diff --git a/hercules_ir/src/dot.rs b/hercules_ir/src/dot.rs
index 8efabd7a99157b9bb8b40e5abe76d2d42513bb79..5ccda9dc6bf39c016ba44b6eb1085679288d5c54 100644
--- a/hercules_ir/src/dot.rs
+++ b/hercules_ir/src/dot.rs
@@ -16,6 +16,7 @@ use crate::*;
pub fn xdot_module(
module: &ir::Module,
reverse_postorders: &Vec<Vec<NodeID>>,
+ typing: Option<&ModuleTyping>,
doms: Option<&Vec<DomTree>>,
fork_join_maps: Option<&Vec<HashMap<NodeID, NodeID>>>,
devices: Option<&Vec<Device>>,
@@ -30,6 +31,7 @@ pub fn xdot_module(
write_dot(
&module,
&reverse_postorders,
+ typing,
doms,
fork_join_maps,
devices,
@@ -53,6 +55,7 @@ pub fn xdot_module(
pub fn write_dot<W: Write>(
module: &ir::Module,
reverse_postorders: &Vec<Vec<NodeID>>,
+ typing: Option<&ModuleTyping>,
doms: Option<&Vec<DomTree>>,
fork_join_maps: Option<&Vec<HashMap<NodeID, NodeID>>>,
devices: Option<&Vec<Device>>,
@@ -89,6 +92,7 @@ pub fn write_dot<W: Write>(
function_id,
color,
module,
+ typing,
&function.schedules[node_id.idx()],
w,
)?;
@@ -249,6 +253,7 @@ fn write_node<W: Write>(
function_id: FunctionID,
color: &str,
module: &Module,
+ typing: Option<&ModuleTyping>,
schedules: &Vec<Schedule>,
w: &mut W,
) -> std::fmt::Result {
@@ -331,30 +336,37 @@ fn write_node<W: Write>(
} else {
format!("{} ({})", node.upper_case_name(), suffix)
};
+ let xlabel = format!("{}", node_id.idx());
+ let mut tylabel = String::new();
+ if let Some(ty) = typing.map(|typing| typing[function_id.idx()][node_id.idx()]) {
+ module.write_type(ty, &mut tylabel)?;
+ }
let mut iter = schedules.into_iter();
if let Some(first) = iter.next() {
- let subtitle = iter.fold(format!("{:?}", first), |b, i| format!("{}, {:?}", b, i));
+ let schedules = iter.fold(format!("{:?}", first), |b, i| format!("{}, {:?}", b, i));
write!(
w,
- "{}_{}_{} [xlabel={}, label=<{}<BR /><FONT POINT-SIZE=\"8\">{}</FONT>>, color={}];\n",
+ "{}_{}_{} [xlabel={}, label=<{}<BR /><FONT POINT-SIZE=\"8\">{}</FONT><BR /><FONT POINT-SIZE=\"8\">{}</FONT>>, color={}];\n",
node.lower_case_name(),
function_id.idx(),
node_id.idx(),
- node_id.idx(),
+ xlabel,
label,
- subtitle,
+ tylabel,
+ schedules,
color
)?;
} else {
write!(
w,
- "{}_{}_{} [xlabel={}, label=\"{}\", color={}];\n",
+ "{}_{}_{} [xlabel={}, label=<{}<BR /><FONT POINT-SIZE=\"8\">{}</FONT>>, color={}];\n",
node.lower_case_name(),
function_id.idx(),
node_id.idx(),
- node_id.idx(),
+ xlabel,
label,
+ tylabel,
color
)?;
}
diff --git a/hercules_ir/src/ir.rs b/hercules_ir/src/ir.rs
index f62c00c15f9e8715254c515834d8fe63c2715539..bf7806dcc371112419bbf7e21ef3b206df3a2e32 100644
--- a/hercules_ir/src/ir.rs
+++ b/hercules_ir/src/ir.rs
@@ -628,6 +628,7 @@ pub fn dynamic_constants_bottom_up(
dynamic_constants: &Vec<DynamicConstant>,
) -> impl Iterator<Item = DynamicConstantID> + '_ {
let mut visited = bitvec![u8, Lsb0; 0; dynamic_constants.len()];
+ let mut invalid = bitvec![u8, Lsb0; 0; dynamic_constants.len()];
let mut stack = (0..dynamic_constants.len())
.map(DynamicConstantID::new)
.collect::<Vec<DynamicConstantID>>();
@@ -647,13 +648,16 @@ pub fn dynamic_constants_bottom_up(
// We have to yield the children of this node before
// this node itself. We keep track of which nodes have
// yielded using visited.
