rpcs3_version: Bump to 0.0.25

sys_fs: Minor fixup (#12913 )
Stubbed sys_fs_newfs
2022-11-01 21:50:08 +00:00 · 2022-11-01 19:49:50 +03:00 · 2022-11-01 19:08:55 +03:00 · 2022-10-31 18:53:09 +03:00 · 2022-10-31 18:53:09 +03:00 · 2022-10-31 18:53:09 +03:00
566 changed files with 34326 additions and 12611 deletions
@@ -1,9 +1,9 @@
 #!/bin/sh -ex

 # Setup Qt variables
-export QT_BASE_DIR=/opt/qt${QTVERMIN}
-export PATH=$QT_BASE_DIR/bin:$PATH
-export LD_LIBRARY_PATH=$QT_BASE_DIR/lib/x86_64-linux-gnu:$QT_BASE_DIR/lib
+export QT_BASE_DIR=/opt/qt"${QTVERMIN}"
+export PATH="$QT_BASE_DIR"/bin:"$PATH"
+export LD_LIBRARY_PATH="$QT_BASE_DIR"/lib/x86_64-linux-gnu:"$QT_BASE_DIR"/lib

 if [ -z "$CIRRUS_CI" ]; then
   cd rpcs3 || exit 1
@@ -23,19 +23,19 @@ mkdir build && cd build || exit 1

 if [ "$COMPILER" = "gcc" ]; then
    # These are set in the dockerfile
-    export CC=${GCC_BINARY}
-    export CXX=${GXX_BINARY}
+    export CC="${GCC_BINARY}"
+    export CXX="${GXX_BINARY}"
    export LINKER=gold
    # We need to set the following variables for LTO to link properly
-    export AR=/usr/bin/gcc-ar-$GCCVER
-    export RANLIB=/usr/bin/gcc-ranlib-$GCCVER
+    export AR=/usr/bin/gcc-ar-"$GCCVER"
+    export RANLIB=/usr/bin/gcc-ranlib-"$GCCVER"
    export CFLAGS="-fuse-linker-plugin"
 else
-    export CC=${CLANG_BINARY}
-    export CXX=${CLANGXX_BINARY}
+    export CC="${CLANG_BINARY}"
+    export CXX="${CLANGXX_BINARY}"
    export LINKER=lld
-    export AR=/usr/bin/llvm-ar-$LLVMVER
-    export RANLIB=/usr/bin/llvm-ranlib-$LLVMVER
+    export AR=/usr/bin/llvm-ar-"$LLVMVER"
+    export RANLIB=/usr/bin/llvm-ranlib-"$LLVMVER"
 fi

 export CFLAGS="$CFLAGS -fuse-ld=${LINKER}"
@@ -13,7 +13,7 @@ if [ "$DEPLOY_APPIMAGE" = "true" ]; then
    ls ./appdir/usr/lib/
    rm -r ./appdir/usr/share/doc
    cp "$(readlink -f /lib/x86_64-linux-gnu/libnsl.so.1)" ./appdir/usr/lib/libnsl.so.1
-    export PATH=/rpcs3/build/squashfs-root/usr/bin/:${PATH}
+    export PATH=/rpcs3/build/squashfs-root/usr/bin/:"${PATH}"

    # Embed newer libstdc++ for distros that don't come with it (ubuntu 16.04)
    mkdir -p appdir/usr/optional/ ; mkdir -p appdir/usr/optional/libstdc++/
@@ -14,7 +14,7 @@ echo "AVVER=$AVVER" >> ../.ci/ci-vars.env

 cd bin
 mkdir "rpcs3.app/Contents/lib/"
-cp "/usr/local/opt/llvm@13/lib/libc++abi.1.0.dylib" "rpcs3.app/Contents/lib/libc++abi.1.dylib"
+cp "/usr/local/opt/llvm@13/lib/c++/libc++abi.1.0.dylib" "rpcs3.app/Contents/lib/libc++abi.1.dylib"
 rm -rf "rpcs3.app/Contents/Frameworks/QtPdf.framework" \
 "rpcs3.app/Contents/Frameworks/QtQml.framework" \
 "rpcs3.app/Contents/Frameworks/QtQmlModels.framework" \
@@ -61,7 +61,7 @@ windows_task:

 linux_task:
  container:
-    image: rpcs3/rpcs3-ci-bionic:1.3
+    image: rpcs3/rpcs3-ci-bionic:1.5
    cpu: 4
    memory: 16G
  env:
@@ -108,7 +108,7 @@ freebsd_task:
  matrix:
    - name: Cirrus FreeBSD
      freebsd_instance:
-        image_family: freebsd-13-0
+        image_family: freebsd-13-1
        cpu: 8
        memory: 8G
  env:
@@ -21,7 +21,7 @@ Cpp11BracedListStyle: true
 IndentCaseLabels: false
 SortIncludes: false
 ReflowComments: true
-AlignConsecutiveAssignments: true
+AlignConsecutiveAssignments: false
 AlignTrailingComments: true
 AlignAfterOpenBracket: DontAlign
 ConstructorInitializerAllOnOneLineOrOnePerLine: false
@@ -1,66 +0,0 @@
---
-name: Regression report
-about: If something used to work before, but now it doesn't
-title: ''
-labels: ''
-assignees: ''
-
---
-
-## Please do not ask for help or report compatibility regressions here, use [RPCS3 Discord server](https://discord.me/RPCS3) or [forums](https://forums.rpcs3.net/) instead.
-
-## Quick summary
-Please briefly describe what has stopped working correctly.
-
-## Details
-Please describe the regression as accurately as possible.
-
-#### 0. Make sure you're running with settings as close to default as possible
-* **Do NOT enable any emulator game patches when reporting issues**
-* Only change settings that are required for the game to work
-
-#### 1. Please provide _exact_ build (or commit) information that introduced the regression you're reporting.
-* Please see [How to find the build that caused a regression](https://wiki.rpcs3.net/index.php?title=Help:Using_different_versions_of_RPCS3#How_to_find_the_build_that_caused_a_regression.3F) in our wiki.
-* You can find [History of RPCS3 builds](https://rpcs3.net/compatibility?b) here.
-
-#### 2. Please attach TWO logs:
-* One for the build with regression.
-* One for the build that works as expected.
-
-##### How to obtain RPCS3's log:
-* Run the game until you find the regression.
-* Completely close RPCS3, or move to the next step in case it has crashed.
-* Locate RPCS3's log file:
-	+ ```RPCS3.log``` (It can show up as just RPCS3 and have a notepad icon)
-	or
-	+ ```RPCS3.log.gz``` (It can show up as RPCS3.log and have a zip or rar icon)
-![image](https://user-images.githubusercontent.com/44116740/84433247-aa15fa80-ac36-11ea-913e-6fe25a1d040e.png)
-	On Linux it will be in ```~/.cache/rpcs3/```.
-	On Windows it will be near the executable.
-* Attach the log:
-	+ Drag and drop your log file into the issue.
-	+ Or upload it to the cloud, such as Dropbox, Mega etc.
-
-#### 3. If you describe graphical regression, please provide an RSX capture and a RenderDoc capture that demonstrate it.
-* To create an RSX capture, use _Create_ _RSX_ _Capture_ under _Utilities_.
-Captures will be stored in RPCS3 folder → captures.
-	+ Compress your capture with 7z, Rar etc.
-	And drag and drop it into the issue.
-	+ Or upload it to the cloud, such as Dropbox, Mega etc.
-* To create a RenderDoc capture, please refer to [RenderDoc's documentation](https://renderdoc.org/docs/how/how_capture_frame.html).
-
-#### 4. Please attach screenshots of your problem.
-* Enable performance overlay with at least medium level of detail.
-You can find it in _Emulator_ tab in _Settings_.
-
-#### 5. Please provide comparison with real PS3.
-
-#### 6. Please provide your system configuration:
-* OS
-* CPU
-* GPU
-* Driver version
-* etc.
-
-#### Please include.
-* Anything else you deem to be important
@@ -0,0 +1,93 @@
+name: Regression report
+description: If something used to work before, but now it doesn't
+title: "[Regression] Enter a title here"
+labels: []
+body:
+  - type: markdown
+    attributes:
+      value: |
+        # Summary
+        Please do not ask for help or report compatibility regressions here, use [RPCS3 Discord server](https://discord.me/RPCS3) or [forums](https://forums.rpcs3.net/) instead.
+  - type: textarea
+    id: quick-summary
+    attributes:
+      label: Quick summary
+      description: Please briefly describe what has stopped working correctly.
+    validations:
+      required: true
+  - type: textarea
+    id: details
+    attributes:
+      label: Details
+      description: Please describe the regression as accurately as possible. Include screenshots if neccessary.
+    validations:
+      required: false
+  - type: markdown
+    attributes:
+      value: |
+        # Identify the regressed build
+        Please provide the _exact_ build (or commit) information that introduced the regression you're reporting.
+        * Please see [How to find the build that caused a regression](https://wiki.rpcs3.net/index.php?title=Help:Using_different_versions_of_RPCS3#How_to_find_the_build_that_caused_a_regression.3F) in our wiki.
+        * You can find [History of RPCS3 builds](https://rpcs3.net/compatibility?b) here.
+        
+        Make sure you're running with settings as close to default as possible 
+        * **Do NOT enable any emulator game patches when reporting issues**
+        * Only change settings that are required for the game to work
+  - type: input
+    id: regressed-build
+    attributes:
+      label: Build with regression
+      description: Provide _exact_ build (or commit) information that introduced the regression you're reporting.
+      placeholder: v0.0.23-13948-31df99f7
+    validations:
+      required: true
+  - type: markdown
+    attributes:
+      value: |
+        # Log files
+        Obtaining the log file:
+        * Run the game until you find the regression.
+        * Completely close RPCS3 and locate the log file.
+        
+        RPCS3's Log file will be ```RPCS3.log.gz``` (sometimes shows as RPCS3.log with zip icon) or ```RPCS3.log``` (sometimes shows as RPCS3 wtih notepad icon).
+        * On Windows it will be in the RPCS3 directory near the executable
+        * On Linux it will be in ```~/.cache/rpcs3/```
+        * On MacOS it will be in ```~/Library/Caches/rpcs3```. If you're unable to locate it copy paste the path in Spotlight and hit enter.
+  - type: textarea
+    id: logs
+    attributes:
+      label: Attach two log files
+      description: |
+        Attach one file for the build with regression and another for a build that works.
+        Drag & drop the files into this input field, or upload them to another service (f.ex. Dropbox, Mega) and provide a link.
+    validations:
+      required: true
+  - type: markdown
+    attributes:
+      value: |
+        # Other details
+        If you describe a graphical regression, please provide an RSX capture and/or a RenderDoc capture that demonstrate it.
+        * To create an RSX capture, use _Create_ _RSX_ _Capture_ under _Utilities_.
+        * Captures will be stored in RPCS3 folder → captures.
+        * To create a RenderDoc capture, please refer to [RenderDoc's documentation](https://renderdoc.org/docs/how/how_capture_frame.html).
+  - type: textarea 
+    id: captures
+    attributes:
+      label: Attach capture files for visual issues
+      description: Compress your capture with 7z, Rar etc. and attach it here, or upload it to the cloud (Dropbox, Mega etc) and add a link to it.
+    validations:
+      required: false
+  - type: textarea 
+    id: system-info
+    attributes:
+      label: System configuration
+      description: Provide information about your system, such as operating system, CPU and GPU model, GPU driver version and other information that describes your system configuration.
+    validations:
+      required: false
+  - type: textarea 
+    id: other-details
+    attributes:
+      label: Other details
+      description: Include anything else you deem to be important.
+    validations:
+      required: false
@@ -1,59 +0,0 @@
---
-name: Bug report
-about: If something doesn't work correctly in RPCS3
-title: ''
-labels: ''
-assignees: ''
-
---
-
-## Please do not ask for help or report compatibility regressions here, use [RPCS3 Discord server](https://discord.me/RPCS3) or [forums](https://forums.rpcs3.net/) instead.
-
-## Quick summary
-Please briefly describe what is not working correctly.
-
-## Details
-
-Please describe the problem as accurately as possible.
-
-#### 0. Make sure you're running with settings as close to default as possible
-* **Do NOT enable any emulator game patches when reporting issues**
-* Only change settings that are required for the game to work
-
-#### 1. Please attach RPCS3's log.
-* Run the game until you find the issue.
-* Completely close RPCS3, or move to the next step in case it has crashed.
-* Locate RPCS3's log file:
-	+ ```RPCS3.log``` (It can show up as just RPCS3 and have a notepad icon)
-	or
-	+ ```RPCS3.log.gz``` (It can show up as RPCS3.log and have a zip or rar icon)
-![image](https://user-images.githubusercontent.com/44116740/84433247-aa15fa80-ac36-11ea-913e-6fe25a1d040e.png)
-	On Linux it will be in ```~/.cache/rpcs3/```.
-	On Windows it will be near the executable.
-* Attach the log:
-	+ Drag and drop your log file into the issue.
-	+ Or upload it to the cloud, such as Dropbox, Mega etc.
-
-#### 2. If you describe graphical issue, please provide an RSX capture and a RenderDoc capture that demonstrate it.
-* To create an RSX capture, use _Create_ _RSX_ _Capture_ under _Utilities_.
-Captures will be stored in RPCS3 folder → captures.
-	+ Compress your capture with 7z, Rar etc.
-	And	drag and drop it into the issue.
-	+ Or upload it to the cloud, such as Dropbox, Mega etc.
-* To create a RenderDoc capture, please refer to [RenderDoc's documentation](https://renderdoc.org/docs/how/how_capture_frame.html).
-
-#### 3. Please attach screenshots of your problem.
-* Enable performance overlay with at least medium level of detail.
-You can find it in _Emulator_ tab in _Settings_.
-
-#### 4. Please provide comparison with real PS3.
-
-#### 5. Please provide your system configuration:
-* OS
-* CPU
-* GPU
-* Driver version
-* etc.
-
-#### Please include.
-* Anything else you deem to be important
@@ -0,0 +1,78 @@
+name: Bug report
+description: If something doesn't work correctly in RPCS3
+title: "Enter a title here"
+labels: []
+body:
+  - type: markdown
+    attributes:
+      value: |
+        # Summary
+        Please do not ask for help or report compatibility regressions here, use [RPCS3 Discord server](https://discord.me/RPCS3) or [forums](https://forums.rpcs3.net/) instead.
+  - type: textarea
+    id: quick-summary
+    attributes:
+      label: Quick summary
+      description: Please briefly describe what is not working correctly.
+    validations:
+      required: true
+  - type: textarea
+    id: details
+    attributes:
+      label: Details
+      description: |
+        Please describe the problem as accurately as possible.
+        Provide a comparison with a real PS3, if possible.
+    validations:
+      required: false
+  - type: markdown
+    attributes:
+      value: |
+        # Log files
+        
+        Provide a log file that includes the bug you're reporting.
+        
+        Obtaining the log file:
+        * Run the game until you find the regression.
+        * Completely close RPCS3 and locate the log file.
+        
+        RPCS3's Log file will be ```RPCS3.log.gz``` (sometimes shows as RPCS3.log with zip icon) or ```RPCS3.log``` (sometimes shows as RPCS3 wtih notepad icon).
+        * On Windows it will be in the RPCS3 directory near the executable
+        * On Linux it will be in ```~/.cache/rpcs3/```
+        * On MacOS it will be in ```~/Library/Caches/rpcs3```. If you're unable to locate it copy paste the path in Spotlight and hit enter.
+  - type: textarea
+    id: logs
+    attributes:
+      label: Attach a log file
+      description: |
+        Drag & drop the files into this input field, or upload them to another service (f.ex. Dropbox, Mega) and provide a link.
+    validations:
+      required: true
+  - type: markdown
+    attributes:
+      value: |
+        # Other details
+        If you describe a graphical issue, please provide an RSX capture and/or a RenderDoc capture that demonstrate it.
+        * To create an RSX capture, use _Create_ _RSX_ _Capture_ under _Utilities_.
+        * Captures will be stored in RPCS3 folder → captures.
+        * To create a RenderDoc capture, please refer to [RenderDoc's documentation](https://renderdoc.org/docs/how/how_capture_frame.html).
+  - type: textarea 
+    id: captures
+    attributes:
+      label: Attach capture files for visual issues
+      description: Compress your capture with 7z, Rar etc. and attach it here, or upload it to the cloud (Dropbox, Mega etc) and add a link to it.
+    validations:
+      required: false
+  - type: textarea 
+    id: system-info
+    attributes:
+      label: System configuration
+      description: Provide information about your system, such as operating system, CPU and GPU model, GPU driver version and other information that describes your system configuration.
+    validations:
+      required: false
+  - type: textarea 
+    id: other-details
+    attributes:
+      label: Other details
+      description: Include anything else you deem to be important.
+    validations:
+      required: false
@@ -1,41 +0,0 @@
---
-name: Feature Request
-about: If RPCS3 lacks a feature you would like to see
-title: '[Feature Request] Enter Feature Title Here'
-labels: ''
-assignees: ''
-
---
-
-## Please do not ask for help or report compatibility regressions here, use [RPCS3 Discord server](https://discord.me/RPCS3) or [forums](https://forums.rpcs3.net/) instead.
-
-## Quick summary
-Please briefly describe what feature you would like RPCS3 to have.
-
-## Details
-
-Please describe the feature as accurately as possible.
-
-#### 1. Please describe, what part of RPCS3 would be affected by your feature:
-* Gameplay
-* Debugging
-* UI
-* Patches
-* Installation
-* Github
-* CI
-* etc.
-
-#### 2. Please tell us, why your feature is important to RPCS3.
-
-#### 3. Please attach screenshots of the feature implemented in other projects.
-
-#### 4. Please provide your system configuration:
-* OS
-* CPU
-* GPU
-* Driver version
-* etc.
-
-#### Please include.
-* Anything else you deem to be important
@@ -0,0 +1,36 @@
+name: Feature request
+description: If RPCS3 lacks a feature you would like to see
+title: "[Feature request] Enter a title here"
+labels: []
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Please do not ask for help or report compatibility regressions here, use [RPCS3 Discord server](https://discord.me/RPCS3) or [forums](https://forums.rpcs3.net/) instead.
+  - type: textarea
+    id: quick-summary
+    attributes:
+      label: Quick summary
+      description: Please briefly describe what feature you would like RPCS3 to have.
+    validations:
+      required: true
+  - type: textarea
+    id: details
+    attributes:
+      label: Details
+      description: Please describe the feature as accurately as possible.
+    validations:
+      required: false
+  - type: markdown
+    attributes:
+      value: |
+        Please include the following information:
+        * What part of RPCS3 would be affected by your feature? (Gameplay, Debugging, UI, Patches, Installation, CI etc.)
+        * Why your feature is important to RPCS3.
+        * If the feature is implemented in other projects, attach screenshots.
+        * If this feature is something that a game is trying to use, upload a log file for it.
+        
+        RPCS3's Log file will be ```RPCS3.log.gz``` (sometimes shows as RPCS3.log with zip icon) or ```RPCS3.log``` (sometimes shows as RPCS3 wtih notepad icon).
+        * On Windows it will be in the RPCS3 directory near the executable
+        * On Linux it will be in ```~/.cache/rpcs3/```
+        * On MacOS it will be in ```~/Library/Caches/rpcs3```. If you're unable to locate it copy paste the path in Spotlight and hit enter.
@@ -0,0 +1,14 @@
+---
+name: Advanced
+about: For developers and experienced users only
+title: ''
+labels: ''
+assignees: ''
+
+---
+
+## Please do not ask for help or report compatibility regressions here, use [RPCS3 Discord server](https://discord.me/RPCS3) or [forums](https://forums.rpcs3.net/) instead.
+
+You're using the advanced template. You're expected to know what to write in order to fill in all the required information for proper report.
+
+If you're unsure on what to do, please return back to the issue type selection and choose different category.
@@ -1,5 +1,5 @@
 /* 7zTypes.h -- Basic types
-2021-12-25 : Igor Pavlov : Public domain */
+2022-04-01 : Igor Pavlov : Public domain */

 #ifndef __7Z_TYPES_H
 #define __7Z_TYPES_H
@@ -133,10 +133,6 @@ typedef int WRes;
 #define MY__E_ERROR_NEGATIVE_SEEK MY_HRESULT_FROM_errno_CONST_ERROR(EINVAL)
 */

-// gcc / clang : (sizeof(long) == sizeof(void*)) in 32/64 bits
-typedef          long INT_PTR;
-typedef unsigned long UINT_PTR;
-
 #define TEXT(quote) quote

 #define FILE_ATTRIBUTE_READONLY       0x0001
@@ -520,6 +516,14 @@ struct ISzAlloc

 #endif

+#define k_PropVar_TimePrec_0        0
+#define k_PropVar_TimePrec_Unix     1
+#define k_PropVar_TimePrec_DOS      2
+#define k_PropVar_TimePrec_HighPrec 3
+#define k_PropVar_TimePrec_Base     16
+#define k_PropVar_TimePrec_100ns (k_PropVar_TimePrec_Base + 7)
+#define k_PropVar_TimePrec_1ns   (k_PropVar_TimePrec_Base + 9)
+
 EXTERN_C_END

 #endif
@@ -1,7 +1,7 @@
-#define MY_VER_MAJOR 21
-#define MY_VER_MINOR 07
+#define MY_VER_MAJOR 22
+#define MY_VER_MINOR 01
 #define MY_VER_BUILD 0
-#define MY_VERSION_NUMBERS "21.07"
+#define MY_VERSION_NUMBERS "22.01"
 #define MY_VERSION MY_VERSION_NUMBERS

 #ifdef MY_CPU_NAME
@@ -10,12 +10,12 @@
  #define MY_VERSION_CPU MY_VERSION
 #endif

-#define MY_DATE "2021-12-26"
+#define MY_DATE "2022-07-15"
 #undef MY_COPYRIGHT
 #undef MY_VERSION_COPYRIGHT_DATE
 #define MY_AUTHOR_NAME "Igor Pavlov"
 #define MY_COPYRIGHT_PD "Igor Pavlov : Public domain"
-#define MY_COPYRIGHT_CR "Copyright (c) 1999-2021 Igor Pavlov"
+#define MY_COPYRIGHT_CR "Copyright (c) 1999-2022 Igor Pavlov"

 #ifdef USE_COPYRIGHT_CR
  #define MY_COPYRIGHT MY_COPYRIGHT_CR
@@ -1,5 +1,5 @@
 /* CpuArch.h -- CPU specific code
-2021-07-13 : Igor Pavlov : Public domain */
+2022-07-15 : Igor Pavlov : Public domain */

 #ifndef __CPU_ARCH_H
 #define __CPU_ARCH_H
@@ -123,12 +123,15 @@ MY_CPU_64BIT means that processor can work with 64-bit registers.
 #endif


-#if  defined(__sparc64__)
-  #define MY_CPU_NAME "sparc64"
-  #define MY_CPU_64BIT
-#elif defined(__sparc__)
-  #define MY_CPU_NAME "sparc"
-  /* #define MY_CPU_32BIT */
+#if  defined(__riscv) \
+  || defined(__riscv__)
+  #if __riscv_xlen == 32
+    #define MY_CPU_NAME "riscv32"
+  #elif __riscv_xlen == 64
+    #define MY_CPU_NAME "riscv64"
+  #else
+    #define MY_CPU_NAME "riscv"
+  #endif
 #endif


