Changeset 2af1943 for libcfa/src/concurrency/kernel

Timestamp:

Jul 28, 2022, 11:59:02 AM (22 months ago)

Author:

Thierry Delisle <tdelisle@…>

Branches:

ADT, ast-experimental, master, pthread-emulation

Children:

c4c8571

Parents:

13d326ec

Message:

Small improvements to some of the alignment requirements in CFA runtime.

Location:

libcfa/src/concurrency/kernel

Files:

• cluster.cfa (modified) (6 diffs)
• cluster.hfa (modified) (2 diffs)
• fwd.hfa (modified) (1 diff)
• private.hfa (modified) (1 diff)
• startup.cfa (modified) (1 diff)

libcfa/src/concurrency/kernel/cluster.cfa

@@ -229 +229 @@
                         for( idx ; lanes_count ) {
                                 __intrusive_lane_t & sl = readyQ.data[idx];
-                                assert(!readyQ.data[idx].lock);
+                                assert(!readyQ.data[idx].l.lock);
 
                                         if(is_empty(sl)) {
-                                                assert( sl.anchor.next == 0p );
-                                                assert( sl.anchor.ts   == MAX );
-                                                assert( mock_head(sl)  == sl.prev );
+                                                assert( sl.l.anchor.next == 0p );
+                                                assert( sl.l.anchor.ts   == MAX );
+                                                assert( mock_head(sl)  == sl.l.prev );
                                         } else {
-                                                assert( sl.anchor.next != 0p );
-                                                assert( sl.anchor.ts   != MAX );
-                                                assert( mock_head(sl)  != sl.prev );
+                                                assert( sl.l.anchor.next != 0p );
+                                                assert( sl.l.anchor.ts   != MAX );
+                                                assert( mock_head(sl)  != sl.l.prev );
                                         }
                         }
@@ -249 +249 @@
 static inline void fix(__intrusive_lane_t & ll) {
         if(is_empty(ll)) {
-                verify(ll.anchor.next == 0p);
-                ll.prev = mock_head(ll);
+                verify(ll.l.anchor.next == 0p);
+                ll.l.prev = mock_head(ll);
         }
 }
@@ -299 +299 @@
         tscs = alloc(count, tscs`realloc);
         for(i; count) {
-                tscs[i].tv = rdtscl();
-                tscs[i].ma = 0;
+                tscs[i].t.tv = rdtscl();
+                tscs[i].t.ma = 0;
         }
 }
@@ -400 +400 @@
                 for( idx; ncount ~ ocount) {
                         // Lock is not strictly needed but makes checking invariants much easier
-                        __attribute__((unused)) bool locked = __atomic_try_acquire(&readyQ.data[idx].lock);
+                        __attribute__((unused)) bool locked = __atomic_try_acquire(&readyQ.data[idx].l.lock);
                         verify(locked);
 
@@ -418 +418 @@
 
                         // Unlock the lane
-                        __atomic_unlock(&readyQ.data[idx].lock);
+                        __atomic_unlock(&readyQ.data[idx].l.lock);
 
                         // TODO print the queue statistics here
@@ -467 +467 @@
 }
 
+#define nested_offsetof(type, field) ((off_t)(&(((type*)0)-> field)))
+
 // Ctor
 void ?{}( __intrusive_lane_t & this ) {
-        this.lock = false;
-        this.prev = mock_head(this);
-        this.anchor.next = 0p;
-        this.anchor.ts   = MAX;
+        this.l.lock = false;
+        this.l.prev = mock_head(this);
+        this.l.anchor.next = 0p;
+        this.l.anchor.ts   = MAX;
         #if !defined(__CFA_NO_STATISTICS__)
-                this.cnt  = 0;
+                this.l.cnt  = 0;
         #endif
 