- if left.idx() >= visited.len() || right.idx() >= visited.len() {
+ if left.idx() >= visited.len()
+ || right.idx() >= visited.len()
+ || invalid[left.idx()]
+ || invalid[right.idx()]
+ {
// This is an invalid dynamic constant and should be
// skipped.
+ invalid.set(id.idx(), true);
continue;
- }
- let can_yield = visited[left.idx()] && visited[right.idx()];
- if can_yield {
+ } else if visited[left.idx()] && visited[right.idx()] {
visited.set(id.idx(), true);
yield id;
} else {
diff --git a/hercules_opt/src/device_placement.rs b/hercules_opt/src/device_placement.rs
deleted file mode 100644
index 2badd69df428db12a0f3a46ac4aee05f8154d171..0000000000000000000000000000000000000000
--- a/hercules_opt/src/device_placement.rs
+++ /dev/null
@@ -1,3 +0,0 @@
-use hercules_ir::ir::*;
-
-use crate::*;
diff --git a/hercules_opt/src/gcm.rs b/hercules_opt/src/gcm.rs
index f919acc7caa822d5693e940a3a4fae05d7900d4a..462d10871565bfed9b351d2627c44af8ca778ffc 100644
--- a/hercules_opt/src/gcm.rs
+++ b/hercules_opt/src/gcm.rs
@@ -1163,11 +1163,12 @@ fn object_allocation(
Node::Call {
control: _,
function: callee,
- dynamic_constants: _,
+ ref dynamic_constants,
args: _,
} => {
+ let dynamic_constants = dynamic_constants.clone();
for device in BACKED_DEVICES {
- if let Some(callee_backing_size) = backing_allocations[&callee]
+ if let Some(mut callee_backing_size) = backing_allocations[&callee]
.get(&device)
.map(|(callee_total, _)| *callee_total)
{
@@ -1177,6 +1178,27 @@ fn object_allocation(
// in the callee, so just assume the largest alignment.
*total = align(&mut edit, *total, LARGEST_ALIGNMENT);
offsets.insert(id, *total);
+ // Substitute the dynamic constant parameters in the
+ // callee's backing size.
+ let first_dc = edit.num_dynamic_constants() + 10000;
+ for (p_idx, dc_n) in zip(0..dynamic_constants.len(), first_dc..) {
+ let dc_a =
+ edit.add_dynamic_constant(DynamicConstant::Parameter(p_idx));
+ callee_backing_size = substitute_dynamic_constants(
+ dc_a,
+ DynamicConstantID::new(dc_n),
+ callee_backing_size,
+ &mut edit,
+ );
+ }
+ for (dc_n, dc_b) in zip(first_dc.., dynamic_constants.iter()) {
+ callee_backing_size = substitute_dynamic_constants(
+ DynamicConstantID::new(dc_n),
+ *dc_b,
+ callee_backing_size,
+ &mut edit,
+ );
+ }
*total = edit.add_dynamic_constant(DynamicConstant::Add(
*total,
callee_backing_size,
diff --git a/hercules_opt/src/inline.rs b/hercules_opt/src/inline.rs
index 064e3d73a1d9604ca5b284fe52b8c2e8c5a0339e..1d2bac97d848ace910d614a42743c6ea5fe3aa9e 100644
--- a/hercules_opt/src/inline.rs
+++ b/hercules_opt/src/inline.rs
@@ -179,7 +179,7 @@ fn inline_func(
// as the new references we just made in the first step. We
// actually want to institute all the updates
// *simultaneously*, hence the two step maneuver.
