intrinsics: Eliminate recv intrinsic

This intrinsic existed just to get ahold of the return pointer. I replaced it
with a call_with_retptr intrinsic that grabs the return pointer and passes it
to another Rust function, thereby eliminating the need to call C functions on
the Rust stack.

4 files changed, 75 insertions, 59 deletions

diff --git a/src/rt/intrinsics/intrinsics.cpp b/src/rt/intrinsics/intrinsics.cpp
index 145a41c5a6d..23f4cf50672 100644
--- a/src/rt/intrinsics/intrinsics.cpp
+++ b/src/rt/intrinsics/intrinsics.cpp
@@ -53,12 +53,22 @@ rust_intrinsic_addr_of(void **retptr,
     *retptr = valptr;
 }
 
+struct rust_fn {
+    uintptr_t *fn;
+    rust_box *env;
+};
+
+typedef void (*retptr_fn)(void **retptr,
+			   void *env,
+			   void **dptr);
+// FIXME (1185): This exists just to get access to the return pointer
 extern "C" void
-rust_intrinsic_recv(void **retptr,
-                    void *env,
-                    type_desc *ty,
-                    rust_port *port) {
-    port_recv((uintptr_t*)retptr, port);
+rust_intrinsic_call_with_retptr(void **retptr,
+				void *env,
+				type_desc *ty,
+				rust_fn *recvfn) {
+    retptr_fn fn = ((retptr_fn)(recvfn->fn));
+    ((retptr_fn)(*fn))(NULL, recvfn->env, retptr);
 }
 
