[RustamKuramshin]

Problem

Loading any Audio Unit that exposes a side-chain (or any extra) input bus crashed with SIGSEGV inside the plugin constructor on macOS. Reproduced with FabFilter Pro-C 2, TDR Kotelnikov GE, Softube FET Compressor Mk II, UAD 1176LN and Pulsar P11 Abyss (.component), all with pedalboard 0.9.26 from PyPI; the same plugins load fine as VST3.

from pedalboard import load_plugin
load_plugin("/Library/Audio/Plug-Ins/Components/TDR Kotelnikov GE.component")
# Fatal Python error: Segmentation fault

Cause

ExternalPlugin::process() builds the AudioBuffer handed to processBlock() with getTotalNumOutputChannels() channels. Plugins with extra input buses have more total input channels than that, and JUCE's AudioUnitPluginInstance::renderGetInput() locates every input bus by its channel offset inside this very buffer (getBusBuffer(*currentBuffer, true, busIndex)). For a stereo effect with a stereo side-chain the AU asks for bus 1 at offset 2, the buffer has only two channels, AudioBuffer::allocateChannels reads null pointers past its end and the plugin's render callback does memmove to address 0.

Backtrace (debug build):

juce::AudioBuffer<float>::allocateChannels          juce_AudioSampleBuffer.h:1273  (jassert dataToReferTo[i] != nullptr)
juce::AudioProcessor::getBusBuffer                   juce_AudioProcessor.h:649
juce::AudioUnitPluginInstance::renderGetInput        juce_AudioUnitPluginFormat.mm:1821 (busIndex=1, channelOffset=2)
<plugin> -> AudioUnitRender
juce::AudioUnitPluginInstance::processAudio          juce_AudioUnitPluginFormat.mm:1090
Pedalboard::ExternalPlugin<AudioUnitPluginFormat>::process         ExternalPlugin.h:1213
Pedalboard::ExternalPlugin<AudioUnitPluginFormat>::detectReloadType  ExternalPlugin.h:1038
Pedalboard::ExternalPlugin<AudioUnitPluginFormat>::reinstantiatePlugin

The VST3 wrapper in JUCE checks bus bounds itself, which is why only AU crashed.

Fix

Allocate max(getTotalNumInputChannels(), getTotalNumOutputChannels()) channels. The extra channels are zero-filled scratch, exactly like the extra output channels already were. Behaviour for plugins without extra buses is unchanged.

Verification

Debug build on macOS 26.7 / arm64 / Python 3.13: the five AUs above and their VST3 counterparts all load and process audio, with identical reported latency and output level between the AU and VST3 build of each plugin. A guard inside JUCE's renderGetInput was tried as well and turned out to be unnecessary once the buffer has the right size, so the JUCE submodule is untouched.

[Manishram-ai]

This matches JUCE's docs for processBlock, which say the buffer needs at least as many channels as the larger of total inputs and total outputs, including side-chain buses. Before this, a plugin with a side-chain input got a buffer sized to its outputs only. The extra channels now get the zeroed dummy buffers from the loop below, so the side-chain just sees silence. I couldn't test it without one of those plugins, but the change looks right to me.