@@ -1,5 +1,5 @@
 /* DllSecur.c -- DLL loading security
-2021-12-25 : Igor Pavlov : Public domain */
+2022-07-15 : Igor Pavlov : Public domain */

 #include "Precomp.h"

@@ -11,6 +11,10 @@

 #ifndef UNDER_CE

+#if defined(__GNUC__) && (__GNUC__ >= 8)
+  #pragma GCC diagnostic ignored "-Wcast-function-type"
+#endif
+
 typedef BOOL (WINAPI *Func_SetDefaultDllDirectories)(DWORD DirectoryFlags);

 #define MY_LOAD_LIBRARY_SEARCH_USER_DIRS 0x400
@@ -34,7 +38,7 @@ static const char * const g_Dlls =
 #endif

 // #define MY_CAST_FUNC  (void(*)())
-#define MY_CAST_FUNC  
+#define MY_CAST_FUNC

 void My_SetDefaultDllDirectories()
 {
@@ -1,5 +1,5 @@
 /* LzmaEnc.c -- LZMA Encoder
-2021-11-18: Igor Pavlov : Public domain */
+2022-07-15: Igor Pavlov : Public domain */

 #include "Precomp.h"

@@ -2970,6 +2970,7 @@ const Byte *LzmaEnc_GetCurBuf(CLzmaEncHandle pp)
 }


+// (desiredPackSize == 0) is not allowed
 SRes LzmaEnc_CodeOneMemBlock(CLzmaEncHandle pp, BoolInt reInit,
    Byte *dest, size_t *destLen, UInt32 desiredPackSize, UInt32 *unpackSize)
 {
@@ -2990,14 +2991,10 @@ SRes LzmaEnc_CodeOneMemBlock(CLzmaEncHandle pp, BoolInt reInit,
  if (reInit)
    LzmaEnc_Init(p);
  LzmaEnc_InitPrices(p);
-
-  nowPos64 = p->nowPos64;
  RangeEnc_Init(&p->rc);
  p->rc.outStream = &outStream.vt;
-
-  if (desiredPackSize == 0)
-    return SZ_ERROR_OUTPUT_EOF;
-
+  nowPos64 = p->nowPos64;
+  
  res = LzmaEnc_CodeOneBlock(p, desiredPackSize, *unpackSize);
  
  *unpackSize = (UInt32)(p->nowPos64 - nowPos64);
@@ -153,6 +153,20 @@ typedef int64_t                 khronos_int64_t;
 typedef uint64_t                khronos_uint64_t;
 #define KHRONOS_SUPPORT_INT64   1
 #define KHRONOS_SUPPORT_FLOAT   1
+/*
+ * To support platform where unsigned long cannot be used interchangeably with
+ * inptr_t (e.g. CHERI-extended ISAs), we can use the stdint.h intptr_t.
+ * Ideally, we could just use (u)intptr_t everywhere, but this could result in
+ * ABI breakage if khronos_uintptr_t is changed from unsigned long to
+ * unsigned long long or similar (this results in different C++ name mangling).
+ * To avoid changes for existing platforms, we restrict usage of intptr_t to
+ * platforms where the size of a pointer is larger than the size of long.
+ */
+#if defined(__SIZEOF_LONG__) && defined(__SIZEOF_POINTER__)
+#if __SIZEOF_POINTER__ > __SIZEOF_LONG__
+#define KHRONOS_USE_INTPTR_T
+#endif
+#endif

 #elif defined(__VMS ) || defined(__sgi)

@@ -235,14 +249,21 @@ typedef unsigned short int     khronos_uint16_t;
 * pointers are 64 bits, but 'long' is still 32 bits. Win64 appears
 * to be the only LLP64 architecture in current use.
 */
-#ifdef _WIN64
+#ifdef KHRONOS_USE_INTPTR_T
+typedef intptr_t               khronos_intptr_t;
+typedef uintptr_t              khronos_uintptr_t;
+#elif defined(_WIN64)
 typedef signed   long long int khronos_intptr_t;
 typedef unsigned long long int khronos_uintptr_t;
-typedef signed   long long int khronos_ssize_t;
-typedef unsigned long long int khronos_usize_t;
 #else
 typedef signed   long  int     khronos_intptr_t;
 typedef unsigned long  int     khronos_uintptr_t;
+#endif
+
+#if defined(_WIN64)
+typedef signed   long long int khronos_ssize_t;
+typedef unsigned long long int khronos_usize_t;
+#else
 typedef signed   long  int     khronos_ssize_t;
 typedef unsigned long  int     khronos_usize_t;
 #endif
@@ -32,9 +32,9 @@ extern "C" {
 #define GLAPI extern
 #endif

-#define GL_GLEXT_VERSION 20210420
+#define GL_GLEXT_VERSION 20220530

-#include "khrplatform.h"
+#include <KHR/khrplatform.h>

 /* Generated C header for:
 * API: gl
@@ -8214,6 +8214,10 @@ GLAPI void APIENTRY glMemoryBarrierEXT (GLbitfield barriers);
 #define GL_EXT_shader_integer_mix 1
 #endif /* GL_EXT_shader_integer_mix */

+#ifndef GL_EXT_shader_samples_identical
+#define GL_EXT_shader_samples_identical 1
+#endif /* GL_EXT_shader_samples_identical */
+
 #ifndef GL_EXT_shadow_funcs
 #define GL_EXT_shadow_funcs 1
 #endif /* GL_EXT_shadow_funcs */
@@ -8625,6 +8629,36 @@ GLAPI void APIENTRY glTextureNormalEXT (GLenum mode);
 #define GL_RGBA_SNORM                     0x8F93
 #endif /* GL_EXT_texture_snorm */

+#ifndef GL_EXT_texture_storage
+#define GL_EXT_texture_storage 1
+#define GL_TEXTURE_IMMUTABLE_FORMAT_EXT   0x912F
+#define GL_RGBA32F_EXT                    0x8814
+#define GL_RGB32F_EXT                     0x8815
+#define GL_ALPHA32F_EXT                   0x8816
+#define GL_LUMINANCE32F_EXT               0x8818
+#define GL_LUMINANCE_ALPHA32F_EXT         0x8819
+#define GL_RGBA16F_EXT                    0x881A
+#define GL_RGB16F_EXT                     0x881B
+#define GL_ALPHA16F_EXT                   0x881C
+#define GL_LUMINANCE16F_EXT               0x881E
+#define GL_LUMINANCE_ALPHA16F_EXT         0x881F
+#define GL_BGRA8_EXT                      0x93A1
+#define GL_R8_EXT                         0x8229
+#define GL_RG8_EXT                        0x822B
+#define GL_R32F_EXT                       0x822E
+#define GL_RG32F_EXT                      0x8230
+#define GL_R16F_EXT                       0x822D
+#define GL_RG16F_EXT                      0x822F
+typedef void (APIENTRYP PFNGLTEXSTORAGE1DEXTPROC) (GLenum target, GLsizei levels, GLenum internalformat, GLsizei width);
+typedef void (APIENTRYP PFNGLTEXSTORAGE2DEXTPROC) (GLenum target, GLsizei levels, GLenum internalformat, GLsizei width, GLsizei height);
+typedef void (APIENTRYP PFNGLTEXSTORAGE3DEXTPROC) (GLenum target, GLsizei levels, GLenum internalformat, GLsizei width, GLsizei height, GLsizei depth);
+#ifdef GL_GLEXT_PROTOTYPES
+GLAPI void APIENTRY glTexStorage1DEXT (GLenum target, GLsizei levels, GLenum internalformat, GLsizei width);
+GLAPI void APIENTRY glTexStorage2DEXT (GLenum target, GLsizei levels, GLenum internalformat, GLsizei width, GLsizei height);
+GLAPI void APIENTRY glTexStorage3DEXT (GLenum target, GLsizei levels, GLenum internalformat, GLsizei width, GLsizei height, GLsizei depth);
+#endif
+#endif /* GL_EXT_texture_storage */
+
 #ifndef GL_EXT_texture_swizzle
 #define GL_EXT_texture_swizzle 1
 #define GL_TEXTURE_SWIZZLE_R_EXT          0x8E42
@@ -747,12 +747,12 @@ AF1 sharpness){
  // Immediate constants for peak range.
  AF2 peakC=AF2(1.0,-1.0*4.0);
  // Limiters, these need to be high precision RCPs.
-  AF1 hitMinR=mn4R*ARcpF1(AF1_(4.0)*mx4R);
-  AF1 hitMinG=mn4G*ARcpF1(AF1_(4.0)*mx4G);
-  AF1 hitMinB=mn4B*ARcpF1(AF1_(4.0)*mx4B);
-  AF1 hitMaxR=(peakC.x-mx4R)*ARcpF1(AF1_(4.0)*mn4R+peakC.y);
-  AF1 hitMaxG=(peakC.x-mx4G)*ARcpF1(AF1_(4.0)*mn4G+peakC.y);
-  AF1 hitMaxB=(peakC.x-mx4B)*ARcpF1(AF1_(4.0)*mn4B+peakC.y);
+  AF1 hitMinR=min(mn4R,eR)*ARcpF1(AF1_(4.0)*mx4R);
+  AF1 hitMinG=min(mn4G,eG)*ARcpF1(AF1_(4.0)*mx4G);
+  AF1 hitMinB=min(mn4B,eB)*ARcpF1(AF1_(4.0)*mx4B);
+  AF1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpF1(AF1_(4.0)*mn4R+peakC.y);
+  AF1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpF1(AF1_(4.0)*mn4G+peakC.y);
+  AF1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpF1(AF1_(4.0)*mn4B+peakC.y);
  AF1 lobeR=max(-hitMinR,hitMaxR);
  AF1 lobeG=max(-hitMinG,hitMaxG);
  AF1 lobeB=max(-hitMinB,hitMaxB);
@@ -845,12 +845,12 @@ AF1 sharpness){
  // Immediate constants for peak range.
  AH2 peakC=AH2(1.0,-1.0*4.0);
  // Limiters, these need to be high precision RCPs.
-  AH1 hitMinR=mn4R*ARcpH1(AH1_(4.0)*mx4R);
-  AH1 hitMinG=mn4G*ARcpH1(AH1_(4.0)*mx4G);
-  AH1 hitMinB=mn4B*ARcpH1(AH1_(4.0)*mx4B);
-  AH1 hitMaxR=(peakC.x-mx4R)*ARcpH1(AH1_(4.0)*mn4R+peakC.y);
-  AH1 hitMaxG=(peakC.x-mx4G)*ARcpH1(AH1_(4.0)*mn4G+peakC.y);
-  AH1 hitMaxB=(peakC.x-mx4B)*ARcpH1(AH1_(4.0)*mn4B+peakC.y);
+  AH1 hitMinR=min(mn4R,eR)*ARcpH1(AH1_(4.0)*mx4R);
+  AH1 hitMinG=min(mn4G,eG)*ARcpH1(AH1_(4.0)*mx4G);
+  AH1 hitMinB=min(mn4B,eB)*ARcpH1(AH1_(4.0)*mx4B);
+  AH1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH1(AH1_(4.0)*mn4R+peakC.y);
+  AH1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH1(AH1_(4.0)*mn4G+peakC.y);
+  AH1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH1(AH1_(4.0)*mn4B+peakC.y);
  AH1 lobeR=max(-hitMinR,hitMaxR);
  AH1 lobeG=max(-hitMinG,hitMaxG);
  AH1 lobeB=max(-hitMinB,hitMaxB);
@@ -963,12 +963,12 @@ AF1 sharpness){
  // Immediate constants for peak range.
  AH2 peakC=AH2(1.0,-1.0*4.0);
  // Limiters, these need to be high precision RCPs.
-  AH2 hitMinR=mn4R*ARcpH2(AH2_(4.0)*mx4R);
-  AH2 hitMinG=mn4G*ARcpH2(AH2_(4.0)*mx4G);
-  AH2 hitMinB=mn4B*ARcpH2(AH2_(4.0)*mx4B);
-  AH2 hitMaxR=(peakC.x-mx4R)*ARcpH2(AH2_(4.0)*mn4R+peakC.y);
-  AH2 hitMaxG=(peakC.x-mx4G)*ARcpH2(AH2_(4.0)*mn4G+peakC.y);
-  AH2 hitMaxB=(peakC.x-mx4B)*ARcpH2(AH2_(4.0)*mn4B+peakC.y);
+  AH2 hitMinR=min(mn4R,eR)*ARcpH2(AH2_(4.0)*mx4R);
+  AH2 hitMinG=min(mn4G,eG)*ARcpH2(AH2_(4.0)*mx4G);
+  AH2 hitMinB=min(mn4B,eB)*ARcpH2(AH2_(4.0)*mx4B);
+  AH2 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH2(AH2_(4.0)*mn4R+peakC.y);
+  AH2 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH2(AH2_(4.0)*mn4G+peakC.y);
+  AH2 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH2(AH2_(4.0)*mn4B+peakC.y);
  AH2 lobeR=max(-hitMinR,hitMaxR);
  AH2 lobeG=max(-hitMinG,hitMaxG);
  AH2 lobeB=max(-hitMinB,hitMaxB);
@@ -4,7 +4,7 @@ include(ExternalProject)

 ExternalProject_Add(moltenvk
 	GIT_REPOSITORY https://github.com/KhronosGroup/MoltenVK.git
-	GIT_TAG 1236d2f
+	GIT_TAG 260bad4
 	BUILD_IN_SOURCE 1
 	SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/MoltenVK
 	CONFIGURE_COMMAND "${CMAKE_CURRENT_SOURCE_DIR}/MoltenVK/fetchDependencies" --macos
@@ -1,7 +0,0 @@
-diff --git a/ExternalRevisions/SPIRV-Cross_repo_revision b/ExternalRevisions/SPIRV-Cross_repo_revision
-index 777346a0..c1971b45 100644
--- a/ExternalRevisions/SPIRV-Cross_repo_revision
-+++ b/ExternalRevisions/SPIRV-Cross_repo_revision
-@@ -1 +1 @@
-0d4ce028bf8b8a94d325dc1e1c20446153ba19c4
-+44c3333a1c315ead00c24f7aef5fa8a7ccf49299
@@ -39,28 +39,28 @@
  <PropertyGroup Label="UserMacros">
    <CmakeGenerator>"Visual Studio $(VisualStudioVersion.Substring(0,2))"</CmakeGenerator>
    <CmakeReleaseCLI>call vsdevcmd.bat -arch=amd64
-      cmake -G $(CmakeGenerator) -A x64 -DCMAKE_BUILD_TYPE="Release" -DSPIRV_WERROR=OFF -DSPIRV-Headers_SOURCE_DIR=$([System.IO.Path]::GetFullPath('$(MSBuildThisFileDirectory)/SPIRV-Headers').Replace('\', '/')) SPIRV-Tools -B build</CmakeReleaseCLI>
+      cmake -G $(CmakeGenerator) -A x64 -DCMAKE_BUILD_TYPE="Release" -DSPIRV_WERROR=OFF -DSPIRV-Headers_SOURCE_DIR="$([System.IO.Path]::GetFullPath('$(MSBuildThisFileDirectory)/SPIRV-Headers').Replace('\', '/'))" SPIRV-Tools -B build</CmakeReleaseCLI>
    <CmakeDebugCLI>call vsdevcmd.bat -arch=amd64
-      cmake -G $(CmakeGenerator) -A x64 -DCMAKE_BUILD_TYPE="Debug" -DSPIRV_WERROR=OFF -DSPIRV-Headers_SOURCE_DIR=$([System.IO.Path]::GetFullPath('$(MSBuildThisFileDirectory)/SPIRV-Headers').Replace('\', '/')) -S SPIRV-Tools -B build</CmakeDebugCLI>
+      cmake -G $(CmakeGenerator) -A x64 -DCMAKE_BUILD_TYPE="Debug" -DSPIRV_WERROR=OFF -DSPIRV-Headers_SOURCE_DIR="$([System.IO.Path]::GetFullPath('$(MSBuildThisFileDirectory)/SPIRV-Headers').Replace('\', '/'))" -S SPIRV-Tools -B build</CmakeDebugCLI>
    <PropsAbsPath>$([System.IO.Path]::GetFullPath('$(MSBuildThisFileDirectory)..\..\buildfiles\msvc\common_default.props'))</PropsAbsPath>
  </PropertyGroup>
  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
    <NMakeBuildCommandLine>$(CmakeReleaseCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Release /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m</NMakeBuildCommandLine>
+      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Release /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m</NMakeBuildCommandLine>
    <NMakeReBuildCommandLine>$(CmakeReleaseCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Release /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m</NMakeReBuildCommandLine>
+      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Release /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m</NMakeReBuildCommandLine>
    <NMakeCleanCommandLine>$(CmakeReleaseCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Release /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m</NMakeCleanCommandLine>
+      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Release /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m</NMakeCleanCommandLine>
  </PropertyGroup>
  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
    <NMakeBuildCommandLine>$(CmakeDebugCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Debug /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m
+      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Debug /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m
    </NMakeBuildCommandLine>
    <NMakeReBuildCommandLine>$(CmakeDebugCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Debug /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m
+      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Debug /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m
    </NMakeReBuildCommandLine>
    <NMakeCleanCommandLine>$(CmakeDebugCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Debug /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m
+      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Debug /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m
    </NMakeCleanCommandLine>
  </PropertyGroup>
  <ItemDefinitionGroup>
@@ -68,4 +68,4 @@
  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
  <ImportGroup Label="ExtensionTargets">
  </ImportGroup>
-</Project>
+</Project>
@@ -27,7 +27,7 @@ target_compile_definitions(soundtouch PUBLIC
 	SOUNDTOUCH_FLOAT_SAMPLES;
 )

-target_compile_options(soundtouch PUBLIC "-w")
+target_compile_options(soundtouch PRIVATE "-w")

