Question 1

How much faster is FastMM4-AVX compared to the original FastMM4?

Accepted Answer

In multi-threaded benchmarks, FastMM4-AVX can be over twice as fast, with tests showing up to 69.75% speed improvement in NexusDB scenarios. However, single-threaded performance may have minimal gains or even slowdowns due to AVX overhead.

Question 2

Does FastMM4-AVX work with FreePascal on Linux?

Accepted Answer

Yes, it supports FreePascal/Lazarus on Linux 64-bit, with version history noting fixes for compilation and AVX-512 support. It has been tested with Lazarus 1.6.4 and later versions for cross-platform compatibility.

Question 3

How to disable AVX in FastMM4-AVX to avoid slowdowns?

Accepted Answer

Define DisableAVX, DisableAVX1, DisableAVX2, or DisableAVX512 in your project to turn off AVX instructions. This is recommended if your CPU has transition penalties or reduced turbo frequency, as per the README's advice.

Question 4

What's the difference between FastMM4-AVX and FastMM5?

Accepted Answer

FastMM4-AVX focuses on AVX optimizations and efficient synchronization for multi-threaded gains, while FastMM5 may perform better in some single-threaded or high-contention scenarios. Benchmarks in the README show mixed results, with FastMM5 sometimes faster in multi-threaded tests.

Question 5

Is FastMM4-AVX safe for production use in critical applications?

Accepted Answer

It has performance benefits but security gaps: no safe-unlinking and double-free detection only in debug mode. For high-security needs, enable FullDebugMode during testing and consider additional safeguards like CheckHeapForCorruption in production.

Question 6

Can I use FastMM4-AVX in a DLL without issues?

Accepted Answer

Yes, but avoid defining UseReleaseStack, as it causes deadlocks during DLL unload on modern Windows. The README explicitly warns against this, recommending alternative synchronization methods or feature disablement.

Question 7

What alignment should I use for best performance with AVX?

Accepted Answer

With AVX enabled, blocks are aligned by 32 bytes by default, but this increases memory overhead. If your CPU supports FSRM, you can disable AVX and use 8-byte alignment for better efficiency, as suggested in the README.

FastMM4-AVX

What is FastMM4-AVX?

Overview

Use Cases

Best For

Related Projects

Found a gem we're missing?

Not Ideal For

Pros & Cons

Pros

Cons

Frequently Asked Questions