 extern "C" void
diff --git a/src/rt/intrinsics/intrinsics.i386.ll.in b/src/rt/intrinsics/intrinsics.i386.ll.in
index effc6005972..10267ba1700 100644
--- a/src/rt/intrinsics/intrinsics.i386.ll.in
+++ b/src/rt/intrinsics/intrinsics.i386.ll.in
@@ -8,11 +8,20 @@ target triple = "@CFG_TARGET_TRIPLE@"
 %struct.UT_hash_table = type { %struct.UT_hash_bucket*, i32, i32, i32, %struct.UT_hash_handle*, i32, i32, i32, i32, i32 }
 %struct.UT_hash_bucket = type { %struct.UT_hash_handle*, i32, i32 }
 %struct.rust_vec = type { i32, i32, [0 x i8] }
-%class.rust_port = type { i32, i32, %class.rust_kernel*, %struct.rust_task*, i32, %class.circular_buffer, %class.lock_and_signal }
-%class.rust_kernel = type { i32 (...)**, %class.memory_region, %class.rust_log, %class.rust_srv*, %class.lock_and_signal, %class.array_list.0, %struct.randctx, i32, %class.hash_map, i32, i32, i32, %struct.rust_env* }
-%class.memory_region = type { i32 (...)**, %class.rust_srv*, %class.memory_region*, i32, %class.array_list, i8, i8, %class.lock_and_signal }
+%struct.rust_fn = type { i32*, %struct.rust_box* }
+%struct.rust_box = type opaque
+%struct.rust_task = type { %struct.rust_task_user, i32, %class.context, %struct.stk_seg*, i32, %struct.rust_scheduler*, %class.rust_crate_cache*, %class.rust_kernel*, i8*, %class.rust_task_list*, %struct.rust_cond*, i8*, %struct.rust_task*, i32, i32, %class.timer, i32*, i32, i32, %class.memory_region, i8, i8, i8, %class.lock_and_signal, %class.hash_map.4, %class.rust_obstack, %"class.std::map", i32, %"class.debug::task_debug_info" }
+%struct.rust_task_user = type { i32, i32, %struct.chan_handle, i32 }
+%struct.chan_handle = type { i32, i32 }
+%class.context = type { %struct.registers_t, %class.context* }
+%struct.registers_t = type { i32, i32, i32, i32, i32, i32, i32, i32, i16, i16, i16, i16, i16, i16, i32, i32 }
+%struct.stk_seg = type { %struct.stk_seg*, i32, i32, i32, [0 x i8] }
+%struct.rust_scheduler = type { %class.rust_thread, i32, i32, %class.rust_log, i32, %class.rust_srv*, i8*, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_crate_cache, %struct.randctx, %class.rust_kernel*, i32, i32, %class.lock_and_signal, i32, %union.pthread_attr_t, %struct.rust_env*, %class.context }
+%class.rust_thread = type { i32 (...)**, i8, i32 }
+%class.rust_log = type { i32 (...)**, %class.rust_srv*, %struct.rust_scheduler*, i8 }
 %class.rust_srv = type { i32 (...)**, %struct.rust_env*, %class.memory_region }
 %struct.rust_env = type { i32, i32, i8*, i8, i8, i8* }
+%class.memory_region = type { i32 (...)**, %class.rust_srv*, %class.memory_region*, i32, %class.array_list, i8, i8, %class.lock_and_signal }
 %class.array_list = type { i32, %"struct.memory_region::alloc_header"**, i32 }
 %"struct.memory_region::alloc_header" = type { i8 }
 %class.lock_and_signal = type { i32 (...)**, %union.pthread_cond_t, %union.pthread_mutex_t, i32, i8, i8 }
@@ -21,19 +30,16 @@ target triple = "@CFG_TARGET_TRIPLE@"
 %union.pthread_mutex_t = type { %"struct.<anonymous union>::__pthread_mutex_s" }
 %"struct.<anonymous union>::__pthread_mutex_s" = type { i32, i32, i32, i32, i32, %union.anon }
 %union.anon = type { i32 }
-%class.rust_log = type { i32 (...)**, %class.rust_srv*, %struct.rust_scheduler*, i8 }
-%struct.rust_scheduler = type { %class.rust_thread, i32, i32, %class.rust_log, i32, %class.rust_srv*, i8*, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_crate_cache, %struct.randctx, %class.rust_kernel*, i32, i32, %class.lock_and_signal, i32, %union.pthread_attr_t, %struct.rust_env*, %class.context }
-%class.rust_thread = type { i32 (...)**, i8, i32 }
 %class.rust_task_list = type { %class.indexed_list, %struct.rust_scheduler*, i8* }
-%class.indexed_list = type { i32 (...)**, %class.array_list.2 }
-%class.array_list.2 = type { i32, %struct.rust_task**, i32 }
-%struct.rust_task = type { %struct.rust_task_user, i32, %class.context, %struct.stk_seg*, i32, %struct.rust_scheduler*, %class.rust_crate_cache*, %class.rust_kernel*, i8*, %class.rust_task_list*, %struct.rust_cond*, i8*, %struct.rust_task*, i32, i32, %class.timer, i32*, i32, i32, %class.memory_region, i8, i8, i8, %class.lock_and_signal, %class.hash_map.4, %class.rust_obstack, %"class.std::map", i32, %"class.debug::task_debug_info" }
-%struct.rust_task_user = type { i32, i32, %struct.chan_handle, i32 }
-%struct.chan_handle = type { i32, i32 }
-%class.context = type { %struct.registers_t, %class.context* }
-%struct.registers_t = type { i32, i32, i32, i32, i32, i32, i32, i32, i16, i16, i16, i16, i16, i16, i32, i32 }
-%struct.stk_seg = type { %struct.stk_seg*, i32, i32, i32, [0 x i8] }
+%class.indexed_list = type { i32 (...)**, %class.array_list.1 }
+%class.array_list.1 = type { i32, %struct.rust_task**, i32 }
 %class.rust_crate_cache = type { %struct.type_desc*, %struct.rust_scheduler*, i32 }
+%struct.randctx = type { i32, [256 x i32], [256 x i32], i32, i32, i32 }
+%class.rust_kernel = type { i32 (...)**, %class.memory_region, %class.rust_log, %class.rust_srv*, %class.lock_and_signal, %class.array_list.3, %struct.randctx, i32, %class.hash_map, i32, i32, i32, %struct.rust_env* }
+%class.array_list.3 = type { i32, %struct.rust_scheduler**, i32 }
+%class.hash_map = type { %"struct.hash_map<int, rust_task *>::map_entry"* }
+%"struct.hash_map<int, rust_task *>::map_entry" = type opaque
+%union.pthread_attr_t = type { i32, [32 x i8] }
 %struct.rust_cond = type { i8 }
 %class.timer = type { i32 (...)**, i64, i64 }
 %class.hash_map.4 = type { %"struct.hash_map<int, rust_port *>::map_entry"* }
@@ -49,12 +55,6 @@ target triple = "@CFG_TARGET_TRIPLE@"
 %"class.std::map.5" = type { %"class.std::_Rb_tree.6" }
 %"class.std::_Rb_tree.6" = type { %"struct.std::_Rb_tree<void *, std::pair<void *const, std::basic_string<char> >, std::_Select1st<std::pair<void *const, std::basic_string<char> > >, std::less<void *>, std::allocator<std::pair<void *const, std::basic_string<char> > > >::_Rb_tree_impl" }
 %"struct.std::_Rb_tree<void *, std::pair<void *const, std::basic_string<char> >, std::_Select1st<std::pair<void *const, std::basic_string<char> > >, std::less<void *>, std::allocator<std::pair<void *const, std::basic_string<char> > > >::_Rb_tree_impl" = type { %"struct.std::less", %"struct.std::_Rb_tree_node_base", i32 }
-%struct.randctx = type { i32, [256 x i32], [256 x i32], i32, i32, i32 }
-%union.pthread_attr_t = type { i32, [32 x i8] }
-%class.array_list.0 = type { i32, %struct.rust_scheduler**, i32 }
-%class.hash_map = type { %"struct.hash_map<int, rust_task *>::map_entry"* }
-%"struct.hash_map<int, rust_task *>::map_entry" = type opaque
-%class.circular_buffer = type { %class.rust_kernel*, i32, i32, i32, i32, i8* }
 