 if (CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86|X86|amd64|AMD64|em64t|EM64T)")
 	target_compile_definitions(soundtouch PUBLIC
@@ -62,12 +62,13 @@
    <ClCompile Include="asmjit\src\asmjit\arm\a64builder.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64compiler.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64emithelper.cpp" />
+    <ClCompile Include="asmjit\src\asmjit\arm\a64formatter.cpp" />
+    <ClCompile Include="asmjit\src\asmjit\arm\a64func.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64instapi.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64instdb.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64operand.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64rapass.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\armformatter.cpp" />
-    <ClCompile Include="asmjit\src\asmjit\arm\armfunc.cpp" />
  </ItemGroup>
  <ItemGroup>
    <ClInclude Include="asmjit\src\asmjit\core\api-build_p.h" />
@@ -134,11 +135,14 @@
    <ClInclude Include="asmjit\src\asmjit\x86\x86opcode_p.h" />
    <ClInclude Include="asmjit\src\asmjit\x86\x86operand.h" />
    <ClInclude Include="asmjit\src\asmjit\x86\x86rapass_p.h" />
+    <ClInclude Include="asmjit\src\asmjit\arm\a64archtraits_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64assembler.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64builder.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64compiler.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64emithelper_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64emitter.h" />
+    <ClInclude Include="asmjit\src\asmjit\arm\a64formatter_p.h" />
+    <ClInclude Include="asmjit\src\asmjit\arm\a64func_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64globals.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64instapi_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64instdb.h" />
@@ -146,9 +150,7 @@
    <ClInclude Include="asmjit\src\asmjit\arm\a64operand.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64rapass_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64utils.h" />
-    <ClInclude Include="asmjit\src\asmjit\arm\armarchtraits_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\armformatter_p.h" />
-    <ClInclude Include="asmjit\src\asmjit\arm\armfunc_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\armglobals.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\armoperand.h" />
  </ItemGroup>
@@ -52,12 +52,13 @@
    <ClCompile Include="asmjit\src\asmjit\arm\a64builder.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64compiler.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64emithelper.cpp" />
+    <ClCompile Include="asmjit\src\asmjit\arm\a64formatter.cpp" />
+    <ClCompile Include="asmjit\src\asmjit\arm\a64func.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64instapi.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64instdb.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64operand.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\a64rapass.cpp" />
    <ClCompile Include="asmjit\src\asmjit\arm\armformatter.cpp" />
-    <ClCompile Include="asmjit\src\asmjit\arm\armfunc.cpp" />
  </ItemGroup>
  <ItemGroup>
    <ClInclude Include="asmjit\src\asmjit\core\api-build_p.h" />
@@ -124,11 +125,14 @@
    <ClInclude Include="asmjit\src\asmjit\x86\x86opcode_p.h" />
    <ClInclude Include="asmjit\src\asmjit\x86\x86operand.h" />
    <ClInclude Include="asmjit\src\asmjit\x86\x86rapass_p.h" />
+    <ClInclude Include="asmjit\src\asmjit\arm\a64archtraits_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64assembler.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64builder.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64compiler.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64emithelper_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64emitter.h" />
+    <ClInclude Include="asmjit\src\asmjit\arm\a64formatter_p.h" />
+    <ClInclude Include="asmjit\src\asmjit\arm\a64func_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64globals.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64instapi_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64instdb.h" />
@@ -136,9 +140,7 @@
    <ClInclude Include="asmjit\src\asmjit\arm\a64operand.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64rapass_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\a64utils.h" />
-    <ClInclude Include="asmjit\src\asmjit\arm\armarchtraits_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\armformatter_p.h" />
-    <ClInclude Include="asmjit\src\asmjit\arm\armfunc_p.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\armglobals.h" />
    <ClInclude Include="asmjit\src\asmjit\arm\armoperand.h" />
  </ItemGroup>
@@ -12,6 +12,7 @@ else()
 	option(CURL_USE_WOLFSSL "enable wolfSSL for SSL/TLS" ON)
 	option(HTTP_ONLY "disables all protocols except HTTP (This overrides all CURL_DISABLE_* options)" ON)
 	option(CURL_ZLIB "Set to ON to enable building curl with zlib support." OFF)
+	option(USE_LIBIDN2 "Use libidn2 for IDN support" OFF) # Disabled because MacOs CI doesn't work otherwise
 	set(CURL_CA_PATH "none" CACHE STRING "Location of default CA path. Set 'none' to disable or 'auto' for auto-detection. Defaults to 'auto'.")
 	if(USE_MSVC_STATIC_CRT)
 		set(CURL_STATIC_CRT ON CACHE BOOL "Use static crt to build curl" FORCE)
@@ -98,6 +98,7 @@
    <ClCompile Include="curl\lib\escape.c" />
    <ClCompile Include="curl\lib\file.c" />
    <ClCompile Include="curl\lib\fileinfo.c" />
+    <ClCompile Include="curl\lib\fopen.c" />
    <ClCompile Include="curl\lib\formdata.c" />
    <ClCompile Include="curl\lib\ftp.c" />
    <ClCompile Include="curl\lib\ftplistparser.c" />
@@ -279,9 +280,11 @@
    <ClInclude Include="curl\lib\dynbuf.h" />
    <ClInclude Include="curl\lib\easyif.h" />
    <ClInclude Include="curl\lib\easyoptions.h" />
+    <ClInclude Include="curl\lib\easy_lock.h" />
    <ClInclude Include="curl\lib\escape.h" />
    <ClInclude Include="curl\lib\file.h" />
    <ClInclude Include="curl\lib\fileinfo.h" />
+    <ClInclude Include="curl\lib\fopen.h" />
    <ClInclude Include="curl\lib\formdata.h" />
    <ClInclude Include="curl\lib\ftp.h" />
    <ClInclude Include="curl\lib\ftplistparser.h" />
@@ -480,6 +480,9 @@
    <ClCompile Include="curl\lib\timediff.c">
      <Filter>Source Files</Filter>
    </ClCompile>
+    <ClCompile Include="curl\lib\fopen.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
  </ItemGroup>
  <ItemGroup>
    <ClInclude Include="curl\include\curl\curl.h">
@@ -965,6 +968,12 @@
    <ClInclude Include="curl\lib\timediff.h">
      <Filter>Header Files</Filter>
    </ClInclude>
+    <ClInclude Include="curl\lib\easy_lock.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
+    <ClInclude Include="curl\lib\fopen.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
  </ItemGroup>
  <ItemGroup>
    <ResourceCompile Include="curl\lib\libcurl.rc">
@@ -46,21 +46,21 @@
  </PropertyGroup>
  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
    <NMakeBuildCommandLine>$(CmakeReleaseCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Release /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m</NMakeBuildCommandLine>
+      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Release /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m</NMakeBuildCommandLine>
    <NMakeReBuildCommandLine>$(CmakeReleaseCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Release /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m</NMakeReBuildCommandLine>
+      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Release /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m</NMakeReBuildCommandLine>
    <NMakeCleanCommandLine>$(CmakeReleaseCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Release /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m</NMakeCleanCommandLine>
+      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Release /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m</NMakeCleanCommandLine>
  </PropertyGroup>
  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
    <NMakeBuildCommandLine>$(CmakeDebugCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Debug /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m
+      msbuild.exe build\ALL_BUILD.vcxproj /t:build /p:Configuration=Debug /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m
    </NMakeBuildCommandLine>
    <NMakeReBuildCommandLine>$(CmakeDebugCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Debug /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m
+      msbuild.exe build\ALL_BUILD.vcxproj /t:rebuild /p:Configuration=Debug /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m
    </NMakeReBuildCommandLine>
    <NMakeCleanCommandLine>$(CmakeDebugCLI)
-      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Debug /p:ForceImportBeforeCppTargets=$(PropsAbsPath) /m
+      msbuild.exe build\ALL_BUILD.vcxproj /t:clean /p:Configuration=Debug /p:ForceImportBeforeCppTargets="$(PropsAbsPath)" /m
    </NMakeCleanCommandLine>
  </PropertyGroup>
  <ItemDefinitionGroup>
@@ -68,4 +68,4 @@
  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
  <ImportGroup Label="ExtensionTargets">
  </ImportGroup>
-</Project>
+</Project>
@@ -265,9 +265,8 @@ namespace pine
 				case MsgVersion:
 				{
 					char version[256] = {};
-					sprintf(version, "RPCS3 %s", Impl::get_version_and_branch().c_str());
+					snprintf(version, sizeof(version), "RPCS3 %s", Impl::get_version_and_branch().c_str());
 					const u32 size = strlen(version) + 1;
-					version[size] = 0x00;
 					if (!SafetyChecks(buf_cnt, 0, ret_cnt, size + 4, buf_size))
 						goto error;
 					ToArray(ret_buffer, size, ret_cnt);
@@ -291,8 +290,7 @@ namespace pine
 					const auto title_string = Impl::get_title();
 					const auto size = title_string.size() + 1;
 					char* title = new char[size];
-					sprintf(title, "%s", title_string.c_str());
-					title[size] = 0x00;
+					snprintf(title, size, "%s", title_string.c_str());
 					if (!SafetyChecks(buf_cnt, 0, ret_cnt, size + 4, buf_size))
 						goto error;
 					ToArray(ret_buffer, size, ret_cnt);
@@ -307,8 +305,7 @@ namespace pine
 					const auto title_id_string = Impl::get_title_ID();
 					const auto size = title_id_string.size() + 1;
 					char* title_id = new char[size];
-					sprintf(title_id, "%s", title_id_string.c_str());
-					title_id[size] = 0x00;
+					snprintf(title_id, size, "%s", title_id_string.c_str());
 					if (!SafetyChecks(buf_cnt, 0, ret_cnt, size + 4, buf_size))
 						goto error;
 					ToArray(ret_buffer, size, ret_cnt);
@@ -323,8 +320,7 @@ namespace pine
 					const auto hash_string = Impl::get_executable_hash();
 					const auto size = hash_string.size() + 1;
 					char* hash = new char[size];
-					sprintf(hash, "%s", hash_string.c_str());
-					hash[size] = 0x00;
+					snprintf(hash, size, "%s", hash_string.c_str());
 					if (!SafetyChecks(buf_cnt, 0, ret_cnt, size + 4, buf_size))
 						goto error;
 					ToArray(ret_buffer, size, ret_cnt);
@@ -339,8 +335,7 @@ namespace pine
 					const auto game_version_string = Impl::get_app_version();
 					const auto size = game_version_string.size() + 1;
 					char* game_version = new char[size];
-					sprintf(game_version, "%s", game_version_string.c_str());
-					game_version[size] = 0x00;
+					snprintf(game_version, size, "%s", game_version_string.c_str());
 					if (!SafetyChecks(buf_cnt, 0, ret_cnt, size, buf_size))
 						goto error;
 					ToArray(ret_buffer, size, ret_cnt);
@@ -5,9 +5,10 @@ if(USE_SYSTEM_WOLFSSL)
 	target_link_libraries(wolfssl INTERFACE PkgConfig::WolfSSL)
 else()
 	# TODO(cjj19970505@live.cn)
-	# OPENSSL_EXTRA, WOLFSSL_DES_ECB and HAVE_SNI are unconfigurable from CMake cache.
+	# OPENSSL_EXTRA, and HAVE_SNI are unconfigurable from CMake cache.
 	# but they do have it in a TODO list (wolfssl/CMakeList, 1021)
 	add_compile_definitions(OPENSSL_EXTRA WOLFSSL_DES_ECB HAVE_SNI HAVE_WRITE_DUP)
+	add_compile_definitions(FP_MAX_BITS=8192)

 	set(WOLFSSL_TLS13 "no" CACHE INTERNAL "")
 	set(WOLFSSL_SHA224 "yes" CACHE INTERNAL "")
@@ -19,7 +20,7 @@ else()
 	set(WOLFSSL_CHACHA "yes" CACHE INTERNAL "")
 	set(WOLFSSL_FILESYSTEM "yes" CACHE INTERNAL "")
 	set(WOLFSSL_PWDBASED "yes" CACHE INTERNAL "")
-	set(WOLFSSL_FAST_MATH "no" CACHE INTERNAL "")
+	set(WOLFSSL_FAST_MATH "yes" CACHE INTERNAL "")
 	set(WOLFSSL_EXAMPLES "no" CACHE INTERNAL "")
 	set(WOLFSSL_CRYPT_TESTS "no" CACHE INTERNAL "")
 	set(WOLFSSL_ASYNC_THREADS "no" CACHE INTERNAL "")
@@ -43,6 +43,8 @@
 #define HAVE_ECC
 #define ECC_SHAMIR
 #define ECC_TIMING_RESISTANT
+#define USE_FAST_MATH
+#define FP_MAX_BITS 8192

 /* UTF-8 aware filesystem functions for Windows */
 #define WOLFSSL_USER_FILESYSTEM
@@ -102,9 +102,10 @@ git submodule update --init

 #### Configuring the Qt plugin (if used)

-1) Go to the Qt5 menu and edit Qt5 options.
-2) Add the path to your Qt installation with compiler e.g. `<QtInstallFolder>\5.15.2\msvc2019_64`.
-3) While selecting the rpcs3qt project, go to Qt5->Project Setting and select the version you added.
+1) Go to `Extensions->Qt VS Tools->Qt Versions`.
+2) Add the path to your Qt installation with compiler e.g. `<QtInstallFolder>\5.15.2\msvc2019_64`, version will fill in automatically.
+3) Go to `Extensions->Qt VS Tools->Options->Legacy Project Format`.
+4) Set `Build: Run pre-build setup` to `true`.

 #### Building the projects

@@ -104,9 +104,9 @@ if(WIN32)
    add_definitions(-D_WIN32_WINNT=0x0602)
 endif()

-if(APPLE)
-    include_directories(/opt/local/include)
-    link_directories(/opt/local/lib)
+if(APPLE AND CMAKE_OSX_ARCHITECTURES STREQUAL "arm64")
+    include_directories(/opt/homebrew/include)
+    link_directories(/opt/homebrew/lib)
 endif()

 # Warnings are silenced for 3rdparty code
@@ -5,10 +5,10 @@ RPCS3
 [![Cirrus CI - Base Branch Build Status](https://img.shields.io/cirrus/github/RPCS3/rpcs3?label=Cirrus%20CI%20(FreeBSD)&logo=cirrus-ci)](https://cirrus-ci.com/github/RPCS3/rpcs3)
 [![RPCS3 discord server](https://img.shields.io/discord/272035812277878785?color=5865F2&label=RPCS3%20Discord&logo=discord&logoColor=white)](https://discord.me/rpcs3)

-The world's first free and open-source PlayStation 3 emulator/debugger, written in C++ for Windows and Linux.
+The world's first free and open-source PlayStation 3 emulator/debugger, written in C++ for Windows, Linux, macOS and FreeBSD.

 You can find some basic information on our [**website**](https://rpcs3.net/). Game info is being populated on the [**Wiki**](https://wiki.rpcs3.net/).
-For discussion about this emulator, PS3 emulation, and game compatibility reports, please visit our [**forums**](https://forums.rpcs3.net) and our [**Discord server**](https://discord.me/RPCS3).
+For discussion about this emulator, PS3 emulation, and game compatibility reports, please visit our [**forums**](https://forums.rpcs3.net) and our [**Discord server**](https://discord.gg/RPCS3).

 [**Support Lead Developers Nekotekina and kd-11 on Patreon**](https://www.patreon.com/Nekotekina)

@@ -60,9 +60,15 @@ std::vector<std::string> cfg::make_int_range(s64 min, s64 max)

 bool try_to_int64(s64* out, std::string_view value, s64 min, s64 max)
 {
+	if (value.empty())
+	{
+		if (out) cfg_log.error("cfg::try_to_uint64(): called with an empty string");
+		return false;
+	}
+
 	s64 result;
-	const char* start = &value.front();
-	const char* end = &value.back() + 1;
+	const char* start = value.data();
+	const char* end = start + value.size();
 	int base = 10;
 	int sign = +1;

@@ -72,7 +78,7 @@ bool try_to_int64(s64* out, std::string_view value, s64 min, s64 max)
 		start += 1;
 	}

-	if (start[0] == '0' && (start[1] == 'x' || start[1] == 'X'))
+	if (start[0] == '0' && value.size() >= 2 && (start[1] == 'x' || start[1] == 'X'))
 	{
 		// Limited hex support
 		base = 16;
@@ -106,12 +112,18 @@ std::vector<std::string> cfg::make_uint_range(u64 min, u64 max)

 bool try_to_uint64(u64* out, std::string_view value, u64 min, u64 max)
 {
+	if (value.empty())
+	{
+		if (out) cfg_log.error("cfg::try_to_uint64(): called with an empty string");
+		return false;
+	}
+
 	u64 result;
-	const char* start = &value.front();
-	const char* end = &value.back() + 1;
+	const char* start = value.data();
+	const char* end = start + value.size();
 	int base = 10;

-	if (start[0] == '0' && (start[1] == 'x' || start[1] == 'X'))
+	if (start[0] == '0' && value.size() >= 2 && (start[1] == 'x' || start[1] == 'X'))
 	{
 		// Limited hex support
 		base = 16;
@@ -136,6 +148,42 @@ bool try_to_uint64(u64* out, std::string_view value, u64 min, u64 max)
 	return true;
 }

+std::vector<std::string> cfg::make_float_range(f64 min, f64 max)
+{
+	return {std::to_string(min), std::to_string(max)};
+}
+
+bool try_to_float(f64* out, std::string_view value, f64 min, f64 max)
+{
+	if (value.empty())
+	{
+		if (out) cfg_log.error("cfg::try_to_float(): called with an empty string");
+		return false;
+	}
+
+	// std::from_chars float is yet to be implemented on Xcode it seems
+	// And strtod doesn't support ranged view so we need to ensure it meets a null terminator
+	const std::string str = std::string{value};
+
+	char* end_check{};
+	const double result = std::strtod(str.data(), &end_check);
+
+	if (end_check != str.data() + str.size())
+	{
+		if (out) cfg_log.error("cfg::try_to_float('%s'): invalid float", value);
+		return false;
+	}
+
+	if (result < min || result > max)
+	{
+		if (out) cfg_log.error("cfg::try_to_float('%s'): out of bounds (%f..%f)", value, min, max);
+		return false;
+	}
+
+	if (out) *out = result;
+	return true;
+}
+
 bool cfg::try_to_enum_value(u64* out, decltype(&fmt_class_string<int>::format) func, std::string_view value)
 {
 	u64 max = umax;
@@ -162,8 +210,8 @@ bool cfg::try_to_enum_value(u64* out, decltype(&fmt_class_string<int>::format) f
 	}

 	u64 result;
-	const char* start = &value.front();
-	const char* end = &value.back() + 1;
+	const char* start = value.data();
+	const char* end = start + value.size();
 	int base = 10;

 	if (start[0] == '0' && (start[1] == 'x' || start[1] == 'X'))
@@ -20,6 +20,9 @@ namespace cfg
 	// Format min and max unsigned values
 	std::vector<std::string> make_uint_range(u64 min, u64 max);

+	// Format min and max float values
+	std::vector<std::string> make_float_range(f64 min, f64 max);
+
 	// Internal hack
 	bool try_to_enum_value(u64* out, decltype(&fmt_class_string<int>::format) func, std::string_view);

@@ -309,6 +312,72 @@ namespace cfg
 		}
 	};

+	// Float entry with custom Min/Max range.
+	template <s32 Min, s32 Max>
+	class _float final : public _base
+	{
+		static_assert(Min < Max, "Invalid cfg::_float range");
+
+		using float_type = f64;
+		atomic_t<float_type> m_value;
+
+	public:
+		float_type def;
+
+		// Expose range
+		static constexpr float_type max = Max;
+		static constexpr float_type min = Min;
+
+		_float(node* owner, const std::string& name, float_type def = std::min<float_type>(Max, std::max<float_type>(Min, 0)), bool dynamic = false)
+			: _base(type::_int, owner, name, dynamic)
+			, m_value(def)
+			, def(def)
+		{
+		}
+
+		operator float_type() const
+		{
+			return m_value;
+		}
+
+		float_type get() const
+		{
+			return m_value;
+		}
+
+		void from_default() override
+		{
+			m_value = def;
+		}
+
+		std::string to_string() const override
+		{
+			return std::to_string(m_value);
+		}
+
+		bool from_string(std::string_view value, bool /*dynamic*/ = false) override
+		{
+			f64 result;
+			if (try_to_float(&result, value, Min, Max))
+			{
+				m_value = static_cast<float_type>(result);
+				return true;
+			}
+
+			return false;
+		}
+
+		void set(const f64& value)
+		{
+			m_value = static_cast<float_type>(value);
+		}
+
+		std::vector<std::string> to_list() const override
+		{
+			return make_float_range(Min, Max);
+		}
+	};
+
 	// Alias for 32 bit int
 	using int32 = _int<s32{smin}, s32{smax}>;

@@ -71,6 +71,10 @@ namespace fs
 		s64 atime;
 		s64 mtime;
 		s64 ctime;
+
+		using enable_bitcopy = std::true_type;
+
+		constexpr bool operator==(const stat_t&) const = default;
 	};

 	// Helper, layout is equal to iovec struct
@@ -105,6 +109,8 @@ namespace fs
 			: stat_t{}
 		{
 		}
+
+		using enable_bitcopy = std::false_type;
 	};

 	// Directory handle base
@@ -681,9 +687,10 @@ namespace fs
 		T obj;
 		u64 pos;

-		container_stream(T&& obj)
+		container_stream(T&& obj, const stat_t& init_stat = {})
 			: obj(std::forward<T>(obj))
 			, pos(0)
+			, m_stat(init_stat)
 		{
 		}

@@ -694,6 +701,7 @@ namespace fs
 		bool trunc(u64 length) override
 		{
 			obj.resize(length);
+			update_time(true);
 			return true;
 		}

@@ -708,6 +716,7 @@ namespace fs
 				{
 					std::copy(obj.cbegin() + pos, obj.cbegin() + pos + max, static_cast<value_type*>(buffer));
 					pos = pos + max;
+					update_time();
 					return max;
 				}
 			}
@@ -743,6 +752,7 @@ namespace fs
 			obj.insert(obj.end(), src + overlap, src + size);
 			pos += size;

+			if (size) update_time(true);
 			return size;
 		}

@@ -767,13 +777,34 @@ namespace fs
 		{
 			return obj.size();
 		}
+
+		stat_t stat() override
+		{
+			return m_stat;
+		}
+
+	private:
+		stat_t m_stat{};
+
+		void update_time(bool write = false)
+		{
+			// TODO: Accurate timestamps
+			m_stat.atime++;
+
+			if (write)
+			{
+				m_stat.mtime++;
+				m_stat.mtime = std::max(m_stat.atime, m_stat.mtime);
+				m_stat.ctime = m_stat.mtime;
+			}
+		}
 	};

 	template <typename T>
-	file make_stream(T&& container = T{})
+	file make_stream(T&& container = T{}, const stat_t& stat = stat_t{})
 	{
 		file result;
-		result.reset(std::make_unique<container_stream<T>>(std::forward<T>(container)));
+		result.reset(std::make_unique<container_stream<T>>(std::forward<T>(container), stat));
 		return result;
 	}

@@ -7,6 +7,8 @@
 #include "mutex.h"
 #include "util/vm.hpp"
 #include "util/asm.hpp"
+#include "util/v128.hpp"
+#include "util/simd.hpp"
 #include <charconv>
 #include <zlib.h>

@@ -18,6 +20,34 @@ LOG_CHANNEL(jit_log, "JIT");

 void jit_announce(uptr func, usz size, std::string_view name)
 {
+#ifdef __linux__
+	static const struct tmp_perf_map
+	{
+		std::string name{fmt::format("/tmp/perf-%d.map", getpid())};
+		fs::file data{name, fs::rewrite + fs::append};
+
+		tmp_perf_map() = default;
+		tmp_perf_map(const tmp_perf_map&) = delete;
+		tmp_perf_map& operator=(const tmp_perf_map&) = delete;
+
+		~tmp_perf_map()
+		{
+			fs::remove_file(name);
+		}
+	} s_map;
+
+	if (size && name.size())
+	{
+		s_map.data.write(fmt::format("%x %x %s\n", func, size, name));
+	}
+
+	if (!func && !size && !name.size())
+	{
+		fs::remove_file(s_map.name);
+		return;
+	}
+#endif
+
 	if (!size)
 	{
 		jit_log.error("Empty function announced: %s (%p)", name, func);
@@ -81,12 +111,6 @@ void jit_announce(uptr func, usz size, std::string_view name)
 			dump.write(reinterpret_cast<uchar*>(func), size);
 		}
 	}
-
-#ifdef __linux__
-	static const fs::file s_map(fmt::format("/tmp/perf-%d.map", getpid()), fs::rewrite + fs::append);
-
-	s_map.write(fmt::format("%x %x %s\n", func, size, name));
-#endif
 }

 static u8* get_jit_memory()
@@ -351,6 +375,473 @@ asmjit::inline_runtime::~inline_runtime()
 	utils::memory_protect(m_data, m_size, utils::protection::rx);
 }

