[libc] Reduce reference counter to a 32-bit integer

[jhuber6]

Summary:
This reference counter tracks how many threads are using a given slab.
Currently it's a 64-bit integer, this patch reduces it to a 32-bit
integer. The benefit of this is that we save a few registers now that we
no longer need to use two for these operations. This increases the risk
of overflow, but given that the largest value we accept for a single
slab is ~131,000 it is a long way off of the maximum of four billion or
so. Obviously we can oversubscribe the reference count by having threads
attempt to claim the lock and then try to free it, but I assert that it
is exceedingly unlikely that we will somehow have over four billion GPU
threads stalled in the same place.

A later optimization could be done to split the reference counter and
pointers into a struct of arrays, that will save 128 KiB of static
memory (as we currently use 512 KiB for the slab array).

[llvmbot]

@llvm/pr-subscribers-libc

Author: Joseph Huber (jhuber6)

Changes

Summary:
This reference counter tracks how many threads are using a given slab.
Currently it's a 64-bit integer, this patch reduces it to a 32-bit
integer. The benefit of this is that we save a few registers now that we
no longer need to use two for these operations. This increases the risk
of overflow, but given that the largest value we accept for a single
slab is ~131,000 it is a long way off of the maximum of four billion or
so. Obviously we can oversubscribe the reference count by having threads
attempt to claim the lock and then try to free it, but I assert that it
is exceedingly unlikely that we will somehow have over four billion GPU
threads stalled in the same place.

A later optimization could be done to split the reference counter and
pointers into a struct of arrays, that will save 128 KiB of static
memory (as we currently use 512 KiB for the slab array).

---

Full diff: https://github.com/llvm/llvm-project/pull/150961.diff

1 Files Affected:

• (modified) libc/src/__support/GPU/allocator.cpp (+23-21)

diff --git a/libc/src/__support/GPU/allocator.cpp b/libc/src/__support/GPU/allocator.cpp
index b9c86c5ab73c9..d588ee388ad55 100644
--- a/libc/src/__support/GPU/allocator.cpp
+++ b/libc/src/__support/GPU/allocator.cpp
@@ -40,7 +40,8 @@ constexpr static uint32_t MIN_ALIGNMENT = MIN_SIZE - 1;
 constexpr static uint32_t MAX_TRIES = 1024;
 
 // A sentinel used to indicate an invalid but non-null pointer value.
-constexpr static uint64_t SENTINEL = cpp::numeric_limits<uint64_t>::max();
+template <typename T>
+constexpr static T SENTINEL = cpp::numeric_limits<T>::max();
 
 static_assert(!(ARRAY_SIZE & (ARRAY_SIZE - 1)), "Must be a power of two");
 
@@ -343,20 +344,20 @@ struct GuardPtr {
 private:
   struct RefCounter {
     // Indicates that the object is in its deallocation phase and thus invalid.
-    static constexpr uint64_t INVALID = uint64_t(1) << 63;
+    static constexpr uint32_t INVALID = uint32_t(1) << 31;
 
     // If a read preempts an unlock call we indicate this so the following
     // unlock call can swap out the helped bit and maintain exclusive ownership.
-    static constexpr uint64_t HELPED = uint64_t(1) << 62;
+    static constexpr uint32_t HELPED = uint32_t(1) << 30;
 
     // Resets the reference counter, cannot be reset to zero safely.
-    void reset(uint32_t n, uint64_t &count) {
+    void reset(uint32_t n, uint32_t &count) {
       counter.store(n, cpp::MemoryOrder::RELAXED);
       count = n;
     }
 
     // Acquire a slot in the reference counter if it is not invalid.
-    bool acquire(uint32_t n, uint64_t &count) {
+    bool acquire(uint32_t n, uint32_t &count) {
       count = counter.fetch_add(n, cpp::MemoryOrder::RELAXED) + n;
       return (count & INVALID) == 0;
     }
@@ -369,7 +370,7 @@ struct GuardPtr {
       // another thread resurrected the counter and we quit, or a parallel read
       // helped us invalidating it. For the latter, claim that flag and return.
       if (counter.fetch_sub(n, cpp::MemoryOrder::RELAXED) == n) {
-        uint64_t expected = 0;
+        uint32_t expected = 0;
         if (counter.compare_exchange_strong(expected, INVALID,
                                             cpp::MemoryOrder::RELAXED,
                                             cpp::MemoryOrder::RELAXED))
@@ -392,28 +393,28 @@ struct GuardPtr {
       return (val & INVALID) ? 0 : val;
     }
 
-    cpp::Atomic<uint64_t> counter{0};
+    cpp::Atomic<uint32_t> counter{0};
   };
 
-  cpp::Atomic<Slab *> ptr{nullptr};
-  RefCounter ref{};
+  cpp::Atomic<Slab *> ptr;
+  RefCounter ref;
 
