OCLP-AMD-Patch

AMD Sequoia Graphics Patcher: Development Working Notes & Historical Log

Historical notes, research logs, and development timeline for macOS 15 Sequoia on legacy AMD GCN GPUs.

[!NOTE] For the primary technical documentation and reverse-engineering guide, see README.md (or README.txt).


Table of Contents

  1. Symptoms: What Broke on macOS 15 Sequoia
  2. Supported Hardware & Cross-GPU Compatibility
  3. Phase 1: Spindump & Crash Log Forensic Analysis
  4. Phase 2: Isolating the Shader (xdr::convert_image_to_image_loop)
  5. Phase 3: Bug 1 — Compiler Crash at 0x7DC05 (libAMDIL902.dylib)
  6. Phase 4: Bug 2 — Reflection Crash at 0x910E0 (MetalOld.dylib)
  7. The Core Problem: Unexpected Struct Layouts & Container Envelopes
  8. Phase 5: Bug 3 — iStat Menus 7 & SceneKit Render Pipeline Crash
  9. Phase 6: Bug 4 — Safari SIGFPE Divide-by-Zero in HDRImageConverter_Metal
  10. Phase 7: Bug 5 — Safari SIGSEGV at 0x60 in HDRImageConverter_Metal encodeInputImage:
  11. Phase 8: Boot Safety & The 25% Apple Logo Hang
  12. Phase 9: The Solution — In-Driver Companion (libAMDFix.dylib)
  13. How to Verify Offline (test_end_to_end.m)
  14. How to Use the Patcher

1. Symptoms: What Broke on macOS 15 Sequoia

When running macOS 15 Sequoia on legacy AMD GCN graphics via OCLP, several major issues appeared:

  1. Safari & Preview Freezes: Opening wide-color (Display P3) websites or viewing .HEIC photos in Preview triggered a permanent beachball freeze.
  2. Safari Crash on Specific Websites: Visiting websites like theanimecommunity.com or loading WebKit touch icons caused Safari to instantly crash with EXC_ARITHMETIC (SIGFPE).
  3. Safari Crash on Touch Icons / Favicons: Processing web icons in WebKit triggered EXC_BAD_ACCESS (SIGSEGV) at address 0x0000000000000060.
  4. iStat Menus 7 Launch Crash: Crashed with SIGSEGV immediately on startup inside SceneKit / MTLRenderPipelineReflectionInternal.

2. Supported Hardware & Cross-GPU Compatibility

Does this patch fix other AMD Macs with different GPUs?

Yes, absolutely.

In macOS, Apple does not compile individual graphics drivers for every distinct GPU model. Instead, drivers are compiled for broad GPU architectural generations:

Because OpenCore Legacy Patcher (OCLP) installs the exact same driver bundle (AMDMTLBronzeDriver.bundle and libAMDIL902.dylib) for all GCN 1–3 Macs running macOS Sequoia, the exact same compiler null-dereference bug (0x7DC05), reflection container mismatch (0x910E0), and ImageIO argument encoder crashes exist identically across all of these models:

Mac Model GPU Architecture GPU Model Driver Bundle Used
MacBook Pro (Retina, 15”, Mid 2015) GCN 2 (Sea Islands) AMD Radeon R9 M370X AMDMTLBronzeDriver + libAMDIL902
Mac Pro (Late 2013, “Trash Can”) GCN 1 (Southern Islands) Dual AMD FirePro D300 / D500 / D700 AMDMTLBronzeDriver + libAMDIL902
iMac (Retina 5K, 27”, Late 2014 & Mid 2015) GCN 2 (Sea Islands) AMD Radeon R9 M290X / M295X AMDMTLBronzeDriver + libAMDIL902
iMac (Retina 5K, 27”, Late 2015) GCN 3 (Volcanic Islands) AMD Radeon R9 M380 / M390 / M395 / M395X AMDMTLBronzeDriver + libAMDIL902
iMac (27”, Late 2012 & Late 2013) GCN 1 (Southern Islands) AMD Radeon HD 7750M / 7800M series AMDMTLBronzeDriver + libAMDIL902
Mac Pro (2010–2012) / Hackintoshes GCN 1–3 AMD HD 7870, 7950, 7970, R9 280, 290, 380, 390 AMDMTLBronzeDriver + libAMDIL902

Because this patch operates directly on the shared AMDMTLBronzeDriver and libAMDIL902.dylib binaries, applying it repairs the Sequoia graphics stack across all of the above Mac models.

