language-servers/roc
1
0
Fork 0
mirror of https://github.com/roc-lang/roc.git synced 2025-09-29 06:44:46 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 7

Merge pull request #5020 from roc-lang/dev-unbox-u32

Browse source 
dev backend: reading and writing 32, 16 and 8-bit values
This commit is contained in:
Folkert de Vries 2023-02-21 17:33:21 +01:00 • committed by GitHub
commit 34340de60c
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
18 changed files with 1862 additions and 706 deletions
Download patch file Download diff file Expand all files Collapse all files

456
crates/compiler/gen_dev/src/generic64/mod.rs
View file

@ -2,7 +2,7 @@ use crate::{
single_register_floats, single_register_int_builtins, single_register_integers, Backend, Env,
Relocation,
};
use bumpalo::collections::Vec;
use bumpalo::collections::{CollectIn, Vec};
use roc_builtins::bitcode::{self, FloatWidth, IntWidth};
use roc_collections::all::MutMap;
use roc_error_macros::internal_error;
@ -236,22 +236,67 @@ pub trait Assembler<GeneralReg: RegTrait, FloatReg: RegTrait>: Sized + Copy {
// base32 is similar to stack based instructions but they reference the base/frame pointer.
fn mov_freg64_base32(buf: &mut Vec<'_, u8>, dst: FloatReg, offset: i32);
fn mov_reg64_base32(buf: &mut Vec<'_, u8>, dst: GeneralReg, offset: i32);
fn mov_base32_freg64(buf: &mut Vec<'_, u8>, offset: i32, src: FloatReg);
fn mov_base32_reg64(buf: &mut Vec<'_, u8>, offset: i32, src: GeneralReg);
fn mov_reg64_base32(buf: &mut Vec<'_, u8>, dst: GeneralReg, offset: i32);
fn mov_reg32_base32(buf: &mut Vec<'_, u8>, dst: GeneralReg, offset: i32);
fn mov_reg16_base32(buf: &mut Vec<'_, u8>, dst: GeneralReg, offset: i32);
fn mov_reg8_base32(buf: &mut Vec<'_, u8>, dst: GeneralReg, offset: i32);
fn mov_base32_freg64(buf: &mut Vec<'_, u8>, offset: i32, src: FloatReg);
fn mov_base32_reg64(buf: &mut Vec<'_, u8>, offset: i32, src: GeneralReg);
fn mov_base32_reg32(buf: &mut Vec<'_, u8>, offset: i32, src: GeneralReg);
fn mov_base32_reg16(buf: &mut Vec<'_, u8>, offset: i32, src: GeneralReg);
fn mov_base32_reg8(buf: &mut Vec<'_, u8>, offset: i32, src: GeneralReg);
// move from memory (a pointer) to register
fn mov_reg64_mem64_offset32(
buf: &mut Vec<'_, u8>,
dst: GeneralReg,
src: GeneralReg,
offset: i32,
);
fn mov_reg32_mem32_offset32(
buf: &mut Vec<'_, u8>,
dst: GeneralReg,
src: GeneralReg,
offset: i32,
);
fn mov_reg16_mem16_offset32(
buf: &mut Vec<'_, u8>,
dst: GeneralReg,
src: GeneralReg,
offset: i32,
);
fn mov_reg8_mem8_offset32(buf: &mut Vec<'_, u8>, dst: GeneralReg, src: GeneralReg, offset: i32);
// move from register to memory
fn mov_mem64_offset32_reg64(
buf: &mut Vec<'_, u8>,
dst: GeneralReg,
offset: i32,
src: GeneralReg,
);
fn mov_mem32_offset32_reg32(
buf: &mut Vec<'_, u8>,
dst: GeneralReg,
offset: i32,
src: GeneralReg,
);
fn mov_mem16_offset32_reg16(
buf: &mut Vec<'_, u8>,
dst: GeneralReg,
offset: i32,
src: GeneralReg,
);
fn mov_mem8_offset32_reg8(buf: &mut Vec<'_, u8>, dst: GeneralReg, offset: i32, src: GeneralReg);
fn movesd_mem64_offset32_freg64(
buf: &mut Vec<'_, u8>,
ptr: GeneralReg,
offset: i32,
src: FloatReg,
);
/// Sign extends the data at `offset` with `size` as it copies it to `dst`
/// size must be less than or equal to 8.
