M-Labs
/
nac3
8
0
Fork 5
Code Issues 87 Pull Requests 3 Releases Wiki Activity

Fix behavior of with parallel #347

Merged
sb10q merged 5 commits from issue-298 into master 2023-11-01 13:19:52 +08:00
Conversation 0 Commits 5 Files Changed 3
3 changed files with 194 additions and 102 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
Cargo.lock generated
View File

@ -4,9 +4,9 @@ version = 3
[[package]]
name = "ahash"
version = "0.7.6"
version = "0.7.7"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "fcb51a0695d8f838b1ee009b3fbf66bda078cd64590202a864a8f3e8c4315c47"
checksum = "5a824f2aa7e75a0c98c5a504fceb80649e9c35265d44525b5f94de4771a395cd"
dependencies = [
"getrandom",
"once_cell",
@ -15,14 +15,15 @@ dependencies = [
[[package]]
name = "ahash"
version = "0.8.3"
version = "0.8.5"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "2c99f64d1e06488f620f932677e24bc6e2897582980441ae90a671415bd7ec2f"
checksum = "cd7d5a2cecb58716e47d67d5703a249964b14c7be1ec3cad3affc295b2d1c35d"
dependencies = [
"cfg-if",
"getrandom",
"once_cell",
"version_check",
"zerocopy",
]
[[package]]
@ -387,7 +388,7 @@ version = "0.11.2"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "ab5ef0d4909ef3724cc8cce6ccc8572c5c817592e9285f5464f8e86f8bd3726e"
dependencies = [
"ahash 0.7.6",
"ahash 0.7.7",
]
[[package]]
@ -674,7 +675,7 @@ dependencies = [
name = "nac3parser"
version = "0.1.2"
dependencies = [
"ahash 0.8.3",
"ahash 0.8.5",
"insta",
"lalrpop",
"lalrpop-util",
@ -1547,3 +1548,23 @@ checksum = "56c1936c4cc7a1c9ab21a1ebb602eb942ba868cbd44a99cb7cdc5892335e1c85"
dependencies = [
"linked-hash-map",
]
[[package]]
name = "zerocopy"
version = "0.7.11"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "4c19fae0c8a9efc6a8281f2e623db8af1db9e57852e04cde3e754dd2dc29340f"
dependencies = [
"zerocopy-derive",
]
[[package]]
name = "zerocopy-derive"
version = "0.7.11"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "fc56589e9ddd1f1c28d4b4b5c773ce232910a6bb67a70133d61c9e347585efe9"
dependencies = [
"proc-macro2",
"quote",
"syn 2.0.38",
]

163
nac3artiq/src/codegen.rs
View File

@ -28,9 +28,17 @@ use std::{
pub struct ArtiqCodeGenerator<'a> {
name: String,
/// The size of a `size_t` variable in bits.
size_t: u32,
/// Monotonic counter for naming `start`/`stop` variables used by `with parallel` blocks.
name_counter: u32,
/// Variable for tracking the start of a `with parallel` block.
start: Option<Expr<Option<Type>>>,
/// Variable for tracking the end of a `with parallel` block.
end: Option<Expr<Option<Type>>>,
timeline: &'a (dyn TimeFns + Sync),
}
@ -42,7 +50,81 @@ impl<'a> ArtiqCodeGenerator<'a> {
timeline: &'a (dyn TimeFns + Sync),
) -> ArtiqCodeGenerator<'a> {
assert!(size_t == 32 || size_t == 64);
ArtiqCodeGenerator { name, size_t, name_counter: 0, start: None, end: None, timeline }
ArtiqCodeGenerator {
name,
size_t,
name_counter: 0,
start: None,
end: None,
timeline,
}
}
/// If the generator is currently in a direct-`parallel` block context, emits IR that resets the
/// position of the timeline to the initial timeline position before entering the `parallel`
/// block.
///
/// Direct-`parallel` block context refers to when the generator is generating statements whose
/// closest parent `with` statement is a `with parallel` block.
fn timeline_reset_start<'ctx, 'b>(
&mut self,
ctx: &mut CodeGenContext<'ctx, 'b>
) -> Result<(), String> {
if let Some(start) = self.start.clone() {
let start_val = self.gen_expr(ctx, &start)?
.unwrap()
.to_basic_value_enum(ctx, self, start.custom.unwrap())?;
self.timeline.emit_at_mu(ctx, start_val);
}
Ok(())
}
/// If the generator is currently in a `parallel` block context, emits IR that updates the
/// maximum end position of the `parallel` block as specified by the timeline `end` value.