(Note: Modern AMD GPUs like Polaris RX 480/580, Vega 56/64, and Navi RX 5000/6000 use AMDRadeonX4000/X5000/X6000 drivers with modern Metal 3 runtimes and do not use MetalOld.dylib or AMDMTLBronzeDriver, so they never experienced these legacy GCN bugs).


3. Phase 1: Spindump & Crash Log Forensic Analysis

How We Knew Safari Was Stuck in semaphore_wait_trap

When Safari froze on an HDR image, macOS system diagnostics (spindump) captured the exact call stack of Safari’s main thread:

Thread 0x14a0    Thread name "caulk.messenger.shared:17"
37  thread_start + 15 (libsystem_pthread.dylib + 6231) [0x7ff804dfe857]
  37  _pthread_start + 115 (libsystem_pthread.dylib + 24069) [0x7ff804e02e05]
    37  caulk::concurrent::details::worker_thread::run() + 30 (caulk + 6500)
      37  semaphore_wait_trap + 10 (libsystem_kernel.dylib + 2758) [0x7ff804dc0ac6]

This proved that Safari was not stuck in an infinite CPU loop or rendering glitch. It was blocked in semaphore_wait_trap, waiting synchronously for a reply from Apple’s background compiler daemon: com.apple.MTLCompilerService.

The Silent Death of com.apple.MTLCompilerService

Checking the crash logs (DiagnosticReports and system log) revealed why the semaphore never woke up: MTLCompilerService had crashed silently in the background:

Launchd throttled processes:
  MTLCompilerService throttled after SIGSEGV: throttled samples 1-25

Because MTLCompilerService died while compiling, the XPC port died with it. Safari and Preview were left waiting forever on an answer that would never arrive—causing the indefinite beachball freeze.


4. Phase 2: Isolating the Shader (xdr::convert_image_to_image_loop)

How We Found the Exact Shader

Tracing Safari’s image decoding pipeline revealed: CGImageDestinationAddImage ➔ IIOImageDestination::addImage ➔ HDRImageConverter_Metal convertImage:toImage:.

Inside /System/Library/Frameworks/ImageIO.framework/Versions/A/Resources/default.metallib, we discovered Apple’s precompiled Metal compute shaders used for color-space transformations. The primary HDR compute function is: xdr::convert_image_to_image_loop (with constant specialization xdr::kSamplingSize = {1, 1}).

Differential GPU Isolation Test

To isolate whether this failure was caused by the system-wide Metal compiler or the AMD driver specifically, we compiled xdr::convert_image_to_image_loop across both GPUs present on the dual-GPU MacBook Pro:

This differential test conclusively isolated the breakdown to the AMD compiler plugin (libAMDIL902.dylib), confirming that the fault only occurs when translating this specific compute kernel for AMD GCN hardware.


5. Phase 3: Bug 1 — Compiler Crash at 0x7DC05 (libAMDIL902.dylib)

Where Did 0x7DC05 Come From? Was It Arbitrary?

