diff --git a/Cargo.lock b/Cargo.lock
index 37216ee3fbaf095900ca648fcbcad447ef3af34c..1f9d1747a6ee3734825aeb18d792164cf6f99f19 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -712,6 +712,16 @@ dependencies = [
"with_builtin_macros",
]
+[[package]]
+name = "juno_casts_and_intrinsics"
+version = "0.1.0"
+dependencies = [
+ "async-std",
+ "hercules_rt",
+ "juno_build",
+ "with_builtin_macros",
+]
+
[[package]]
name = "juno_frontend"
version = "0.1.0"
diff --git a/Cargo.toml b/Cargo.toml
index c6b2468fb9d41b467d48221d6621d9fcbb197034..a34845f8b26aeb97d076e7ebfcf8125efa43473c 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -20,6 +20,7 @@ members = [
"juno_scheduler",
"juno_build",
- "juno_samples/matmul",
"juno_samples/simple3",
+ "juno_samples/matmul",
+ "juno_samples/casts_and_intrinsics",
]
diff --git a/hercules_cg/src/cpu.rs b/hercules_cg/src/cpu.rs
index eaa7374b7a8f6c3df66a9a5085d179a143b81d34..8437e9e44b2c7792b54f563f86f59ee5fa6bd23a 100644
--- a/hercules_cg/src/cpu.rs
+++ b/hercules_cg/src/cpu.rs
@@ -597,7 +597,54 @@ impl<'a> CPUContext<'a> {
write!(w, ", -1")?;
}
}
- SUnaryOperator::Cast(_) => todo!(),
+ SUnaryOperator::Cast(dst_ty) => {
+ let src_ty = &self.svalue_types[input];
+ if src_ty.is_integer()
+ && dst_ty.is_integer()
+ && src_ty.num_bits() > dst_ty.num_bits()
+ {
+ write!(w, "trunc ")?;
+ } else if src_ty.is_signed()
+ && dst_ty.is_integer()
+ && src_ty.num_bits() < dst_ty.num_bits()
+ {
+ write!(w, "sext ")?;
+ } else if src_ty.is_integer()
+ && dst_ty.is_integer()
+ && src_ty.num_bits() < dst_ty.num_bits()
+ {
+ write!(w, "zext ")?;
+ } else if src_ty.is_integer() && dst_ty.is_integer() {
+ // A no-op.
+ write!(w, "bitcast ")?;
+ } else if src_ty.is_float()
+ && dst_ty.is_float()
+ && src_ty.num_bits() > dst_ty.num_bits()
+ {
+ write!(w, "fptrunc ")?;
+ } else if src_ty.is_float()
+ && dst_ty.is_float()
+ && src_ty.num_bits() < dst_ty.num_bits()
+ {
+ write!(w, "fpext ")?;
+ } else if src_ty.is_float() && dst_ty.is_float() {
+ // A no-op.
+ write!(w, "bitcast ")?;
+ } else if src_ty.is_float() && dst_ty.is_signed() {
+ write!(w, "fptosi ")?;
+ } else if src_ty.is_float() && dst_ty.is_integer() {
+ write!(w, "fptoui ")?;
+ } else if src_ty.is_signed() && dst_ty.is_float() {
+ write!(w, "sitofp ")?;
+ } else if src_ty.is_integer() && dst_ty.is_float() {
+ write!(w, "uitofp ")?;
+ } else {
+ panic!("PANIC: Invalid cast type combination.");
+ }
+ self.emit_svalue(input, true, w)?;
+ write!(w, " to ")?;
+ self.emit_type(dst_ty, w)?;
+ }
}
}
SInst::Binary { left, right, op } => {
@@ -677,6 +724,24 @@ impl<'a> CPUContext<'a> {
}
}
}
+ SInst::IntrinsicCall { intrinsic, args } => {
+ emit_assign(w)?;
+ write!(w, "call ")?;
+ self.emit_type(&self.svalue_types[&self_svalue], w)?;
+ write!(
+ w,
+ " @llvm.{}.{}(",
+ // TODO: make lower case names conform to LLVM expectations.
