remove Ref allocation on task switch

Author: JeffBezanson

base/task.jl

@@ -615,6 +615,8 @@ function yield()
     end
 end
 
+@inline set_next_task(t::Task) = ccall(:jl_set_next_task, Cvoid, (Any,), t)
+
 """
     yield(t::Task, arg = nothing)
 
@@ -624,7 +626,8 @@ immediately yields to `t` before calling the scheduler.
 function yield(t::Task, @nospecialize(x=nothing))
     t.result = x
     enq_work(current_task())
-    return try_yieldto(ensure_rescheduled, Ref(t))
+    set_next_task(t)
+    return try_yieldto(ensure_rescheduled)
 end
 
 """
@@ -637,14 +640,15 @@ or scheduling in any way. Its use is discouraged.
 """
 function yieldto(t::Task, @nospecialize(x=nothing))
     t.result = x
-    return try_yieldto(identity, Ref(t))
+    set_next_task(t)
+    return try_yieldto(identity)
 end
 
-function try_yieldto(undo, reftask::Ref{Task})
+function try_yieldto(undo)
     try
-        ccall(:jl_switchto, Cvoid, (Any,), reftask)
+        ccall(:jl_switch, Cvoid, ())
     catch
-        undo(reftask[])
+        undo(ccall(:jl_get_next_task, Ref{Task}, ()))
         rethrow()
     end
     ct = current_task()
@@ -696,18 +700,19 @@ function trypoptask(W::StickyWorkqueue)
     return t
 end
 
-@noinline function poptaskref(W::StickyWorkqueue)
+@noinline function poptask(W::StickyWorkqueue)
     task = trypoptask(W)
     if !(task isa Task)
         task = ccall(:jl_task_get_next, Ref{Task}, (Any, Any), trypoptask, W)
     end
-    return Ref(task)
+    set_next_task(task)
+    nothing
 end
 
 function wait()
     W = Workqueues[Threads.threadid()]
-    reftask = poptaskref(W)
-    result = try_yieldto(ensure_rescheduled, reftask)
+    poptask(W)
+    result = try_yieldto(ensure_rescheduled)
     process_events()
     # return when we come out of the queue
     return result

src/ccall.cpp

@@ -1619,6 +1619,16 @@ static jl_cgval_t emit_ccall(jl_codectx_t &ctx, jl_value_t **args, size_t nargs)
             tbaa_decorate(tbaa_const, ctx.builder.CreateLoad(pct)),
             retboxed, rt, unionall, static_rt);
     }
+    else if (is_libjulia_func(jl_set_next_task)) {
+        assert(lrt == T_void);
+        assert(!isVa && !llvmcall && nccallargs == 1);
+        JL_GC_POP();
+        Value *ptls_pv = emit_bitcast(ctx, ctx.ptlsStates, T_ppjlvalue);
+        const int nt_offset = offsetof(jl_tls_states_t, next_task);
+        Value *pnt = ctx.builder.CreateGEP(ptls_pv, ConstantInt::get(T_size, nt_offset / sizeof(void*)));
+        ctx.builder.CreateStore(emit_pointer_from_objref(ctx, boxed(ctx, argv[0])), pnt);
+        return ghostValue(jl_nothing_type);
+    }
     else if (is_libjulia_func(jl_sigatomic_begin)) {
         assert(lrt == T_void);
         assert(!isVa && !llvmcall && nccallargs == 0);

src/gc.c

@@ -2645,6 +2645,8 @@ static void jl_gc_queue_thread_local(jl_gc_mark_cache_t *gc_cache, jl_gc_mark_sp
 {
     gc_mark_queue_obj(gc_cache, sp, ptls2->current_task);
     gc_mark_queue_obj(gc_cache, sp, ptls2->root_task);
+    if (ptls2->next_task)
+        gc_mark_queue_obj(gc_cache, sp, ptls2->next_task);
     if (ptls2->previous_exception)
         gc_mark_queue_obj(gc_cache, sp, ptls2->previous_exception);
 }

src/julia_internal.h

@@ -1011,6 +1011,7 @@ JL_DLLEXPORT int jl_array_isassigned(jl_array_t *a, size_t i);
 
 JL_DLLEXPORT uintptr_t jl_object_id_(jl_value_t *tv, jl_value_t *v) JL_NOTSAFEPOINT;
 JL_DLLEXPORT jl_value_t *jl_get_current_task(void);
+JL_DLLEXPORT void jl_set_next_task(jl_task_t *task);
 
 // -- synchronization utilities -- //
 

src/julia_threads.h

@@ -185,6 +185,7 @@ struct _jl_tls_states_t {
     uv_cond_t wake_signal;
     volatile sig_atomic_t defer_signal;
     struct _jl_task_t *current_task;
+    struct _jl_task_t *next_task;
 #ifdef MIGRATE_TASKS
     struct _jl_task_t *previous_task;
 #endif

src/task.c

@@ -248,10 +248,21 @@ JL_DLLEXPORT void julia_init(JL_IMAGE_SEARCH rel)
     _julia_init(rel);
 }
 
+JL_DLLEXPORT void jl_set_next_task(jl_task_t *task)
+{
+    jl_get_ptls_states()->next_task = task;
+}
+
+JL_DLLEXPORT jl_task_t *jl_get_next_task(void)
+{
+    return jl_get_ptls_states()->next_task;
+}
+
 void jl_release_task_stack(jl_ptls_t ptls, jl_task_t *task);
 
-static void ctx_switch(jl_ptls_t ptls, jl_task_t **pt)
+static void ctx_switch(jl_ptls_t ptls)
 {
+    jl_task_t **pt = &ptls->next_task;
     jl_task_t *t = *pt;
     assert(t != ptls->current_task);
     jl_task_t *lastt = ptls->current_task;
@@ -366,10 +377,10 @@ static jl_ptls_t NOINLINE refetch_ptls(void)
     return jl_get_ptls_states();
 }
 
-JL_DLLEXPORT void jl_switchto(jl_task_t **pt)
+JL_DLLEXPORT void jl_switch(void)
 {
     jl_ptls_t ptls = jl_get_ptls_states();
-    jl_task_t *t = *pt;
+    jl_task_t *t = ptls->next_task;
     jl_task_t *ct = ptls->current_task;
     if (t == ct) {
         return;
@@ -401,7 +412,7 @@ JL_DLLEXPORT void jl_switchto(jl_task_t **pt)
         jl_timing_block_stop(blk);
 #endif
 
-    ctx_switch(ptls, pt);
+    ctx_switch(ptls);
 
 #ifdef MIGRATE_TASKS
     ptls = refetch_ptls();
@@ -432,6 +443,12 @@ JL_DLLEXPORT void jl_switchto(jl_task_t **pt)
         jl_sigint_safepoint(ptls);
 }
 
+JL_DLLEXPORT void jl_switchto(jl_task_t **pt)
+{
+    jl_set_next_task(*pt);
+    jl_switch();
+}
+
 JL_DLLEXPORT JL_NORETURN void jl_no_exc_handler(jl_value_t *e)
 {
     jl_printf(JL_STDERR, "fatal: error thrown and no exception handler available.\n");

src/threading.c

@@ -285,6 +285,7 @@ void jl_init_threadtls(int16_t tid)
     ptls->bt_data = bt_data;
     ptls->sig_exception = NULL;
     ptls->previous_exception = NULL;
+    ptls->next_task = NULL;
 #ifdef _OS_WINDOWS_
     ptls->needs_resetstkoflw = 0;
 #endif