It was not arbitrary. It came directly from the spindump and crash logs:

5  llvm::amdMtlPlugin::ILPluginHelper::compileShader(...) (libAMDIL902.dylib + 602840) [0x5e64b62d8]
5  llvm::legacy::PassManagerImpl::run(llvm::Module&) (libAMDIL902.dylib + 11340297) [0x5e6ef3a09]
5  llvm::ILMetadataPropagation::runAnalysis(llvm::Module&) (libAMDIL902.dylib + 515582) [0x5e64a0dfe]
5  llvm::ILMetadataPropagation::initializeIABArgTypeSet(llvm::Argument*) + 73 (libAMDIL902.dylib + 515077) [0x5e64a0c05]

In the spindump, libAMDIL902.dylib was loaded at base address 0x5e6423000. The crash occurred at instruction pointer 0x5e64a0c05. Subtracting the library base from the instruction pointer:

0x5e64a0c05 - 0x5e6423000 = 0x7DC05 (or 515077 in decimal)

0x7DC05 is the exact file offset in libAMDIL902.dylib where the CPU encountered the crash.

Reverse-Engineering the Machine Code at 0x7DC05

Disassembling initializeIABArgTypeSet using otool -tvV:

0x7DBFE:  leaq    -0x48(%rbp), %r15
0x7DC02:  movq    0x80(%rax), %rax       ## Reads internal argument metadata pointer
0x7DC05:  movq    (%rax), %rdx           ## DEREFERENCE! If rax is NULL -> CRASH!
0x7DC08:  movq    %r15, %rdi
0x7DC0B:  movq    %r12, %rsi
0x7DC0E:  callq   0x7DC82                ## insert into SmallPtrSet

Why Was 0x80(%rax) NULL?

In LLVM / AMDIL compiler IR, each shader argument is represented by an llvm::Argument*. At offset 0x80, the compiler looks for a pointer to the type descriptor.

The Surgical Fix: Null-Filter Code Cave

We redirected 0x7DBFE to an assembly code cave at 0xAD0430:

cave:
    leaq    -0x48(%rbp), %r15
    leaq    -0xe8(%rbp), %r12
    movq    0x80(%rax), %rax
    testq   %rax, %rax             ## Check if rax is NULL
    je      skip                   ## If NULL, safely skip dereference and insert!
    movq    (%rax), %rdx           ## Safe to dereference
    movq    %r15, %rdi
    movq    %r12, %rsi
    callq   0x7DC82                ## insert()
skip:
    jmp     0x7DC1E                ## Resume function cleanly

With this fix, MTLCompilerService compiled the shader in 0.2s without crashing!


6. Phase 4: Bug 2 — Reflection Crash at 0x910E0 (MetalOld.dylib)

Once Bug 1 was fixed, MTLCompilerService survived. But when apps inspected the compiled shader reflection, stock MetalOld.dylib crashed!

Where Did 0x910E0 Come From?

It was reported directly in the crash log (safari crash 2.txt):

Thread 0 Crashed::  Dispatch queue: com.apple.main-thread
0   MetalOld.dylib   0x10b7f70e0 MTLInputStageReflectionReader::deserialize(...) + 222
1   MetalOld.dylib   0x10b84e83f -[MTLFunctionReflectionInternal initWithDevice:reflectionData:functionType:options:] + 573
2   MetalOld.dylib   0x10b8592e4 __53-[_MTLFunction reflectionWithOptions:binaryArchives:]_block_invoke + 74

The base address of MetalOld.dylib was 0x10b766000. 0x10b7f70e0 - 0x10b766000 = 0x910E0. 0x910E0 is the exact file offset in MetalOld.dylib where the crash occurred.

Disassembly of MTLInputStageReflectionReader::deserialize

0x9109E:  movabsq $0x4e494253504c544d, %rcx  ## ASCII "MTLPSBIN" (Metal Pipeline State Binary)
0x910A8:  cmpq    %rcx, (%rax)               ## Does byte 0 start with "MTLPSBIN"?
0x910AB:  jne     0x910C3                    ## If NOT, jump to legacy fallback!
...
0x910C7:  callq   deserializeUint32()        ## Reads byte 0 as argument count!
0x910E0:  movb    $0x1, -0x50(%rbp, %rax)    ## CRASH: SIGSEGV out-of-bounds stack write!

