New Multi-thread Friendly Memory Manager for FPC written in x86_64 assembly

As a gift to the FPC community, I just committed a new Memory Manager for FPC.
Check mormot.core.fpcx64mm.pas in our mORMot2 repository.
This is a stand-alone unit for FPC only.

It targets Windows and Linux multi-threaded Service applications - typically mORMot daemons.
It is written in almost pure x86_64 assembly, and some unique tricks in the Delphi/FPC Memory Manager world.

It is based on FastMM4 (not FastMM5), and we didn't follow the path of the FastMM4-AVX version - instead of AVX, we use plain good (non-temporal) SSE2 opcode, and we rely on the mremap API on Linux for very efficient reallocation. Using mremap is perhaps the biggest benefit of this memory manager - it leverages a killer feature of the Linux kernel for sure. By the way, we directly call the Kernel without the need of the libc.

We tuned our x86_64 assembly a lot, and made it cross-platform (Windows and POSIX). We profiled the multi-threading, especially by adding some additional small blocks for GetMem (which is a less expensive notion of "arenas" as used in FastMM5 and most C allocators), introducing an innovatice and very efficient round-robin of tiny blocks (<128 bytes), and proper spinning for FreeMem and medium blocks.

It runs all our regression tests with huge performance and stability - including multi-threaded tests with almost no slow down: sleep is reported as less than 1 ms during a 1 minute test. It has also been validated on some demanding multi-threaded tasks.

We are proud to offer a new Multi-thread Friendly Memory Manager for FPC written in x86_64 assembly, implemented as such:

targetting Linux (and Windows) multi-threaded Services
only for FPC on the x86_64 target - use the RTL MM on Delphi or ARM
based on FastMM4 proven algorithms by Pierre le Riche
code has been reduced to the only necessary featureset for production
deep asm refactoring for cross-platform, compactness and efficiency
can report detailed statistics (with threads contention and memory leaks)
mremap() makes large block ReallocMem a breeze on Linux
inlined SSE2 movaps loop is more efficient that subfunction(s)
lockless round-robin of tiny blocks (<=128/256 bytes) for better scaling
optional lockless bin list to avoid freemem() thread contention
three app modes: default mono-thread friendly, FPCMM_SERVER or FPCMM_BOOST

Usage: include this unit as the very first in your FPC project uses clause.

Why another Memory Manager on FPC?

The built-in heap.inc is well written and cross-platform and cross-CPU, but its threadvar arena for small blocks tends to consume a lot of memory on multi-threaded servers, and has suboptimal allocation performance
C memory managers (glibc, Intel TBB, jemalloc) have a very high RAM consumption (especially Intel TBB) and panic/SIGKILL on any GPF
Pascal alternatives (FastMM4,ScaleMM2,BrainMM) are Windows+Delphi specific
Our lockess round-robin of tiny blocks is a unique algorithm among Memory Managers AFAIK
It was so fun deeping into SSE2 x86_64 assembly and Pierre's insight
Resulting code is still easy to understand and maintain, and performs very well
It tends to have very low fragmentation, and consume less memory than FPC alternatives, especially for multi-threaded projects
It is really Open Source (MPL/GPL/LGPL) and may be used on production for FPC x86_64 apps just as much as FastMM4 was for Delphi

Feedback is welcome!

Synopse Open Source

New Multi-thread Friendly Memory Manager for FPC written in x86_64 assembly