 define void @rust_intrinsic_vec_len(i32* nocapture %retptr, i8* nocapture %env, %struct.type_desc* nocapture %ty, %struct.rust_vec** nocapture %vp) nounwind {
   %1 = load %struct.rust_vec** %vp, align 4, !tbaa !0
@@ -90,14 +90,17 @@ define void @rust_intrinsic_addr_of(i8** nocapture %retptr, i8* nocapture %env,
   ret void
 }
 
-define void @rust_intrinsic_recv(i8** %retptr, i8* nocapture %env, %struct.type_desc* nocapture %ty, %class.rust_port* %port) {
-  %1 = bitcast i8** %retptr to i32*
-  tail call void @port_recv(i32* %1, %class.rust_port* %port)
+define void @rust_intrinsic_call_with_retptr(i8** %retptr, i8* nocapture %env, %struct.type_desc* nocapture %ty, %struct.rust_fn* nocapture %recvfn) {
+  %1 = getelementptr inbounds %struct.rust_fn* %recvfn, i32 0, i32 0
+  %2 = load i32** %1, align 4, !tbaa !0
+  %3 = bitcast i32* %2 to void (i8**, i8*, i8**)*
+  %4 = getelementptr inbounds %struct.rust_fn* %recvfn, i32 0, i32 1
+  %5 = load %struct.rust_box** %4, align 4, !tbaa !0
+  %6 = bitcast %struct.rust_box* %5 to i8*
+  tail call void %3(i8** null, i8* %6, i8** %retptr)
   ret void
 }
 