7. The Core Problem: Unexpected Struct Layouts & Container Envelopes

Based on the forensic disassembly and data structures observed above, this failure occurs because of how the legacy driver and macOS 15 interact: the AMD driver expects arguments and reflection buffers in a fixed legacy format, whereas modern macOS 15 passes complex new argument structures and wraps reflection data inside an unexpected outer container envelope.

Here is how these two structural changes cause downstream failures:

1. The Argument Mismatch in the Compiler (libAMDIL902.dylib)


2. The “Unexpected Length & 4 Billion Arguments” Issue in MetalOld.dylib


8. Phase 5: Bug 3 — iStat Menus 7 & SceneKit Render Pipeline Crash

When iStat Menus 7 launches, it initializes 3D status meters using Apple’s SceneKit framework. SceneKit creates a Render Pipeline with vertex and fragment shaders.

The Crash

Thread 1 Crashed:: Dispatch queue: com.apple.root.default-qos
0   MetalOld.dylib   0x1d97d00e0 MTLInputStageReflectionReader::deserialize(...) + 222
1   MetalOld.dylib   0x1d9856605 -[MTLRenderPipelineReflectionInternal initWithVertexData:fragmentData:serializedVertexDescriptor:device:options:flags:] + 207
2   MetalOld.dylib   0x1d984a0e1 newRenderPipeline(...) + 590
3   SceneKit         -[SCNMTLResourceManager _createPipelineStateWithDescriptor:desc:pipeline:] + 629

Why It Happened

Just like compute shaders, macOS 15 wraps both vertexData and fragmentData reflection streams in the 0xef13c710 envelope. When SceneKit called initWithVertexData:fragmentData:..., MetalOld.dylib tried to deserialize both raw envelopes and crashed at +222 (0x910E0).

The Fix

libAMDFix.dylib intercepts -[MTLRenderPipelineReflectionInternal initWithVertexData:fragmentData:...] (as well as initWithTileData:... and initWithObjectData:...) and -[MTLComputePipelineReflectionInternal initWithSerializedData:...], transparently slicing off the 80-byte header so that MetalOld.dylib receives pristine "MTLPSBIN" data for all render and compute pipelines.


9. Phase 6: Bug 4 — Safari SIGFPE Divide-by-Zero in HDRImageConverter_Metal

When browsing websites like theanimecommunity.com or loading web icons, Safari crashed with:

Exception Type:   EXC_ARITHMETIC (SIGFPE)
Exception Codes:  0x0000000000000001, 0x0000000000000000
Exception Note:   EXC_CORPSE_NOTIFY
Termination Signal: Floating point exception: 8

Disassembly of the Crash

ImageIO`-[HDRImageConverter_Metal convertImage:transform:toImage:transform:]:
    0x7ff8137acf4b <+437>: movq   %r14, %rdi             ; r14 is encoder for buffer 1
    0x7ff8137acf4e <+440>: movq   %r13, %rsi             ; @selector(alignment)
    0x7ff8137acf54 <+446>: callq  *0x3229ac8e(%rip)      ; [encoder alignment]
    ...
    0x7ff8137acf8c <+502>: movq   %rax, %rcx             ; %rcx = alignment!
    0x7ff8137acf8f <+505>: leaq   (%rbx,%r14), %rax      ; %rax = (len + align)
    0x7ff8137acf93 <+509>: decq   %rax                   ; %rax = (len + align - 1)
    0x7ff8137acf96 <+512>: xorl   %edx, %edx             ; %rdx = 0
    0x7ff8137acf98 <+514>: divq   %rcx                   ; CPU DIVIDE BY %rcx!

The Root Cause: Unchecked Nil Call & Division by Zero

In HDRImageConverter_Metal, ImageIO calls: id enc = [fn newArgumentEncoderWithBufferIndex: 1]; and then calculates aligned memory size with: aligned_size = ((enc.encodedLength + enc.alignment - 1) / enc.alignment) * enc.alignment;

If an argument buffer index is missing in the shader (e.g. buffer 2 or buffer 7 in xdr::compute_luma_gain_histogram_8x8), newArgumentEncoderWithBufferIndex: returns nil. Apple’s ImageIO code never checks for nil. In Objective-C, sending any message to nil returns 0: [nil alignment] ➔ 0! ImageIO then executes divq %rcx where %rcx = 0, causing an immediate CPU divide-by-zero hardware trap (SIGFPE)!