///
/// In general the `end` parameter should be set to `self.end` for updating the maximum end
/// position for the current `parallel` block. Other values can be passed in to update the
/// maximum end position for other `parallel` blocks.
///
/// `parallel`-block context refers to when the generator is generating statements within a
/// (possibly indirect) `parallel` block.
///
/// * `store_name` - The LLVM value name for the pointer to `end`. `.addr` will be appended to
/// the end of the provided value name.
fn timeline_update_end_max<'ctx, 'b>(
&mut self,
ctx: &mut CodeGenContext<'ctx, 'b>,
end: Option<Expr<Option<Type>>>,
store_name: Option<&str>,
) -> Result<(), String> {
if let Some(end) = end {
let old_end = self.gen_expr(ctx, &end)?
.unwrap()
.to_basic_value_enum(ctx, self, end.custom.unwrap())?;
let now = self.timeline.emit_now_mu(ctx);
let smax = ctx.module.get_function("llvm.smax.i64").unwrap_or_else(|| {
let i64 = ctx.ctx.i64_type();
ctx.module.add_function(
"llvm.smax.i64",
i64.fn_type(&[i64.into(), i64.into()], false),
None,
)
});
let max = ctx
.builder
.build_call(smax, &[old_end.into(), now.into()], "smax")
.try_as_basic_value()
.left()
.unwrap();
let end_store = self.gen_store_target(
ctx,
&end,
store_name.map(|name| format!("{name}.addr")).as_deref())?;
ctx.builder.build_store(end_store, max);
}
Ok(())
}
}
@ -67,30 +149,10 @@ impl<'b> CodeGenerator for ArtiqCodeGenerator<'b> {
params: Vec<(Option<StrRef>, ValueEnum<'ctx>)>,
) -> Result<Option<BasicValueEnum<'ctx>>, String> {
let result = gen_call(self, ctx, obj, fun, params)?;
if let Some(end) = self.end.clone() {
let old_end = self.gen_expr(ctx, &end)?.unwrap().to_basic_value_enum(ctx, self, end.custom.unwrap())?;
let now = self.timeline.emit_now_mu(ctx);
let smax = ctx.module.get_function("llvm.smax.i64").unwrap_or_else(|| {
let i64 = ctx.ctx.i64_type();
ctx.module.add_function(
"llvm.smax.i64",
i64.fn_type(&[i64.into(), i64.into()], false),
None,
)
});
let max = ctx
.builder
.build_call(smax, &[old_end.into(), now.into()], "smax")
.try_as_basic_value()
.left()
.unwrap();
let end_store = self.gen_store_target(ctx, &end, Some("end_store.addr"))?;
ctx.builder.build_store(end_store, max);
}
if let Some(start) = self.start.clone() {
let start_val = self.gen_expr(ctx, &start)?.unwrap().to_basic_value_enum(ctx, self, start.custom.unwrap())?;
self.timeline.emit_at_mu(ctx, start_val);
}
self.timeline_update_end_max(ctx, self.end.clone(), Some("end"))?;
self.timeline_reset_start(ctx)?;
Ok(result)
}
@ -102,6 +164,7 @@ impl<'b> CodeGenerator for ArtiqCodeGenerator<'b> {
if let StmtKind::With { items, body, .. } = &stmt.node {
if items.len() == 1 && items[0].optional_vars.is_none() {
let item = &items[0];
// Behavior of parallel and sequential:
// Each function call (indirectly, can be inside a sequential block) within a parallel
// block will update the end variable to the maximum now_mu in the block.
@ -119,11 +182,15 @@ impl<'b> CodeGenerator for ArtiqCodeGenerator<'b> {
if id == &"parallel".into() {
let old_start = self.start.take();
let old_end = self.end.take();
let now = if let Some(old_start) = &old_start {
self.gen_expr(ctx, old_start)?.unwrap().to_basic_value_enum(ctx, self, old_start.custom.unwrap())?
self.gen_expr(ctx, old_start)?
.unwrap()
.to_basic_value_enum(ctx, self, old_start.custom.unwrap())?