+ intrinsic.lower_case_name(),
+ Self::intrinsic_type_str(&self.svalue_types[&self_svalue])
+ )?;
+ self.emit_svalue(&args[0], true, w)?;
+ for idx in 1..args.len() {
+ write!(w, ", ")?;
+ self.emit_svalue(&args[idx], true, w)?;
+ }
+ write!(w, ")")?;
+ }
SInst::ProductExtract { product, indices } => {
emit_assign(w)?;
write!(w, "extractvalue ")?;
diff --git a/hercules_cg/src/sched_gen.rs b/hercules_cg/src/sched_gen.rs
index e5378b9cca4ae553ff8a760202db0526b13f7211..70a898b6c259b8bf7543b5ff4d1202132c71cf9b 100644
--- a/hercules_cg/src/sched_gen.rs
+++ b/hercules_cg/src/sched_gen.rs
@@ -1071,6 +1071,17 @@ impl<'a> FunctionContext<'a> {
self.stypes[self.typing[id.idx()].idx()].clone().unwrap(),
));
}
+ Node::IntrinsicCall {
+ intrinsic,
+ ref args,
+ } => {
+ let args = args.iter().map(|id| get_svalue(*id).clone()).collect();
+ block.insts.push(SInst::IntrinsicCall { intrinsic, args });
+ block.virt_regs.push((
+ self_virt_reg(),
+ self.stypes[self.typing[id.idx()].idx()].clone().unwrap(),
+ ));
+ }
Node::Read {
collect,
diff --git a/hercules_cg/src/sched_ir.rs b/hercules_cg/src/sched_ir.rs
index 3ceee621f4469127c40453e364b87a230a947e85..6a482443eaf39c6f3b87c7b19ec06adbe67ff5ad 100644
--- a/hercules_cg/src/sched_ir.rs
+++ b/hercules_cg/src/sched_ir.rs
@@ -213,6 +213,28 @@ impl SType {
}
}
+ pub fn is_signed(&self) -> bool {
+ match self {
+ SType::Integer8 | SType::Integer16 | SType::Integer32 | SType::Integer64 => true,
+ _ => false,
+ }
+ }
+
+ pub fn is_integer(&self) -> bool {
+ self.is_unsigned() || self.is_signed() || *self == SType::Boolean
+ }
+
+ pub fn num_bits(&self) -> u8 {
+ match self {
+ SType::Boolean => 1,
+ SType::Integer8 | SType::UnsignedInteger8 => 8,
+ SType::Integer16 | SType::UnsignedInteger16 => 16,
+ SType::Integer32 | SType::UnsignedInteger32 | SType::Float32 => 32,
+ SType::Integer64 | SType::UnsignedInteger64 | SType::Float64 => 64,
+ _ => panic!(),
+ }
+ }
+
pub fn try_product(&self) -> Option<&[SType]> {
if let SType::Product(fields) = self {
Some(fields)
@@ -334,6 +356,10 @@ pub enum SInst {
third: SValue,
op: STernaryOperator,
},
+ IntrinsicCall {
+ intrinsic: Intrinsic,
+ args: Box<[SValue]>,
+ },
ProductExtract {
product: SValue,
indices: Box<[usize]>,
@@ -487,6 +513,7 @@ impl SInst {
third: _,
op,
} => op.upper_case_name(),
+ SInst::IntrinsicCall { intrinsic, args: _ } => intrinsic.upper_case_name(),
SInst::ProductExtract {
product: _,
indices: _,
diff --git a/hercules_cg/src/sched_schedule.rs b/hercules_cg/src/sched_schedule.rs
index 16720bbc60380866cd2dbd868ebf9f6cbd5a91fe..4ba407f0e316e6611e311ff095bff6121a922c5a 100644
--- a/hercules_cg/src/sched_schedule.rs
+++ b/hercules_cg/src/sched_schedule.rs
@@ -44,7 +44,7 @@ pub fn sched_get_uses(inst: &SInst) -> Box<dyn Iterator<Item = &SValue> + '_> {
false_target: _,
true_target: _,
} => Box::new(once(cond)),
- SInst::PartitionExit { data_outputs } => Box::new(data_outputs.iter().map(|svalue| svalue)),
+ SInst::PartitionExit { data_outputs } => Box::new(data_outputs.iter()),
SInst::Return { value } => Box::new(once(value)),
SInst::Unary { input, op: _ } => Box::new(once(input)),
