Changeset 37ba662

Timestamp:

Jun 16, 2020, 3:32:00 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:

04b5cef

Parents:

2073d207

Message:

Minor improvements to alignments and memory layout

Location:

libcfa/src/concurrency

Files:

• invoke.h (modified) (1 diff)
• kernel.hfa (modified) (5 diffs)
• kernel_private.hfa (modified) (2 diffs)
• ready_queue.cfa (modified) (1 diff)
• stats.hfa (modified) (1 diff)

libcfa/src/concurrency/invoke.h

@@ -182 +182 @@
                 //SKULLDUGGERY errno is not save in the thread data structure because returnToKernel appears to be the only function to require saving and restoring it
 
-                // coroutine body used to store context
-                struct $coroutine  self_cor;
-
-                // current active context
-                struct $coroutine * curr_cor;
-
-                // monitor body used for mutual exclusion
-                struct $monitor    self_mon;
-
-                // pointer to monitor with sufficient lifetime for current monitors
-                struct $monitor *  self_mon_p;
-
                 // pointer to the cluster on which the thread is running
                 struct cluster * curr_cluster;
-
-                // monitors currently held by this thread
-                struct __monitor_group_t monitors;
 
                 // Link lists fields
                 // instrusive link field for threads
                 struct __thread_desc_link link;
+
+                // coroutine body used to store context
+                struct $coroutine  self_cor;
+
+                // current active context
+                struct $coroutine * curr_cor;
+
+                // monitor body used for mutual exclusion
+                struct $monitor    self_mon;
+
+                // pointer to monitor with sufficient lifetime for current monitors
+                struct $monitor *  self_mon_p;
+
+                // monitors currently held by this thread
+                struct __monitor_group_t monitors;
 
                 struct {

libcfa/src/concurrency/kernel.hfa

@@ -61 +61 @@
 
 // Wrapper around kernel threads
-struct processor {
+struct __attribute__((aligned(128))) processor {
+        // Main state
         inline __processor_id_t;
 
-        // Main state
+        // Cluster from which to get threads
+        struct cluster * cltr;
+
+        // Set to true to notify the processor should terminate
+        volatile bool do_terminate;
+
         // Coroutine ctx who does keeps the state of the processor
         struct processorCtx_t runner;
-
-        // Cluster from which to get threads
-        struct cluster * cltr;
 
         // Name of the processor
@@ -90 +93 @@
         // Idle lock (kernel semaphore)
         __bin_sem_t idle;
-
-        // Termination
-        // Set to true to notify the processor should terminate
-        volatile bool do_terminate;
 
         // Termination synchronisation (user semaphore)
@@ -153 +152 @@
 //TODO adjust cache size to ARCHITECTURE
 // Structure holding the relaxed ready queue
-struct __attribute__((aligned(128))) __ready_queue_t {
+struct __ready_queue_t {
         // Data tracking how many/which lanes are used
         // Aligned to 128 for cache locality
@@ -162 +161 @@
         // used can change on each push/pop but this data
         // only changes on shrink/grow
-        struct __attribute__((aligned(64))) {
+        struct {
                 // Arary of lanes
                 __intrusive_lane_t * volatile data;
@@ -176 +175 @@
 //-----------------------------------------------------------------------------
 // Cluster
-struct cluster {
+struct __attribute__((aligned(128))) cluster {
         // Ready queue for threads
         __ready_queue_t ready_queue;

libcfa/src/concurrency/kernel_private.hfa

@@ -26 +26 @@
 // Scheduler
 
-struct __attribute__((aligned(64))) __scheduler_lock_id_t;
+struct __attribute__((aligned(128))) __scheduler_lock_id_t;
 
 extern "C" {
@@ -111 +111 @@
 // Cells use by the reader writer lock
 // while not generic it only relies on a opaque pointer
-struct __attribute__((aligned(64))) __scheduler_lock_id_t {
+struct __attribute__((aligned(128))) __scheduler_lock_id_t {
         __processor_id_t * volatile handle;
         volatile bool lock;

libcfa/src/concurrency/ready_queue.cfa

@@ -120 +120 @@
         // Return new spot.
         /*paranoid*/ verify(n < ready);
-        /*paranoid*/ verify(__alignof__(data[n]) == cache_line_size);
+        /*paranoid*/ verify(__alignof__(data[n]) == (2 * cache_line_size));
         /*paranoid*/ verify((((uintptr_t)&data[n]) % cache_line_size) == 0);
         return n;

libcfa/src/concurrency/stats.hfa

@@ -66 +66 @@
         #endif
 
-        struct __stats_t {
+        struct __attribute__((aligned(128))) __stats_t {
                 __stats_readQ_t ready;
                 #if defined(HAVE_LINUX_IO_URING_H)