- let first_dc = edit.num_dynamic_constants() + 100;
+ let first_dc = edit.num_dynamic_constants() + 10000;
for (dc_a, dc_n) in zip(dcs_a, first_dc..) {
substitute_dynamic_constants_in_node(
*dc_a,
diff --git a/hercules_opt/src/interprocedural_sroa.rs b/hercules_opt/src/interprocedural_sroa.rs
index 49fbcbbd712fad2977c64fd2ca1d9643bd95d74c..f597cd80347d94a7c927d6fe085d80f843e280eb 100644
--- a/hercules_opt/src/interprocedural_sroa.rs
+++ b/hercules_opt/src/interprocedural_sroa.rs
@@ -320,7 +320,7 @@ fn compress_return_products(editors: &mut Vec<FunctionEditor>, all_callsites_edi
let mut substituted = old_return_type_ids[function_id.idx()];
assert_eq!(old_dcs.len(), new_dcs.len());
- let first_dc = edit.num_dynamic_constants() + 100;
+ let first_dc = edit.num_dynamic_constants() + 10000;
for (dc_a, dc_n) in zip(old_dcs, first_dc..) {
substituted = substitute_dynamic_constants_in_type(
dc_a,
@@ -424,7 +424,7 @@ fn remove_return_singletons(editors: &mut Vec<FunctionEditor>, all_callsites_edi
if singleton_removed[function.idx()] {
let edit_successful = editor.edit(|mut edit| {
let mut substituted = old_return_type_ids[function.idx()];
- let first_dc = edit.num_dynamic_constants() + 100;
+ let first_dc = edit.num_dynamic_constants() + 10000;
let dc_params: Vec<_> = (0..dc_args.len())
.map(|param_idx| {
edit.add_dynamic_constant(DynamicConstant::Parameter(param_idx))
diff --git a/hercules_opt/src/lib.rs b/hercules_opt/src/lib.rs
index 01ae1c99ad3613e826801afebdb0e15376ae1377..e3cca1612354fee6b5544264b97a37f1352bbafd 100644
--- a/hercules_opt/src/lib.rs
+++ b/hercules_opt/src/lib.rs
@@ -4,7 +4,6 @@ pub mod ccp;
pub mod crc;
pub mod dce;
pub mod delete_uncalled;
-pub mod device_placement;
pub mod editor;
pub mod float_collections;
pub mod fork_concat_split;
@@ -30,7 +29,6 @@ pub use crate::ccp::*;
pub use crate::crc::*;
pub use crate::dce::*;
pub use crate::delete_uncalled::*;
-pub use crate::device_placement::*;
pub use crate::editor::*;
pub use crate::float_collections::*;
pub use crate::fork_concat_split::*;
diff --git a/juno_scheduler/src/pm.rs b/juno_scheduler/src/pm.rs
index 76e81ee9f0147caed9db3714515bd0362bbb3ae4..8b3e90502128df6fb8962692fffd5dc67e1aabb8 100644
--- a/juno_scheduler/src/pm.rs
+++ b/juno_scheduler/src/pm.rs
@@ -1761,12 +1761,14 @@ fn run_pass(
pm.make_reverse_postorders();
if force_analyses {
+ pm.make_typing();
pm.make_doms();
pm.make_fork_join_maps();
pm.make_devices();
}
let reverse_postorders = pm.reverse_postorders.take().unwrap();
+ let typing = pm.typing.take();
let doms = pm.doms.take();
let fork_join_maps = pm.fork_join_maps.take();
let devices = pm.devices.take();
@@ -1775,6 +1777,7 @@ fn run_pass(
xdot_module(
module,
&reverse_postorders,
+ typing.as_ref(),
doms.as_ref(),
fork_join_maps.as_ref(),
devices.as_ref(),