Fix DC param ID identfication

.gitignore

 /target
 *.dot
+!paper_resources/*.dot
 *.bc
 *.out
 *.ll

hercules_opt/src/inline.rs

@@ -24,21 +24,19 @@ pub fn inline(editors: &mut [FunctionEditor], callgraph: &CallGraph) {
         .map(|editor| collapse_returns(editor))
         .collect();
 
-    // Step 3: verify that each possible dynamic constant parameter index has a
-    // single unique dynamic constant ID. If this isn't true, dynamic constant
-    // substitution won't work, and this should be true anyway!
-    let mut found_idxs = HashMap::new();
-    for id in editors[0].dynamic_constant_ids() {
-        let dc = editors[0].get_dynamic_constant(id);
-        if let DynamicConstant::Parameter(idx) = *dc {
-            assert!(!found_idxs.contains_key(&idx));
-            found_idxs.insert(idx, id);
-        }
-    }
-    let mut dc_param_idx_to_dc_id = vec![];
-    for idx in 0..found_idxs.len() {
-        dc_param_idx_to_dc_id.push(found_idxs[&idx]);
-    }
+    // Step 3: get dynamic constant IDs for parameters.
+    let max_num_dc_params = editors
+        .iter()
+        .map(|editor| editor.func().num_dynamic_constants)
+        .max()
+        .unwrap();
+    let mut dc_args = vec![];
+    editors[0].edit(|mut edit| {
+        dc_args = (0..max_num_dc_params as usize)
+            .map(|i| edit.add_dynamic_constant(DynamicConstant::Parameter(i)))
+            .collect();
+        Ok(edit)
+    });
 
     // Step 4: run inlining on each function individually. Iterate the functions
     // in topological order.
@@ -49,12 +47,7 @@ pub fn inline(editors: &mut [FunctionEditor], callgraph: &CallGraph) {
         // 2. Shared references to all of the functions called by that function.
         let callees = callgraph.get_callees(to_inline_id);
         let editor_refs = get_mut_and_immuts(editors, to_inline_id, callees);
-        inline_func(
-            editor_refs.0,
-            editor_refs.1,
-            &single_return_nodes,
-            &dc_param_idx_to_dc_id,
-        );
+        inline_func(editor_refs.0, editor_refs.1, &single_return_nodes, &dc_args);
     }
 }
 

hercules_opt/src/interprocedural_sroa.rs

@@ -235,18 +235,18 @@ fn compress_return_products(editors: &mut Vec<FunctionEditor>, all_callsites_edi
     // dc_param_idx_to_dc_id to get the id of the dynamic constants in the function,
     // and then replace dc_param_idx_to_dc_id[i] with call.dynamic_constants[i],
     // for all i.
-    let mut found_idxs = HashMap::new();
-    for id in editors[0].dynamic_constant_ids() {
-        let dc = editors[0].get_dynamic_constant(id);
-        if let DynamicConstant::Parameter(idx) = *dc {
-            assert!(!found_idxs.contains_key(&idx));
-            found_idxs.insert(idx, id);
-        }
-    }
-    let mut dc_param_idx_to_dc_id = vec![];
-    for idx in 0..found_idxs.len() {
-        dc_param_idx_to_dc_id.push(found_idxs[&idx]);
-    }
+    let max_num_dc_params = editors
+        .iter()
+        .map(|editor| editor.func().num_dynamic_constants)
+        .max()
+        .unwrap();
+    let mut dc_args = vec![];
+    editors[0].edit(|mut edit| {
+        dc_args = (0..max_num_dc_params as usize)
+            .map(|i| edit.add_dynamic_constant(DynamicConstant::Parameter(i)))
+            .collect();
+        Ok(edit)
+    });
 
     // Step 2. Modify the return type of all editors corresponding to a function
     // for which we can edit every callsite, and the return type is a product.
@@ -314,7 +314,7 @@ fn compress_return_products(editors: &mut Vec<FunctionEditor>, all_callsites_edi
             // a better abstraction around bulk replacement.
 