         // We add a boat-load of assertions here because the anchor code is very fragile
-        /* paranoid */ _Static_assert( offsetof( thread$, link ) == offsetof(__intrusive_lane_t, anchor) );
-        /* paranoid */ verify( offsetof( thread$, link ) == offsetof(__intrusive_lane_t, anchor) );
-        /* paranoid */ verify( ((uintptr_t)( mock_head(this) ) + offsetof( thread$, link )) == (uintptr_t)(&this.anchor) );
-        /* paranoid */ verify( &mock_head(this)->link.next == &this.anchor.next );
-        /* paranoid */ verify( &mock_head(this)->link.ts   == &this.anchor.ts   );
+        /* paranoid */ _Static_assert( offsetof( thread$, link ) == nested_offsetof(__intrusive_lane_t, l.anchor) );
+        /* paranoid */ verify( offsetof( thread$, link ) == nested_offsetof(__intrusive_lane_t, l.anchor) );
+        /* paranoid */ verify( ((uintptr_t)( mock_head(this) ) + offsetof( thread$, link )) == (uintptr_t)(&this.l.anchor) );
+        /* paranoid */ verify( &mock_head(this)->link.next == &this.l.anchor.next );
+        /* paranoid */ verify( &mock_head(this)->link.ts   == &this.l.anchor.ts   );
         /* paranoid */ verify( mock_head(this)->link.next == 0p );
         /* paranoid */ verify( mock_head(this)->link.ts   == MAX );
-        /* paranoid */ verify( mock_head(this) == this.prev );
-        /* paranoid */ verify( __alignof__(__intrusive_lane_t) == 128 );
-        /* paranoid */ verify( __alignof__(this) == 128 );
-        /* paranoid */ verifyf( ((intptr_t)(&this) % 128) == 0, "Expected address to be aligned %p %% 128 == %zd", &this, ((intptr_t)(&this) % 128) );
-}
+        /* paranoid */ verify( mock_head(this) == this.l.prev );
+        /* paranoid */ verify( __alignof__(__intrusive_lane_t) == 64 );
+        /* paranoid */ verify( __alignof__(this) == 64 );
+        /* paranoid */ verifyf( ((intptr_t)(&this) % 64) == 0, "Expected address to be aligned %p %% 64 == %zd", &this, ((intptr_t)(&this) % 64) );
+}
+
+#undef nested_offsetof
 
 // Dtor is trivial
 void ^?{}( __intrusive_lane_t & this ) {
         // Make sure the list is empty
-        /* paranoid */ verify( this.anchor.next == 0p );
-        /* paranoid */ verify( this.anchor.ts   == MAX );
-        /* paranoid */ verify( mock_head(this)  == this.prev );
+        /* paranoid */ verify( this.l.anchor.next == 0p );
+        /* paranoid */ verify( this.l.anchor.ts   == MAX );
+        /* paranoid */ verify( mock_head(this)    == this.l.prev );
 }
 

libcfa/src/concurrency/kernel/cluster.hfa

@@ -39 +39 @@
         if (ts_next == ULLONG_MAX) return;
         unsigned long long now = rdtscl();
-        unsigned long long pma = __atomic_load_n(&tscs[ idx ].ma, __ATOMIC_RELAXED);
-        __atomic_store_n(&tscs[ idx ].tv, ts_next, __ATOMIC_RELAXED);
-        __atomic_store_n(&tscs[ idx ].ma, moving_average(now, ts_prev, pma), __ATOMIC_RELAXED);
+        unsigned long long pma = __atomic_load_n(&tscs[ idx ].t.ma, __ATOMIC_RELAXED);
+        __atomic_store_n(&tscs[ idx ].t.tv, ts_next, __ATOMIC_RELAXED);
+        __atomic_store_n(&tscs[ idx ].t.ma, moving_average(now, ts_prev, pma), __ATOMIC_RELAXED);
 }
 
@@ -61 +61 @@
                 if(ptsc != ULLONG_MAX) {
                         /* paranoid */ verify( start + i < count );
-                        unsigned long long tsc = moving_average(ctsc, ptsc, tscs[start + i].ma);
+                        unsigned long long tsc = moving_average(ctsc, ptsc, tscs[start + i].t.ma);
                         if(tsc > max) max = tsc;
                 }

libcfa/src/concurrency/kernel/fwd.hfa

@@ -35 +35 @@
 extern "C" {
         extern "Cforall" {
-                extern __attribute__((aligned(128))) thread_local struct KernelThreadData {
+                extern __attribute__((aligned(64))) thread_local struct KernelThreadData {
                         struct thread$          * volatile this_thread;
                         struct processor        * volatile this_processor;

libcfa/src/concurrency/kernel/private.hfa

@@ -88 +88 @@
 #elif defined(CFA_HAVE_LINUX_RSEQ_H)
         extern "Cforall" {
-                extern __attribute__((aligned(128))) thread_local volatile struct rseq __cfaabi_rseq;
+                extern __attribute__((aligned(64))) thread_local volatile struct rseq __cfaabi_rseq;
         }
 #else

libcfa/src/concurrency/kernel/startup.cfa

@@ -152 +152 @@
 #elif defined(CFA_HAVE_LINUX_RSEQ_H)
         extern "Cforall" {
-                __attribute__((aligned(128))) thread_local volatile struct rseq __cfaabi_rseq @= {
+                __attribute__((aligned(64))) thread_local volatile struct rseq __cfaabi_rseq @= {
                         .cpu_id : RSEQ_CPU_ID_UNINITIALIZED,
                 };