} else {
self.timeline.emit_now_mu(ctx)
};
// Emulate variable allocation, as we need to use the CodeGenContext
// HashMap to store our variable due to lifetime limitation
// Note: we should be able to store variables directly if generic
@ -157,8 +224,11 @@ impl<'b> CodeGenerator for ArtiqCodeGenerator<'b> {
ctx.builder.build_store(end, now);
self.end = Some(end_expr);
self.name_counter += 1;
gen_block(self, ctx, body.iter())?;
let current = ctx.builder.get_insert_block().unwrap();
// if the current block is terminated, move before the terminator
// we want to set the timeline before reaching the terminator
// TODO: This may be unsound if there are multiple exit paths in the
@ -172,6 +242,7 @@ impl<'b> CodeGenerator for ArtiqCodeGenerator<'b> {
} else {
false
};
// set duration
let end_expr = self.end.take().unwrap();
let end_val = self
@ -183,49 +254,31 @@ impl<'b> CodeGenerator for ArtiqCodeGenerator<'b> {
if old_start.is_none() {
self.timeline.emit_at_mu(ctx, end_val);
}
// inside a parallel block, should update the outer max now_mu
if let Some(old_end) = &old_end {
let outer_end_val = self
.gen_expr(ctx, old_end)?
.unwrap()
.to_basic_value_enum(ctx, self, old_end.custom.unwrap())?;
let smax =
ctx.module.get_function("llvm.smax.i64").unwrap_or_else(|| {
let i64 = ctx.ctx.i64_type();
ctx.module.add_function(
"llvm.smax.i64",
i64.fn_type(&[i64.into(), i64.into()], false),
None,
)
});
let max = ctx
.builder
.build_call(smax, &[end_val.into(), outer_end_val.into()], "smax")
.try_as_basic_value()
.left()
.unwrap();
let outer_end = self.gen_store_target(ctx, old_end, Some("outer_end.addr"))?;
ctx.builder.build_store(outer_end, max);
}
self.timeline_update_end_max(ctx, old_end.clone(), Some("outer.end"))?;
self.start = old_start;
self.end = old_end;
if reset_position {
ctx.builder.position_at_end(current);
}
return Ok(());
} else if id == &"sequential".into() {
let start = self.start.take();
for stmt in body.iter() {
self.gen_stmt(ctx, stmt)?;
if ctx.is_terminated() {
break;
}
}
gen_block(self, ctx, body.iter())?;
self.start = start;
// Reset the timeline when we are exiting the sequential block
self.timeline_reset_start(ctx)?;
return Ok(());
}
}
}
// not parallel/sequential
gen_with(self, ctx, stmt)
} else {

100
nac3artiq/src/timeline.rs
View File

@ -1,9 +1,16 @@
use inkwell::{values::BasicValueEnum, AddressSpace, AtomicOrdering};
use nac3core::codegen::CodeGenContext;
/// Functions for manipulating the timeline.
pub trait TimeFns {
/// Emits LLVM IR for `now_mu`.
fn emit_now_mu<'ctx, 'a>(&self, ctx: &mut CodeGenContext<'ctx, 'a>) -> BasicValueEnum<'ctx>;
/// Emits LLVM IR for `at_mu`.
fn emit_at_mu<'ctx, 'a>(&self, ctx: &mut CodeGenContext<'ctx, 'a>, t: BasicValueEnum<'ctx>);
/// Emits LLVM IR for `delay_mu`.