@ -958,6 +1003,30 @@ impl<
}
}
fn build_num_sub_checked(
&mut self,
dst: &Symbol,
src1: &Symbol,
src2: &Symbol,
num_layout: &InLayout<'a>,
return_layout: &InLayout<'a>,
) {
let function_name = match self.interner().get(*num_layout) {
Layout::Builtin(Builtin::Int(width)) => &bitcode::NUM_SUB_CHECKED_INT[width],
Layout::Builtin(Builtin::Float(width)) => &bitcode::NUM_SUB_CHECKED_FLOAT[width],
Layout::Builtin(Builtin::Decimal) => bitcode::DEC_SUB_WITH_OVERFLOW,
x => internal_error!("NumSubChecked is not defined for {:?}", x),
};
self.build_fn_call(
dst,
function_name.to_string(),
&[*src1, *src2],
&[*num_layout, *num_layout],
return_layout,
)
}
fn build_num_mul(&mut self, dst: &Symbol, src1: &Symbol, src2: &Symbol, layout: &InLayout<'a>) {
use Builtin::Int;
@ -1112,6 +1181,16 @@ impl<
.load_to_general_reg(&mut self.buf, src2);
ASM::eq_reg64_reg64_reg64(&mut self.buf, dst_reg, src1_reg, src2_reg);
}
Layout::STR => {
// use a zig call
self.build_fn_call(
dst,
bitcode::STR_EQUAL.to_string(),
&[*src1, *src2],
&[Layout::STR, Layout::STR],
&Layout::BOOL,
)
}
x => todo!("NumEq: layout, {:?}", x),
}
}
@ -1128,6 +1207,23 @@ impl<
.load_to_general_reg(&mut self.buf, src2);
ASM::neq_reg64_reg64_reg64(&mut self.buf, dst_reg, src1_reg, src2_reg);
}
Layout::Builtin(Builtin::Str) => {
self.build_fn_call(
dst,
bitcode::STR_EQUAL.to_string(),
&[*src1, *src2],
&[Layout::STR, Layout::STR],
&Layout::BOOL,
);
// negate the result
let tmp = &Symbol::DEV_TMP;
let tmp_reg = self.storage_manager.claim_general_reg(&mut self.buf, tmp);
ASM::mov_reg64_imm64(&mut self.buf, tmp_reg, 164);
let dst_reg = self.storage_manager.load_to_general_reg(&mut self.buf, dst);
ASM::neq_reg64_reg64_reg64(&mut self.buf, dst_reg, dst_reg, tmp_reg);
}
x => todo!("NumNeq: layout, {:?}", x),
}
}
@ -1567,43 +1663,14 @@ impl<
ASM::add_reg64_reg64_reg64(buf, tmp, tmp, list_ptr);
let element_ptr = tmp;
match *ret_layout {
single_register_integers!() if ret_stack_size == 8 => {
let dst_reg = storage_manager.claim_general_reg(buf, dst);
ASM::mov_reg64_mem64_offset32(buf, dst_reg, element_ptr, 0);
}
single_register_floats!() => {
let dst_reg = storage_manager.claim_float_reg(buf, dst);
ASM::mov_freg64_freg64(buf, dst_reg, CC::FLOAT_RETURN_REGS[0]);
}
Layout::STR => {
// the `list_ptr` register is now unused, and we can use it as scratch space
let tmp_reg = list_ptr;
Self::unbox_str_or_list(
buf,
storage_manager,
*dst,
element_ptr,
tmp_reg,
);
}
other => {
//
match self.layout_interner.get(other) {
Layout::Boxed(_) => {
let dst_reg = storage_manager.claim_general_reg(buf, dst);
ASM::mov_reg64_reg64(buf, dst_reg, CC::GENERAL_RETURN_REGS[0]);
}
_ => {
todo!(
"cannot load {} from the heap yet",
self.layout_interner.dbg(other)
);
}
}
}
}
Self::ptr_read(
buf,
storage_manager,
self.layout_interner,
element_ptr,
*ret_layout,