SInst::Binary { left, right, op: _ } => Box::new(once(left).chain(once(right))),
@@ -54,6 +54,7 @@ pub fn sched_get_uses(inst: &SInst) -> Box<dyn Iterator<Item = &SValue> + '_> {
third,
op: _,
} => Box::new(once(first).chain(once(second)).chain(once(third))),
+ SInst::IntrinsicCall { intrinsic: _, args } => Box::new(args.iter()),
SInst::ProductExtract {
product,
indices: _,
diff --git a/hercules_ir/src/gcm.rs b/hercules_ir/src/gcm.rs
index 3180a88b215dfcc36817abef816795561e03f4da..19df2d6bb3a42c4c989e9d04eb84b37caa6cf4a0 100644
--- a/hercules_ir/src/gcm.rs
+++ b/hercules_ir/src/gcm.rs
@@ -18,7 +18,6 @@ pub fn gcm(
antideps: &Vec<(NodeID, NodeID)>,
loops: &LoopTree,
fork_join_map: &HashMap<NodeID, NodeID>,
- partial_partition: &mut Vec<Option<PartitionID>>,
) -> Vec<NodeID> {
let mut bbs: Vec<Option<NodeID>> = vec![None; function.nodes.len()];
@@ -186,30 +185,12 @@ pub fn gcm(
// #[feature(iter_collect_into)]
// Look between the LCA and the schedule early location to place the
- // node. If a data node can't be scheduled to any control nodes in its
- // partition (this may happen if all of the control nodes in a partition
- // got deleted, for example), then the data node can be scheduled into a
- // control node in a different partition.
+ // node.
let schedule_early = schedule_early[id.idx()].unwrap();
- let need_to_repartition = !dom
- .chain(lca.unwrap_or(schedule_early), schedule_early)
- .any(|dominator| {
- partial_partition[id.idx()].is_none()
- || partial_partition[dominator.idx()] == partial_partition[id.idx()]
- });
- if need_to_repartition {
- partial_partition[id.idx()] = None;
- }
let mut chain = dom
// If the node has no users, then it doesn't really matter where we
// place it - just place it at the early placement.
- .chain(lca.unwrap_or(schedule_early), schedule_early)
- // Filter the dominator chain by control nodes in the same partition
- // as this data node, if the data node is in a partition already.
- .filter(|dominator| {
- partial_partition[id.idx()].is_none()
- || partial_partition[dominator.idx()] == partial_partition[id.idx()]
- });
+ .chain(lca.unwrap_or(schedule_early), schedule_early);
let mut location = chain.next().unwrap();
while let Some(control_node) = chain.next() {
// If the next node further up the dominator tree is in a shallower
diff --git a/hercules_opt/src/editor.rs b/hercules_opt/src/editor.rs
index 3286fa884663450ab0b15915467e1e82cfc428ff..a75ea38aa3b7703e7efecc952b54d9ccb8df45c6 100644
--- a/hercules_opt/src/editor.rs
+++ b/hercules_opt/src/editor.rs
@@ -687,16 +687,9 @@ pub fn repair_plan(plan: &mut Plan, new_function: &Function, edits: &[Edit]) {
&antideps,
&loops,
&fork_join_map,
- &mut new_partitions,
);
- let added_and_to_repartition_data_nodes: Vec<NodeID> = new_partitions
- .iter()
- .enumerate()
- .filter(|(_, part)| part.is_none())
- .map(|(idx, _)| NodeID::new(idx))
- .collect();
- for data_id in added_and_to_repartition_data_nodes {
- new_partitions[data_id.idx()] = new_partitions[bbs[data_id.idx()].idx()];
+ for data_idx in 0..new_function.nodes.len() {
+ new_partitions[data_idx] = new_partitions[bbs[data_idx].idx()];
}
// Step 6: create a solitary gravestone partition. This will get removed