fn emit_delay_mu<'ctx, 'a>(&self, ctx: &mut CodeGenContext<'ctx, 'a>, dt: BasicValueEnum<'ctx>);
}
@ -20,23 +27,25 @@ impl TimeFns for NowPinningTimeFns64 {
.get_global("now")
.unwrap_or_else(|| ctx.module.add_global(i64_type, None, "now"));
let now_hiptr =
ctx.builder.build_bitcast(now, i32_type.ptr_type(AddressSpace::default()), "now_hiptr");
ctx.builder.build_bitcast(now, i32_type.ptr_type(AddressSpace::default()), "now.hi.addr");
if let BasicValueEnum::PointerValue(now_hiptr) = now_hiptr {
let now_loptr = unsafe {
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(2, false)], "now_gep")
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(2, false)], "now.lo.addr")
};
if let (BasicValueEnum::IntValue(now_hi), BasicValueEnum::IntValue(now_lo)) = (
ctx.builder.build_load(now_hiptr, "now_hi"),
ctx.builder.build_load(now_loptr, "now_lo"),
ctx.builder.build_load(now_hiptr, "now.hi"),
ctx.builder.build_load(now_loptr, "now.lo"),
) {
let zext_hi = ctx.builder.build_int_z_extend(now_hi, i64_type, "now_zext_hi");
let zext_hi = ctx.builder.build_int_z_extend(now_hi, i64_type, "");
let shifted_hi = ctx.builder.build_left_shift(
zext_hi,
i64_type.const_int(32, false),
"now_shifted_zext_hi",
"",
);
let zext_lo = ctx.builder.build_int_z_extend(now_lo, i64_type, "now_zext_lo");
ctx.builder.build_or(shifted_hi, zext_lo, "now_or").into()
let zext_lo = ctx.builder.build_int_z_extend(now_lo, i64_type, "");
ctx.builder.build_or(shifted_hi, zext_lo, "now_mu").into()
} else {
unreachable!();
}
@ -48,14 +57,15 @@ impl TimeFns for NowPinningTimeFns64 {
fn emit_at_mu<'ctx, 'a>(&self, ctx: &mut CodeGenContext<'ctx, 'a>, t: BasicValueEnum<'ctx>) {
let i32_type = ctx.ctx.i32_type();
let i64_type = ctx.ctx.i64_type();
let i64_32 = i64_type.const_int(32, false);
if let BasicValueEnum::IntValue(time) = t {
let time_hi = ctx.builder.build_int_truncate(
ctx.builder.build_right_shift(time, i64_32, false, "now_lshr"),
ctx.builder.build_right_shift(time, i64_32, false, "time.hi"),
i32_type,
"now_trunc",
"",
);
let time_lo = ctx.builder.build_int_truncate(time, i32_type, "now_trunc");
let time_lo = ctx.builder.build_int_truncate(time, i32_type, "time.lo");
let now = ctx
.module
.get_global("now")
@ -63,11 +73,12 @@ impl TimeFns for NowPinningTimeFns64 {
let now_hiptr = ctx.builder.build_bitcast(
now,
i32_type.ptr_type(AddressSpace::default()),
"now_bitcast",
"now.hi.addr",
);
if let BasicValueEnum::PointerValue(now_hiptr) = now_hiptr {
let now_loptr = unsafe {
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(2, false)], "now_gep")
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(2, false)], "now.lo.addr")
};
ctx.builder
.build_store(now_hiptr, time_hi)
@ -97,41 +108,43 @@ impl TimeFns for NowPinningTimeFns64 {
.get_global("now")
.unwrap_or_else(|| ctx.module.add_global(i64_type, None, "now"));
let now_hiptr =
ctx.builder.build_bitcast(now, i32_type.ptr_type(AddressSpace::default()), "now_hiptr");
ctx.builder.build_bitcast(now, i32_type.ptr_type(AddressSpace::default()), "now.hi.addr");
if let BasicValueEnum::PointerValue(now_hiptr) = now_hiptr {
let now_loptr = unsafe {
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(2, false)], "now_loptr")
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(2, false)], "now.lo.addr")
};
if let (
BasicValueEnum::IntValue(now_hi),
BasicValueEnum::IntValue(now_lo),
BasicValueEnum::IntValue(dt),
) = (
ctx.builder.build_load(now_hiptr, "now_hi"),
ctx.builder.build_load(now_loptr, "now_lo"),
ctx.builder.build_load(now_hiptr, "now.hi"),
ctx.builder.build_load(now_loptr, "now.lo"),
dt,
) {
let zext_hi = ctx.builder.build_int_z_extend(now_hi, i64_type, "now_zext_hi");
let zext_hi = ctx.builder.build_int_z_extend(now_hi, i64_type, "");
let shifted_hi = ctx.builder.build_left_shift(
zext_hi,
i64_type.const_int(32, false),
"now_shifted_zext_hi",
"",
);
let zext_lo = ctx.builder.build_int_z_extend(now_lo, i64_type, "now_zext_lo");
let now_val = ctx.builder.build_or(shifted_hi, zext_lo, "now_or");
