Wren + arm64 => wrong float properties

[p0nce]

Also reported by QA: Cubase 12.0.30 + macOS => crash

[p0nce]

nativeWindowResize gets called but _window is null

[p0nce]

Now the question is: is this still crashing for Cubase 12.0.30 + macOS

[p0nce]

Reproduced in Cubase Pro 12, before fix, crashes at instantiation while displaying a non-Wren-styled UI (or perhaps, partially styled)

• Does the above fix help? => No
• Same in VST2? No, no problem in VST2. (EDIT: makes sense, since VST2 are Rosetta on Cubase I think)

[p0nce]

Log from Cubase

Application Specific Information:
Cubase 12(2030,0x30f79c000) malloc: Heap corruption detected, free list is damaged at 0x600002f099e0
*** Incorrect guard value: 7233174018435146591
abort() called


Thread 0::  Dispatch queue: com.apple.main-thread
0   QuartzCore                    	    0x7ff81bb1e642 get_malloc_zone() + 0
1   QuartzCore                    	    0x7ff81b9b8053 CA::Layer::setter(unsigned int, _CAValueType, void const*) + 291
2   QuartzCore                    	    0x7ff81b9c11b9 -[CALayer setContentsCenter:] + 27
3   CoreUI                        	    0x7ff81bd0ce19 -[CUIThemeFacet updateLayer:effects:] + 770
4   CoreUI                        	    0x7ff81bd978cc CUICoreThemeRenderer::CreateOrUpdatePushButtonOrBevelButtonRoundLayer(CUIRenditionKey*, CUIDescriptor const*, CALayer**) + 370
5   CoreUI                        	    0x7ff81bd0a79c CUICoreThemeRenderer::CreateOrUpdateLayer(CUIDescriptor const*, CALayer**) + 886
6   CoreUI                        	    0x7ff81bd0a3a8 CUIRenderer::CreateOrUpdateLayer(__CFDictionary const*, CALayer**) + 200
7   AppKit                        	    0x7ff816e55bce -[NSCompositeAppearance _callCoreUIWithBlock:options:requireBezelTintColor:] + 398
8   AppKit                        	    0x7ff816e70395 -[NSAppearance _createOrUpdateLayer:options:] + 79
9   AppKit                        	    0x7ff816f17bb7 _withOverlaidDictionary + 128
10  AppKit                        	    0x7ff81734e352 -[NSWidgetView _updateWidgetLayers] + 219
11  AppKit                        	    0x7ff816f187fb -[NSWidgetView layout] + 17
12  AppKit                        	    0x7ff816ea2b8f _NSViewLayout + 564
13  AppKit                        	    0x7ff816ea2661 -[NSView _layoutSubtreeWithOldSize:] + 352
14  AppKit                        	    0x7ff816ea27bc -[NSView _layoutSubtreeWithOldSize:] + 699
15  AppKit                        	    0x7ff816ea27bc -[NSView _layoutSubtreeWithOldSize:] + 699

[p0nce]

Mmmmm, bug is release-only (EDIT: no), memory-corruption.

• Does Panagement 2.4 + VST3 fail in the same way => no
• Disable Wren to see => no crash indeed ...

[p0nce]

Repro script:

dplug-build -a all -c VST3-FULL --installer -b release-debug
open builds/XXX-FULL-1.0.0.pkg
lldb -- /Applications/Cubase\ 12.app

[p0nce]

Perhaps, quite simply, we never tested Wren on arm64

• reproduce with Distort => OK, central knob not positioned in VST3 + Cubase, it's working in REAPER Rosetta though
  Note: a workaround would probably be to not use Wren and use classic reflow... :(

• Test workaround of removing Wren in Distort => nope, same problem with the drive knob => EDIT, bad test, not sure

[p0nce]

Perhaps another bug, but probably the same
Knob correctly positioned, but oddly not drawn properly.
Also in debug.
printf doesn't work unfortunately, but lldb and running non-notarized builds now work.
Sounds like a graphic primitive not working.

• Same in x86_64? => no, it works in x86_64

[p0nce]

A matrix for testing if the center knob of Distort is properly drawn:

• Logic + arm64 + AU => yes

• Logic + x86_64 (Rosetta) + AU => yes

• REAPER + x86_64 (Rosetta) + VST3 => yes

• Cubase + arm64 + VST3 => no

• Cubase + arm64 + VST2 => no

• Cubase + x86_64 (Rosetta) + VST3 => yes

• Cubase + x86_64 (Rosetta) + VST2 => yes

• FLStudio + arm64 + VST2 => no

• FLStudio + x86_64 (Rosetta) + VST2 => yes

So yeah apart from the Logic mystery the Distort bug seems related to arm64 and aaSoftDisc.