-declare void @port_recv(i32*, %class.rust_port*)
-
 define void @rust_intrinsic_get_type_desc(i8** nocapture %retptr, i8* nocapture %env, %struct.type_desc* %ty) nounwind {
   %ty.c = bitcast %struct.type_desc* %ty to i8*
   store i8* %ty.c, i8** %retptr, align 4, !tbaa !0
diff --git a/src/rt/intrinsics/intrinsics.x86_64.ll.in b/src/rt/intrinsics/intrinsics.x86_64.ll.in
index aeec58a0fac..849acd5b3e0 100644
--- a/src/rt/intrinsics/intrinsics.x86_64.ll.in
+++ b/src/rt/intrinsics/intrinsics.x86_64.ll.in
@@ -8,11 +8,20 @@ target triple = "@CFG_TARGET_TRIPLE@"
 %struct.UT_hash_table = type { %struct.UT_hash_bucket*, i32, i32, i32, %struct.UT_hash_handle*, i64, i32, i32, i32, i32 }
 %struct.UT_hash_bucket = type { %struct.UT_hash_handle*, i32, i32 }
 %struct.rust_vec = type { i64, i64, [0 x i8] }
-%class.rust_port = type { i64, i64, %class.rust_kernel*, %struct.rust_task*, i64, %class.circular_buffer, %class.lock_and_signal }
-%class.rust_kernel = type { i32 (...)**, %class.memory_region, %class.rust_log, %class.rust_srv*, %class.lock_and_signal, %class.array_list.0, %struct.randctx, i64, %class.hash_map, i64, i32, i32, %struct.rust_env* }
-%class.memory_region = type { i32 (...)**, %class.rust_srv*, %class.memory_region*, i32, %class.array_list, i8, i8, %class.lock_and_signal }
+%struct.rust_fn = type { i64*, %struct.rust_box* }
+%struct.rust_box = type opaque
+%struct.rust_task = type { %struct.rust_task_user, i64, %class.context, %struct.stk_seg*, i64, %struct.rust_scheduler*, %class.rust_crate_cache*, %class.rust_kernel*, i8*, %class.rust_task_list*, %struct.rust_cond*, i8*, %struct.rust_task*, i32, i64, %class.timer, i64*, i32, i32, %class.memory_region, i8, i8, i8, %class.lock_and_signal, %class.hash_map.4, %class.rust_obstack, %"class.std::map", i32, %"class.debug::task_debug_info" }
+%struct.rust_task_user = type { i64, i64, %struct.chan_handle, i64 }
+%struct.chan_handle = type { i64, i64 }
+%class.context = type { %struct.registers_t, %class.context* }
+%struct.registers_t = type { [22 x i64] }
+%struct.stk_seg = type { %struct.stk_seg*, i64, i32, [0 x i8] }
+%struct.rust_scheduler = type { %class.rust_thread, i64, i64, %class.rust_log, i32, %class.rust_srv*, i8*, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_crate_cache, %struct.randctx, %class.rust_kernel*, i32, i32, %class.lock_and_signal, i64, %union.pthread_attr_t, %struct.rust_env*, %class.context }
+%class.rust_thread = type { i32 (...)**, i8, i64 }
+%class.rust_log = type { i32 (...)**, %class.rust_srv*, %struct.rust_scheduler*, i8 }
 %class.rust_srv = type { i32 (...)**, %struct.rust_env*, %class.memory_region }
 %struct.rust_env = type { i64, i64, i8*, i8, i8, i8* }
+%class.memory_region = type { i32 (...)**, %class.rust_srv*, %class.memory_region*, i32, %class.array_list, i8, i8, %class.lock_and_signal }
 %class.array_list = type { i64, %"struct.memory_region::alloc_header"**, i64 }
 %"struct.memory_region::alloc_header" = type { i8 }
 %class.lock_and_signal = type { i32 (...)**, %union.pthread_cond_t, %union.pthread_mutex_t, i64, i8, i8 }
@@ -21,19 +30,16 @@ target triple = "@CFG_TARGET_TRIPLE@"
 %union.pthread_mutex_t = type { %"struct.<anonymous union>::__pthread_mutex_s" }
 %"struct.<anonymous union>::__pthread_mutex_s" = type { i32, i32, i32, i32, i32, i32, %struct.__pthread_internal_list }
 %struct.__pthread_internal_list = type { %struct.__pthread_internal_list*, %struct.__pthread_internal_list* }
