From 5fba111604add7eb04b00d64d923fd47611ca140 Mon Sep 17 00:00:00 2001 From: inode64 overlay Date: Thu, 27 Aug 2026 12:00:00 +0200 Subject: [PATCH] rocclr: support over-aligned reference-counted objects ReferenceCountedObject provides class-specific allocation functions. C++ allocation-function lookup therefore does not reach the alignment-aware global overload for an over-aligned derived class when the class-specific set lacks an std::align_val_t overload. VirtualGPU is such a class: its barrier packets are alignas(64). An allocation returned with only the default alignment violates the class invariant and permits the compiler to issue aligned vector stores to an unaligned address. GCC 15 does exactly that when building with -march=native on AVX-512 hosts, causing a general-protection fault in the VirtualGPU constructor. Forward alignment-aware allocation and deallocation to the matching global operators so all over-aligned ReferenceCountedObject subclasses are allocated correctly. --- rocclr/include/top.hpp | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/rocclr/include/top.hpp b/rocclr/include/top.hpp index 8a1a5d2..8994051 100644 --- a/rocclr/include/top.hpp +++ b/rocclr/include/top.hpp @@ -164,6 +164,12 @@ class ReferenceCountedObject { void* operator new(size_t size) { return ::operator new(size); } void operator delete(void* p) { return ::operator delete(p); } + void* operator new(size_t size, std::align_val_t alignment) { + return ::operator new(size, alignment); + } + void operator delete(void* p, std::align_val_t alignment) { + return ::operator delete(p, alignment); + } void* operator new(size_t size, size_t extSize) { return ReferenceCountedObject::operator new(size + extSize); }; -- 2.51.0