[p0nce]

Distort's problem: knobRadiusPx is small. because knobRadius is set to zero
Every Wren float property is assigned zero, but not int properties

[p0nce]

Wren instructions trace

 131: 0004  CONSTANT             0 '1'
(fiber 0x600006a3c460) [closure 0x6000011433f0] | [class Plugin 0x6000025df980] | [class Plugin 0x6000025df980] | 620 | 330 | 1 | [class Plugin 0x6000025df980] | [instance 0x600001143840] | [instance 0x600001143870] | [instance 0x6000011438a0] | [instance 0x6000011438a0] | 0 | [foreign 0x6000011438d0] | 1 | 
 131: 0007  CONSTANT             1 '0'
(fiber 0x600006a3c460) [closure 0x6000011433f0] | [class Plugin 0x6000025df980] | [class Plugin 0x6000025df980] | 620 | 330 | 1 | [class Plugin 0x6000025df980] | [instance 0x600001143840] | [instance 0x600001143870] | [instance 0x6000011438a0] | [instance 0x6000011438a0] | 0 | [foreign 0x6000011438d0] | 1 | 0 | 
 131: 0010  LOAD_LOCAL_1
(fiber 0x600006a3c460) [closure 0x6000011433f0] | [class Plugin 0x6000025df980] | [class Plugin 0x6000025df980] | 620 | 330 | 1 | [class Plugin 0x6000025df980] | [instance 0x600001143840] | [instance 0x600001143870] | [instance 0x6000011438a0] | [instance 0x6000011438a0] | 0 | [foreign 0x6000011438d0] | 1 | 0 | 0 | 
 131: 0011  CALL_3             210 'setProp_(_,_,_)'
assign property knobRadius
 * float value 0,000000

The Wren properties:

 knobRadius=(x){
    e.setProp_(1,0,x)
  }

The styling code:

driveKnob.knobRadius = 0.65

so... why would the local hold zero instead of 0.65?

[p0nce]

The wren line:

 ($"_driveKnob").knobRadius(0.65) // does nothing yet, but an UIKnob is returned
                       .numLEDs(15)
                       .LEDRadiusMin(0.06)
                       .LEDRadiusMax(0.06)

compiles to instructions:

  36: 0208  CONSTANT             9 '_driveKnob'
      0211  CALL_0             160 '$'
      0214  CONSTANT             8 '0'
      0217  CALL_1             190 'knobRadius(_)'
  37: 0220  CONSTANT            16 '15'
      0223  CALL_1             191 'numLEDs(_)'
  38: 0226  CONSTANT             8 '0'
      0229  CALL_1             192 'LEDRadiusMin(_)'
  39: 0232  CONSTANT             8 '0'
      0235  CALL_1             193 'LEDRadiusMax(_)'

which is wrong since the literals just get truncated after the dot...

[p0nce]

In Rosetta, correctly compiled to:

 36: 0208  CONSTANT             9 '_driveKnob'
     0211  CALL_0             160 '$'
     0214  CONSTANT            16 '0.65'
     0217  CALL_1             190 'knobRadius(_)'
 37: 0220  CONSTANT            17 '15'
     0223  CALL_1             191 'numLEDs(_)'
 38: 0226  CONSTANT            18 '0.06'
     0229  CALL_1             192 'LEDRadiusMin(_)'
 39: 0232  CONSTANT            18 '0.06'
     0235  CALL_1             193 'LEDRadiusMax(_)'
     0238  POP

[p0nce]

Numbers are parsed with strtod and strtod is locale dependent, would it be possible that in arm64 C runtime uses a different locale?

[p0nce]

strtod("0.65) // does nothing yet, but an UIKnob is returned
                       .numLEDs(15)
                       .LEDRadiusMin(0.06)
                       .LEDRadiusMax(0.06)
    }

    static reflow() { 
        var W = UI.width
        var H = UI.height
        var S = W / UI.defaultWidth
        ($"_imageKnob").position = Rectangle.new(517, 176, 46, 46).scaleByFactor(S)
        ($"_inputSlider").position = Rectangle.new(190, 132, 30, 130).scaleByFactor(S)
        ($"_outputSlider").position = Rectangle.new(410, 132, 30, 130).scaleByFactor(S)
        ($"_onOffSwitch").position = Rectangle.new(90, 177, 30, 40).scaleByFactor(S)
        ($"_driveKnob").position = Rectangle.new(250, 140, 120, 120).scaleByFactor(S)
        ($"_inputLevel").position = Rectangle.new(150, 132, 30, 130).scaleByFactor(S)
        ($"_outputLevel").position = Rectangle.new(450, 132, 30, 130).scaleByFactor(S)
        ($"_colorCorrection").position = Rectangle.new(0, 0, W, H)
        ($"_resizer").position = Rectangle.new(W-30, H-30, 30, 30)
        setupEverything()
    }
}")

parsed number literal = 0

[p0nce]

FLStudio and Cubase on arm64 set the C locale to french???

0.65 is parsed = 0
0,65 is parsed = 0,65

and Wren uses strtod

which means many other arm64 things will break. Shame.

[p0nce]

• Does scanf perform better? => no, is the same. This is rather scary for the other string to number conversions.

[p0nce]

Need scanf and strtod replacement for parsing float without locale interaction => #711

[p0nce]

Should be okay now if you use Dplug v12.7.22+ and wrep-port v1.1.4+, just need to test that.

[p0nce]

Fixed.