             let new_dcs = (*dynamic_constants).to_vec();
-            let old_dcs = dc_param_idx_to_dc_id[..new_dcs.len()].to_vec();
+            let old_dcs = dc_args[..new_dcs.len()].to_vec();
             assert_eq!(old_dcs.len(), new_dcs.len());
             let substs = old_dcs
                 .into_iter()

paper_resources/arr_sum_ir.dot

+digraph "Module" {
+margin=0;
+subgraph sum {
+label="sum<1>"
+style=invis
+cluster=true
+ranksep=1;
+nodesep=1;
+start_0_0 [xlabel=0, label=<Start<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+parameter_0_1 [xlabel=1, label=<Parameter (#0)<BR /><FONT POINT-SIZE="8">Array(Float32, #0)</FONT>>, color=darkblue];
+dynamic_constant_0_2 [xlabel=2, label=<DynamicConstant (#0)<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+projection_0_3 [xlabel=3, label=<Projection<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+add_0_4 [xlabel=4, label=<Add<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+add_0_5 [xlabel=5, label=<Add<BR /><FONT POINT-SIZE="8">Float32</FONT>>, color=darkblue];
+read_0_6 [xlabel=6, label=<Read<BR /><FONT POINT-SIZE="8">Float32</FONT>>, color=darkblue];
+lt_0_7 [xlabel=7, label=<LT<BR /><FONT POINT-SIZE="8">Boolean</FONT>>, color=darkblue];
+if_0_8 [xlabel=8, label=<If<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+projection_0_9 [xlabel=9, label=<Projection<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+return_0_10 [xlabel=10, label=<Return>, color=darkred];
+constant_0_11 [xlabel=11, label=<Constant (0)<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+constant_0_12 [xlabel=12, label=<Constant (1)<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+constant_0_13 [xlabel=13, label=<Constant (0)<BR /><FONT POINT-SIZE="8">Float32</FONT>>, color=darkblue];
+phi_0_14 [xlabel=14, label=<Phi<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+phi_0_15 [xlabel=15, label=<Phi<BR /><FONT POINT-SIZE="8">Float32</FONT>>, color=darkblue];
+region_0_16 [xlabel=16, label=<Region<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+if_0_8 -> projection_0_3 [color=black, style="dashed"];
+phi_0_14 -> add_0_4 [color=black, style=""];
+constant_0_12 -> add_0_4 [color=black, style=""];
+phi_0_15 -> add_0_5 [color=black, style=""];
+read_0_6 -> add_0_5 [color=black, style=""];
+parameter_0_1 -> read_0_6 [color=black, style=""];
+phi_0_14 -> read_0_6 [color=black, style=""];
+add_0_4 -> lt_0_7 [color=black, style=""];
+dynamic_constant_0_2 -> lt_0_7 [color=black, style=""];
+if_0_8 -> region_0_16 [dir=back, color=black, style="dashed"];
+lt_0_7 -> if_0_8 [color=black, style="dotted"];
+if_0_8 -> projection_0_9 [color=black, style="dashed"];
+projection_0_9 -> return_0_10 [color=black, style="dashed"];
+add_0_5 -> return_0_10 [color=black, style="dotted"];
+constant_0_11 -> phi_0_14 [color=black, style=""];
+phi_0_14 -> add_0_4 [dir=back, color=black, style=""];
+region_0_16 -> phi_0_14 [color=black, style="dotted"];
+constant_0_13 -> phi_0_15 [color=black, style=""];
+add_0_5 -> phi_0_15 [color=black, style=""];
+region_0_16 -> phi_0_15 [color=black, style="dotted"];
+start_0_0 -> region_0_16 [color=black, style="dashed"];
+projection_0_3 -> region_0_16 [color=black, style="dashed"];
+}
+}