The Fix: Automatic Safe Fallback Encoder

In libAMDFix.dylib, we hook newArgumentEncoderWithBufferIndex: on _MTLFunction. If a requested buffer index does not exist or returns nil, libAMDFix automatically returns an AMDFallbackEncoder:

When ImageIO performs the alignment math: ((0 + 16 - 1) / 16) * 16 = 0! Zero bytes added, zero division by zero, and Safari renders the webpage cleanly without crashing!


10. Phase 7: Bug 5 — Safari SIGSEGV at 0x60 in HDRImageConverter_Metal encodeInputImage:

Even after preventing the divide-by-zero (SIGFPE) trap, visiting websites like theanimecommunity.com or loading crash_reproducer_mirruro_based.html still crashed Safari with a segmentation fault:

Exception Type:   EXC_BAD_ACCESS (SIGSEGV)
Exception Subtype: KERN_INVALID_ADDRESS at 0x0000000000000060
Termination Signal: Segmentation fault: 11
Faulting Thread:  0
0  ImageIO  -[HDRImageConverter_Metal encodeInputImage:transform:toBuffer:offset:withArgumentEncoder:computeEncoder:index:textures:] + 398
1  ImageIO  -[HDRImageConverter_Metal convertImage:transform:toImage:transform:] + 982
2  ImageIO  -[HDRImageConverter convertImage:toImage:] + 1066
3  ImageIO  IIOConvertHDRData + 235
4  ImageIO  IIOCallConvertHDRData + 227
5  ImageIO  IIOImageDestination::createPixelDataProviderForExtendedRange(...) + 2737
6  ImageIO  IIOImageDestination::addImage(...) + 97
7  ImageIO  CGImageDestinationAddImage + 1262
8  WebCore  WebCore::createIconDataFromBitmaps(...) + 221

Forensic Disassembly of the Crash

Disassembling -[HDRImageConverter_Metal encodeInputImage:...] at offset +353 to +413:

0x7ff820fafad2 <+353>: movq   0x3349344f(%rip), %rsi    ; "constantDataAtIndex:"
0x7ff820fafad9 <+360>: movl   $0x3, %edx              ; index = 3
0x7ff820fafade <+365>: movq   -0x30(%rbp), %rdi       ; withArgumentEncoder
0x7ff820fafae2 <+369>: callq  *0x3229c100(%rip)       ; [argumentEncoder constantDataAtIndex:3]
0x7ff820fafae8 <+375>: movq   -0x58(%rbp), %rcx       ; pointer to 128-byte 4x4 matrix
0x7ff820fafaec <+379>: vmovups (%rcx), %ymm0
0x7ff820fafaf0 <+383>: vmovups 0x20(%rcx), %ymm1
0x7ff820fafaf5 <+388>: vmovups 0x40(%rcx), %ymm2
0x7ff820fafafa <+393>: vmovups 0x60(%rcx), %ymm3
0x7ff820fafaff <+398>: vmovups %ymm3, 0x60(%rax)      ; CRASH! %rax is NULL (0x0)! 0x0 + 0x60 = 0x60!
0x7ff820fafb04 <+403>: vmovups %ymm2, 0x40(%rax)
0x7ff820fafb09 <+408>: vmovups %ymm1, 0x20(%rax)
0x7ff820fafb0e <+413>: vmovups %ymm0, (%rax)