*dst,
);
});
},
);
@ -1871,10 +1938,11 @@ impl<
fn create_array(
&mut self,
sym: &Symbol,
element_layout: &InLayout<'a>,
elements: &'a [ListLiteralElement<'a>],
element_in_layout: &InLayout<'a>,
elements: &[ListLiteralElement<'a>],
) {
let element_width = self.layout_interner.stack_size(*element_layout) as u64;
let element_layout = self.layout_interner.get(*element_in_layout);
let element_width = self.layout_interner.stack_size(*element_in_layout) as u64;
// load the total size of the data we want to store (excludes refcount)
let data_bytes_symbol = Symbol::DEV_TMP;
@ -1904,54 +1972,34 @@ impl<
.load_to_general_reg(&mut self.buf, &Symbol::DEV_TMP3);
// Copy everything into output array.
let mut elem_offset = 0;
let mut element_offset = 0;
for elem in elements {
// TODO: this could be a lot faster when loading large lists
// if we move matching on the element layout to outside this loop.
// It also greatly bloats the code here.
// Refactor this and switch to one external match.
// We also could make loadining indivitual literals much faster
let elem_sym = match elem {
ListLiteralElement::Symbol(sym) => sym,
let element_symbol = match elem {
ListLiteralElement::Symbol(sym) => *sym,
ListLiteralElement::Literal(lit) => {
self.load_literal(&Symbol::DEV_TMP, element_layout, lit);
&Symbol::DEV_TMP
self.load_literal(&Symbol::DEV_TMP, element_in_layout, lit);
Symbol::DEV_TMP
}
};
// TODO: Expand to all types.
match self.layout_interner.get(*element_layout) {
Layout::Builtin(Builtin::Int(IntWidth::I64 | IntWidth::U64) | Builtin::Bool) => {
let sym_reg = self
.storage_manager
.load_to_general_reg(&mut self.buf, elem_sym);
ASM::mov_mem64_offset32_reg64(&mut self.buf, ptr_reg, elem_offset, sym_reg);
}
_ if element_width == 0 => {}
_ if element_width > 8 => {
let (from_offset, size) = self.storage_manager.stack_offset_and_size(elem_sym);
debug_assert!(from_offset % 8 == 0);
debug_assert!(size % 8 == 0);
debug_assert_eq!(size as u64, element_width);
self.storage_manager.with_tmp_general_reg(
&mut self.buf,
|_storage_manager, buf, tmp_reg| {
for i in (0..size as i32).step_by(8) {
ASM::mov_reg64_base32(buf, tmp_reg, from_offset + i);
ASM::mov_mem64_offset32_reg64(
buf,
ptr_reg,
elem_offset + i,
tmp_reg,
);
}
},
);
}
x => todo!("copying data to list with layout, {:?}", x),
}
elem_offset += element_width as i32;
if elem_sym == &Symbol::DEV_TMP {
self.free_symbol(elem_sym);
Self::ptr_write(
&mut self.buf,
&mut self.storage_manager,
ptr_reg,
element_offset,
element_width,
element_layout,
element_symbol,
);
element_offset += element_width as i32;
if element_symbol == Symbol::DEV_TMP {
self.free_symbol(&element_symbol);
}
}
@ -2049,38 +2097,15 @@ impl<
let element_width = self.layout_interner.stack_size(element_layout) as u64;
let element_offset = 0;
// TODO: Expand to all types.