-%class.rust_log = type { i32 (...)**, %class.rust_srv*, %struct.rust_scheduler*, i8 }
-%struct.rust_scheduler = type { %class.rust_thread, i64, i64, %class.rust_log, i32, %class.rust_srv*, i8*, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_task_list, %class.rust_crate_cache, %struct.randctx, %class.rust_kernel*, i32, i32, %class.lock_and_signal, i64, %union.pthread_attr_t, %struct.rust_env*, %class.context }
-%class.rust_thread = type { i32 (...)**, i8, i64 }
 %class.rust_task_list = type { %class.indexed_list, %struct.rust_scheduler*, i8* }
-%class.indexed_list = type { i32 (...)**, %class.array_list.2 }
-%class.array_list.2 = type { i64, %struct.rust_task**, i64 }
-%struct.rust_task = type { %struct.rust_task_user, i64, %class.context, %struct.stk_seg*, i64, %struct.rust_scheduler*, %class.rust_crate_cache*, %class.rust_kernel*, i8*, %class.rust_task_list*, %struct.rust_cond*, i8*, %struct.rust_task*, i32, i64, %class.timer, i64*, i32, i32, %class.memory_region, i8, i8, i8, %class.lock_and_signal, %class.hash_map.4, %class.rust_obstack, %"class.std::map", i32, %"class.debug::task_debug_info" }
-%struct.rust_task_user = type { i64, i64, %struct.chan_handle, i64 }
-%struct.chan_handle = type { i64, i64 }
-%class.context = type { %struct.registers_t, %class.context* }
-%struct.registers_t = type { [22 x i64] }
-%struct.stk_seg = type { %struct.stk_seg*, i64, i32, [0 x i8] }
+%class.indexed_list = type { i32 (...)**, %class.array_list.1 }
+%class.array_list.1 = type { i64, %struct.rust_task**, i64 }
 %class.rust_crate_cache = type { %struct.type_desc*, %struct.rust_scheduler*, i64 }
+%struct.randctx = type { i64, [256 x i64], [256 x i64], i64, i64, i64 }
+%class.rust_kernel = type { i32 (...)**, %class.memory_region, %class.rust_log, %class.rust_srv*, %class.lock_and_signal, %class.array_list.3, %struct.randctx, i64, %class.hash_map, i64, i32, i32, %struct.rust_env* }
+%class.array_list.3 = type { i64, %struct.rust_scheduler**, i64 }
+%class.hash_map = type { %"struct.hash_map<long, rust_task *>::map_entry"* }
+%"struct.hash_map<long, rust_task *>::map_entry" = type opaque
+%union.pthread_attr_t = type { i64, [48 x i8] }
 %struct.rust_cond = type { i8 }
 %class.timer = type { i32 (...)**, i64, i64 }
 %class.hash_map.4 = type { %"struct.hash_map<long, rust_port *>::map_entry"* }
@@ -49,12 +55,6 @@ target triple = "@CFG_TARGET_TRIPLE@"
 %"class.std::map.5" = type { %"class.std::_Rb_tree.6" }
 %"class.std::_Rb_tree.6" = type { %"struct.std::_Rb_tree<void *, std::pair<void *const, std::basic_string<char> >, std::_Select1st<std::pair<void *const, std::basic_string<char> > >, std::less<void *>, std::allocator<std::pair<void *const, std::basic_string<char> > > >::_Rb_tree_impl" }
 %"struct.std::_Rb_tree<void *, std::pair<void *const, std::basic_string<char> >, std::_Select1st<std::pair<void *const, std::basic_string<char> > >, std::less<void *>, std::allocator<std::pair<void *const, std::basic_string<char> > > >::_Rb_tree_impl" = type { %"struct.std::less", %"struct.std::_Rb_tree_node_base", i64 }
-%struct.randctx = type { i64, [256 x i64], [256 x i64], i64, i64, i64 }
-%union.pthread_attr_t = type { i64, [48 x i8] }
-%class.array_list.0 = type { i64, %struct.rust_scheduler**, i64 }
-%class.hash_map = type { %"struct.hash_map<long, rust_task *>::map_entry"* }
-%"struct.hash_map<long, rust_task *>::map_entry" = type opaque
-%class.circular_buffer = type { %class.rust_kernel*, i64, i64, i64, i64, i8* }
 
