));
a.emit_mov(Size::S64, Location::GPR(tmpg1), Location::XMM(tmp1));
tmp1
}
_ => unreachable!(),
};
let src2 = match src2 {
Location::XMM(x) => x,
Location::GPR(_) | Location::Memory(_, _) => {
a.emit_mov(Size::S64, src2, Location::XMM(tmp2));
tmp2
}
Location::Imm32(_) => {
a.emit_mov(Size::S32, src2, Location::GPR(tmpg1));
a.emit_mov(Size::S32, Location::GPR(tmpg1), Location::XMM(tmp2));
tmp2
}
Location::Imm64(_) => {
a.emit_mov(Size::S64, src2, Location::GPR(tmpg1));
a.emit_mov(Size::S64, Location::GPR(tmpg1), Location::XMM(tmp2));
tmp2
}
_ => unreachable!(),
};
let tmp_xmm1 = XMM::XMM8;
let tmp_xmm2 = XMM::XMM9;
let tmp_xmm3 = XMM::XMM10;
static NEG_ZERO: u128 = 0x8000_0000_0000_0000;
static CANONICAL_NAN: u128 = 0x7FF8_0000_0000_0000;
a.emit_mov(Size::S64, Location::XMM(src1), Location::GPR(tmpg1));
a.emit_mov(Size::S64, Location::XMM(src2), Location::GPR(tmpg2));
a.emit_cmp(Size::S64, Location::GPR(tmpg2), Location::GPR(tmpg1));
a.emit_vminsd(src1, XMMOrMemory::XMM(src2), tmp_xmm1);
let label1 = a.get_label();
let label2 = a.get_label();
a.emit_jmp(Condition::NotEqual, label1);
a.emit_vmovapd(XMMOrMemory::XMM(tmp_xmm1), XMMOrMemory::XMM(tmp_xmm2));
a.emit_jmp(Condition::None, label2);
a.emit_label(label1);
// load float -0.0
a.emit_mov(
Size::S64,
Location::Imm64((&NEG_ZERO as *const u128) as u64),
Location::GPR(tmpg1),
);
a.emit_mov(
Size::S64,
Location::Memory(tmpg1, 0),
Location::XMM(tmp_xmm2),
);
a.emit_label(label2);
a.emit_vcmpeqsd(src1, XMMOrMemory::XMM(src2), tmp_xmm3);
a.emit_vblendvpd(tmp_xmm3, XMMOrMemory::XMM(tmp_xmm2), tmp_xmm1, tmp_xmm1);
a.emit_vcmpunordsd(src1, XMMOrMemory::XMM(src2), src1);
// load float canonical nan
a.emit_mov(
Size::S64,
Location::Imm64((&CANONICAL_NAN as *const u128) as u64),
Location::GPR(tmpg1),
);
a.emit_mov(Size::S64, Location::Memory(tmpg1, 0), Location::XMM(src2));
a.emit_vblendvpd(src1, XMMOrMemory::XMM(src2), tmp_xmm1, src1);
match ret {
Location::XMM(x) => {
a.emit_vmovaps(XMMOrMemory::XMM(src1), XMMOrMemory::XMM(x));
}
Location::Memory(_, _) | Location::GPR(_) => {
a.emit_mov(Size::S64, Location::XMM(src1), ret);
}
_ => unreachable!(),
}
self.machine.release_temp_gpr(tmpg2);
self.machine.release_temp_gpr(tmpg1);
self.machine.release_temp_xmm(tmp2);
self.machine.release_temp_xmm(tmp1);
}
Operator::F64Eq => Self::emit_fp_cmpop_avx(
a,
&mut self.machine,