match self.layout_interner.get(element_layout) {
Layout::Builtin(Builtin::Int(IntWidth::I64 | IntWidth::U64)) => {
let sym_reg = self
.storage_manager
.load_to_general_reg(&mut self.buf, &value);
ASM::mov_mem64_offset32_reg64(&mut self.buf, ptr_reg, element_offset, sym_reg);
}
_ if element_width == 0 => {}
_ if element_width > 8 => {
let (from_offset, size) = self.storage_manager.stack_offset_and_size(&value);
debug_assert!(from_offset % 8 == 0);
debug_assert!(size % 8 == 0);
debug_assert_eq!(size as u64, element_width);
self.storage_manager.with_tmp_general_reg(
&mut self.buf,
|_storage_manager, buf, tmp_reg| {
// a crude memcpy
for i in (0..size as i32).step_by(8) {
ASM::mov_reg64_base32(buf, tmp_reg, from_offset + i);
ASM::mov_mem64_offset32_reg64(
buf,
ptr_reg,
element_offset + i,
tmp_reg,
);
}
},
);
}
x => todo!("copying data to list with layout, {:?}", x),
}
Self::ptr_write(
&mut self.buf,
&mut self.storage_manager,
ptr_reg,
element_offset,
element_width,
self.layout_interner.get(element_layout),
value,
);
if value == Symbol::DEV_TMP {
self.free_symbol(&value);
@ -2098,25 +2123,14 @@ impl<
.storage_manager
.load_to_general_reg(&mut self.buf, &ptr);
let ret_stack_size = self.layout_interner.stack_size(element_layout);
match element_layout {
single_register_integers!() if ret_stack_size == 8 => {
let dst_reg = self.storage_manager.claim_general_reg(&mut self.buf, &dst);
ASM::mov_reg64_mem64_offset32(&mut self.buf, dst_reg, ptr_reg, 0);
}
Layout::STR => {
self.storage_manager.with_tmp_general_reg(
&mut self.buf,
|storage_manager, buf, tmp_reg| {
Self::unbox_str_or_list(buf, storage_manager, dst, ptr_reg, tmp_reg);
},
);
}
_ => {
todo!("unboxing of {:?}", self.layout_interner.dbg(element_layout))
}
}
Self::ptr_read(
&mut self.buf,
&mut self.storage_manager,
self.layout_interner,
ptr_reg,
element_layout,
dst,
);
}
fn get_tag_id(&mut self, sym: &Symbol, structure: &Symbol, union_layout: &UnionLayout<'a>) {
@ -2165,6 +2179,11 @@ impl<
let val = *x;
ASM::mov_reg64_imm64(&mut self.buf, reg, i128::from_ne_bytes(val) as i64);
}
(Literal::Byte(x), Layout::Builtin(Builtin::Int(IntWidth::U8 | IntWidth::I8))) => {
let reg = self.storage_manager.claim_general_reg(&mut self.buf, sym);
let val = *x;
ASM::mov_reg64_imm64(&mut self.buf, reg, val as i64);
}
(Literal::Bool(x), Layout::Builtin(Builtin::Bool)) => {
let reg = self.storage_manager.claim_general_reg(&mut self.buf, sym);
let val = [*x as u8; 16];
@ -2180,33 +2199,45 @@ impl<
let val = *x as f32;
ASM::mov_freg32_imm32(&mut self.buf, &mut self.relocs, reg, val);
}
(Literal::Str(x), Layout::Builtin(Builtin::Str)) if x.len() < 24 => {
// Load small string.
self.storage_manager.with_tmp_general_reg(
&mut self.buf,
|storage_manager, buf, reg| {
let base_offset = storage_manager.claim_stack_area(sym, 24);
let mut bytes = [0; 24];
bytes[..x.len()].copy_from_slice(x.as_bytes());
bytes[23] = (x.len() as u8) | 0b1000_0000;
(Literal::Str(x), Layout::Builtin(Builtin::Str)) => {
if x.len() < 24 {
// Load small string.
