Changeset 320ec6fc for libcfa/src/concurrency

Timestamp:

Jul 24, 2020, 1:11:39 PM (4 years ago)

Author:

Thierry Delisle <tdelisle@…>

Branches:

ADT, arm-eh, ast-experimental, enum, forall-pointer-decay, jacob/cs343-translation, master, new-ast, new-ast-unique-expr, pthread-emulation, qualifiedEnum

Children:

39fc03e

Parents:

f4ec4a90

Message:

Changed ready_queue_(grow/shrink) to take a target instead of going incrementing

Location:

libcfa/src/concurrency

Files:

• kernel.cfa (modified) (5 diffs)
• kernel_private.hfa (modified) (1 diff)
• ready_queue.cfa (modified) (5 diffs)
• snzi.hfa (modified) (2 diffs)

libcfa/src/concurrency/kernel.cfa

@@ -242 +242 @@
         #endif
 
-        __atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
+        int target = __atomic_add_fetch( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
 
         id = doregister((__processor_id_t*)&this);
@@ -250 +250 @@
 
                 // Adjust the ready queue size
-                ready_queue_grow( cltr );
+                ready_queue_grow( cltr, target );
 
         // Unlock the RWlock
@@ -260 +260 @@
 // Not a ctor, it just preps the destruction but should not destroy members
 void deinit(processor & this) {
+
+        int target = __atomic_sub_fetch( &this.cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
+
         // Lock the RWlock so no-one pushes/pops while we are changing the queue
         uint_fast32_t last_size = ready_mutate_lock();
 
                 // Adjust the ready queue size
-                ready_queue_shrink( this.cltr );
+                ready_queue_shrink( this.cltr, target );
 
                 // Make sure we aren't on the idle queue
@@ -305 +308 @@
 
         deinit( this );
-
-        __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
 }
 
@@ -936 +937 @@
 
                 /* paranoid */ verify( this.do_terminate == true );
-                __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
                 __cfaabi_dbg_print_safe("Kernel : destroyed main processor context %p\n", &runner);
         }

libcfa/src/concurrency/kernel_private.hfa

@@ -274 +274 @@
 //-----------------------------------------------------------------------
 // Increase the width of the ready queue (number of lanes) by 4
-void ready_queue_grow  (struct cluster * cltr);
+void ready_queue_grow  (struct cluster * cltr, int target);
 
 //-----------------------------------------------------------------------
 // Decrease the width of the ready queue (number of lanes) by 4
-void ready_queue_shrink(struct cluster * cltr);
+void ready_queue_shrink(struct cluster * cltr, int target);
 
 //-----------------------------------------------------------------------

libcfa/src/concurrency/ready_queue.cfa

@@ -37 +37 @@
 #endif
 
-#define BIAS 64
+#define BIAS 16
 
 // returns the maximum number of processors the RWLock support
@@ -500 +500 @@
 
 // Grow the ready queue
-void ready_queue_grow  (struct cluster * cltr) {
+void ready_queue_grow  (struct cluster * cltr, int target) {
         /* paranoid */ verify( ready_mutate_islocked() );
         __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
@@ -515 +515 @@
                 // increase count
                 ncount += 4;
+                /* paranoid */ verify( ncount == target * 4 || target < 2 );
 
                 // Allocate new array (uses realloc and memcpies the data)
@@ -550 +551 @@
 
 // Shrink the ready queue
-void ready_queue_shrink(struct cluster * cltr) {
+void ready_queue_shrink(struct cluster * cltr, int target) {
         /* paranoid */ verify( ready_mutate_islocked() );
         __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
@@ -566 +567 @@
                 // reduce the actual count so push doesn't use the old queues
                 lanes.count -= 4;
-                verify(ocount > lanes.count);
+                /* paranoid */ verify( ocount > lanes.count );
+                /* paranoid */ verify( lanes.count == target * 4 || target < 2 );
 
                 // for printing count the number of displaced threads

libcfa/src/concurrency/snzi.hfa

@@ -148 +148 @@
 
 static inline void arrive( __snzi_t & this, int idx) {
+        idx >>= 2;
         idx &= this.mask;
         arrive( this.nodes[idx] );
@@ -153 +154 @@
 
 static inline void depart( __snzi_t & this, int idx) {
+        idx >>= 2;
         idx &= this.mask;
         depart( this.nodes[idx] );