let zext_lo = ctx.builder.build_int_z_extend(now_lo, i64_type, "");
let now_val = ctx.builder.build_or(shifted_hi, zext_lo, "now");
let time = ctx.builder.build_int_add(now_val, dt, "now_add");
let time = ctx.builder.build_int_add(now_val, dt, "time");
let time_hi = ctx.builder.build_int_truncate(
ctx.builder.build_right_shift(
time,
i64_type.const_int(32, false),
false,
"now_lshr",
"",
),
i32_type,
"now_trunc",
"time.hi",
);
let time_lo = ctx.builder.build_int_truncate(time, i32_type, "now_trunc");
let time_lo = ctx.builder.build_int_truncate(time, i32_type, "time.lo");
ctx.builder
.build_store(now_hiptr, time_hi)
@ -162,11 +175,12 @@ impl TimeFns for NowPinningTimeFns {
.get_global("now")
.unwrap_or_else(|| ctx.module.add_global(i64_type, None, "now"));
let now_raw = ctx.builder.build_load(now.as_pointer_value(), "now");
if let BasicValueEnum::IntValue(now_raw) = now_raw {
let i64_32 = i64_type.const_int(32, false);
let now_lo = ctx.builder.build_left_shift(now_raw, i64_32, "now_shl");
let now_hi = ctx.builder.build_right_shift(now_raw, i64_32, false, "now_lshr");
ctx.builder.build_or(now_lo, now_hi, "now_or").into()
let now_lo = ctx.builder.build_left_shift(now_raw, i64_32, "now.lo");
let now_hi = ctx.builder.build_right_shift(now_raw, i64_32, false, "now.hi");
ctx.builder.build_or(now_lo, now_hi, "now_mu").into()
} else {
unreachable!();
}
@ -176,11 +190,12 @@ impl TimeFns for NowPinningTimeFns {
let i32_type = ctx.ctx.i32_type();
let i64_type = ctx.ctx.i64_type();
let i64_32 = i64_type.const_int(32, false);
if let BasicValueEnum::IntValue(time) = t {
let time_hi = ctx.builder.build_int_truncate(
ctx.builder.build_right_shift(time, i64_32, false, "now_lshr"),
ctx.builder.build_right_shift(time, i64_32, false, ""),
i32_type,
"now_trunc",
"time.hi",
);
let time_lo = ctx.builder.build_int_truncate(time, i32_type, "now_trunc");
let now = ctx
@ -190,11 +205,12 @@ impl TimeFns for NowPinningTimeFns {
let now_hiptr = ctx.builder.build_bitcast(
now,
i32_type.ptr_type(AddressSpace::default()),
"now_bitcast",
"now.hi.addr",
);
if let BasicValueEnum::PointerValue(now_hiptr) = now_hiptr {
let now_loptr = unsafe {
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(1, false)], "now_gep")
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(1, false)], "now.lo.addr")
};
ctx.builder
.build_store(now_hiptr, time_hi)
@ -224,26 +240,28 @@ impl TimeFns for NowPinningTimeFns {
.module
.get_global("now")
.unwrap_or_else(|| ctx.module.add_global(i64_type, None, "now"));
let now_raw = ctx.builder.build_load(now.as_pointer_value(), "now");
let now_raw = ctx.builder.build_load(now.as_pointer_value(), "");
if let (BasicValueEnum::IntValue(now_raw), BasicValueEnum::IntValue(dt)) = (now_raw, dt) {
let now_lo = ctx.builder.build_left_shift(now_raw, i64_32, "now_shl");
let now_hi = ctx.builder.build_right_shift(now_raw, i64_32, false, "now_lshr");
let now_val = ctx.builder.build_or(now_lo, now_hi, "now_or");
let time = ctx.builder.build_int_add(now_val, dt, "now_add");
let now_lo = ctx.builder.build_left_shift(now_raw, i64_32, "now.lo");
let now_hi = ctx.builder.build_right_shift(now_raw, i64_32, false, "now.hi");
let now_val = ctx.builder.build_or(now_lo, now_hi, "now_val");
let time = ctx.builder.build_int_add(now_val, dt, "time");
let time_hi = ctx.builder.build_int_truncate(
ctx.builder.build_right_shift(time, i64_32, false, "now_lshr"),
ctx.builder.build_right_shift(time, i64_32, false, "time.hi"),
i32_type,
"now_trunc",
);
let time_lo = ctx.builder.build_int_truncate(time, i32_type, "now_trunc");
let time_lo = ctx.builder.build_int_truncate(time, i32_type, "time.lo");
let now_hiptr = ctx.builder.build_bitcast(
now,
i32_type.ptr_type(AddressSpace::default()),
"now_bitcast",
"now.hi.addr",
);
if let BasicValueEnum::PointerValue(now_hiptr) = now_hiptr {
let now_loptr = unsafe {
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(1, false)], "now_gep")
ctx.builder.build_gep(now_hiptr, &[i32_type.const_int(1, false)], "now.lo.addr")
};
ctx.builder
.build_store(now_hiptr, time_hi)