self.storage_manager.with_tmp_general_reg(
&mut self.buf,
|storage_manager, buf, reg| {
let base_offset = storage_manager.claim_stack_area(sym, 24);
let mut bytes = [0; 24];
bytes[..x.len()].copy_from_slice(x.as_bytes());
bytes[23] = (x.len() as u8) | 0b1000_0000;
let mut num_bytes = [0; 8];
num_bytes.copy_from_slice(&bytes[..8]);
let num = i64::from_ne_bytes(num_bytes);
ASM::mov_reg64_imm64(buf, reg, num);
ASM::mov_base32_reg64(buf, base_offset, reg);
let mut num_bytes = [0; 8];
num_bytes.copy_from_slice(&bytes[..8]);
let num = i64::from_ne_bytes(num_bytes);
ASM::mov_reg64_imm64(buf, reg, num);
ASM::mov_base32_reg64(buf, base_offset, reg);
num_bytes.copy_from_slice(&bytes[8..16]);
let num = i64::from_ne_bytes(num_bytes);
ASM::mov_reg64_imm64(buf, reg, num);
ASM::mov_base32_reg64(buf, base_offset + 8, reg);
num_bytes.copy_from_slice(&bytes[8..16]);
let num = i64::from_ne_bytes(num_bytes);
ASM::mov_reg64_imm64(buf, reg, num);
ASM::mov_base32_reg64(buf, base_offset + 8, reg);
num_bytes.copy_from_slice(&bytes[16..]);
let num = i64::from_ne_bytes(num_bytes);
ASM::mov_reg64_imm64(buf, reg, num);
ASM::mov_base32_reg64(buf, base_offset + 16, reg);
},
);
num_bytes.copy_from_slice(&bytes[16..]);
let num = i64::from_ne_bytes(num_bytes);
ASM::mov_reg64_imm64(buf, reg, num);
ASM::mov_base32_reg64(buf, base_offset + 16, reg);
},
);
} else {
// load large string (pretend it's a `List U8`). We should move this data into
// the binary eventually because our RC algorithm won't free this value
let elements: Vec<_> = x
.as_bytes()
.iter()
.map(|b| ListLiteralElement::Literal(Literal::Byte(*b)))
.collect_in(self.storage_manager.env.arena);
self.create_array(sym, &Layout::U8, elements.into_bump_slice())
}
}
x => todo!("loading literal, {:?}", x),
}
@ -2494,6 +2525,115 @@ impl<
ASM::mov_base32_reg64(buf, base_offset + 16, tmp_reg);
}
fn ptr_read(
buf: &mut Vec<'a, u8>,
storage_manager: &mut StorageManager<'a, 'r, GeneralReg, FloatReg, ASM, CC>,
layout_interner: &STLayoutInterner<'a>,
ptr_reg: GeneralReg,
element_in_layout: InLayout<'a>,
dst: Symbol,
) {
match layout_interner.get(element_in_layout) {
Layout::Builtin(builtin) => match builtin {
Builtin::Int(int_width) => match int_width {
IntWidth::I128 | IntWidth::U128 => {
// can we treat this as 2 u64's?