The Root Cause

  1. In WebCore, generating icons (createIconDataFromBitmaps) requests color conversions for extended-range / Display P3 images.
  2. For kernels where certain argument buffers are missing (such as buffer 2 and 7 in xdr::compute_luma_gain_histogram_8x8), the system returns AMDFallbackEncoder.
  3. ImageIO’s encoding methods (encodeInputImage:, encodeOutputImage:, encodeInputColorTransform:, encodeOutputColorTransform:, encodeInputGainTransform:, encodeOutputGainTransform:) immediately invoke: [argumentEncoder constantDataAtIndex: 3] to retrieve a pointer where they can write 128 bytes of transform matrix.
  4. AMDFallbackEncoder previously did not implement constantDataAtIndex:. It forwarded the invocation to a no-op handler, returning 0x0 (NULL) in register %rax.
  5. ImageIO unconditionally executed vmovups %ymm3, 0x60(%rax) without verifying %rax. Writing into memory address 0x0 + 0x60 = 0x60 triggered a hardware segmentation fault (KERN_INVALID_ADDRESS at 0x0000000000000060).

The Surgical Fix: Constant Scratch Buffer & Dual Swizzling

  1. AMDFallbackEncoder:
    • Implements a dedicated 4096-byte internal scratch buffer: char _dummy[4096].
    • Implements - (void *)constantDataAtIndex:(NSUInteger)index { return _dummy; }.
    • Any 128-byte AVX matrix stores write safely into _dummy without crashing.
    • Implements full <MTLArgumentEncoder> protocol methods (setBuffer:, setTexture:, setSamplerState:, etc.) and handles selector introspection (_argInfo:) so forwarding is never needed.
  2. BronzeMtlIndirectArgumentBufferEncoder Swizzle:
    • Swizzles constantDataAtIndex: on the real AMD driver encoder: if the driver returns NULL or throws an exception, it automatically returns dummy[4096].
    • Guarantees that neither real nor fallback encoders can ever return a NULL pointer to ImageIO.
  3. Robust Encoder Creation Hooks:
    • Wraps both newArgumentEncoderWithBufferIndex: and newArgumentEncoderWithBufferIndex:reflection:functionReflection: in @try / @catch, allowing real driver encoders to be used whenever possible, and safely returning AMDFallbackEncoder if the driver returns nil or throws.

11. Phase 8: Boot Safety & The 25% Apple Logo Hang

Isn’t MetalOld.dylib Made by OCLP?

No. MetalOld.dylib is not an open-source driver developed by OCLP. It is Apple’s original, closed-source Metal.framework binary from macOS Monterey/Ventura, extracted and redistributed by Dortania/OCLP so that legacy Metal 1 GPUs can run on modern macOS.

Attempt 9: Modifying MetalOld.dylib on Disk (Initial Speculation vs. Reality)

In Attempt 9, we tested patching MetalOld.dylib on disk to skip the 80-byte header via a code cave at 0x1df000. Result: The Mac failed to boot, hanging permanently at the 25% progress bar.

Initial Speculation:

We initially speculated that macOS or OpenCore enforced boot-time signature or library validation on MetalOld.dylib.

Disassembly & What Actually Happened:

Disassembly of the Attempt 9 binary revealed that no signature check occurred:

  1. The executable code section (__TEXT,__text) ended at 0x14cd7e. Offset 0x1df000 was unallocated padding before __DATA, lacking executable memory page attributes.
  2. The cave executed subq $0x50, -0xf0(%rbp), corrupting the caller’s stack frame.
  3. During early boot (~25% Apple logo), WindowServer started and called Metal reflection. Executing invalid memory caused WindowServer to crash with EXC_BAD_ACCESS (SIGSEGV) in a loop before loginwindow could load.

We abandoned Attempt 9 shortly after to prevent affecting other Macs.

Why We Keep MetalOld.dylib 100% Stock on Disk

Despite crash logs pointing to MetalOld.dylib 0x910E0, we keep it untouched for three reasons:

  1. MetalOld.dylib is not defective: It parses clean MTLPSBIN buffers in milliseconds. Slicing off the 80-byte outer envelope dynamically solves the problem.
  2. Cross-GPU safety: MetalOld.dylib is shared across other legacy GPUs (Nvidia Kepler, Intel HD 4000). Modifying it on disk risks regressions on other Macs.
  3. OCLP update resilience: OCLP root patches overwrite MetalOld.dylib on every system update. Keeping our fix inside AMDMTLBronzeDriver ensures the patch survives.