+#if defined(ARCH_X64)
+asmjit::simd_builder::simd_builder(CodeHolder* ch) noexcept
+	: native_asm(ch)
+{
+	_init(0);
+	consts[~v128()] = this->newLabel();
+}
+
+asmjit::simd_builder::~simd_builder()
+{
+}
+
+void asmjit::simd_builder::_init(uint new_vsize)
+{
+	if ((!new_vsize && utils::has_avx512_icl()) || new_vsize == 64)
+	{
+		v0 = x86::zmm0;
+		v1 = x86::zmm1;
+		v2 = x86::zmm2;
+		v3 = x86::zmm3;
+		v4 = x86::zmm4;
+		v5 = x86::zmm5;
+		vsize = 64;
+	}
+	else if ((!new_vsize && utils::has_avx2()) || new_vsize == 32)
+	{
+		v0 = x86::ymm0;
+		v1 = x86::ymm1;
+		v2 = x86::ymm2;
+		v3 = x86::ymm3;
+		v4 = x86::ymm4;
+		v5 = x86::ymm5;
+		vsize = 32;
+	}
+	else
+	{
+		v0 = x86::xmm0;
+		v1 = x86::xmm1;
+		v2 = x86::xmm2;
+		v3 = x86::xmm3;
+		v4 = x86::xmm4;
+		v5 = x86::xmm5;
+		vsize = new_vsize ? new_vsize : 16;
+	}
+
+	if (utils::has_avx512())
+	{
+		if (!new_vsize)
+			vmask = -1;
+	}
+	else
+	{
+		vmask = 0;
+	}
+}
+
+void asmjit::simd_builder::operator()() noexcept
+{
+	for (auto&& [x, y] : consts)
+	{
+		this->align(AlignMode::kData, 16);
+		this->bind(y);
+		this->embed(&x, 16);
+	}
+}
+
+void asmjit::simd_builder::vec_cleanup_ret()
+{
+	if (utils::has_avx() && vsize > 16)
+		this->vzeroupper();
+	this->ret();
+}
+
+void asmjit::simd_builder::vec_set_all_zeros(const Operand& v)
+{
+	x86::Xmm reg(v.id());
+	if (utils::has_avx())
+		this->vpxor(reg, reg, reg);
+	else
+		this->xorps(reg, reg);
+}
+
+void asmjit::simd_builder::vec_set_all_ones(const Operand& v)
+{
+	x86::Xmm reg(v.id());
+	if (x86::Zmm zr(v.id()); zr == v)
+		this->vpternlogd(zr, zr, zr, 0xff);
+	else if (x86::Ymm yr(v.id()); yr == v)
+		this->vpcmpeqd(yr, yr, yr);
+	else if (utils::has_avx())
+		this->vpcmpeqd(reg, reg, reg);
+	else
+		this->pcmpeqd(reg, reg);
+}
+
+void asmjit::simd_builder::vec_set_const(const Operand& v, const v128& val)
+{
+	if (!val._u)
+		return vec_set_all_zeros(v);
+	if (!~val._u)
+		return vec_set_all_ones(v);
+	else
+	{
+		Label co = consts[val];
+		if (!co.isValid())
+			co = consts[val] = this->newLabel();
+		if (x86::Zmm zr(v.id()); zr == v)
+			this->vbroadcasti32x4(zr, x86::oword_ptr(co));
+		else if (x86::Ymm yr(v.id()); yr == v)
+			this->vbroadcasti128(yr, x86::oword_ptr(co));
+		else if (utils::has_avx())
+			this->vmovaps(x86::Xmm(v.id()), x86::oword_ptr(co));
+		else
+			this->movaps(x86::Xmm(v.id()), x86::oword_ptr(co));
+	}
+}
+
+void asmjit::simd_builder::vec_clobbering_test(u32 esize, const Operand& v, const Operand& rhs)
+{
+	if (esize == 64)
+	{
+		this->emit(x86::Inst::kIdVptestmd, x86::k0, v, rhs);
+		this->ktestw(x86::k0, x86::k0);
+	}
+	else if (esize == 32)
+	{
+		this->emit(x86::Inst::kIdVptest, v, rhs);
+	}
+	else if (esize == 16 && utils::has_avx())
+	{
+		this->emit(x86::Inst::kIdVptest, v, rhs);
+	}
+	else if (esize == 16 && utils::has_sse41())
+	{
+		this->emit(x86::Inst::kIdPtest, v, rhs);
+	}
+	else
+	{
+		if (v != rhs)
+			this->emit(x86::Inst::kIdPand, v, rhs);
+		if (esize == 16)
+			this->emit(x86::Inst::kIdPacksswb, v, v);
+		this->emit(x86::Inst::kIdMovq, x86::rax, v);
+		if (esize == 16 || esize == 8)
+			this->test(x86::rax, x86::rax);
+		else if (esize == 4)
+			this->test(x86::eax, x86::eax);
+		else if (esize == 2)
+			this->test(x86::ax, x86::ax);
+		else if (esize == 1)
+			this->test(x86::al, x86::al);
+		else
+			fmt::throw_exception("Unimplemented");
+	}
+}
+
+void asmjit::simd_builder::vec_broadcast_gpr(u32 esize, const Operand& v, const x86::Gp& r)
+{
+	if (esize == 2)
+	{
+		if (utils::has_avx512())
+			this->emit(x86::Inst::kIdVpbroadcastw, v, r.r32());
+		else if (utils::has_avx())
+		{
+			this->emit(x86::Inst::kIdVmovd, v, r.r32());
+			if (utils::has_avx2())
+				this->emit(x86::Inst::kIdVpbroadcastw, v, v);
+			else
+			{
+				this->emit(x86::Inst::kIdVpunpcklwd, v, v, v);
+				this->emit(x86::Inst::kIdVpshufd, v, v, Imm(0));
+			}
+		}
+		else
+		{
+			this->emit(x86::Inst::kIdMovd, v, r.r32());
+			this->emit(x86::Inst::kIdPunpcklwd, v, v);
+			this->emit(x86::Inst::kIdPshufd, v, v, Imm(0));
+		}
+	}
+	else if (esize == 4)
+	{
+		if (utils::has_avx512())
+			this->emit(x86::Inst::kIdVpbroadcastd, v, r.r32());
+		else if (utils::has_avx())
+		{
+			this->emit(x86::Inst::kIdVmovd, v, r.r32());
+			if (utils::has_avx2())
+				this->emit(x86::Inst::kIdVpbroadcastd, v, v);
+			else
+				this->emit(x86::Inst::kIdVpshufd, v, v, Imm(0));
+		}
+		else
+		{
+			this->emit(x86::Inst::kIdMovd, v, r.r32());
+			this->emit(x86::Inst::kIdPshufd, v, v, Imm(0));
+		}
+	}
+	else
+	{
+		fmt::throw_exception("Unimplemented");
+	}
+}
+
+asmjit::x86::Mem asmjit::simd_builder::ptr_scale_for_vec(u32 esize, const x86::Gp& base, const x86::Gp& index)
+{
+	switch (ensure(esize))
+	{
+	case 1: return x86::ptr(base, index, 0, 0);
+	case 2: return x86::ptr(base, index, 1, 0);
+	case 4: return x86::ptr(base, index, 2, 0);
+	case 8: return x86::ptr(base, index, 3, 0);
+	default: fmt::throw_exception("Bad esize");
+	}
+}
+
+void asmjit::simd_builder::vec_load_unaligned(u32 esize, const Operand& v, const x86::Mem& src)
+{
+	ensure(std::has_single_bit(esize));
+	ensure(std::has_single_bit(vsize));
+
+	if (esize == 2)
+	{
+		ensure(vsize >= 2);
+		if (vsize == 2)
+			vec_set_all_zeros(v);
+		if (vsize == 2 && utils::has_avx())
+			this->emit(x86::Inst::kIdVpinsrw, x86::Xmm(v.id()), x86::Xmm(v.id()), src, Imm(0));
+		else if (vsize == 2)
+			this->emit(x86::Inst::kIdPinsrw, v, src, Imm(0));
+		else if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu16, v, src);
+		else
+			return vec_load_unaligned(vsize, v, src);
+	}
+	else if (esize == 4)
+	{
+		ensure(vsize >= 4);
+		if (vsize == 4 && utils::has_avx())
+			this->emit(x86::Inst::kIdVmovd, x86::Xmm(v.id()), src);
+		else if (vsize == 4)
+			this->emit(x86::Inst::kIdMovd, v, src);
+		else if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu32, v, src);
+		else
+			return vec_load_unaligned(vsize, v, src);
+	}
+	else if (esize == 8)
+	{
+		ensure(vsize >= 8);
+		if (vsize == 8 && utils::has_avx())
+			this->emit(x86::Inst::kIdVmovq, x86::Xmm(v.id()), src);
+		else if (vsize == 8)
+			this->emit(x86::Inst::kIdMovq, v, src);
+		else if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu64, v, src);
+		else
+			return vec_load_unaligned(vsize, v, src);
+	}
+	else if (esize >= 16)
+	{
+		ensure(vsize >= 16);
+		if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu64, v, src); // Not really needed
+		else if (utils::has_avx())
+			this->emit(x86::Inst::kIdVmovdqu, v, src);
+		else
+			this->emit(x86::Inst::kIdMovups, v, src);
+	}
+	else
+	{
+		fmt::throw_exception("Unimplemented");
+	}
+}
+
+void asmjit::simd_builder::vec_store_unaligned(u32 esize, const Operand& v, const x86::Mem& dst)
+{
+	ensure(std::has_single_bit(esize));
+	ensure(std::has_single_bit(vsize));
+
+	if (esize == 2)
+	{
+		ensure(vsize >= 2);
+		if (vsize == 2 && utils::has_avx())
+			this->emit(x86::Inst::kIdVpextrw, dst, x86::Xmm(v.id()), Imm(0));
+		else if (vsize == 2 && utils::has_sse41())
+			this->emit(x86::Inst::kIdPextrw, dst, v, Imm(0));
+		else if (vsize == 2)
+			this->push(x86::rax), this->pextrw(x86::eax, x86::Xmm(v.id()), 0), this->mov(dst, x86::ax), this->pop(x86::rax);
+		else if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu16, dst, v);
+		else
+			return vec_store_unaligned(vsize, v, dst);
+	}
+	else if (esize == 4)
+	{
+		ensure(vsize >= 4);
+		if (vsize == 4 && utils::has_avx())
+			this->emit(x86::Inst::kIdVmovd, dst, x86::Xmm(v.id()));
+		else if (vsize == 4)
+			this->emit(x86::Inst::kIdMovd, dst, v);
+		else if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu32, dst, v);
+		else
+			return vec_store_unaligned(vsize, v, dst);
+	}
+	else if (esize == 8)
+	{
+		ensure(vsize >= 8);
+		if (vsize == 8 && utils::has_avx())
+			this->emit(x86::Inst::kIdVmovq, dst, x86::Xmm(v.id()));
+		else if (vsize == 8)
+			this->emit(x86::Inst::kIdMovq, dst, v);
+		else if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu64, dst, v);
+		else
+			return vec_store_unaligned(vsize, v, dst);
+	}
+	else if (esize >= 16)
+	{
+		ensure(vsize >= 16);
+		if ((vmask && vmask < 8) || vsize >= 64)
+			this->emit(x86::Inst::kIdVmovdqu64, dst, v); // Not really needed
+		else if (utils::has_avx())
+			this->emit(x86::Inst::kIdVmovdqu, dst, v);
+		else
+			this->emit(x86::Inst::kIdMovups, dst, v);
+	}
+	else
+	{
+		fmt::throw_exception("Unimplemented");
+	}
+}
+
+void asmjit::simd_builder::_vec_binary_op(x86::Inst::Id sse_op, x86::Inst::Id vex_op, x86::Inst::Id evex_op, const Operand& dst, const Operand& lhs, const Operand& rhs)
+{
+	if (utils::has_avx())
+	{
+		if (evex_op != x86::Inst::kIdNone && (vex_op == x86::Inst::kIdNone || this->_extraReg.isReg() || vsize >= 64))
+		{
+			this->evex().emit(evex_op, dst, lhs, rhs);
+		}
+		else
+		{
+			this->emit(vex_op, dst, lhs, rhs);
+		}
+	}
+	else if (dst == lhs)
+	{
+		this->emit(sse_op, dst, rhs);
+	}
+	else if (dst == rhs)
+	{
+		fmt::throw_exception("Unimplemented");
+	}
+	else
+	{
+		this->emit(x86::Inst::kIdMovaps, dst, lhs);
+		this->emit(sse_op, dst, rhs);
+	}
+}
+
+void asmjit::simd_builder::vec_umin(u32 esize, const Operand& dst, const Operand& lhs, const Operand& rhs)
+{
+	using enum x86::Inst::Id;
+	if (esize == 2)
+	{
+		if (utils::has_sse41())
+			return _vec_binary_op(kIdPminuw, kIdVpminuw, kIdVpminuw, dst, lhs, rhs);
+	}
+	else if (esize == 4)
+	{
+		if (utils::has_sse41())
+			return _vec_binary_op(kIdPminud, kIdVpminud, kIdVpminud, dst, lhs, rhs);
+	}
+
+	fmt::throw_exception("Unimplemented");
+}
+
+void asmjit::simd_builder::vec_umax(u32 esize, const Operand& dst, const Operand& lhs, const Operand& rhs)
+{
+	using enum x86::Inst::Id;
+	if (esize == 2)
+	{
+		if (utils::has_sse41())
+			return _vec_binary_op(kIdPmaxuw, kIdVpmaxuw, kIdVpmaxuw, dst, lhs, rhs);
+	}
+	else if (esize == 4)
+	{
+		if (utils::has_sse41())
+			return _vec_binary_op(kIdPmaxud, kIdVpmaxud, kIdVpmaxud, dst, lhs, rhs);
+	}
+
+	fmt::throw_exception("Unimplemented");
+}
+
+void asmjit::simd_builder::vec_cmp_eq(u32 esize, const Operand& dst, const Operand& lhs, const Operand& rhs)
+{
+	using enum x86::Inst::Id;
+	if (esize == 2)
+	{
+		if (vsize == 64)
+		{
+			this->evex().emit(kIdVpcmpeqw, x86::k0, lhs, rhs);
+			this->evex().emit(kIdVpmovm2w, dst, x86::k0);
+		}
+		else
+		{
+			_vec_binary_op(kIdPcmpeqw, kIdVpcmpeqw, kIdNone, dst, lhs, rhs);
+		}
+	}
+	else if (esize == 4)
+	{
+		if (vsize == 64)
+		{
+			this->evex().emit(kIdVpcmpeqd, x86::k0, lhs, rhs);
+			this->evex().emit(kIdVpmovm2d, dst, x86::k0);
+		}
+		else
+		{
+			_vec_binary_op(kIdPcmpeqw, kIdVpcmpeqw, kIdNone, dst, lhs, rhs);
+		}
+	}
+	else
+	{
+		fmt::throw_exception("Unimplemented");
+	}
+}
+
+void asmjit::simd_builder::vec_extract_high(u32, const Operand& dst, const Operand& src)
+{
+	if (vsize == 32)
+		this->vextracti32x8(x86::Ymm(dst.id()), x86::Zmm(src.id()), 1);
+	else if (vsize == 16)
+		this->vextracti128(x86::Xmm(dst.id()), x86::Ymm(src.id()), 1);
+	else
+	{
+		if (utils::has_avx())
+			this->vpsrldq(x86::Xmm(dst.id()), x86::Xmm(src.id()), vsize);
+		else
+		{
+			this->movdqa(x86::Xmm(dst.id()), x86::Xmm(src.id()));
+			this->psrldq(x86::Xmm(dst.id()), vsize);
+		}
+	}
+}
+
+void asmjit::simd_builder::vec_extract_gpr(u32 esize, const x86::Gp& dst, const Operand& src)
+{
+	if (esize == 8 && utils::has_avx())
+		this->vmovq(dst.r64(), x86::Xmm(src.id()));
+	else if (esize == 8)
+		this->movq(dst.r64(), x86::Xmm(src.id()));
+	else if (esize == 4 && utils::has_avx())
+		this->vmovd(dst.r32(), x86::Xmm(src.id()));
+	else if (esize == 4)
+		this->movd(dst.r32(), x86::Xmm(src.id()));
+	else if (esize == 2 && utils::has_avx())
+		this->vpextrw(dst.r32(), x86::Xmm(src.id()), 0);
+	else if (esize == 2)
+		this->pextrw(dst.r32(), x86::Xmm(src.id()), 0);
+	else
+		fmt::throw_exception("Unimplemented");
+}
+
+#endif /* X86 */
+
 #ifdef LLVM_AVAILABLE

 #include <unordered_map>
@@ -996,6 +1487,11 @@ bool jit_compiler::check(const std::string& path)
 	return false;
 }

+void jit_compiler::update_global_mapping(const std::string& name, u64 addr)
+{
+	m_engine->updateGlobalMapping(name, addr);
+}
+
 void jit_compiler::fin()
 {
 	m_engine->finalizeObject();
@@ -42,6 +42,7 @@
 #include <string>
 #include <string_view>
 #include <unordered_map>
+#include <util/v128.hpp>

 #if defined(ARCH_X64)
 using native_asm = asmjit::x86::Assembler;
@@ -211,40 +212,182 @@ namespace asmjit
 	}

 #if defined(ARCH_X64)
-	template <uint Size>
-	struct native_vec;
-
-	template <>
-	struct native_vec<16> { using type = x86::Xmm; };
-
-	template <>
-	struct native_vec<32> { using type = x86::Ymm; };
-
-	template <>
-	struct native_vec<64> { using type = x86::Zmm; };
-
-	template <uint Size>
-	using native_vec_t = typename native_vec<Size>::type;
-
-	// if (count > step) { for (; ctr < (count - step); ctr += step) {...} count -= ctr; }
-	inline void build_incomplete_loop(native_asm& c, auto ctr, auto count, u32 step, auto&& build)
+	struct simd_builder : native_asm
 	{
-		asmjit::Label body = c.newLabel();
-		asmjit::Label exit = c.newLabel();
+		std::unordered_map<v128, Label> consts;

-		ensure((step & (step - 1)) == 0);
-		c.cmp(count, step);
-		c.jbe(exit);
-		c.sub(count, step);
-		c.align(asmjit::AlignMode::kCode, 16);
-		c.bind(body);
-		build();
-		c.add(ctr, step);
-		c.sub(count, step);
-		c.ja(body);
-		c.add(count, step);
-		c.bind(exit);
-	}
+		Operand v0, v1, v2, v3, v4, v5;
+
+		uint vsize = 16;
+		uint vmask = 0;
+
+		simd_builder(CodeHolder* ch) noexcept;
+		~simd_builder();
+
+		void operator()() noexcept;
+
+		void _init(uint new_vsize = 0);
+		void vec_cleanup_ret();
+		void vec_set_all_zeros(const Operand& v);
+		void vec_set_all_ones(const Operand& v);
+		void vec_set_const(const Operand& v, const v128& value);
+		void vec_clobbering_test(u32 esize, const Operand& v, const Operand& rhs);
+		void vec_broadcast_gpr(u32 esize, const Operand& v, const x86::Gp& r);
+
+		// return x86::ptr(base, ctr, X, 0) where X is set for esize accordingly
+		x86::Mem ptr_scale_for_vec(u32 esize, const x86::Gp& base, const x86::Gp& index);
+
+		void vec_load_unaligned(u32 esize, const Operand& v, const x86::Mem& src);
+		void vec_store_unaligned(u32 esize, const Operand& v, const x86::Mem& dst);
+		void vec_partial_move(u32 esize, const Operand& dst, const Operand& src);
+
+		void _vec_binary_op(x86::Inst::Id sse_op, x86::Inst::Id vex_op, x86::Inst::Id evex_op, const Operand& dst, const Operand& lhs, const Operand& rhs);
+
+		void vec_shuffle_xi8(const Operand& dst, const Operand& lhs, const Operand& rhs)
+		{
+			using enum x86::Inst::Id;
+			_vec_binary_op(kIdPshufb, kIdVpshufb, kIdVpshufb, dst, lhs, rhs);
+		}
+
+		void vec_xor(u32, const Operand& dst, const Operand& lhs, const Operand& rhs)
+		{
+			using enum x86::Inst::Id;
+			_vec_binary_op(kIdPxor, kIdVpxor, kIdVpxord, dst, lhs, rhs);
+		}
+
+		void vec_or(u32, const Operand& dst, const Operand& lhs, const Operand& rhs)
+		{
+			using enum x86::Inst::Id;
+			_vec_binary_op(kIdPor, kIdVpor, kIdVpord, dst, lhs, rhs);
+		}
+
+		void vec_andn(u32, const Operand& dst, const Operand& lhs, const Operand& rhs)
+		{
+			using enum x86::Inst::Id;
+			_vec_binary_op(kIdPandn, kIdVpandn, kIdVpandnd, dst, lhs, rhs);
+		}
+
+		void vec_umin(u32 esize, const Operand& dst, const Operand& lhs, const Operand& rhs);
+		void vec_umax(u32 esize, const Operand& dst, const Operand& lhs, const Operand& rhs);
+		void vec_cmp_eq(u32 esize, const Operand& dst, const Operand& lhs, const Operand& rhs);
+
+		void vec_extract_high(u32 esize, const Operand& dst, const Operand& src);
+		void vec_extract_gpr(u32 esize, const x86::Gp& dst, const Operand& src);
+
+		simd_builder& keep_if_not_masked()
+		{
+			if (vmask && vmask < 8)
+			{
+				this->k(x86::KReg(vmask));
+			}
+
+			return *this;
+		}
+
+		simd_builder& zero_if_not_masked()
+		{
+			if (vmask && vmask < 8)
+			{
+				this->k(x86::KReg(vmask));
+				this->z();
+			}
+
+			return *this;
+		}
+
+		void build_loop(u32 esize, const x86::Gp& reg_ctr, const x86::Gp& reg_cnt, auto&& build, auto&& reduce)
+		{
+			ensure((esize & (esize - 1)) == 0);
+			ensure(esize <= vsize);
+
+			Label body = this->newLabel();
+			Label next = this->newLabel();
+			Label exit = this->newLabel();
+
+			const u32 step = vsize / esize;
+
+			this->xor_(reg_ctr.r32(), reg_ctr.r32()); // Reset counter reg
+			this->cmp(reg_cnt, step);
+			this->jb(next); // If count < step, skip main loop body
+			this->align(AlignMode::kCode, 16);
+			this->bind(body);
+			this->sub(reg_cnt, step);
+			build();
+			this->add(reg_ctr, step);
+			this->cmp(reg_cnt, step);
+			this->jae(body);
+			this->bind(next);
+
+			if (vmask)
+			{
+				// Build single last iteration (masked)
+				this->test(reg_cnt, reg_cnt);
+				this->jz(exit);
+
+				if (esize == 1 && vsize == 64)
+				{
+					this->bzhi(reg_cnt.r64(), x86::Mem(consts[~u128()], 0), reg_cnt.r64());
+					this->kmovq(x86::k7, reg_cnt.r64());
+				}
+				else
+				{
+					this->bzhi(reg_cnt.r32(), x86::Mem(consts[~u128()], 0), reg_cnt.r32());
+					this->kmovd(x86::k7, reg_cnt.r32());
+				}
+
+				vmask = 7;
+				build();
+
+				// Rollout reduction step
+				this->bind(exit);
+				while (true)
+				{
+					vsize /= 2;
+					if (vsize < esize)
+						break;
+					this->_init(vsize);
+					reduce();
+				}
+			}
+			else
+			{
+				// Build unrolled loop tail (reduced vector width)
+				while (true)
+				{
+					vsize /= 2;
+					if (vsize < esize)
+						break;
+
+					// Shall not clobber flags
+					this->_init(vsize);
+					reduce();
+
+					if (vsize == esize)
+					{
+						// Last "iteration"
+						this->test(reg_cnt, reg_cnt);
+						this->jz(exit);
+						build();
+					}
+					else
+					{
+						const u32 step = vsize / esize;
+						Label next = this->newLabel();
+						this->cmp(reg_cnt, step);
+						this->jb(next);
+						build();
+						this->add(reg_ctr, step);
+						this->sub(reg_cnt, step);
+						this->bind(next);
+					}
+				}
+
+				this->bind(exit);
+			}
+
+			this->_init(0);
+		}
+	};

 	// for (; count > 0; ctr++, count--)
 	inline void build_loop(native_asm& c, auto ctr, auto count, auto&& build)
@@ -262,19 +405,40 @@ namespace asmjit
 		c.ja(body);
 		c.bind(exit);
 	}
+
+	inline void maybe_flush_lbr(native_asm& c, uint count = 2)
+	{
+		// Workaround for bad LBR callstacks which happen in some situations (mainly TSX) - execute additional RETs
+		Label next = c.newLabel();
+		c.lea(x86::rcx, x86::qword_ptr(next));
+
+		for (u32 i = 0; i < count; i++)
+		{
+			c.push(x86::rcx);
+			c.sub(x86::rcx, 16);
+		}
+
+		for (u32 i = 0; i < count; i++)
+		{
+			c.ret();
+			c.align(asmjit::AlignMode::kCode, 16);
+		}
+
+		c.bind(next);
+	}
 #endif
 }

 // Build runtime function with asmjit::X86Assembler
 template <typename FT, typename Asm = native_asm, typename F>
-inline FT build_function_asm(std::string_view name, F&& builder)
+inline FT build_function_asm(std::string_view name, F&& builder, ::jit_runtime* custom_runtime = nullptr)
 {
 #ifdef __APPLE__
 	pthread_jit_write_protect_np(false);
 #endif
 	using namespace asmjit;

-	auto& rt = get_global_runtime();
+	auto& rt = custom_runtime ? *custom_runtime : get_global_runtime();

 	CodeHolder code;
 	code.init(rt.environment());
@@ -312,6 +476,12 @@ inline FT build_function_asm(std::string_view name, F&& builder)
 		builder(compiler, args);
 	}

+	if constexpr (std::is_invocable_r_v<void, Asm>)
+	{
+		// Finalization
+		compiler();
+	}
+
 	const auto result = rt._add(&code);
 	jit_announce(result, code.codeSize(), name);
 	return reinterpret_cast<FT>(uptr(result));
@@ -362,6 +532,9 @@ public:
 	// Add object (path to obj file)
 	void add(const std::string& path);

+	// Update global mapping for a single value
+	void update_global_mapping(const std::string& name, u64 addr);
+
 	// Check object file
 	static bool check(const std::string& path);

@@ -5,6 +5,8 @@
 #include "util/logs.hpp"
 #include "util/v128.hpp"

