Fix quicksort benchmark

compiler/build/src/program.rs

@@ -1,7 +1,11 @@
 use bumpalo::Bump;
 use inkwell::context::Context;
 use inkwell::module::Linkage;
+use inkwell::targets::{
+    CodeModel, FileType, InitializationConfig, RelocMode, Target, TargetTriple,
+};
 use inkwell::OptimizationLevel;
+use roc_collections::all::default_hasher;
 use roc_gen::layout_id::LayoutIds;
 use roc_gen::llvm::build::{
     build_proc, build_proc_header, get_call_conventions, module_from_builtins, OptLevel,
@@ -9,10 +13,7 @@ use roc_gen::llvm::build::{
 use roc_load::file::LoadedModule;
 use roc_mono::ir::{Env, PartialProc, Procs};
 use roc_mono::layout::LayoutCache;
-
+use std::collections::HashSet;
-use inkwell::targets::{
-    CodeModel, FileType, InitializationConfig, RelocMode, Target, TargetTriple,
-};
 use std::path::{Path, PathBuf};
 use target_lexicon::{Architecture, OperatingSystem, Triple, Vendor};
 
@@ -95,6 +96,12 @@ pub fn gen(
         home,
         ident_ids: &mut ident_ids,
     };
+    let mut exposed_symbols =
+        HashSet::with_capacity_and_hasher(loaded.exposed_vars_by_symbol.len(), default_hasher());
+
+    for (symbol, _) in loaded.exposed_vars_by_symbol {
+        exposed_symbols.insert(symbol);
+    }
 
     // Add modules' decls to Procs
     for (_, mut decls) in decls_by_id
@@ -113,6 +120,34 @@ pub fn gen(
                             Closure(annotation, _, _, loc_args, boxed_body) => {
                                 let (loc_body, ret_var) = *boxed_body;
 
+                                // If this is an exposed symbol, we need to
+                                // register it as such. Otherwise, since it
+                                // never gets called by Roc code, it will never
+                                // get specialized!
+                                if exposed_symbols.contains(&symbol) {
+                                    let mut pattern_vars =
+                                        bumpalo::collections::Vec::with_capacity_in(
+                                            loc_args.len(),
+                                            arena,
+                                        );
+
+                                    for (var, _) in loc_args.iter() {
+                                        pattern_vars.push(*var);
+                                    }
+
+                                    let layout = layout_cache.from_var(mono_env.arena, annotation, mono_env.subs).unwrap_or_else(|err|
+                                        todo!("TODO gracefully handle the situation where we expose a function to the host which doesn't have a valid layout (e.g. maybe the function wasn't monomorphic): {:?}", err)
+                                    );
+
+                                    procs.insert_exposed(
+                                        symbol,
+                                        layout,
+                                        pattern_vars, //: Vec<'a, Variable>,
+                                        annotation,
+                                        ret_var,
+                                    );
+                                }
+
                                 procs.insert_named(
                                     &mut mono_env,
                                     &mut layout_cache,
@@ -197,7 +232,7 @@ pub fn gen(
         // NOTE: This is here to be uncommented in case verification fails.
         // (This approach means we don't have to defensively clone name here.)
         //
-        // println!("\n\nBuilding and then verifying function {}\n\n", name);
+        // println!("\n\nBuilding and then verifying function {:?}\n\n", proc);
         build_proc(&env, &mut layout_ids, proc, fn_val, arg_basic_types);
 
         if fn_val.verify(true) {
@@ -215,15 +250,13 @@ pub fn gen(
         let cc = get_call_conventions(target.default_calling_convention().unwrap());
         let interns = &env.interns;
 
-        for (symbol, _) in loaded.exposed_vars_by_symbol {
+        for symbol in exposed_symbols {
-            let fn_name = format!(
+            // Since it was exposed, it will be monomorphic, so its LLVM name
-                "{}.{}",
+            // will be ___#1 (e.g. "main#1")
-                symbol.module_string(interns),
+            let fn_name = format!("{}#1", symbol.ident_string(interns));
-                symbol.ident_string(interns)
-            );
             let fn_val = env.module.get_function(&fn_name).unwrap();
 
-            fn_val.set_linkage(Linkage::AvailableExternally);
+            fn_val.set_linkage(Linkage::External);
             fn_val.set_call_conventions(cc);
         }
     }

compiler/gen/src/llvm/build.rs

@@ -1365,10 +1365,10 @@ pub fn build_proc_header<'a, 'ctx, 'env>(
     }
 
     let fn_type = get_fn_type(&ret_type, &arg_basic_types);
-
     let fn_name = layout_ids
         .get(symbol, layout)
         .to_symbol_string(symbol, &env.interns);
+
     let fn_val = env
         .module
         .add_function(fn_name.as_str(), fn_type, Some(Linkage::Private));

compiler/mono/src/expr.rs

@@ -108,6 +108,41 @@ impl<'a> Procs<'a> {
         }
     }
 
+    /// Add a named function that will be publicly exposed to the host
+    pub fn insert_exposed(
+        &mut self,
+        name: Symbol,
+        layout: Layout<'a>,
+        pattern_vars: Vec<'a, Variable>,
+        fn_var: Variable,
+        ret_var: Variable,
+    ) {
+        let tuple = (name, layout);
+
+        // If we've already specialized this one, no further work is needed.
+        if self.specialized.contains_key(&tuple) {
+            return;
+        }
+
+        // We're done with that tuple, so move layout back out to avoid cloning it.
+        let (name, layout) = tuple;
+        let pending = PendingSpecialization {
+            pattern_vars,
+            ret_var,
+            fn_var,
+        };
+
+        // This should only be called when pending_specializations is Some.
+        // Otherwise, it's being called in the wrong pass!
+        match &mut self.pending_specializations {
+            Some(pending_specializations) => {
+                // register the pending specialization, so this gets code genned later
+                add_pending(pending_specializations, name, layout, pending)
+            }
+            None => unreachable!("insert_exposed was called after the pending specializations phase had already completed!"),
+        }
+    }
+
     // TODO trim these down
     #[allow(clippy::too_many_arguments)]
     pub fn insert_anonymous(

examples/quicksort/host.rs

@@ -7,7 +7,7 @@ use std::time::SystemTime;
 #[link(name = "roc_app", kind = "static")]
 extern "C" {
     #[allow(improper_ctypes)]
-    #[link_name = "0#1"]
+    #[link_name = "main#1"]
     fn quicksort(list: Box<[i64]>) -> Box<[i64]>;
 }
 
@@ -46,13 +46,19 @@ pub fn main() {
         }
     };
 
+    // TODO FIXME don't truncate! This is just for testing.
+    nums.truncate(1000);
+
+    let nums: Box<[i64]> = nums.into();
+
+    println!("Running Roc quicksort on {} numbers...", nums.len());
     let start_time = SystemTime::now();
-    let answer = unsafe { quicksort(nums) };
+    let _answer = unsafe { quicksort(nums) };
     let end_time = SystemTime::now();
     let duration = end_time.duration_since(start_time).unwrap();
 
     println!(
-        "Roc quicksort took {:?} ms to compute this answer: {:?}",
+        "Roc quicksort took {:.4} ms to compute this answer: {:?}",
         duration.as_secs_f64() * 1000.0,
         list
     );