Strict Rule: MetalOld.dylib on disk must remain 100% UNTOUCHED STOCK.


12. Phase 9: The Solution — In-Driver Companion (libAMDFix.dylib)

Instead of touching MetalOld.dylib, we let the AMD driver handle container unwrapping and encoder safety dynamically in memory:

[Safari / Preview / App]
       │
       ▼
Loads AMDMTLBronzeDriver (Contains injected LC_LOAD_DYLIB)
       │
       ▼
dyld loads libAMDFix.dylib automatically into process memory
       │
       ▼
1. Unwraps 0xef13c710 reflection containers:
   - Compute functions (MTLFunctionReflectionInternal)
   - Compute pipelines (MTLComputePipelineReflectionInternal)
   - Render pipelines  (MTLRenderPipelineReflectionInternal)
       │
       ▼
2. Guards argument encoders:
   - Provides AMDFallbackEncoder (align=16, len=0, constantData=_dummy) if encoder is nil
   - Swizzles Bronze alignment (never 0) and constantDataAtIndex: (never NULL)
   - Eliminates divide-by-zero (SIGFPE) and NULL dereference (SIGSEGV) in ImageIO
       │
       ▼
Hands clean MTLPSBIN & safe encoders to 100% UNTOUCHED STOCK MetalOld.dylib!

13. How to Verify Offline (test_end_to_end.m)

We have included the exact offline verification program inside src/test_end_to_end.m. You can compile and run it anytime without opening Safari:

cd "portable patcher"
clang -framework Foundation -framework Metal src/test_end_to_end.m -o test_end_to_end
./test_end_to_end

Expected Output:

========================================================
  AMD GPU + Stock MetalOld.dylib Offline Verification
========================================================
[+] Active Device: AMD Radeon R9 M370X

[Test 1/4] Verifying Render Pipeline Reflection (iStat Menus fix)...
[+] PASS: Render pipeline created cleanly without crashing! (pipe=0x7f9d3f01b800, refl=0x7f9d3d30faf0)

[Test 2/4] Verifying Compute Pipeline Reflection (HEIC / Preview fix)...
[+] PASS: Compute pipeline & reflection created cleanly! (pipe=0x7f9d3d310760, refl=0x600001ca4420)

[Test 3/4] Verifying Argument Encoder Safety (Safari SIGFPE fix)...
    Buffer 1: enc=BronzeMtlIndirectArgumentBufferEncoder align=16 len=144 -> aligned_size=144
    Buffer 2: enc=AMDFallbackEncoder align=16 len=0 -> aligned_size=0
    Buffer 5: enc=BronzeMtlIndirectArgumentBufferEncoder align=16 len=368 -> aligned_size=368
    Buffer 7: enc=AMDFallbackEncoder align=16 len=0 -> aligned_size=0
[+] PASS: All buffer encoders safe, zero divide-by-zero risk!

[Test 4/4] Verifying Constant Data Write Safety (Safari SIGSEGV at 0x60 fix)...
    Buffer 1: enc=BronzeMtlIndirectArgumentBufferEncoder constantData=0x110209010 -> 128-byte write safe!
    Buffer 2: enc=AMDFallbackEncoder constantData=0x7fa4f781c610 -> 128-byte write safe!
    Buffer 5: enc=BronzeMtlIndirectArgumentBufferEncoder constantData=0x110209014 -> 128-byte write safe!
    Buffer 7: enc=AMDFallbackEncoder constantData=0x7fa4f781c610 -> 128-byte write safe!
[+] PASS: All encoders return non-null constantData, zero crash risk!

========================================================
  VERIFICATION COMPLETE: ALL 4 SUBSYSTEMS PASSED (Exit 0)
========================================================

14. How to Use the Patcher

Double-click Patch.command in Finder. You have two options:

Restart your Mac after applying or reverting!