paper_resources/matmul_ir.dot

+digraph "Module" {
+margin=0;
+subgraph matmul {
+label="matmul<3>"
+style=invis
+cluster=true
+ranksep=1;
+nodesep=1;
+start_0_0 [xlabel=0, label=<Start<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+parameter_0_1 [xlabel=1, label=<Parameter (#0)<BR /><FONT POINT-SIZE="8">Array(Integer32, #0, #1)</FONT>>, color=darkblue];
+parameter_0_2 [xlabel=2, label=<Parameter (#1)<BR /><FONT POINT-SIZE="8">Array(Integer32, #1, #2)</FONT>>, color=darkblue];
+fork_0_3 [xlabel=3, label=<Fork (#0, #2)<BR /><FONT POINT-SIZE="8">Control</FONT><BR /><FONT POINT-SIZE="8">ParallelFork</FONT>>, color=darkred];
+thread_id_0_4 [xlabel=4, label=<ThreadID (0)<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+thread_id_0_5 [xlabel=5, label=<ThreadID (1)<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+fork_0_6 [xlabel=6, label=<Fork (#1)<BR /><FONT POINT-SIZE="8">Control</FONT><BR /><FONT POINT-SIZE="8">Vectorizable</FONT>>, color=darkred];
+thread_id_0_7 [xlabel=7, label=<ThreadID (0)<BR /><FONT POINT-SIZE="8">UnsignedInteger64</FONT>>, color=darkblue];
+join_0_8 [xlabel=8, label=<Join<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+join_0_9 [xlabel=9, label=<Join<BR /><FONT POINT-SIZE="8">Control</FONT>>, color=darkred];
+reduce_0_10 [xlabel=10, label=<Reduce<BR /><FONT POINT-SIZE="8">Array(Integer32, #0, #2)</FONT><BR /><FONT POINT-SIZE="8">ParallelReduce</FONT>>, color=darkblue];
+return_0_11 [xlabel=11, label=<Return>, color=darkred];
+read_0_12 [xlabel=12, label=<Read<BR /><FONT POINT-SIZE="8">Integer32</FONT>>, color=darkblue];
+read_0_13 [xlabel=13, label=<Read<BR /><FONT POINT-SIZE="8">Integer32</FONT>>, color=darkblue];
+mul_0_14 [xlabel=14, label=<Mul<BR /><FONT POINT-SIZE="8">Integer32</FONT>>, color=darkblue];
+reduce_0_15 [xlabel=15, label=<Reduce<BR /><FONT POINT-SIZE="8">Integer32</FONT><BR /><FONT POINT-SIZE="8">TightAssociative</FONT>>, color=darkblue];
+add_0_16 [xlabel=16, label=<Add<BR /><FONT POINT-SIZE="8">Integer32</FONT>>, color=darkblue];
+write_0_17 [xlabel=17, label=<Write<BR /><FONT POINT-SIZE="8">Array(Integer32, #0, #2)</FONT>>, color=darkblue];
+constant_0_18 [xlabel=18, label=<Constant ([])<BR /><FONT POINT-SIZE="8">Array(Integer32, #0, #2)</FONT><BR /><FONT POINT-SIZE="8">NoResetConstant</FONT>>, color=darkblue];
+constant_0_19 [xlabel=19, label=<Constant (0)<BR /><FONT POINT-SIZE="8">Integer32</FONT>>, color=darkblue];
+start_0_0 -> fork_0_3 [color=black, style="dashed"];
+fork_0_3 -> thread_id_0_4 [color=black, style="dotted"];
+fork_0_3 -> thread_id_0_5 [color=black, style="dotted"];
+fork_0_3 -> fork_0_6 [color=black, style="dashed"];
+fork_0_6 -> thread_id_0_7 [color=black, style="dotted"];
+fork_0_6 -> join_0_8 [color=black, style="dashed"];
+join_0_8 -> join_0_9 [color=black, style="dashed"];
+join_0_9 -> reduce_0_10 [color=black, style="dotted"];
+constant_0_18 -> reduce_0_10 [color=black, style=""];
+write_0_17 -> reduce_0_10 [color=black, style=""];
+join_0_9 -> return_0_11 [color=black, style="dashed"];
+reduce_0_10 -> return_0_11 [color=black, style="dotted"];
+parameter_0_1 -> read_0_12 [color=black, style=""];
+thread_id_0_4 -> read_0_12 [color=black, style=""];
+thread_id_0_7 -> read_0_12 [color=black, style=""];
+parameter_0_2 -> read_0_13 [color=black, style=""];
+thread_id_0_7 -> read_0_13 [color=black, style=""];
+thread_id_0_5 -> read_0_13 [color=black, style=""];
+read_0_12 -> mul_0_14 [color=black, style=""];
+read_0_13 -> mul_0_14 [color=black, style=""];
+join_0_8 -> reduce_0_15 [color=black, style="dotted"];
+constant_0_19 -> reduce_0_15 [color=black, style=""];
+add_0_16 -> reduce_0_15 [color=black, style=""];
+mul_0_14 -> add_0_16 [color=black, style=""];
+reduce_0_15 -> add_0_16 [color=black, style=""];
+reduce_0_10 -> write_0_17 [color=black, style=""];
+reduce_0_15 -> write_0_17 [color=black, style=""];
+thread_id_0_4 -> write_0_17 [color=black, style=""];
+thread_id_0_5 -> write_0_17 [color=black, style=""];
+parameter_0_1 -> thread_id_0_4 [style=invis];
+parameter_0_2 -> thread_id_0_5 [style=invis];
+}
+}