+#include <locale>
+#include <codecvt>
 #include <algorithm>
 #include <string_view>
 #include "Thread.h"
@@ -13,8 +15,6 @@
 #include <Windows.h>
 #else
 #include <errno.h>
-#include <locale>
-#include <codecvt>
 #endif

 std::string wchar_to_utf8(std::wstring_view src)
@@ -26,11 +26,17 @@ std::string wchar_to_utf8(std::wstring_view src)
 	WideCharToMultiByte(CP_UTF8, 0, src.data(), src.size(), utf8_string.data(), tmp_size, nullptr, nullptr);
 	return utf8_string;
 #else
-	std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>, wchar_t> converter{};
+	std::wstring_convert<std::codecvt_utf8<wchar_t>, wchar_t> converter{};
 	return converter.to_bytes(src.data());
 #endif
 }

+std::string utf16_to_utf8(std::u16string_view src)
+{
+	std::wstring_convert<std::codecvt_utf8_utf16<char16_t>, char16_t> converter{};
+	return converter.to_bytes(src.data());
+}
+
 std::wstring utf8_to_wchar(std::string_view src)
 {
 #ifdef _WIN32
@@ -609,3 +615,13 @@ bool fmt::match(const std::string& source, const std::string& mask)

 	return true;
 }
+
+std::string get_file_extension(const std::string& file_path)
+{
+	if (usz dotpos = file_path.find_last_of('.'); dotpos != std::string::npos && dotpos + 1 < file_path.size())
+	{
+		return file_path.substr(dotpos + 1);
+	}
+
+	return {};
+}
@@ -10,6 +10,7 @@

 std::wstring utf8_to_wchar(std::string_view src);
 std::string wchar_to_utf8(std::wstring_view src);
+std::string utf16_to_utf8(std::u16string_view src);

 // Copy null-terminated string from a std::string or a char array to a char array with truncation
 template <typename D, typename T>
@@ -26,6 +27,12 @@ bool try_to_int64(s64* out, std::string_view value, s64 min, s64 max);
 // Convert string to unsigned integer
 bool try_to_uint64(u64* out, std::string_view value, u64 min, u64 max);

+// Convert string to float
+bool try_to_float(f64* out, std::string_view value, f64 min, f64 max);
+
+// Get the file extension of a file path ("png", "jpg", etc.)
+std::string get_file_extension(const std::string& file_path);
+
 namespace fmt
 {
 	std::string replace_all(std::string_view src, std::string_view from, std::string_view to, usz count = -1);
@@ -8,7 +8,6 @@
 #include "Thread.h"
 #include "Utilities/JIT.h"
 #include <thread>
-#include <sstream>
 #include <cfenv>

 #ifdef _WIN32
@@ -17,6 +16,9 @@
 #include <process.h>
 #include <sysinfoapi.h>
 #else
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
 #ifdef __APPLE__
 #define _XOPEN_SOURCE
 #define __USE_GNU
@@ -96,14 +98,6 @@ extern thread_local std::string(*g_tls_log_prefix)();
 // Report error and call std::abort(), defined in main.cpp
 [[noreturn]] void report_fatal_error(std::string_view);

-template <>
-void fmt_class_string<std::thread::id>::format(std::string& out, u64 arg)
-{
-	std::ostringstream ss;
-	ss << get_object(arg);
-	out += ss.str();
-}
-
 std::string dump_useful_thread_info()
 {
 	thread_local volatile bool guard = false;
@@ -1248,16 +1242,18 @@ bool handle_access_violation(u32 addr, bool is_writing, ucontext_t* context) noe

 	const auto cpu = get_current_cpu_thread();

-	if (rsx::g_access_violation_handler)
+	if (addr < RAW_SPU_BASE_ADDR && vm::check_addr(addr) && rsx::g_access_violation_handler)
 	{
+		bool state_changed = false;
+
 		if (cpu)
 		{
-			vm::temporary_unlock(*cpu);
+			state_changed = vm::temporary_unlock(*cpu);
 		}

 		bool handled = rsx::g_access_violation_handler(addr, is_writing);

-		if (cpu && (cpu->state += cpu_flag::temp, cpu->test_stopped()))
+		if (state_changed && (cpu->state += cpu_flag::temp, cpu->test_stopped()))
 		{
 			//
 		}
@@ -1536,14 +1532,14 @@ bool handle_access_violation(u32 addr, bool is_writing, ucontext_t* context) noe
 			// If we fail due to being busy, wait a bit and try again.
 			while (static_cast<u32>(sending_error) == CELL_EBUSY)
 			{
+				thread_ctrl::wait_for(1000);
+				sending_error = sys_event_port_send(pf_port_id, data1, data2, data3);
+
 				if (cpu->is_stopped())
 				{
 					sending_error = {};
 					break;
 				}
-
-				thread_ctrl::wait_for(1000);
-				sending_error = sys_event_port_send(pf_port_id, data1, data2, data3);
 			}

 			if (sending_error)
@@ -1644,7 +1640,7 @@ static void append_thread_name(std::string& msg)
 	}
 	else
 	{
-		fmt::append(msg, "Thread id = %s.\n", std::this_thread::get_id());
+		fmt::append(msg, "Thread id = %u.\n", thread_ctrl::get_tid());
 	}
 }

@@ -3139,6 +3135,8 @@ u64 thread_ctrl::get_tid()
 {
 #ifdef _WIN32
 	return GetCurrentThreadId();
+#elif defined(__linux__)
+	return syscall(SYS_gettid);
 #else
 	return reinterpret_cast<u64>(pthread_self());
 #endif
@@ -635,13 +635,21 @@ public:
 	// Join thread by thread_state::finished
 	named_thread& operator=(thread_state s)
 	{
+		bool notify_sync = false;
+
+		if (s >= thread_state::aborting && thread::m_sync.fetch_op([](u64& v){ return !(v & 3) && (v |= 1); }).second)
+		{
+			notify_sync = true;
+		}
+
 		if constexpr (std::is_assignable_v<Context&, thread_state>)
 		{
 			static_cast<Context&>(*this) = s;
 		}

-		if (s >= thread_state::aborting && thread::m_sync.fetch_op([](u64& v){ return !(v & 3) && (v |= 1); }).second)
+		if (notify_sync)
 		{
+			// Notify after context abortion has been made so all conditions for wake-up be satisfied by the time of notification
 			thread::m_sync.notify_one(1);
 		}

@@ -1,37 +1,36 @@
 #pragma once

 #include "util/types.hpp"
+#include "util/vm.hpp"
 #include "StrFmt.h"
 #include <vector>
 #include <algorithm>

-
 namespace utils
 {
 	class address_range_vector;

-
 	/**
-	 * Constexprs
+	 * Helpers
 	 */
-	constexpr inline u32 page_start(u32 addr)
+	static inline u32 page_start(u32 addr)
 	{
-		return addr & ~4095u;
+		return addr & ~(c_page_size - 1);
 	}

-	constexpr inline u32 next_page(u32 addr)
+	static inline u32 next_page(u32 addr)
 	{
-		return page_start(addr + 4096u);
+		return page_start(addr) + c_page_size;
 	}

-	constexpr inline u32 page_end(u32 addr)
+	static inline u32 page_end(u32 addr)
 	{
 		return next_page(addr) - 1;
 	}

-	constexpr inline u32 is_page_aligned(u32 addr)
+	static inline u32 is_page_aligned(u32 val)
 	{
-		return page_start(addr) == addr;
+		return (val & (c_page_size - 1)) == 0;
 	}


@@ -186,7 +185,7 @@ namespace utils

 		bool is_page_range() const
 		{
-			return (valid() && start % 4096u == 0 && length() % 4096u == 0);
+			return (valid() && is_page_aligned(start) && is_page_aligned(length()));
 		}

 		address_range to_page_range() const
@@ -939,7 +939,7 @@ std::basic_string<u32> patch_engine::apply(const std::string& name, u8* dst, u32
 	}

 	std::basic_string<u32> applied_total;
-	const auto& container = m_map.at(name);
+	const auto& container = ::at32(m_map, name);
 	const auto& serial = Emu.GetTitleID();
 	const auto& app_version = Emu.GetAppVersion();

@@ -975,7 +975,7 @@ std::basic_string<u32> patch_engine::apply(const std::string& name, u8* dst, u32
 				continue;
 			}

-			const auto& app_versions = serials.at(found_serial);
+			const auto& app_versions = ::at32(serials, found_serial);
 			std::string found_app_version;

 			if (app_versions.find(app_version) != app_versions.end())
@@ -988,7 +988,7 @@ std::basic_string<u32> patch_engine::apply(const std::string& name, u8* dst, u32
 				is_all_versions = true;
 			}

-			if (!found_app_version.empty() && app_versions.at(found_app_version))
+			if (!found_app_version.empty() && ::at32(app_versions, found_app_version))
 			{
 				// This patch is enabled
 				if (is_all_serials)
@@ -1073,7 +1073,7 @@ void patch_engine::unload(const std::string& name)
 		return;
 	}

-	const auto& container = m_map.at(name);
+	const auto& container = ::at32(m_map, name);

 	for (const auto& [description, patch] : container.patch_info_map)
 	{
@@ -371,12 +371,12 @@ public:
 	bool bit_test_reset(uint bit) = delete;
 	bool bit_test_invert(uint bit) = delete;

-	bool all_of(bs_t arg)
+	bool all_of(bs_t arg) const
 	{
 		return base::load().all_of(arg);
 	}

-	bool none_of(bs_t arg)
+	bool none_of(bs_t arg) const
 	{
 		return base::load().none_of(arg);
 	}
@@ -700,6 +700,12 @@ template<typename T>
 struct size3_base
 {
 	T width, height, depth;
+
+	template<typename NT>
+	explicit constexpr operator size3_base<NT>() const
+	{
+		return{ static_cast<NT>(width), static_cast<NT>(height), static_cast<NT>(depth) };
+	}
 };

 template<typename T>
@@ -6,14 +6,27 @@ void shared_mutex::imp_lock_shared(u32 val)
 {
 	ensure(val < c_err); // "shared_mutex underflow"

+	// Try to steal the notification bit
+	if (val & c_sig && m_value.compare_exchange(val, val - c_sig + 1))
+	{
+		return;
+	}
+
 	for (int i = 0; i < 10; i++)
 	{
-		busy_wait();
-
 		if (try_lock_shared())
 		{
 			return;
 		}
+
+		const u32 old = m_value;
+
+		if (old & c_sig && m_value.compare_and_swap_test(old, old - c_sig + 1))
+		{
+			return;
+		}
+
+		busy_wait();
 	}

 	// Acquire writer lock and downgrade
@@ -75,11 +88,24 @@ void shared_mutex::imp_lock(u32 val)
 {
 	ensure(val < c_err); // "shared_mutex underflow"

+	// Try to steal the notification bit
+	if (val & c_sig && m_value.compare_exchange(val, val - c_sig + c_one))
+	{
+		return;
+	}
+
 	for (int i = 0; i < 10; i++)
 	{
 		busy_wait();

-		if (!m_value && try_lock())
+		const u32 old = m_value;
+
+		if (!old && try_lock())
+		{
+			return;
+		}
+
+		if (old & c_sig && m_value.compare_and_swap_test(old, old - c_sig + c_one))
 		{
 			return;
 		}
@@ -171,6 +171,11 @@ public:
 	{
 		return m_value.load() < c_one - 1;
 	}
+
+	bool has_waiters() const
+	{
+		return m_value.load() > c_one;
+	}
 };

 // Simplified shared (reader) lock implementation.
@@ -0,0 +1,144 @@
+#include "stdafx.h"
+#include "stack_trace.h"
+
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#include <Windows.h>
+#define DBGHELP_TRANSLATE_TCHAR
+#include <DbgHelp.h>
+#include <codecvt>
+#else
+#include <execinfo.h>
+#endif
+
+namespace utils
+{
+#ifdef _WIN32
+	std::string wstr_to_utf8(LPWSTR data, int str_len)
+	{
+		if (!str_len)
+		{
+			return {};
+		}
+
+		// Calculate size
+		const auto length = WideCharToMultiByte(CP_UTF8, 0, data, str_len, NULL, 0, NULL, NULL);
+
+		// Convert
+		std::vector<char> out(length + 1, 0);
+		WideCharToMultiByte(CP_UTF8, 0, data, str_len, out.data(), length, NULL, NULL);
+		return out.data();
+	}
+
+	std::vector<void*> get_backtrace(int max_depth)
+	{
+		std::vector<void*> result = {};
+
+		const auto hProcess = ::GetCurrentProcess();
+		const auto hThread = ::GetCurrentThread();
+
+		CONTEXT context{};
+		RtlCaptureContext(&context);
+
+		STACKFRAME64 stack = {};
+		stack.AddrPC.Offset = context.Rip;
+		stack.AddrPC.Mode = AddrModeFlat;
+		stack.AddrStack.Offset = context.Rsp;
+		stack.AddrStack.Mode = AddrModeFlat;
+		stack.AddrFrame.Offset = context.Rbp;
+		stack.AddrFrame.Mode = AddrModeFlat;
+
+		while (max_depth--)
+		{
+			if (!StackWalk64(
+				IMAGE_FILE_MACHINE_AMD64,
+				hProcess,
+				hThread,
+				&stack,
+				&context,
+				NULL,
+				SymFunctionTableAccess64,
+				SymGetModuleBase64,
+				NULL))
+			{
+				break;
+			}
+
+			result.push_back(reinterpret_cast<void*>(stack.AddrPC.Offset));
+		}
+
+		return result;
+	}
+
+	std::vector<std::string> get_backtrace_symbols(const std::vector<void*>& stack)
+	{
+		std::vector<std::string> result = {};
+		std::vector<u8> symbol_buf(sizeof(SYMBOL_INFO) + sizeof(TCHAR) * 256);
+
+		const auto hProcess = ::GetCurrentProcess();
+
+		auto sym = reinterpret_cast<SYMBOL_INFO*>(symbol_buf.data());
+		sym->SizeOfStruct = sizeof(SYMBOL_INFO);
+		sym->MaxNameLen = 256;
+
+		IMAGEHLP_LINEW64 line_info{};
+		line_info.SizeOfStruct = sizeof(IMAGEHLP_LINEW64);
+
+		SymInitialize(hProcess, NULL, TRUE);
+		SymSetOptions(SYMOPT_LOAD_LINES);
+
+		for (const auto& pointer : stack)
+		{
+			DWORD64 unused;
+			SymFromAddr(hProcess, reinterpret_cast<DWORD64>(pointer), &unused, sym);
+
+			if (sym->NameLen)
+			{
+				const auto function_name = wstr_to_utf8(sym->Name, static_cast<int>(sym->NameLen));
+
+				// Attempt to get file and line information if available
+				DWORD unused2;
+				if (SymGetLineFromAddrW(hProcess, reinterpret_cast<DWORD64>(pointer), &unused2, &line_info))
+				{
+					const auto full_path = fmt::format("%s:%u %s", wstr_to_utf8(line_info.FileName, -1), line_info.LineNumber, function_name);
+					result.push_back(full_path);
+				}
+				else
+				{
+					result.push_back(function_name);
+				}
+			}
+			else
+			{
+				result.push_back(fmt::format("rpcs3@0xp", pointer));
+			}
+		}
+
+		return result;
+	}
+#else
+	std::vector<void*> get_backtrace(int max_depth)
+	{
+		std::vector<void*> result(max_depth);
+		int depth = backtrace(result.data(), max_depth);
+
+		result.resize(depth);
+		return result;
+	}
+
+	std::vector<std::string> get_backtrace_symbols(const std::vector<void*>& stack)
+	{
+		std::vector<std::string> result;
+		result.reserve(stack.size());
+
+		const auto symbols = backtrace_symbols(stack.data(), static_cast<int>(stack.size()));
+		for (usz i = 0; i < stack.size(); ++i)
+		{
+			result.push_back(symbols[i]);
+		}
+
+		free(symbols);
+		return result;
+	}
+#endif
+}
@@ -0,0 +1,20 @@
+#pragma once
+#include <util/types.hpp>
+#include <util/logs.hpp>
+
+namespace utils
+{
+	std::vector<void*> get_backtrace(int max_depth = 255);
+	std::vector<std::string> get_backtrace_symbols(const std::vector<void*>& stack);
+
+	FORCE_INLINE void print_trace(logs::channel& logger, int max_depth = 255)
+	{
+		const auto trace = get_backtrace(max_depth);
+		const auto lines = get_backtrace_symbols(trace);
+
+		for (const auto& line : lines)
+		{
+			logger.error("%s", line);
+		}
+	}
+}
@@ -99,13 +99,13 @@ inline int futex(volatile void* uaddr, int futex_op, uint val, const timespec* t

 				if (!timeout)
 				{
-					rec.cv.wait(lock, [&] { return !rec.mask; });
+					rec.cv.wait(lock, FN(!rec.mask));
 				}
 				else if (futex_op == FUTEX_WAIT)
 				{
 					const auto nsec = std::chrono::nanoseconds(timeout->tv_nsec + timeout->tv_sec * 1000000000ull);

-					if (!rec.cv.wait_for(lock, nsec, [&] { return !rec.mask; }))
+					if (!rec.cv.wait_for(lock, nsec, FN(!rec.mask)))
 					{
 						res = -1;
 						errno = ETIMEDOUT;
@@ -30,7 +30,8 @@ namespace utils

 		if (lo())
 		{
-			version += "." + std::to_string(lo());
+			version += '.';
+			version += std::to_string(lo());
 		}

 		if (type() != version_type::release)
@@ -40,7 +41,8 @@ namespace utils
 				version += "-" + postfix();
 			}

-			version += " " + utils::to_string(type());
+			version += ' ';
+			version += utils::to_string(type());

 			if (type_index() > 1)
 			{
@@ -3,6 +3,7 @@ YoRHa Theme for RPCS3
 by Ani @ https://github.com/AniLeo
 r1 (2018.02.27)
 r2 (2021.08.28)
+r3 (2022.08.20)
 */

 /*
@@ -26,7 +27,7 @@ aea993

 /* Every widget */
 QWidget {
-	font-family: SCE-PS3 Rodin LATIN;
+	font-family: "SCE-PS3 Rodin LATIN";
 	font-size: 9.00pt;

 	color: #292929;
@@ -98,6 +99,9 @@ QToolTip {
 }

 /* Table headers */
+QHeaderView {
+	color: #b3ac98;
+}
 QHeaderView::section {
 	text-transform: uppercase;
 	background: #4d4940;
@@ -109,22 +113,32 @@ QHeaderView::section {
 	border: none;
 }

-/* All other Tabs */
+/* All Tabs */
 QTabBar {
+	border-bottom: 0.05em solid #4d4940;
 	text-transform: uppercase;
 }
 QTabBar::tab {
 	background: transparent;
+	width: 4.80em;
 	padding-left: 0.50em;
 	padding-right: 0.50em;
-	padding-top: 0.25em;
-	padding-bottom: 0.25em;
+	padding-top: 0.65em;
+	padding-bottom: 0.65em;
 	margin-right: 0.25em;
+	font-size: 10.5pt;
+	font-weight: 550;
+}
+QTabBar::tab:last {
+	margin-right: 0em;
+}
+QTabBar::tab:!selected:hover {
+	background: transparent;
+	color: #4d4940;
 }
 QTabBar::tab::selected {
 	background: #4d4940;
 	color: #aea993;
-	border-bottom-style: solid;
 }

 /* Checkboxes */
@@ -153,29 +167,35 @@ QCheckBox::indicator::disabled, QRadioButton::indicator::disabled {
 	background-color: #828790; /* Gray */
 }

-/* Combo Boxes */
-QComboBox {
+/* Combo Boxes, Datetime dropdown */
+QComboBox, QDateTimeEdit, QLineEdit {
 	background: transparent;
 	color: #4d4940;
 	border: 0.05em solid #1e1e1e;
 	border-radius: 0.15em;
-	padding-bottom: 0.2em;
-	padding-left: 0.4em;
+	padding-top: 0.1em;
+	padding-bottom: 0.1em;
 }
 QComboBox QAbstractItemView {
 	background: #b3ac98;
 }
-QComboBox::disabled {
+QComboBox::disabled, QDateTimeEdit::disabled, QLineEdit::disabled {
 	background: #828790;
 	color: #4d4940;
 }
+/* Text boxes need less left padding than combo boxes */
+QComboBox {
+	padding-left: 0.4em;
+}
+QLineEdit {
+	padding-left: 0.2em;
+}

 /* Group Boxes (Settings Dialog) */
 QGroupBox {
 	margin-top: 1em;
 	border: 0.05em solid #4d4940;
 	text-transform: uppercase;
-	font-size: 9.00pt;
 }
 QGroupBox::title {
 	subcontrol-origin: margin;
@@ -236,6 +256,10 @@ QDockWidget::close-button, QDockWidget::float-button {
 QTabWidget::pane {
 	border: 0em solid #4d4940;
 }
+/* Center tabs */
+QTabWidget::tab-bar {
+	alignment: center;
+}

 /* Top menu bar */
 QMenuBar {
@@ -284,6 +308,12 @@ QToolButton::hover {
 	background-color: #b3ac98;
 }

+/* Calendar widget */
+QCalendarWidget QWidget{
+	background: #b3ac98;
+	color: #4d4940;
+}
+

 /*** RPCS3 Specifics ***/

@@ -303,40 +333,24 @@ QWidget#trophy_manager, QWidget#cg_disasm, QWidget#log_viewer, QMainWindow#main_
 	border: 0.05em solid #4d4940;
 	color: #292929;
 	font-family: none;
-	font-size: 9.50pt;
+	font-size: 9.5pt;
 }

-/* Pad Settings, Emulator Settings Dialogs: Tabs */
-#pad_settings_dialog QTabWidget::tab-bar {
-	alignment: center;
+/* Align the log widget tabs to the left */
+QTabWidget#tab_widget_log::tab-bar {
+	alignment: left;
 }
-QTabBar#tab_bar_settings, #pad_settings_dialog QTabBar {
-	border-bottom: 0.05em solid #4d4940;
-	text-transform: uppercase;
+
+/* Restore log section QTabBar to small sizes */
+QTabBar#tab_bar_log {
+	border-bottom: 0;
 }
-QTabBar::tab#tab_bar_settings, #pad_settings_dialog QTabBar::tab {
-	background: transparent;
-	width: 5.20em;
-	padding-left: 0.50em;
-	padding-right: 0.50em;
-	padding-top: 0.65em;
-	padding-bottom: 0.65em;
-	margin-right: 0.25em;
-	font-size: 10.5pt;
-	font-weight: 550;
-}
-QTabBar::tab:last#tab_bar_settings, #pad_settings_dialog QTabBar::tab:last {
-	margin-right: 0em;
-}
-QTabBar::tab:!selected:hover#tab_bar_settings, #pad_settings_dialog QTabBar::tab:!selected:hover {
-	background: transparent;
-	color: #4d4940;
-}
-QTabBar::tab::selected#tab_bar_settings, #pad_settings_dialog QTabBar::tab::selected {
-	background: #4d4940;
-	color: #aea993;
-	border-bottom-style: solid;
-	margin-top: 0.15em;
+QTabBar#tab_bar_log::tab {
+	width: 2.00em;
+	padding: 0.25em;
+	font-size: 9.0pt;
+	font-weight: 400;
+	margin-right: 0;
 }

 /* Pad Settings: Controller Image */
@@ -362,6 +376,7 @@ QLineEdit#mw_searchbar {
 	margin-left: 0.7em;
 	color: #4d4940;
 	font-size: 10.25pt;
+	border: 0em;
 }