diff --git a/hercules_opt/src/pass.rs b/hercules_opt/src/pass.rs
index 03aff10ade7e4aca8ecf58d94127a405640eb7e3..f5e0a9ad5ac9a081efe3016cd0dcb6430c688bc6 100644
--- a/hercules_opt/src/pass.rs
+++ b/hercules_opt/src/pass.rs
@@ -275,7 +275,6 @@ impl PassManager {
antideps,
loops,
fork_join_map,
- &mut vec![None; function.nodes.len()],
)
},
)
diff --git a/juno_samples/casts_and_intrinsics/Cargo.toml b/juno_samples/casts_and_intrinsics/Cargo.toml
new file mode 100644
index 0000000000000000000000000000000000000000..f49797969012f5195a0338b7b14fa04414dddb03
--- /dev/null
+++ b/juno_samples/casts_and_intrinsics/Cargo.toml
@@ -0,0 +1,18 @@
+[package]
+name = "juno_casts_and_intrinsics"
+version = "0.1.0"
+authors = ["Russel Arbore <rarbore2@illinois.edu>"]
+edition = "2021"
+
+[[bin]]
+name = "juno_casts_and_intrinsics"
+path = "src/main.rs"
+
+[build-dependencies]
+juno_build = { path = "../../juno_build" }
+
+[dependencies]
+juno_build = { path = "../../juno_build" }
+hercules_rt = { path = "../../hercules_rt" }
+with_builtin_macros = "0.1.0"
+async-std = "*"
diff --git a/juno_samples/casts_and_intrinsics/build.rs b/juno_samples/casts_and_intrinsics/build.rs
new file mode 100644
index 0000000000000000000000000000000000000000..8422f7e60619b48aca2d5e783eaed45e70be71c3
--- /dev/null
+++ b/juno_samples/casts_and_intrinsics/build.rs
@@ -0,0 +1,12 @@
+extern crate juno_build;
+use juno_build::JunoCompiler;
+
+fn main() {
+ JunoCompiler::new()
+ .file_in_src("casts_and_intrinsics.jn")
+ .unwrap()
+ .schedule_in_src("casts_and_intrinsics.sch")
+ .unwrap()
+ .build()
+ .unwrap();
+}
diff --git a/juno_samples/casts_and_intrinsics/src/casts_and_intrinsics.jn b/juno_samples/casts_and_intrinsics/src/casts_and_intrinsics.jn
new file mode 100644
index 0000000000000000000000000000000000000000..5261063353496496e3e994a91620073438528ccd
--- /dev/null
+++ b/juno_samples/casts_and_intrinsics/src/casts_and_intrinsics.jn
@@ -0,0 +1,5 @@
+#[entry]
+fn casts_and_intrinsics(input : f32) -> i32 {
+ let sqrt = sqrt!::<f32>(input);
+ return sqrt as i32;
+}
diff --git a/juno_samples/casts_and_intrinsics/src/casts_and_intrinsics.sch b/juno_samples/casts_and_intrinsics/src/casts_and_intrinsics.sch
new file mode 100644
index 0000000000000000000000000000000000000000..80ec2766eac8850ff736cd8a1f52dbe87bf24553
--- /dev/null
+++ b/juno_samples/casts_and_intrinsics/src/casts_and_intrinsics.sch
@@ -0,0 +1,2 @@
+function casts_and_intrinsics {
+}
diff --git a/juno_samples/casts_and_intrinsics/src/main.rs b/juno_samples/casts_and_intrinsics/src/main.rs
new file mode 100644
index 0000000000000000000000000000000000000000..344168e0e1995becff3b4580f4d957002a9ee2a5
--- /dev/null
+++ b/juno_samples/casts_and_intrinsics/src/main.rs
@@ -0,0 +1,22 @@
+#![feature(future_join)]
+
+extern crate async_std;
+extern crate juno_build;
+extern crate hercules_rt;
+
+juno_build::juno!("casts_and_intrinsics");
+
+fn main() {
+ async_std::task::block_on(async {
+ let output = unsafe {
+ casts_and_intrinsics(16.0).await
+ };
+ println!("{}", output);
+ assert_eq!(output, 4);
+ });
+}
+
+#[test]
+fn casts_and_intrinsics_test() {
+ main();
+}