Performance regression in Frida Gum 16.2.0 affecting fuzzing speed

[vnc0]

Summary

I observed a performance degradation (20-35%) when upgrading Frida GUM from version 16.1.11 to 16.2.0 while running AFL++ in a fuzzing environment (PR 2290). Speed is essential for fuzzing.

Notably, this degradation continues to affect the latest Frida GUM version (16.6.6).

Further investigation is required to identify the root causes and to determine if any adjustments in the integration or configuration could mitigate this issue.

Environment and Setup

• Repository: AFLplusplus
• Build Configuration: Compiled with the flag CFLAGS="-DUSEMMAP=1"
• Platform: AFL++ 4.31a running on MacOS

Reproduction Steps

• Clone the AFLplusplus repository.
• Run CFLAGS="-DUSEMMAP=1" make in the repository root.
• Change into the frida_mode directory.
• Edit the GNUmakefile to set the appropriate GUM_DEVKIT_VERSION.
• Build frida_mode using CFLAGS="-DUSEMMAP=1" make.
• Navigate to test/png and execute make frida.

Observed Results

Measured execution speeds for different Frida GUM versions were recorded as follows:

• Frida GUM 16.1.5: 1586 executions per second
• Frida GUM 16.1.11: 1603 executions per second
• Frida GUM 16.2.0: 1298 executions per second
• Frida GUM 16.3.0: 1311 executions per second
• Frida GUM 16.6.6: 1309 executions per second

The regression becomes apparent shortly after version 16.1.11, with speeds dropping significantly from 1603 to 1298 executions per second in version 16.2.0 and remaining similarly low in later releases.

[vanhauser-thc]

Thank you vnc0 for identifying the problem!
Let's hope there is an easy fix.