 /* Set Theme UI colors */
@@ -440,3 +455,11 @@ QWidget#trophy_notification_frame {
 QSlider#sizeSlider::groove:horizontal {
 	height: 1.5em;
 }
+
+/* 
+	On Patch Creator, it's the patch validation label
+	since we can't avoid repainting color, let's bold it instead
+*/
+QLabel#validLabel {
+	font-weight: 600;
+}
@@ -26,11 +26,17 @@ struct loaded_npdrm_keys
 	}

 	// TODO: Check if correct for ELF files usage
-	u128 last_key() const
+	u128 last_key(usz backwards = 0) const
 	{
+		backwards++;
 		const usz pos = dec_keys_pos;
-		return pos ? dec_keys[(pos - 1) % std::size(dec_keys)].load() : u128{};
+		return pos >= backwards ? dec_keys[(pos - backwards) % std::size(dec_keys)].load() : u128{};
 	}
+
+	SAVESTATE_INIT_POS(2);
+	loaded_npdrm_keys() = default;
+	loaded_npdrm_keys(utils::serial& ar);
+	void save(utils::serial& ar);
 };

 struct NPD_HEADER
@@ -53,11 +53,13 @@
  </ItemGroup>
  <ItemGroup>
    <ClInclude Include="Emu\Audio\Cubeb\CubebBackend.h" />
+    <ClInclude Include="Emu\Audio\Cubeb\cubeb_enumerator.h" />
  </ItemGroup>
  <ItemGroup>
    <ClCompile Include="Emu\Audio\Cubeb\CubebBackend.cpp" />
+    <ClCompile Include="Emu\Audio\Cubeb\cubeb_enumerator.cpp" />
  </ItemGroup>
  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
  <ImportGroup Label="ExtensionTargets">
  </ImportGroup>
-</Project>
+</Project>
@@ -10,10 +10,16 @@
    <ClCompile Include="Emu\Audio\Cubeb\CubebBackend.cpp">
      <Filter>Source Files</Filter>
    </ClCompile>
+    <ClCompile Include="Emu\Audio\Cubeb\cubeb_enumerator.cpp">
+      <Filter>Source Files</Filter>
+    </ClCompile>
  </ItemGroup>
  <ItemGroup>
    <ClInclude Include="Emu\Audio\Cubeb\CubebBackend.h">
      <Filter>Source Files</Filter>
    </ClInclude>
+    <ClInclude Include="Emu\Audio\Cubeb\cubeb_enumerator.h">
+      <Filter>Source Files</Filter>
+    </ClInclude>
  </ItemGroup>
 </Project>
@@ -6,10 +6,16 @@

 AudioBackend::AudioBackend() {}

-void AudioBackend::SetErrorCallback(std::function<void()> cb)
+void AudioBackend::SetWriteCallback(std::function<u32(u32 /* byte_cnt */, void* /* buffer */)> cb)
 {
-	std::lock_guard lock(m_error_cb_mutex);
-	m_error_callback = cb;
+	std::lock_guard lock(m_cb_mutex);
+	m_write_callback = cb;
+}
+
+void AudioBackend::SetStateCallback(std::function<void(AudioStateEvent)> cb)
+{
+	std::lock_guard lock(m_state_cb_mutex);
+	m_state_callback = cb;
 }

 /*
@@ -105,7 +111,7 @@ std::pair<AudioChannelCnt, AudioChannelCnt> AudioBackend::get_channel_count_and_
 	audio_out_configuration& audio_out_cfg = g_fxo->get<audio_out_configuration>();
 	std::lock_guard lock(audio_out_cfg.mtx);
 	ensure(device_index < audio_out_cfg.out.size());
-	const audio_out_configuration::audio_out& out = audio_out_cfg.out.at(device_index);
+	const audio_out_configuration::audio_out& out = ::at32(audio_out_cfg.out, device_index);
 	return out.get_channel_count_and_downmixer();
 }

@@ -114,7 +120,7 @@ AudioChannelCnt AudioBackend::get_max_channel_count(u32 device_index)
 	audio_out_configuration& audio_out_cfg = g_fxo->get<audio_out_configuration>();
 	std::lock_guard lock(audio_out_cfg.mtx);
 	ensure(device_index < audio_out_cfg.out.size());
-	const audio_out_configuration::audio_out& out = audio_out_cfg.out.at(device_index);
+	const audio_out_configuration::audio_out& out = ::at32(audio_out_cfg.out, device_index);

 	AudioChannelCnt count = AudioChannelCnt::STEREO;

@@ -134,3 +140,24 @@ AudioChannelCnt AudioBackend::get_max_channel_count(u32 device_index)

 	return count;
 }
+
+AudioChannelCnt AudioBackend::convert_channel_count(u64 raw)
+{
+	switch (raw)
+	{
+	default:
+	case 8:
+		return AudioChannelCnt::SURROUND_7_1;
+	case 7:
+	case 6:
+		return AudioChannelCnt::SURROUND_5_1;
+	case 5:
+	case 4:
+	case 3:
+	case 2:
+	case 1:
+		return AudioChannelCnt::STEREO;
+	case 0:
+		fmt::throw_exception("Usupported channel count");
+	}
+}
@@ -37,6 +37,12 @@ enum class AudioChannelCnt : u32
 	SURROUND_7_1 = 8,
 };

+enum class AudioStateEvent : u32
+{
+	UNSPECIFIED_ERROR,
+	DEFAULT_DEVICE_MAYBE_CHANGED,
+};
+
 class AudioBackend
 {
 public:
@@ -60,19 +66,21 @@ public:
 	virtual std::string_view GetName() const = 0;

 	// (Re)create output stream with new parameters. Blocks until data callback returns.
+	// If dev_id is empty, then default device will be selected.
+	// May override channel count if device has smaller number of channels.
 	// Should return 'true' on success.
-	virtual bool Open(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) = 0;
+	virtual bool Open(std::string_view dev_id, AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) = 0;

 	// Reset backend state. Blocks until data callback returns.
 	virtual void Close() = 0;

 	// Sets write callback. It's called when backend requests new data to be sent.
 	// Callback should return number of submitted bytes. Calling other backend functions from callback is unsafe.
-	virtual void SetWriteCallback(std::function<u32(u32 /* byte_cnt */, void* /* buffer */)> cb) = 0;
+	virtual void SetWriteCallback(std::function<u32(u32 /* byte_cnt */, void* /* buffer */)> cb);

-	// Sets error callback. It's called when backend detects uncorrectable error condition in audio chain.
+	// Sets error callback. It's called when backend detects event in audio chain that needs immediate attention.
 	// Calling other backend functions from callback is unsafe.
-	virtual void SetErrorCallback(std::function<void()> cb);
+	virtual void SetStateCallback(std::function<void(AudioStateEvent)> cb);

 	/*
 	 * All functions below require that Open() was called prior.
@@ -101,6 +109,11 @@ public:
 	 */
 	virtual bool Operational() { return true; }

+	/*
+	 * This virtual method should be reimplemented if backend can report device changes
+	 */
+	virtual bool DefaultDeviceChanged() { return false; }
+
 	/*
 	 * Helper methods
 	 */
@@ -145,6 +158,11 @@ public:
 	 */
 	static AudioChannelCnt get_max_channel_count(u32 device_index);

+	/*
+	 * Converts raw channel count to value usable by backends
+	 */
+	static AudioChannelCnt convert_channel_count(u64 raw);
+
 	/*
 	 * Downmix audio stream.
 	 */
@@ -208,8 +226,11 @@ protected:
 	AudioFreq       m_sampling_rate = AudioFreq::FREQ_48K;
 	AudioChannelCnt m_channels = AudioChannelCnt::STEREO;

-	shared_mutex m_error_cb_mutex{};
-	std::function<void()> m_error_callback{};
+	std::timed_mutex m_cb_mutex{};
+	std::function<u32(u32, void *)> m_write_callback{};
+
+	shared_mutex m_state_cb_mutex{};
+	std::function<void(AudioStateEvent)> m_state_callback{};

 	bool m_playing = false;

@@ -2,6 +2,7 @@

 #include <algorithm>
 #include "util/logs.hpp"
+#include "Emu/Audio/audio_device_enumerator.h"

 #ifdef _WIN32
 #include <Windows.h>
@@ -15,27 +16,46 @@ CubebBackend::CubebBackend()
 {
 #ifdef _WIN32
 	// Cubeb requires COM to be initialized on the thread calling cubeb_init on Windows
-	HRESULT hr = CoInitializeEx(nullptr, COINIT_MULTITHREADED);
-	if (SUCCEEDED(hr))
+	if (HRESULT hr = CoInitializeEx(nullptr, COINIT_MULTITHREADED); SUCCEEDED(hr))
 	{
 		m_com_init_success = true;
 	}
 #endif

-	if (int err = cubeb_init(&m_ctx, "RPCS3", nullptr))
+	cubeb *ctx{};
+	if (int err = cubeb_init(&ctx, "RPCS3", nullptr))
 	{
-		Cubeb.error("cubeb_init() failed: 0x%08x", err);
-		m_ctx = nullptr;
+		Cubeb.error("cubeb_init() failed: %i", err);
 		return;
 	}

-	Cubeb.notice("Using backend %s", cubeb_get_backend_id(m_ctx));
+	if (int err = cubeb_register_device_collection_changed(ctx, CUBEB_DEVICE_TYPE_OUTPUT, device_collection_changed_cb, this))
+	{
+		Cubeb.error("cubeb_register_device_collection_changed() failed: %i", err);
+	}
+	else
+	{
+		m_dev_collection_cb_enabled = true;
+	}
+
+	Cubeb.notice("Using backend %s", cubeb_get_backend_id(ctx));
+
+	std::lock_guard cb_lock{m_state_cb_mutex};
+	m_ctx = ctx;
 }

 CubebBackend::~CubebBackend()
 {
 	Close();

+	if (m_dev_collection_cb_enabled)
+	{
+		if (int err = cubeb_register_device_collection_changed(m_ctx, CUBEB_DEVICE_TYPE_OUTPUT, nullptr, nullptr))
+		{
+			Cubeb.error("cubeb_register_device_collection_changed() failed: %i", err);
+		}
+	}
+
 	if (m_ctx)
 	{
 		cubeb_destroy(m_ctx);
@@ -56,11 +76,27 @@ bool CubebBackend::Initialized()

 bool CubebBackend::Operational()
 {
-	std::lock_guard lock(m_error_cb_mutex);
-	return m_stream != nullptr && !m_reset_req;
+	return m_stream != nullptr && !m_reset_req.observe();
 }

-bool CubebBackend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt)
+bool CubebBackend::DefaultDeviceChanged()
+{
+	if (m_default_device.empty() || m_reset_req.observe())
+	{
+		return false;
+	}
+
+	device_handle device = GetDevice();
+	if (!device.handle)
+	{
+		Cubeb.error("Selected device not found. Trying alternative approach...");
+		device = GetDefaultDeviceAlt(m_sampling_rate, m_sample_size, m_channels);
+	}
+
+	return !device.handle || device.id != m_default_device;
+}
+
+bool CubebBackend::Open(std::string_view dev_id, AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt)
 {
 	if (!Initialized())
 	{
@@ -68,12 +104,49 @@ bool CubebBackend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChanne
 		return false;
 	}

-	std::lock_guard lock(m_cb_mutex);
-	CloseUnlocked();
+	Close();
+	std::lock_guard lock{m_cb_mutex};
+
+	const bool use_default_device = dev_id.empty() || dev_id == audio_device_enumerator::DEFAULT_DEV_ID;
+
+	if (use_default_device) Cubeb.notice("Trying to open default device");
+	else Cubeb.notice("Trying to open device with dev_id='%s'", dev_id);
+
+	device_handle device = GetDevice(use_default_device ? "" : dev_id);
+
+	if (!device.handle)
+	{
+		if (use_default_device)
+		{
+			device = GetDefaultDeviceAlt(freq, sample_size, ch_cnt);
+
+			if (!device.handle)
+			{
+				Cubeb.error("Cannot detect default device. Channel count detection unavailable.");
+			}
+		}
+		else
+		{
+			Cubeb.error("Device with id=%s not found", dev_id);
+			return false;
+		}
+	}
+
+	if (device.ch_cnt == 0)
+	{
+		Cubeb.error("Device reported invalid channel count, using stereo instead");
+		device.ch_cnt = 2;
+	}
+
+	if (use_default_device)
+	{
+		std::lock_guard lock{m_state_cb_mutex};
+		m_default_device = device.id;
+	}

 	m_sampling_rate = freq;
 	m_sample_size = sample_size;
-	m_channels = ch_cnt;
+	m_channels = static_cast<AudioChannelCnt>(std::min(static_cast<u32>(convert_channel_count(device.ch_cnt)), static_cast<u32>(ch_cnt)));
 	full_sample_size = get_channels() * get_sample_size();

 	cubeb_stream_params stream_param{};
@@ -82,70 +155,69 @@ bool CubebBackend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChanne
 	stream_param.channels = get_channels();
 	stream_param.layout = [&]()
 	{
-		switch (ch_cnt)
+		switch (m_channels)
 		{
 		case AudioChannelCnt::STEREO:       return CUBEB_LAYOUT_STEREO;
 		case AudioChannelCnt::SURROUND_5_1: return CUBEB_LAYOUT_3F2_LFE;
 		case AudioChannelCnt::SURROUND_7_1: return CUBEB_LAYOUT_3F4_LFE;
 		default:
-			ensure(false);
-			return CUBEB_LAYOUT_UNDEFINED;
+			fmt::throw_exception("Invalid audio channel count");
 		}
 	}();
+	stream_param.prefs = m_dev_collection_cb_enabled && device.handle ? CUBEB_STREAM_PREF_DISABLE_DEVICE_SWITCHING : CUBEB_STREAM_PREF_NONE;

 	u32 min_latency{};
 	if (int err = cubeb_get_min_latency(m_ctx, &stream_param, &min_latency))
 	{
-		Cubeb.error("cubeb_get_min_latency() failed: 0x%08x", err);
+		Cubeb.error("cubeb_get_min_latency() failed: %i", err);
 		min_latency = 0;
 	}

 	const u32 stream_latency = std::max(static_cast<u32>(AUDIO_MIN_LATENCY * get_sampling_rate()), min_latency);
-	if (int err = cubeb_stream_init(m_ctx, &m_stream, "Main stream", nullptr, nullptr, nullptr, &stream_param, stream_latency, data_cb, state_cb, this))
+
+	if (int err = cubeb_stream_init(m_ctx, &m_stream, "Main stream", nullptr, nullptr, device.handle, &stream_param, stream_latency, data_cb, state_cb, this))
 	{
-		Cubeb.error("cubeb_stream_init() failed: 0x%08x", err);
+		Cubeb.error("cubeb_stream_init() failed: %i", err);
 		m_stream = nullptr;
-	}
-	else if (int err = cubeb_stream_start(m_stream))
-	{
-		Cubeb.error("cubeb_stream_start() failed: 0x%08x", err);
-		CloseUnlocked();
-	}
-	else if (int err = cubeb_stream_set_volume(m_stream, 1.0))
-	{
-		Cubeb.error("cubeb_stream_set_volume() failed: 0x%08x", err);
+		return false;
 	}

-	if (m_stream == nullptr)
+	if (int err = cubeb_stream_start(m_stream))
 	{
-		Cubeb.error("Failed to open audio backend. Make sure that no other application is running that might block audio access (e.g. Netflix).");
+		Cubeb.error("cubeb_stream_start() failed: %i", err);
+		Close();
 		return false;
 	}

+	if (int err = cubeb_stream_set_volume(m_stream, 1.0))
+	{
+		Cubeb.error("cubeb_stream_set_volume() failed: %i", err);
+	}
+
 	return true;
 }

-void CubebBackend::CloseUnlocked()
+void CubebBackend::Close()
 {
 	if (m_stream != nullptr)
 	{
 		if (int err = cubeb_stream_stop(m_stream))
 		{
-			Cubeb.error("cubeb_stream_stop() failed: 0x%08x", err);
+			Cubeb.error("cubeb_stream_stop() failed: %i", err);
 		}

 		cubeb_stream_destroy(m_stream);
-		m_stream = nullptr;
 	}

-	m_playing = false;
-	m_last_sample.fill(0);
-}
+	{
+		std::lock_guard lock{m_cb_mutex};
+		m_stream = nullptr;
+		m_playing = false;
+		m_last_sample.fill(0);
+	}

-void CubebBackend::Close()
-{
-	std::lock_guard lock(m_cb_mutex);
-	CloseUnlocked();
+	std::lock_guard lock{m_state_cb_mutex};
+	m_default_device.clear();
 }

 void CubebBackend::Play()
@@ -177,12 +249,6 @@ void CubebBackend::Pause()
 	m_last_sample.fill(0);
 }

-void CubebBackend::SetWriteCallback(std::function<u32(u32, void *)> cb)
-{
-	std::lock_guard lock(m_cb_mutex);
-	m_write_callback = cb;
-}
-
 f64 CubebBackend::GetCallbackFrameLen()
 {
 	if (m_stream == nullptr)
@@ -194,20 +260,187 @@ f64 CubebBackend::GetCallbackFrameLen()
 	u32 stream_latency{};
 	if (int err = cubeb_stream_get_latency(m_stream, &stream_latency))
 	{
-		Cubeb.error("cubeb_stream_get_latency() failed: 0x%08x", err);
+		Cubeb.error("cubeb_stream_get_latency() failed: %i", err);
 		stream_latency = 0;
 	}

 	return std::max<f64>(AUDIO_MIN_LATENCY, static_cast<f64>(stream_latency) / get_sampling_rate());
 }

-long CubebBackend::data_cb(cubeb_stream* /* stream */, void* user_ptr, void const* /* input_buffer */, void* output_buffer, long nframes)
+CubebBackend::device_handle CubebBackend::GetDevice(std::string_view dev_id)
 {
+	const bool default_dev = dev_id.empty();
+
+	if (default_dev) Cubeb.notice("Searching for default device");
+	else Cubeb.notice("Searching for device with dev_id='%s'", dev_id);
+
+	cubeb_device_collection dev_collection{};
+	if (int err = cubeb_enumerate_devices(m_ctx, CUBEB_DEVICE_TYPE_OUTPUT, &dev_collection))
+	{
+		Cubeb.error("cubeb_enumerate_devices() failed: %i", err);
+		return {};
+	}
+
+	if (dev_collection.count == 0)
+	{
+		Cubeb.error("No output devices available");
+		if (int err = cubeb_device_collection_destroy(m_ctx, &dev_collection))
+		{
+			Cubeb.error("cubeb_device_collection_destroy() failed: %i", err);
+		}
+		return {};
+	}
+
+	Cubeb.notice("Found %d possible output devices", dev_collection.count);
+
+	device_handle result{};
+
+	for (u64 dev_idx = 0; dev_idx < dev_collection.count; dev_idx++)
+	{
+		const cubeb_device_info& dev_info = dev_collection.device[dev_idx];
+
+		device_handle device{};
+		device.handle = dev_info.devid;
+		device.ch_cnt = dev_info.max_channels;
+
+		if (dev_info.device_id)
+		{
+			device.id = dev_info.device_id;
+		}
+
+		if (device.id.empty())
+		{
+			Cubeb.error("device_id is missing from device");
+			continue;
+		}
+
+		if (default_dev)
+		{
+			if (dev_info.preferred & CUBEB_DEVICE_PREF_MULTIMEDIA)
+			{
+				result = std::move(device);
+				break;
+			}
+		}
+		else if (device.id == dev_id)
+		{
+			result = std::move(device);
+			break;
+		}
+	}
+
+	if (result.handle)
+	{
+		Cubeb.notice("Found device '%s' with %d channels", result.id, result.ch_cnt);
+	}
+	else
+	{
+		Cubeb.notice("No device found for dev_id='%s'", dev_id);
+	}
+
+	if (int err = cubeb_device_collection_destroy(m_ctx, &dev_collection))
+	{
+		Cubeb.error("cubeb_device_collection_destroy() failed: %i", err);
+	}
+
+	return result;
+};
+
+CubebBackend::device_handle CubebBackend::GetDefaultDeviceAlt(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt)
+{
+	Cubeb.notice("Starting alternative search for default device with freq=%d, sample_size=%d and ch_cnt=%d", static_cast<u32>(freq), static_cast<u32>(sample_size), static_cast<u32>(ch_cnt));
+
+	cubeb_stream_params param =
+	{
+		.format   = sample_size == AudioSampleSize::S16 ? CUBEB_SAMPLE_S16NE : CUBEB_SAMPLE_FLOAT32NE,
+		.rate     = static_cast<u32>(freq),
+		.channels = static_cast<u32>(ch_cnt),
+		.layout   = CUBEB_LAYOUT_UNDEFINED,
+		.prefs    = CUBEB_STREAM_PREF_DISABLE_DEVICE_SWITCHING
+	};
+
+	u32 min_latency{};
+	if (int err = cubeb_get_min_latency(m_ctx, &param, &min_latency))
+	{
+		Cubeb.error("cubeb_get_min_latency() failed: %i", err);
+		min_latency = 100;
+	}
+
+	cubeb_stream* tmp_stream{};
+	static auto dummy_data_cb = [](cubeb_stream*, void*, void const*, void*, long) -> long { return 0; };
+	static auto dummy_state_cb = [](cubeb_stream*, void*, cubeb_state) {};
+
+	if (int err = cubeb_stream_init(m_ctx, &tmp_stream, "Default device detector", nullptr, nullptr, nullptr, &param, min_latency, dummy_data_cb, dummy_state_cb, nullptr))
+	{
+		Cubeb.error("cubeb_stream_init() failed: %i", err);
+		return {};
+	}
+
+	cubeb_device* crnt_dev{};
+
+	if (int err = cubeb_stream_get_current_device(tmp_stream, &crnt_dev); err != CUBEB_OK || !crnt_dev)
+	{
+		Cubeb.error("cubeb_stream_get_current_device() failed: err=%i, crnt_dev=%d", err, !!crnt_dev);
+		cubeb_stream_destroy(tmp_stream);
+		return {};
+	}
+
+	std::string out_dev_name;
+
+	if (crnt_dev->output_name)
+	{
+		out_dev_name = crnt_dev->output_name;
+	}
+
+	if (int err = cubeb_stream_device_destroy(tmp_stream, crnt_dev))
+	{
+		Cubeb.error("cubeb_stream_device_destroy() failed: %i", err);
+	}
+
+	cubeb_stream_destroy(tmp_stream);
+
+	if (out_dev_name.empty())
+	{
+		Cubeb.notice("No default device available");
+		return {};
+	}
+
+	return GetDevice(out_dev_name);
+}
+
+long CubebBackend::data_cb(cubeb_stream* stream, void* user_ptr, void const* /* input_buffer */, void* output_buffer, long nframes)
+{
+	if (nframes <= 0)
+	{
+		Cubeb.error("data_cb called with nframes=%d", nframes);
+		return 0;
+	}
+
+	if (!output_buffer)
+	{
+		Cubeb.error("data_cb called with invalid output_buffer");
+		return CUBEB_ERROR;
+	}
+
+	if (!stream)
+	{
+		Cubeb.error("data_cb called with invalid stream");
+		return CUBEB_ERROR;
+	}
+
 	CubebBackend* const cubeb = static_cast<CubebBackend*>(user_ptr);
+	ensure(cubeb);
+
 	std::unique_lock lock(cubeb->m_cb_mutex, std::defer_lock);