todo!()
}
IntWidth::I64 | IntWidth::U64 => {
let dst_reg = storage_manager.claim_general_reg(buf, &dst);
ASM::mov_reg64_mem64_offset32(buf, dst_reg, ptr_reg, 0);
}
IntWidth::I32 | IntWidth::U32 => {
let dst_reg = storage_manager.claim_general_reg(buf, &dst);
ASM::mov_reg32_mem32_offset32(buf, dst_reg, ptr_reg, 0);
}
IntWidth::I16 | IntWidth::U16 => {
let dst_reg = storage_manager.claim_general_reg(buf, &dst);
ASM::mov_reg16_mem16_offset32(buf, dst_reg, ptr_reg, 0);
}
IntWidth::I8 | IntWidth::U8 => {
let dst_reg = storage_manager.claim_general_reg(buf, &dst);
ASM::mov_reg8_mem8_offset32(buf, dst_reg, ptr_reg, 0);
}
},
Builtin::Float(_) => {
let dst_reg = storage_manager.claim_float_reg(buf, &dst);
ASM::mov_freg64_freg64(buf, dst_reg, CC::FLOAT_RETURN_REGS[0]);
}
Builtin::Bool => {
// the same as an 8-bit integer
let dst_reg = storage_manager.claim_general_reg(buf, &dst);
ASM::mov_reg8_mem8_offset32(buf, dst_reg, ptr_reg, 0);
}
Builtin::Decimal => {
// same as 128-bit integer
}
Builtin::Str | Builtin::List(_) => {
storage_manager.with_tmp_general_reg(buf, |storage_manager, buf, tmp_reg| {
Self::unbox_str_or_list(buf, storage_manager, dst, ptr_reg, tmp_reg);
});
}
},
Layout::Boxed(_) => {
// the same as 64-bit integer (for 64-bit targets)
let dst_reg = storage_manager.claim_general_reg(buf, &dst);
ASM::mov_reg64_mem64_offset32(buf, dst_reg, ptr_reg, 0);
}
_ => todo!("unboxing of {:?}", layout_interner.dbg(element_in_layout)),
}
}
fn ptr_write(
buf: &mut Vec<'a, u8>,
storage_manager: &mut StorageManager<'a, 'r, GeneralReg, FloatReg, ASM, CC>,
ptr_reg: GeneralReg,
element_offset: i32,
element_width: u64,
element_layout: Layout<'a>,
value: Symbol,
) {
match element_layout {
Layout::Builtin(Builtin::Int(IntWidth::I64 | IntWidth::U64)) => {
let sym_reg = storage_manager.load_to_general_reg(buf, &value);
ASM::mov_mem64_offset32_reg64(buf, ptr_reg, element_offset, sym_reg);
}
Layout::Builtin(Builtin::Int(IntWidth::I32 | IntWidth::U32)) => {
let sym_reg = storage_manager.load_to_general_reg(buf, &value);
ASM::mov_mem32_offset32_reg32(buf, ptr_reg, element_offset, sym_reg);
}
Layout::Builtin(Builtin::Int(IntWidth::I16 | IntWidth::U16)) => {
let sym_reg = storage_manager.load_to_general_reg(buf, &value);
ASM::mov_mem16_offset32_reg16(buf, ptr_reg, element_offset, sym_reg);
}
Layout::Builtin(Builtin::Int(IntWidth::I8 | IntWidth::U8) | Builtin::Bool) => {
let sym_reg = storage_manager.load_to_general_reg(buf, &value);
ASM::mov_mem8_offset32_reg8(buf, ptr_reg, element_offset, sym_reg);
}
Layout::Builtin(Builtin::Float(FloatWidth::F64 | FloatWidth::F32)) => {
let sym_reg = storage_manager.load_to_float_reg(buf, &value);
ASM::movesd_mem64_offset32_freg64(buf, ptr_reg, element_offset, sym_reg);
}
_ if element_width == 0 => {}
_ if element_width > 8 => {
let (from_offset, size) = storage_manager.stack_offset_and_size(&value);
debug_assert!(from_offset % 8 == 0);
debug_assert!(size % 8 == 0);
debug_assert_eq!(size as u64, element_width);
storage_manager.with_tmp_general_reg(buf, |_storage_manager, buf, tmp_reg| {
// a crude memcpy
for i in (0..size as i32).step_by(8) {
ASM::mov_reg64_base32(buf, tmp_reg, from_offset + i);
ASM::mov_mem64_offset32_reg64(buf, ptr_reg, element_offset + i, tmp_reg);
}
});
}
x => todo!("copying data to list with layout, {:?}", x),
}
}
/// Updates a jump instruction to a new offset and returns the number of bytes written.
fn update_jmp_imm32_offset(
&mut self,