Occlusion queries: wrap detection uses stale baseline, crashes when visibility buffer starts near end of pass

[BXYMartin]

I am running MeloNX on iOS with MoltenVK 1.4.1. Random crashes (SIGSEGV) are observed in-game consistently and reproducible across multiple devices, while calling for QueueSubmit. In the debugger, it crashed at

AGXMetal13_3`-[AGXG13GFamilyRenderContext setVisibilityResultMode:offset:]:
    0x11a383bf8 <+0>:   pacibsp 
    0x11a383bfc <+4>:   stp    x20, x19, [sp, #-0x20]!
    0x11a383c00 <+8>:   stp    x29, x30, [sp, #0x10]
    0x11a383c04 <+12>:  add    x29, sp, #0x10
    0x11a383c08 <+16>:  adrp   x8, 1021
    0x11a383c0c <+20>:  ldrsw  x8, [x8, #0x27c]
    0x11a383c10 <+24>:  ldr    x8, [x0, x8]
    0x11a383c14 <+28>:  add    x9, x8, #0x14, lsl #12    ; =0x14000 
    0x11a383c18 <+32>:  add    x19, x9, #0x8c0
    0x11a383c1c <+36>:  mov    w9, #0x88dc               ; =35036 
    0x11a383c20 <+40>:  add    x9, x8, x9
    0x11a383c24 <+44>:  ubfx   x10, x3, #3, #29
->  0x11a383c28 <+48>:  strh   w10, [x9, #0x500]
    0x11a383c2c <+52>:  cmp    w2, #0x1
    0x11a383c30 <+56>:  cset   w11, eq
    0x11a383c34 <+60>:  cmp    w2, #0x0
    0x11a383c38 <+64>:  cset   w12, ne
    0x11a383c3c <+68>:  ldr    w13, [x9]
    0x11a383c40 <+72>:  lsl    w12, w12, #15
    0x11a383c44 <+76>:  and    w13, w13, #0xffff3fff

Further debugging shows the crash happens if a Metal render pass begins while the visibility buffer offset is already near its end (e.g., 262136/262144), the first occlusion query wraps the buffer immediately. Because firstVisibilityResultOffsetInRenderPass is still 0 (only updated after accumulation), the wrap path fires instantly, forcing a render-pass end/restart while the encoder is bound to a near-end offset. ChatGPT suggests that this can fault on AGX, producing duplicated beginMetalRenderPass logs and a crash. Here's the full logs for debugging before it crashed:

[MVK-AGENT]MVKCommandEncoder::beginMetalRenderPass::visibilityBufferSetup hasBuffer:true,offset:262136,size:262144,cmdUse:9,isRestart:false,firstOffsetInPass:0,timestamp:394843302
[MVK-AGENT]MVKCommandEncoder::beginMetalRenderPass::visibilityBufferSetup hasBuffer:true,offset:262136,size:262144,cmdUse:9,isRestart:false,firstOffsetInPass:0,timestamp:394843302
[MVK-AGENT]MVKOcclusionQueryCommandEncoderState::encode::setVisibilityResultMode mode:2,prevMode:0,offset:262136,bufferSize:262144,firstOffsetInPass:0,timestamp:394843302
[MVK-AGENT]MVKOcclusionQueryCommandEncoderState::nextMetalQuery::beforeAdvance offset:262136,size:262144,emptyQueue:false},timestamp:394843302
[MVK-AGENT]MVKVisibilityBuffer::advanceOffset::advance offset:0,halfSize:131072,wrapped:true,timestamp:394843302
[MVK-AGENT]MVKOcclusionQueryCommandEncoderState::nextMetalQuery::afterAdvance offset:0,size:262144,numCopyFences:�,firstOffsetInPass:0,timestamp:394843302
[MVK-AGENT]MVKOcclusionQueryCommandEncoderState::nextMetalQuery::wrapTriggeredStore offset:0,firstOffsetInPass:0,timestamp:394843302
[MVK-AGENT]MVKOcclusionQueryCommandEncoderState::nextMetalQuery::beforeAdvance offset:0,size:262144,emptyQueue:false},timestamp:394843302
[MVK-AGENT]MVKVisibilityBuffer::advanceOffset::advance offset:8,halfSize:131072,wrapped:false,timestamp:394843302
[MVK-AGENT]MVKOcclusionQueryCommandEncoderState::nextMetalQuery::afterAdvance offset:8,size:262144,numCopyFences:�,firstOffsetInPass:0,timestamp:394843302
SIGSEGV

And ChatGPT proposed this fix, and after some test it does fix the segmentation fault in-game

diff --git a/MoltenVK/MoltenVK/Commands/MVKCommandBuffer.mm b/MoltenVK/MoltenVK/Commands/MVKCommandBuffer.mm
index 707c2672..2453603d 100644
--- a/MoltenVK/MoltenVK/Commands/MVKCommandBuffer.mm
+++ b/MoltenVK/MoltenVK/Commands/MVKCommandBuffer.mm
@@ -777,6 +777,9 @@ static MVKBarrierStage commandUseToBarrierStage(MVKCommandUse use) {
                if (!_pEncodingContext->visibilityResultBuffer.buffer()) {
                        _pEncodingContext->visibilityResultBuffer = _device->getVisibilityBuffer();
                }
+               // Track the starting visibility offset for this Metal render pass so wrap detection compares
+               // against the correct baseline even when the buffer was already partially consumed.
+               _pEncodingContext->firstVisibilityResultOffsetInRenderPass = _pEncodingContext->visibilityResultBuffer.offset();
                mtlRPDesc.visibilityResultBuffer = _pEncodingContext->visibilityResultBuffer.buffer();
        }

I am quite new to graphical APIs so I hope someone can confirm the root cause and have it fixed. Thank you!