-	if (nframes && lock.try_lock() && cubeb->m_write_callback && cubeb->m_playing)
+	if (!cubeb->m_reset_req.observe() && lock.try_lock_for(std::chrono::microseconds{50}) && cubeb->m_write_callback && cubeb->m_playing)
 	{
+		if (stream != cubeb->m_stream)
+		{
+			// Cubeb.error("data_cb called with unknown stream");
+			return CUBEB_ERROR;
+		}
+
 		const u32 sample_size = cubeb->full_sample_size.observe();
 		const u32 bytes_req = nframes * sample_size;
 		u32 written = std::min(cubeb->m_write_callback(bytes_req, output_buffer), bytes_req);
@@ -227,26 +460,102 @@ long CubebBackend::data_cb(cubeb_stream* /* stream */, void* user_ptr, void cons
 	{
 		// Stream parameters are modified only after stream_destroy. stream_destroy will return
 		// only after this callback returns, so it's safe to access full_sample_size here.
-		memset(output_buffer, 0, nframes * cubeb->full_sample_size.observe());
+		memset(output_buffer, 0, nframes * cubeb->full_sample_size);
 	}

 	return nframes;
 }

-void CubebBackend::state_cb(cubeb_stream* /* stream */, void* user_ptr, cubeb_state state)
+void CubebBackend::state_cb(cubeb_stream* stream, void* user_ptr, cubeb_state state)
 {
-	CubebBackend* const cubeb = static_cast<CubebBackend*>(user_ptr);
+	if (!stream)
+	{
+		Cubeb.error("state_cb called with invalid stream");
+		return;
+	}

-	if (state == CUBEB_STATE_ERROR)
+	CubebBackend* const cubeb = static_cast<CubebBackend*>(user_ptr);
+	ensure(cubeb);
+
+	std::lock_guard lock(cubeb->m_state_cb_mutex);
+
+	if (stream != cubeb->m_stream)
+	{
+		Cubeb.error("state_cb called with unknown stream");
+		return;
+	}
+
+	switch (state)
+	{
+	case CUBEB_STATE_ERROR:
 	{
 		Cubeb.error("Stream entered error state");

-		std::lock_guard lock(cubeb->m_error_cb_mutex);
-		cubeb->m_reset_req = true;
-
-		if (cubeb->m_error_callback)
+		if (!cubeb->m_reset_req.test_and_set() && cubeb->m_state_callback)
 		{
-			cubeb->m_error_callback();
+			cubeb->m_state_callback(AudioStateEvent::UNSPECIFIED_ERROR);
 		}
+
+		break;
+	}
+	case CUBEB_STATE_STARTED:
+	{
+		Cubeb.notice("Stream started");
+		break;
+	}
+	case CUBEB_STATE_STOPPED:
+	{
+		Cubeb.notice("Stream stopped");
+		break;
+	}
+	case CUBEB_STATE_DRAINED:
+	{
+		Cubeb.notice("Stream drained");
+		break;
+	}
+	default:
+	{
+		Cubeb.notice("Stream entered unknown state %d", static_cast<u32>(state));
+		break;
+	}
+	}
+}
+
+void CubebBackend::device_collection_changed_cb(cubeb* context, void* user_ptr)
+{
+	Cubeb.notice("Device collection changed");
+
+	if (!context)
+	{
+		Cubeb.error("device_collection_changed_cb called with invalid stream");
+		return;
+	}
+
+	CubebBackend* const cubeb = static_cast<CubebBackend*>(user_ptr);
+	ensure(cubeb);
+
+	if (cubeb->m_reset_req.observe())
+	{
+		return;
+	}
+
+	std::lock_guard cb_lock{cubeb->m_state_cb_mutex};
+
+	if (context != cubeb->m_ctx)
+	{
+		Cubeb.error("device_collection_changed_cb called with unkown context");
+		return;
+	}
+
+	// Non default device is used (or default device cannot be detected)
+	if (cubeb->m_default_device.empty())
+	{
+		Cubeb.notice("Skipping default device notification");
+		return;
+	}
+
+	if (cubeb->m_state_callback)
+	{
+		cubeb->m_state_callback(AudioStateEvent::DEFAULT_DEVICE_MAYBE_CHANGED);
 	}
 }
@@ -20,11 +20,11 @@ public:

 	bool Initialized() override;
 	bool Operational() override;
+	bool DefaultDeviceChanged() override;

-	bool Open(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) override;
+	bool Open(std::string_view dev_id, AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) override;
 	void Close() override;

-	void SetWriteCallback(std::function<u32(u32, void *)> cb) override;
 	f64 GetCallbackFrameLen() override;

 	void Play() override;
@@ -39,16 +39,28 @@ private:
 	bool m_com_init_success = false;
 #endif

-	shared_mutex m_cb_mutex{};
-	std::function<u32(u32, void *)> m_write_callback{};
 	std::array<u8, sizeof(float) * static_cast<u32>(AudioChannelCnt::SURROUND_7_1)> m_last_sample{};
 	atomic_t<u8> full_sample_size = 0;

-	bool m_reset_req = false;
+	atomic_t<bool> m_reset_req = false;
+
+	// Protected by callback mutex
+	std::string m_default_device{};
+
+	bool m_dev_collection_cb_enabled = false;

 	// Cubeb callbacks
 	static long data_cb(cubeb_stream* stream, void* user_ptr, void const* input_buffer, void* output_buffer, long nframes);
 	static void state_cb(cubeb_stream* stream, void* user_ptr, cubeb_state state);
+	static void device_collection_changed_cb(cubeb* context, void* user_ptr);

-	void CloseUnlocked();
+	struct device_handle
+	{
+		cubeb_devid handle{};
+		std::string id;
+		u32 ch_cnt{};
+	};
+
+	device_handle GetDevice(std::string_view dev_id = "");
+	device_handle GetDefaultDeviceAlt(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt);
 };
@@ -0,0 +1,113 @@
+#include "Emu/Audio/Cubeb/cubeb_enumerator.h"
+#include "util/logs.hpp"
+#include <algorithm>
+
+#ifdef _WIN32
+#include <Windows.h>
+#include <system_error>
+#endif
+
+LOG_CHANNEL(cubeb_dev_enum);
+
+cubeb_enumerator::cubeb_enumerator() : audio_device_enumerator()
+{
+#ifdef _WIN32
+	// Cubeb requires COM to be initialized on the thread calling cubeb_init on Windows
+	HRESULT hr = CoInitializeEx(nullptr, COINIT_MULTITHREADED);
+	if (SUCCEEDED(hr))
+	{
+		com_init_success = true;
+	}
+#endif
+
+	if (int err = cubeb_init(&ctx, "RPCS3 device enumeration", nullptr))
+	{
+		cubeb_dev_enum.error("cubeb_init() failed: %i", err);
+		ctx = nullptr;
+	}
+}
+
+cubeb_enumerator::~cubeb_enumerator()
+{
+	if (ctx)
+	{
+		cubeb_destroy(ctx);
+	}
+
+#ifdef _WIN32
+	if (com_init_success)
+	{
+		CoUninitialize();
+	}
+#endif
+}
+
+std::vector<audio_device_enumerator::audio_device> cubeb_enumerator::get_output_devices()
+{
+	if (ctx == nullptr)
+	{
+		return {};
+	}
+
+	cubeb_device_collection dev_collection{};
+	if (int err = cubeb_enumerate_devices(ctx, CUBEB_DEVICE_TYPE_OUTPUT, &dev_collection))
+	{
+		cubeb_dev_enum.error("cubeb_enumerate_devices() failed: %i", err);
+		return {};
+	}
+
+	if (dev_collection.count == 0)
+	{
+		cubeb_dev_enum.error("No output devices available");
+		if (int err = cubeb_device_collection_destroy(ctx, &dev_collection))
+		{
+			cubeb_dev_enum.error("cubeb_device_collection_destroy() failed: %i", err);
+		}
+		return {};
+	}
+
+	std::vector<audio_device> device_list{};
+
+	for (u64 dev_idx = 0; dev_idx < dev_collection.count; dev_idx++)
+	{
+		const cubeb_device_info& dev_info = dev_collection.device[dev_idx];
+
+		if (dev_info.state == CUBEB_DEVICE_STATE_UNPLUGGED)
+		{
+			continue;
+		}
+
+		audio_device dev =
+		{
+			.id = std::string{dev_info.device_id},
+			.name = std::string{dev_info.friendly_name},
+			.max_ch = dev_info.max_channels
+		};
+
+		if (dev.id.empty())
+		{
+			cubeb_dev_enum.warning("Empty device id - skipping");
+			continue;
+		}
+
+		if (dev.name.empty())
+		{
+			dev.name = dev.id;
+		}
+
+		cubeb_dev_enum.notice("Found device: id=%s, name=%s, max_ch=%d", dev.id, dev.name, dev.max_ch);
+		device_list.emplace_back(dev);
+	}
+
+	if (int err = cubeb_device_collection_destroy(ctx, &dev_collection))
+	{
+		cubeb_dev_enum.error("cubeb_device_collection_destroy() failed: %i", err);
+	}
+
+	std::sort(device_list.begin(), device_list.end(), [](audio_device_enumerator::audio_device a, audio_device_enumerator::audio_device b)
+	{
+		return a.name < b.name;
+	});
+
+	return device_list;
+}
@@ -0,0 +1,22 @@
+#pragma once
+
+#include "Emu/Audio/audio_device_enumerator.h"
+
+#include "cubeb/cubeb.h"
+
+class cubeb_enumerator final : public audio_device_enumerator
+{
+public:
+
+	cubeb_enumerator();
+	~cubeb_enumerator() override;
+
+	std::vector<audio_device> get_output_devices() override;
+
+private:
+
+	cubeb* ctx{};
+#ifdef _WIN32
+	bool com_init_success = false;
+#endif
+};
@@ -6,6 +6,8 @@
 #include "FAudioBackend.h"
 #include "Emu/system_config.h"
 #include "Emu/System.h"
+#include "Emu/Audio/audio_device_enumerator.h"
+#include "Utilities/StrUtil.h"

 LOG_CHANNEL(FAudio_, "FAudio");

@@ -117,11 +119,10 @@ bool FAudioBackend::Initialized()

 bool FAudioBackend::Operational()
 {
-	std::lock_guard lock(m_error_cb_mutex);
-	return m_source_voice != nullptr && !m_reset_req;
+	return m_source_voice != nullptr && !m_reset_req.observe();
 }

-bool FAudioBackend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt)
+bool FAudioBackend::Open(std::string_view dev_id, AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt)
 {
 	if (!Initialized())
 	{
@@ -132,9 +133,37 @@ bool FAudioBackend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChann
 	std::lock_guard lock(m_cb_mutex);
 	CloseUnlocked();

+	const bool use_default_dev = dev_id.empty() || dev_id == audio_device_enumerator::DEFAULT_DEV_ID;
+	u64 devid{};
+	if (!use_default_dev)
+	{
+		if (!try_to_uint64(&devid, dev_id, 0, UINT32_MAX))
+		{
+			FAudio_.error("Invalid device id - %s", dev_id);
+			return false;
+		}
+	}
+
+	if (u32 res = FAudio_CreateMasteringVoice(m_instance, &m_master_voice, FAUDIO_DEFAULT_CHANNELS, FAUDIO_DEFAULT_SAMPLERATE, 0, static_cast<u32>(devid), nullptr))
+	{
+		FAudio_.error("FAudio_CreateMasteringVoice() failed(0x%08x)", res);
+		m_master_voice = nullptr;
+		return false;
+	}
+
+	FAudioVoiceDetails vd{};
+	FAudioVoice_GetVoiceDetails(m_master_voice, &vd);
+
+	if (vd.InputChannels == 0)
+	{
+		FAudio_.error("Channel count of 0 is invalid");
+		CloseUnlocked();
+		return false;
+	}
+
 	m_sampling_rate = freq;
 	m_sample_size = sample_size;
-	m_channels = ch_cnt;
+	m_channels = static_cast<AudioChannelCnt>(std::min(static_cast<u32>(convert_channel_count(vd.InputChannels)), static_cast<u32>(ch_cnt)));;

 	FAudioWaveFormatEx waveformatex;
 	waveformatex.wFormatTag = get_convert_to_s16() ? FAUDIO_FORMAT_PCM : FAUDIO_FORMAT_IEEE_FLOAT;
@@ -153,43 +182,30 @@ bool FAudioBackend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChann
 	OnLoopEnd = nullptr;
 	OnVoiceError = nullptr;

-	if (u32 res = FAudio_CreateMasteringVoice(m_instance, &m_master_voice, FAUDIO_DEFAULT_CHANNELS, FAUDIO_DEFAULT_SAMPLERATE, 0, 0, nullptr))
-	{
-		FAudio_.error("FAudio_CreateMasteringVoice() failed(0x%08x)", res);
-		m_master_voice = nullptr;
-	}
-	else if (u32 res = FAudio_CreateSourceVoice(m_instance, &m_source_voice, &waveformatex, 0, FAUDIO_DEFAULT_FREQ_RATIO, this, nullptr, nullptr))
+	if (u32 res = FAudio_CreateSourceVoice(m_instance, &m_source_voice, &waveformatex, 0, FAUDIO_DEFAULT_FREQ_RATIO, this, nullptr, nullptr))
 	{
 		FAudio_.error("FAudio_CreateSourceVoice() failed(0x%08x)", res);
 		CloseUnlocked();
+		return false;
 	}
-	else if (u32 res = FAudioSourceVoice_Start(m_source_voice, 0, FAUDIO_COMMIT_NOW))
+
+	if (u32 res = FAudioSourceVoice_Start(m_source_voice, 0, FAUDIO_COMMIT_NOW))
 	{
 		FAudio_.error("FAudioSourceVoice_Start() failed(0x%08x)", res);
 		CloseUnlocked();
+		return false;
 	}
-	else if (u32 res = FAudioVoice_SetVolume(m_source_voice, 1.0f, FAUDIO_COMMIT_NOW))
+
+	if (u32 res = FAudioVoice_SetVolume(m_source_voice, 1.0f, FAUDIO_COMMIT_NOW))
 	{
 		FAudio_.error("FAudioVoice_SetVolume() failed(0x%08x)", res);
 	}

-	if (m_source_voice == nullptr)
-	{
-		FAudio_.error("Failed to open audio backend. Make sure that no other application is running that might block audio access (e.g. Netflix).");
-		return false;
-	}
-
-	m_data_buf.resize(get_sampling_rate() * get_sample_size() * get_channels() * INTERNAL_BUF_SIZE_MS * static_cast<u32>(FAUDIO_DEFAULT_FREQ_RATIO) / 1000);
+	m_data_buf.resize(get_sampling_rate() * get_sample_size() * get_channels() * INTERNAL_BUF_SIZE_MS / 1000);

 	return true;
 }

-void FAudioBackend::SetWriteCallback(std::function<u32(u32, void *)> cb)
-{
-	std::lock_guard lock(m_cb_mutex);
-	m_write_callback = cb;
-}
-
 f64 FAudioBackend::GetCallbackFrameLen()
 {
 	constexpr f64 _10ms = 0.01;
@@ -218,7 +234,7 @@ void FAudioBackend::OnVoiceProcessingPassStart_func(FAudioVoiceCallback *cb_obj,
 	FAudioBackend *faudio = static_cast<FAudioBackend *>(cb_obj);

 	std::unique_lock lock(faudio->m_cb_mutex, std::defer_lock);
-	if (BytesRequired && lock.try_lock() && faudio->m_write_callback && faudio->m_playing)
+	if (BytesRequired && !faudio->m_reset_req.observe() && lock.try_lock_for(std::chrono::microseconds{50}) && faudio->m_write_callback && faudio->m_playing)
 	{
 		ensure(BytesRequired <= faudio->m_data_buf.size(), "FAudio internal buffer is too small. Report to developers!");

@@ -250,11 +266,10 @@ void FAudioBackend::OnCriticalError_func(FAudioEngineCallback *cb_obj, u32 Error

 	FAudioBackend *faudio = static_cast<FAudioBackend *>(cb_obj);

-	std::lock_guard lock(faudio->m_error_cb_mutex);
-	faudio->m_reset_req = true;
+	std::lock_guard lock(faudio->m_state_cb_mutex);

-	if (faudio->m_error_callback)
+	if (!faudio->m_reset_req.test_and_set() && faudio->m_state_callback)
 	{
-		faudio->m_error_callback();
+		faudio->m_state_callback(AudioStateEvent::UNSPECIFIED_ERROR);
 	}
 }
@@ -24,10 +24,9 @@ public:
 	bool Initialized() override;
 	bool Operational() override;

-	bool Open(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) override;
+	bool Open(std::string_view dev_id, AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) override;
 	void Close() override;

-	void SetWriteCallback(std::function<u32(u32, void *)> cb) override;
 	f64 GetCallbackFrameLen() override;

 	void Play() override;
@@ -40,12 +39,10 @@ private:
 	FAudioMasteringVoice* m_master_voice{};
 	FAudioSourceVoice* m_source_voice{};

-	shared_mutex m_cb_mutex{};
-	std::function<u32(u32, void *)> m_write_callback{};
 	std::vector<u8> m_data_buf{};
 	std::array<u8, sizeof(float) * static_cast<u32>(AudioChannelCnt::SURROUND_7_1)> m_last_sample{};

-	bool m_reset_req = false;
+	atomic_t<bool> m_reset_req = false;

 	// FAudio voice callbacks
 	static void OnVoiceProcessingPassStart_func(FAudioVoiceCallback *cb_obj, u32 BytesRequired);
@@ -0,0 +1,90 @@
+#ifndef HAVE_FAUDIO
+#error "FAudio support disabled but still being built."
+#endif
+
+#include "Emu/Audio/FAudio/faudio_enumerator.h"
+#include <array>
+#include <algorithm>
+#include "Utilities/StrUtil.h"
+#include "util/logs.hpp"
+
+LOG_CHANNEL(faudio_dev_enum);
+
+faudio_enumerator::faudio_enumerator() : audio_device_enumerator()
+{
+	FAudio *tmp{};
+
+	if (u32 res = FAudioCreate(&tmp, 0, FAUDIO_DEFAULT_PROCESSOR))
+	{
+		faudio_dev_enum.error("FAudioCreate() failed(0x%08x)", res);
+		return;
+	}
+
+	// All succeeded, "commit"
+	instance = tmp;
+}
+
+faudio_enumerator::~faudio_enumerator()
+{
+	if (instance != nullptr)
+	{
+		FAudio_StopEngine(instance);
+		FAudio_Release(instance);
+	}
+}
+
+std::vector<audio_device_enumerator::audio_device> faudio_enumerator::get_output_devices()
+{
+	if (!instance)
+	{
+		return {};
+	}
+
+	u32 dev_cnt{};
+	if (u32 res = FAudio_GetDeviceCount(instance, &dev_cnt))
+	{
+		faudio_dev_enum.error("FAudio_GetDeviceCount() failed(0x%08x)", res);
+		return {};
+	}
+
+	if (dev_cnt == 0)
+	{
+		faudio_dev_enum.warning("No devices available");
+		return {};
+	}
+
+	std::vector<audio_device> device_list{};
+
+	for (u32 dev_idx = 0; dev_idx < dev_cnt; dev_idx++)
+	{
+		FAudioDeviceDetails dev_info{};
+
+		if (u32 res = FAudio_GetDeviceDetails(instance, dev_idx, &dev_info))
+		{
+			faudio_dev_enum.error("FAudio_GetDeviceDetails() failed(0x%08x)", res);
+			continue;
+		}
+
+		audio_device dev =
+		{
+			.id = std::to_string(dev_idx),
+			.name = utf16_to_utf8(std::bit_cast<char16_t*>(&dev_info.DisplayName[0])),
+			.max_ch = dev_info.OutputFormat.Format.nChannels
+		};
+
+		if (dev.name.empty())
+		{
+			dev.name = "Device " + dev.id;
+		}
+
+		faudio_dev_enum.notice("Found device: id=%s, name=%s, max_ch=%d", dev.id, dev.name, dev.max_ch);
+		device_list.emplace_back(dev);
+	}
+
+	std::sort(device_list.begin(), device_list.end(), [](audio_device_enumerator::audio_device a, audio_device_enumerator::audio_device b)
+	{
+		return a.name < b.name;
+	});
+
+	return device_list;
+}
@@ -0,0 +1,22 @@
+#pragma once
+
+#ifndef HAVE_FAUDIO
+#error "FAudio support disabled but still being built."
+#endif
+
+#include "Emu/Audio/audio_device_enumerator.h"
+#include "FAudio.h"
+
+class faudio_enumerator final : public audio_device_enumerator
+{
+public:
+
+	faudio_enumerator();
+	~faudio_enumerator() override;
+
+	std::vector<audio_device> get_output_devices() override;
+
+private:
+
+	FAudio* instance{};
+};
@@ -10,18 +10,15 @@ public:

 	std::string_view GetName() const override { return "Null"sv; }

-	bool Open(AudioFreq /* freq */, AudioSampleSize /* sample_size */, AudioChannelCnt /* ch_cnt */) override
+	bool Open(std::string_view /* dev_id */, AudioFreq /* freq */, AudioSampleSize /* sample_size */, AudioChannelCnt /* ch_cnt */) override
 	{
 		Close();
 		return true;
 	}
 	void Close() override { m_playing = false; }

-	void SetWriteCallback(std::function<u32(u32, void *)> /* cb */) override {};
 	f64 GetCallbackFrameLen() override { return 0.01; };

-	void SetErrorCallback(std::function<void()> /* cb */) override {};
-
 	void Play() override { m_playing = true; }
 	void Pause() override { m_playing = false; }
 	bool IsPlaying() override { return m_playing; }
@@ -0,0 +1,13 @@
+#pragma once
+
+#include "Emu/Audio/audio_device_enumerator.h"
+
+class null_enumerator final : public audio_device_enumerator
+{
+public:
+
+	null_enumerator() {};
+	~null_enumerator() override {};
+
+	std::vector<audio_device> get_output_devices() override { return {}; }
+};
@@ -5,6 +5,8 @@
 #include <algorithm>
 #include "util/logs.hpp"
 #include "Emu/System.h"
+#include "Emu/Audio/audio_device_enumerator.h"
+#include "Utilities/StrUtil.h"

 #include "XAudio2Backend.h"
 #include <Windows.h>
@@ -14,42 +16,85 @@

 LOG_CHANNEL(XAudio);

+template <>
+void fmt_class_string<ERole>::format(std::string& out, u64 arg)
+{
+	format_enum(out, arg, [](auto value)
+	{
+		switch (value)
+		{
+		case eConsole: return "eConsole";
+		case eMultimedia: return "eMultimedia";
+		case eCommunications: return "eCommunications";
+		}
+
+		return unknown;
+	});
+}
+
+template <>
+void fmt_class_string<EDataFlow>::format(std::string& out, u64 arg)
+{
+	format_enum(out, arg, [](auto value)
+	{
+		switch (value)
+		{
+		case eRender: return "eRender";
+		case eCapture: return "eCapture";
+		case eAll: return "eAll";
+		}
+
+		return unknown;
+	});
+}
+
 XAudio2Backend::XAudio2Backend()
 	: AudioBackend()
 {
-	Microsoft::WRL::ComPtr<IXAudio2> instance;
+	Microsoft::WRL::ComPtr<IXAudio2> instance{};
+	Microsoft::WRL::ComPtr<IMMDeviceEnumerator> enumerator{};