 define void @rust_intrinsic_vec_len(i64* nocapture %retptr, i8* nocapture %env, %struct.type_desc* nocapture %ty, %struct.rust_vec** nocapture %vp) nounwind uwtable {
   %1 = load %struct.rust_vec** %vp, align 8, !tbaa !0
@@ -90,14 +90,17 @@ define void @rust_intrinsic_addr_of(i8** nocapture %retptr, i8* nocapture %env,
   ret void
 }
 
-define void @rust_intrinsic_recv(i8** %retptr, i8* nocapture %env, %struct.type_desc* nocapture %ty, %class.rust_port* %port) uwtable {
-  %1 = bitcast i8** %retptr to i64*
-  tail call void @port_recv(i64* %1, %class.rust_port* %port)
+define void @rust_intrinsic_call_with_retptr(i8** %retptr, i8* nocapture %env, %struct.type_desc* nocapture %ty, %struct.rust_fn* nocapture %recvfn) uwtable {
+  %1 = getelementptr inbounds %struct.rust_fn* %recvfn, i64 0, i32 0
+  %2 = load i64** %1, align 8, !tbaa !0
+  %3 = bitcast i64* %2 to void (i8**, i8*, i8**)*
+  %4 = getelementptr inbounds %struct.rust_fn* %recvfn, i64 0, i32 1
+  %5 = load %struct.rust_box** %4, align 8, !tbaa !0
+  %6 = bitcast %struct.rust_box* %5 to i8*
+  tail call void %3(i8** null, i8* %6, i8** %retptr)
   ret void
 }
 
-declare void @port_recv(i64*, %class.rust_port*)
-
 define void @rust_intrinsic_get_type_desc(i8** nocapture %retptr, i8* nocapture %env, %struct.type_desc* %ty) nounwind uwtable {
   %ty.c = bitcast %struct.type_desc* %ty to i8*
   store i8* %ty.c, i8** %retptr, align 8, !tbaa !0
diff --git a/src/rt/rust_builtin.cpp b/src/rt/rust_builtin.cpp
index 0ae191da0f6..3460d460900 100644
--- a/src/rt/rust_builtin.cpp
+++ b/src/rt/rust_builtin.cpp
@@ -515,10 +515,8 @@ rust_task_sleep(rust_task *task, size_t time_in_us) {
     task->yield(time_in_us);
 }
 
-// This is called by an intrinsic on the Rust stack.
-// Do not call on the C stack.
 extern "C" CDECL void
-port_recv(uintptr_t *dptr, rust_port *port) {
+port_recv(uintptr_t *dptr, rust_port *port, uintptr_t *yield) {
     rust_task *task = rust_scheduler::get_task();
     {
         scoped_lock with(port->lock);
@@ -528,6 +526,7 @@ port_recv(uintptr_t *dptr, rust_port *port) {
             (uintptr_t) port, (uintptr_t) dptr, port->unit_sz);
 
         if (port->receive(dptr)) {
+            *yield = false;
             return;
         }
 
@@ -539,7 +538,8 @@ port_recv(uintptr_t *dptr, rust_port *port) {
         task->rendezvous_ptr = dptr;
         task->block(port, "waiting for rendezvous data");
     }
-    task->yield(3);
+    *yield = true;
+    return;
 }
 
 //