   // Should be called be a single lane for each different pointer.
   template <typename... Args>
-  Slab *try_lock_impl(uint32_t n, uint64_t &count, Args &&...args) {
+  Slab *try_lock_impl(uint32_t n, uint32_t &count, Args &&...args) {
     Slab *expected = ptr.load(cpp::MemoryOrder::RELAXED);
     if (!expected &&
         ptr.compare_exchange_strong(
-            expected, reinterpret_cast<Slab *>(SENTINEL),
+            expected, reinterpret_cast<Slab *>(SENTINEL<uint64_t>),
             cpp::MemoryOrder::RELAXED, cpp::MemoryOrder::RELAXED)) {
-      count = cpp::numeric_limits<uint64_t>::max();
+      count = cpp::numeric_limits<uint32_t>::max();
       void *raw = impl::rpc_allocate(sizeof(Slab));
       if (!raw)
         return nullptr;
       return new (raw) Slab(cpp::forward<Args>(args)...);
     }
 
-    if (!expected || expected == reinterpret_cast<Slab *>(SENTINEL))
+    if (!expected || expected == reinterpret_cast<Slab *>(SENTINEL<uint64_t>))
       return nullptr;
 
     if (!ref.acquire(n, count))
@@ -425,7 +426,7 @@ struct GuardPtr {
 
   // Finalize the associated memory and signal that it is ready to use by
   // resetting the counter.
-  void finalize(Slab *mem, uint32_t n, uint64_t &count) {
+  void finalize(Slab *mem, uint32_t n, uint32_t &count) {
     cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
     ptr.store(mem, cpp::MemoryOrder::RELAXED);
     cpp::atomic_thread_fence(cpp::MemoryOrder::ACQUIRE);
@@ -438,7 +439,7 @@ struct GuardPtr {
   // The uniform mask represents which lanes share the same pointer. For each
   // uniform value we elect a leader to handle it on behalf of the other lanes.
   template <typename... Args>
-  Slab *try_lock(uint64_t lane_mask, uint64_t uniform, uint64_t &count,
+  Slab *try_lock(uint64_t lane_mask, uint64_t uniform, uint32_t &count,
                  Args &&...args) {
     count = 0;
     Slab *result = nullptr;
@@ -453,13 +454,13 @@ struct GuardPtr {
 
     // We defer storing the newly allocated slab until now so that we can use
     // multiple lanes to initialize it and release it for use.
-    if (count == cpp::numeric_limits<uint64_t>::max()) {
+    if (count == SENTINEL<uint32_t>) {
       result->initialize(uniform);
       if (gpu::get_lane_id() == uint32_t(cpp::countr_zero(uniform)))
         finalize(result, cpp::popcount(uniform), count);
     }
 
-    if (count != cpp::numeric_limits<uint64_t>::max())
+    if (count != SENTINEL<uint32_t>)
       count = count - cpp::popcount(uniform) +
               impl::lane_count(uniform, gpu::get_lane_id()) + 1;
 
@@ -515,14 +516,15 @@ static Slab *find_slab(uint32_t chunk_size, uint64_t &uniform) {
     if (!offset ||
         slots[index].use_count() < Slab::available_chunks(chunk_size)) {
       uint64_t lane_mask = gpu::get_lane_mask();
-      uint64_t reserved = 0;
+      uint32_t reserved = 0;
 
       Slab *slab = slots[index].try_lock(lane_mask, uniform & lane_mask,
                                          reserved, chunk_size, index);
 
       // If there is a slab allocation in progress we retry a few times.
       for (uint32_t retries = 0;
-           retries < MAX_TRIES && !slab && reserved != SENTINEL; retries++) {
+           retries < MAX_TRIES && !slab && reserved != SENTINEL<uint32_t>;
+           retries++) {
         uint64_t lane_mask = gpu::get_lane_mask();
         slab = slots[index].try_lock(lane_mask, uniform & lane_mask, reserved,
                                      chunk_size, index);
@@ -542,7 +544,7 @@ static Slab *find_slab(uint32_t chunk_size, uint64_t &uniform) {
                           slab->get_chunk_size() != chunk_size)) {
         slots[index].unlock(gpu::get_lane_mask(),
                             gpu::get_lane_mask() & uniform);
-      } else if (!slab && reserved == SENTINEL) {
+      } else if (!slab && reserved == SENTINEL<uint32_t>) {
         uniform = uniform & gpu::get_lane_mask();
         return nullptr;
       } else {
@@ -575,7 +577,7 @@ void *allocate(uint64_t size) {
   uint32_t chunk_size = impl::get_chunk_size(static_cast<uint32_t>(size));
   uint64_t uniform = gpu::match_any(gpu::get_lane_mask(), chunk_size);
   Slab *slab = find_slab(chunk_size, uniform);
-  if (!slab || slab == reinterpret_cast<Slab *>(SENTINEL))
+  if (!slab || slab == reinterpret_cast<Slab *>(SENTINEL<uint64_t>))
     return nullptr;
 
   uint64_t lane_mask = gpu::get_lane_mask();

[arsenm]

Use a typedef somewhere?

[jhuber6]

Use a typedef somewhere?

Could possibly make it a setting if we need to change it later.