[compiler-rt] [scudo] [MTE] resize stack depot for allocation ring buffer (PR #74515)
Christopher Ferris via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 12 19:06:14 PST 2023
================
@@ -1504,6 +1529,28 @@ class Allocator {
return;
u32 AllocationRingBufferSize =
static_cast<u32>(getFlags()->allocation_ring_buffer_size);
+ // We store alloc and free stacks for each entry.
+ constexpr auto kStacksPerRingBufferEntry = 2;
+ u32 TabSize = static_cast<u32>(roundUpPowerOfTwo(kStacksPerRingBufferEntry *
+ AllocationRingBufferSize));
+ constexpr auto kFramesPerStack = 8;
+ static_assert(isPowerOfTwo(kFramesPerStack));
+ u32 RingSize = static_cast<u32>(TabSize * kFramesPerStack);
+ DCHECK(isPowerOfTwo(RingSize));
+ static_assert(sizeof(StackDepot) % alignof(atomic_u64) == 0);
+
+ StackDepotSize = sizeof(StackDepot) + sizeof(atomic_u64) * RingSize +
----------------
cferris1000 wrote:
Is there any kind of alignment you have to worry about here? For example, if StackDepot winds up ending on an 8 byte aligned value, then atomic_u64 values will be aligned improperly. The next one doesn't matter because you get the same alignment across everything after that.
The alignas on the class, doesn't guarantee that a StackDepot object is aligned, but you might be using some other trick to guarantee the alignment I didn't catch.
https://github.com/llvm/llvm-project/pull/74515
More information about the llvm-commits
mailing list