 	// In order to prevent errors on CreateMasteringVoice, apparently we need CoInitializeEx according to:
 	// https://docs.microsoft.com/en-us/windows/win32/api/xaudio2fx/nf-xaudio2fx-xaudio2createvolumemeter
-	HRESULT hr = CoInitializeEx(nullptr, COINIT_MULTITHREADED);
-	if (SUCCEEDED(hr))
+	if (HRESULT hr = CoInitializeEx(nullptr, COINIT_MULTITHREADED); SUCCEEDED(hr))
 	{
 		m_com_init_success = true;
 	}

-	hr = XAudio2Create(instance.GetAddressOf(), 0, XAUDIO2_USE_DEFAULT_PROCESSOR);
-	if (FAILED(hr))
+	if (HRESULT hr = XAudio2Create(instance.GetAddressOf(), 0, XAUDIO2_USE_DEFAULT_PROCESSOR); FAILED(hr))
 	{
 		XAudio.error("XAudio2Create() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
 		return;
 	}

-
-	hr = instance->RegisterForCallbacks(this);
-	if (FAILED(hr))
+	if (HRESULT hr = instance->RegisterForCallbacks(this); FAILED(hr))
 	{
 		// Some error recovery functionality will be lost, but otherwise backend is operational
 		XAudio.error("RegisterForCallbacks() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
 	}

+	// Try to register a listener for device changes
+	if (HRESULT hr = CoCreateInstance(__uuidof(MMDeviceEnumerator), nullptr, CLSCTX_INPROC_SERVER, IID_PPV_ARGS(enumerator.GetAddressOf())); FAILED(hr))
+	{
+		XAudio.error("CoCreateInstance() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+		return;
+	}
+
 	// All succeeded, "commit"
 	m_xaudio2_instance = std::move(instance);
+	m_device_enumerator = std::move(enumerator);
 }

 XAudio2Backend::~XAudio2Backend()
 {
 	Close();

+	if (m_device_enumerator != nullptr)
+	{
+		m_device_enumerator->UnregisterEndpointNotificationCallback(this);
+		m_device_enumerator = nullptr;
+	}
+
 	if (m_xaudio2_instance != nullptr)
 	{
 		m_xaudio2_instance->StopEngine();
@@ -69,14 +114,13 @@ bool XAudio2Backend::Initialized()

 bool XAudio2Backend::Operational()
 {
-	std::lock_guard lock(m_error_cb_mutex);
+	return m_source_voice != nullptr && !m_reset_req.observe();
+}

-	if (m_dev_listener.output_device_changed())
-	{
-		m_reset_req = true;
-	}
-
-	return m_source_voice != nullptr && !m_reset_req;
+bool XAudio2Backend::DefaultDeviceChanged()
+{
+	std::lock_guard lock{m_state_cb_mutex};
+	return !m_reset_req.observe() && m_default_dev_changed;
 }

 void XAudio2Backend::Play()
@@ -112,8 +156,14 @@ void XAudio2Backend::CloseUnlocked()
 		m_master_voice = nullptr;
 	}

+	m_device_enumerator->UnregisterEndpointNotificationCallback(this);
+
 	m_playing = false;
 	m_last_sample.fill(0);
+
+	std::lock_guard lock(m_state_cb_mutex);
+	m_default_dev_changed = false;
+	m_current_device.clear();
 }

 void XAudio2Backend::Close()
@@ -144,7 +194,7 @@ void XAudio2Backend::Pause()
 	}
 }

-bool XAudio2Backend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt)
+bool XAudio2Backend::Open(std::string_view dev_id, AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt)
 {
 	if (!Initialized())
 	{
@@ -155,9 +205,54 @@ bool XAudio2Backend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChan
 	std::lock_guard lock(m_cb_mutex);
 	CloseUnlocked();

+	const bool use_default_device = dev_id.empty() || dev_id == audio_device_enumerator::DEFAULT_DEV_ID;
+	std::string selected_dev_id{};
+
+	if (use_default_device)
+	{
+		Microsoft::WRL::ComPtr<IMMDevice> default_dev{};
+		if (HRESULT hr = m_device_enumerator->GetDefaultAudioEndpoint(eRender, eConsole, default_dev.GetAddressOf()); FAILED(hr))
+		{
+			XAudio.error("GetDefaultAudioEndpoint() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+			return false;
+		}
+
+		LPWSTR default_id{};
+		if (HRESULT hr = default_dev->GetId(&default_id); FAILED(hr))
+		{
+			XAudio.error("GetId() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+			return false;
+		}
+
+		selected_dev_id = wchar_to_utf8(std::wstring_view{default_id});
+		CoTaskMemFree(default_id);
+		if (selected_dev_id.empty())
+		{
+			XAudio.error("Default device id is empty");
+			return false;
+		}
+	}
+
+	if (HRESULT hr = m_xaudio2_instance->CreateMasteringVoice(&m_master_voice, 0, 0, 0, utf8_to_wchar(use_default_device ? selected_dev_id : dev_id).c_str()); FAILED(hr))
+	{
+		XAudio.error("CreateMasteringVoice() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+		m_master_voice = nullptr;
+		return false;
+	}
+
+	XAUDIO2_VOICE_DETAILS vd{};
+	m_master_voice->GetVoiceDetails(&vd);
+
+	if (vd.InputChannels == 0)
+	{
+		XAudio.error("Channel count 0 is invalid");
+		CloseUnlocked();
+		return false;
+	}
+
 	m_sampling_rate = freq;
 	m_sample_size = sample_size;
-	m_channels = ch_cnt;
+	m_channels = static_cast<AudioChannelCnt>(std::min(static_cast<u32>(convert_channel_count(vd.InputChannels)), static_cast<u32>(ch_cnt)));

 	WAVEFORMATEX waveformatex{};
 	waveformatex.wFormatTag = get_convert_to_s16() ? WAVE_FORMAT_PCM : WAVE_FORMAT_IEEE_FLOAT;
@@ -168,65 +263,63 @@ bool XAudio2Backend::Open(AudioFreq freq, AudioSampleSize sample_size, AudioChan
 	waveformatex.wBitsPerSample = get_sample_size() * 8;
 	waveformatex.cbSize = 0;

-	if (HRESULT hr = m_xaudio2_instance->CreateMasteringVoice(&m_master_voice); FAILED(hr))
-	{
-		XAudio.error("CreateMasteringVoice() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
-		m_master_voice = nullptr;
-	}
-	else if (HRESULT hr = m_xaudio2_instance->CreateSourceVoice(&m_source_voice, &waveformatex, 0, XAUDIO2_DEFAULT_FREQ_RATIO, this); FAILED(hr))
+	if (HRESULT hr = m_xaudio2_instance->CreateSourceVoice(&m_source_voice, &waveformatex, 0, XAUDIO2_DEFAULT_FREQ_RATIO, this); FAILED(hr))
 	{
 		XAudio.error("CreateSourceVoice() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
 		CloseUnlocked();
+		return false;
 	}
-	else if (HRESULT hr = m_source_voice->Start(); FAILED(hr))
+
+	if (HRESULT hr = m_source_voice->Start(); FAILED(hr))
 	{
 		XAudio.error("Start() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
 		CloseUnlocked();
+		return false;
 	}
-	else if (HRESULT hr = m_source_voice->SetVolume(1.0f); FAILED(hr))
+
+	if (HRESULT hr = m_device_enumerator->RegisterEndpointNotificationCallback(this); FAILED(hr))
+	{
+		XAudio.error("RegisterEndpointNotificationCallback() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+		CloseUnlocked();
+		return false;
+	}
+
+	if (HRESULT hr = m_source_voice->SetVolume(1.0f); FAILED(hr))
 	{
 		XAudio.error("SetVolume() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
 	}

-	if (m_source_voice == nullptr)
+	m_data_buf.resize(get_sampling_rate() * get_sample_size() * get_channels() * INTERNAL_BUF_SIZE_MS / 1000);
+
+	if (use_default_device)
 	{
-		XAudio.error("Failed to open audio backend. Make sure that no other application is running that might block audio access (e.g. Netflix).");
-		return false;
+		m_current_device = selected_dev_id;
 	}

-	m_data_buf.resize(get_sampling_rate() * get_sample_size() * get_channels() * INTERNAL_BUF_SIZE_MS * static_cast<u32>(XAUDIO2_DEFAULT_FREQ_RATIO) / 1000);
-
 	return true;
 }

-void XAudio2Backend::SetWriteCallback(std::function<u32(u32, void *)> cb)
-{
-	std::lock_guard lock(m_cb_mutex);
-	m_write_callback = cb;
-}
-
 f64 XAudio2Backend::GetCallbackFrameLen()
 {
 	constexpr f64 _10ms = 0.01;

-	if (m_source_voice == nullptr)
+	if (m_xaudio2_instance == nullptr)
 	{
 		XAudio.error("GetCallbackFrameLen() called uninitialized");
 		return _10ms;
 	}

-	void *ext;
+	Microsoft::WRL::ComPtr<IXAudio2Extension> xaudio_ext{};
 	f64 min_latency{};

-	const HRESULT hr = m_xaudio2_instance->QueryInterface(IID_IXAudio2Extension, &ext);
-	if (FAILED(hr))
+	if (HRESULT hr = m_xaudio2_instance->QueryInterface(IID_IXAudio2Extension, std::bit_cast<void**>(xaudio_ext.GetAddressOf())); FAILED(hr))
 	{
 		XAudio.error("QueryInterface() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
 	}
 	else
 	{
 		u32 samples_per_q = 0, freq = 0;
-		static_cast<IXAudio2Extension *>(ext)->GetProcessingQuantum(&samples_per_q, &freq);
+		xaudio_ext->GetProcessingQuantum(&samples_per_q, &freq);

 		if (freq)
 		{
@@ -240,7 +333,7 @@ f64 XAudio2Backend::GetCallbackFrameLen()
 void XAudio2Backend::OnVoiceProcessingPassStart(UINT32 BytesRequired)
 {
 	std::unique_lock lock(m_cb_mutex, std::defer_lock);
-	if (BytesRequired && lock.try_lock() && m_write_callback && m_playing)
+	if (BytesRequired && !m_reset_req.observe() && lock.try_lock_for(std::chrono::microseconds{50}) && m_write_callback && m_playing)
 	{
 		ensure(BytesRequired <= m_data_buf.size(), "XAudio internal buffer is too small. Report to developers!");

@@ -270,11 +363,53 @@ void XAudio2Backend::OnCriticalError(HRESULT Error)
 {
 	XAudio.error("OnCriticalError() called: %s (0x%08x)", std::system_category().message(Error), static_cast<u32>(Error));

-	std::lock_guard lock(m_error_cb_mutex);
-	m_reset_req = true;
+	std::lock_guard lock(m_state_cb_mutex);

-	if (m_error_callback)
+	if (!m_reset_req.test_and_set() && m_state_callback)
 	{
-		m_error_callback();
+		m_state_callback(AudioStateEvent::UNSPECIFIED_ERROR);
 	}
 }
+
+HRESULT XAudio2Backend::OnDefaultDeviceChanged(EDataFlow flow, ERole role, LPCWSTR new_default_device_id)
+{
+	XAudio.notice("OnDefaultDeviceChanged(flow=%s, role=%s, new_default_device_id=0x%x)", flow, role, new_default_device_id);
+
+	if (!new_default_device_id)
+	{
+		XAudio.notice("OnDefaultDeviceChanged(): new_default_device_id empty");
+		return S_OK;
+	}
+
+	// Listen only for one device role, otherwise we're going to receive more than one notification for flow type
+	if (role != eConsole)
+	{
+		XAudio.notice("OnDefaultDeviceChanged(): we don't care about this device");
+		return S_OK;
+	}
+
+	std::lock_guard lock(m_state_cb_mutex);
+
+	// Non default device is used
+	if (m_current_device.empty())
+	{
+		return S_OK;
+	}
+
+	const std::string new_device_id = wchar_to_utf8(std::wstring_view{new_default_device_id});
+
+	if (flow == eRender || flow == eAll)
+	{
+		if (!m_reset_req.observe() && new_device_id != m_current_device)
+		{
+			m_default_dev_changed = true;
+
+			if (m_state_callback)
+			{
+				m_state_callback(AudioStateEvent::DEFAULT_DEVICE_MAYBE_CHANGED);
+			}
+		}
+	}
+
+	return S_OK;
+}
@@ -7,12 +7,12 @@
 #include <memory>
 #include "Utilities/mutex.h"
 #include "Emu/Audio/AudioBackend.h"
-#include "Emu/Audio/audio_device_listener.h"

 #include <xaudio2redist.h>
 #include <wrl/client.h>
+#include <MMDeviceAPI.h>

-class XAudio2Backend final : public AudioBackend, public IXAudio2VoiceCallback, public IXAudio2EngineCallback
+class XAudio2Backend final : public AudioBackend, public IXAudio2VoiceCallback, public IXAudio2EngineCallback, public IMMNotificationClient
 {
 public:
 	XAudio2Backend();
@@ -25,11 +25,11 @@ public:

 	bool Initialized() override;
 	bool Operational() override;
+	bool DefaultDeviceChanged() override;

-	bool Open(AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) override;
+	bool Open(std::string_view dev_id, AudioFreq freq, AudioSampleSize sample_size, AudioChannelCnt ch_cnt) override;
 	void Close() override;

-	void SetWriteCallback(std::function<u32(u32, void *)> cb) override;
 	f64 GetCallbackFrameLen() override;

 	void Play() override;
@@ -43,14 +43,16 @@ private:
 	IXAudio2SourceVoice* m_source_voice{};
 	bool m_com_init_success = false;

-	shared_mutex m_cb_mutex{};
-	std::function<u32(u32, void *)> m_write_callback{};
+	Microsoft::WRL::ComPtr<IMMDeviceEnumerator> m_device_enumerator{};
+
+	// Protected by state callback mutex
+	std::string m_current_device{};
+	bool m_default_dev_changed = false;
+
 	std::vector<u8> m_data_buf{};
 	std::array<u8, sizeof(float) * static_cast<u32>(AudioChannelCnt::SURROUND_7_1)> m_last_sample{};

-	bool m_reset_req = false;
-
-	audio_device_listener m_dev_listener{};
+	atomic_t<bool> m_reset_req = false;

 	// XAudio voice callbacks
 	void OnVoiceProcessingPassStart(UINT32 BytesRequired) override;
@@ -66,5 +68,15 @@ private:
 	void OnProcessingPassEnd() override {};
 	void OnCriticalError(HRESULT Error) override;

+	// IMMNotificationClient callbacks
+	IFACEMETHODIMP_(ULONG) AddRef() override { return 1; };
+	IFACEMETHODIMP_(ULONG) Release() override { return 1; };
+	IFACEMETHODIMP QueryInterface(REFIID /*iid*/, void** /*object*/) override { return E_NOINTERFACE; };
+	IFACEMETHODIMP OnPropertyValueChanged(LPCWSTR /*device_id*/, const PROPERTYKEY /*key*/) override { return S_OK; };
+	IFACEMETHODIMP OnDeviceAdded(LPCWSTR /*device_id*/) override { return S_OK; };
+	IFACEMETHODIMP OnDeviceRemoved(LPCWSTR /*device_id*/) override { return S_OK; };
+	IFACEMETHODIMP OnDeviceStateChanged(LPCWSTR /*device_id*/, DWORD /*new_state*/) override { return S_OK; };
+	IFACEMETHODIMP OnDefaultDeviceChanged(EDataFlow flow, ERole role, LPCWSTR new_default_device_id) override;
+
 	void CloseUnlocked();
 };
@@ -0,0 +1,153 @@
+#ifndef _WIN32
+#error "XAudio2 can only be built on Windows."
+#endif
+
+#include "Emu/Audio/XAudio2/xaudio2_enumerator.h"
+#include "util/logs.hpp"
+#include "Utilities/StrUtil.h"
+#include <algorithm>
+
+#include <wrl/client.h>
+#include <Windows.h>
+#include <system_error>
+#include <mmdeviceapi.h>
+#include <Functiondiscoverykeys_devpkey.h>
+
+LOG_CHANNEL(xaudio_dev_enum);
+
+xaudio2_enumerator::xaudio2_enumerator() : audio_device_enumerator()
+{
+}
+
+xaudio2_enumerator::~xaudio2_enumerator()
+{
+}
+
+std::vector<audio_device_enumerator::audio_device> xaudio2_enumerator::get_output_devices()
+{
+	Microsoft::WRL::ComPtr<IMMDeviceEnumerator> devEnum{};
+	if (HRESULT hr = CoCreateInstance(__uuidof(MMDeviceEnumerator), nullptr, CLSCTX_INPROC_SERVER, IID_PPV_ARGS(devEnum.GetAddressOf())); FAILED(hr))
+	{
+		xaudio_dev_enum.error("CoCreateInstance() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+		return {};
+	}
+
+	Microsoft::WRL::ComPtr<IMMDeviceCollection> devices{};
+	if (HRESULT hr = devEnum->EnumAudioEndpoints(eRender, DEVICE_STATE_ACTIVE | DEVICE_STATE_DISABLED, &devices); FAILED(hr))
+	{
+		xaudio_dev_enum.error("EnumAudioEndpoints() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+		return {};
+	}
+
+	UINT count = 0;
+	if (HRESULT hr = devices->GetCount(&count); FAILED(hr))
+	{
+		xaudio_dev_enum.error("devices->GetCount() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+		return {};
+	}
+
+	if (count == 0)
+	{
+		xaudio_dev_enum.warning("No devices available");
+		return {};
+	}
+
+	std::vector<audio_device> device_list{};
+
+	for (UINT dev_idx = 0; dev_idx < count; dev_idx++)
+	{
+		Microsoft::WRL::ComPtr<IMMDevice> endpoint{};
+		if (HRESULT hr = devices->Item(dev_idx, endpoint.GetAddressOf()); FAILED(hr))
+		{
+			xaudio_dev_enum.error("devices->Item() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+			continue;
+		}
+
+		LPWSTR id = nullptr;
+		if (HRESULT hr = endpoint->GetId(&id); FAILED(hr))
+		{
+			xaudio_dev_enum.error("endpoint->GetId() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+			continue;
+		}
+
+		if (std::wstring_view{id}.empty())
+		{
+			xaudio_dev_enum.error("Empty device id - skipping");
+			CoTaskMemFree(id);
+			continue;
+		}
+
+		audio_device dev{};
+		dev.id = wchar_to_utf8(id);
+
+		CoTaskMemFree(id);
+
+		Microsoft::WRL::ComPtr<IPropertyStore> props{};
+		if (HRESULT hr = endpoint->OpenPropertyStore(STGM_READ, props.GetAddressOf()); FAILED(hr))
+		{
+			xaudio_dev_enum.error("endpoint->OpenPropertyStore() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+			continue;
+		}
+
+		PROPVARIANT var;
+		PropVariantInit(&var);
+
+		if (HRESULT hr = props->GetValue(PKEY_Device_FriendlyName, &var); FAILED(hr))
+		{
+			xaudio_dev_enum.error("props->GetValue() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+			PropVariantClear(&var);
+			continue;
+		}
+
+		if (var.vt != VT_LPWSTR)
+		{
+			PropVariantClear(&var);
+			continue;
+		}
+
+		dev.name = wchar_to_utf8(var.pwszVal);
+		if (dev.name.empty())
+		{
+			dev.name = dev.id;
+		}
+
+		PropVariantClear(&var);
+		dev.max_ch = 2;
+
+		if (HRESULT hr = props->GetValue(PKEY_AudioEngine_DeviceFormat, &var); SUCCEEDED(hr))
+		{
+			if (var.vt == VT_BLOB)
+			{
+				if (var.blob.cbSize == sizeof(PCMWAVEFORMAT))
+				{
+					const PCMWAVEFORMAT* pcm = std::bit_cast<const PCMWAVEFORMAT*>(var.blob.pBlobData);
+					dev.max_ch = pcm->wf.nChannels;
+				}
+				else if (var.blob.cbSize >= sizeof(WAVEFORMATEX))
+				{
+					const WAVEFORMATEX* wfx = std::bit_cast<const WAVEFORMATEX*>(var.blob.pBlobData);
+					if (var.blob.cbSize >= sizeof(WAVEFORMATEX) + wfx->cbSize || wfx->wFormatTag == WAVE_FORMAT_PCM)
+					{
+						dev.max_ch = wfx->nChannels;
+					}
+				}
+			}
+		}
+		else
+		{
+			xaudio_dev_enum.error("props->GetValue() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
+		}
+
+		PropVariantClear(&var);
+
+		xaudio_dev_enum.notice("Found device: id=%s, name=%s, max_ch=%d", dev.id, dev.name, dev.max_ch);
+		device_list.emplace_back(dev);
+	}
+
+	std::sort(device_list.begin(), device_list.end(), [](audio_device_enumerator::audio_device a, audio_device_enumerator::audio_device b)
+	{
+		return a.name < b.name;
+	});
+
+	return device_list;
+}
@@ -0,0 +1,17 @@
+#pragma once
+
+#ifndef _WIN32
+#error "XAudio2 can only be built on Windows."
+#endif
+
+#include "Emu/Audio/audio_device_enumerator.h"
+
+class xaudio2_enumerator final : public audio_device_enumerator
+{
+public:
+
+	xaudio2_enumerator();
+	~xaudio2_enumerator() override;
+
+	std::vector<audio_device> get_output_devices() override;
+};
@@ -0,0 +1,26 @@
+#pragma once
+
+#include "util/types.hpp"
+#include <vector>
+#include <string>
+
+class audio_device_enumerator
+{
+public:
+
+	static constexpr std::string_view DEFAULT_DEV_ID = "@@@default@@@";
+
+	struct audio_device
+	{
+		std::string id{};
+		std::string name{};
+		usz max_ch{};
+	};
+
+	audio_device_enumerator() {};
+
+	virtual ~audio_device_enumerator() = default;
+
+	// Enumerate available output devices.
+	virtual std::vector<audio_device> get_output_devices() = 0;
+};
@@ -1,147 +0,0 @@
-#include "stdafx.h"
-#include "audio_device_listener.h"
-#include "util/logs.hpp"
-#include "Utilities/StrUtil.h"
-#include "Emu/Cell/Modules/cellAudio.h"
-#include "Emu/IdManager.h"
-
-LOG_CHANNEL(IO);
-
-audio_device_listener::audio_device_listener()
-{
-#ifdef _WIN32
-	HRESULT hr = CoInitializeEx(nullptr, COINIT_MULTITHREADED);
-	if (SUCCEEDED(hr))
-	{
-		m_com_init_success = true;
-	}
-
-	// Try to register a listener for device changes
-	hr = CoCreateInstance(__uuidof(MMDeviceEnumerator), nullptr, CLSCTX_INPROC_SERVER, IID_PPV_ARGS(&m_device_enumerator));
-	if (hr != S_OK)
-	{
-		IO.error("CoCreateInstance() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
-	}
-	else if (m_device_enumerator)
-	{
-		hr = m_device_enumerator->RegisterEndpointNotificationCallback(this);
-		if (FAILED(hr))
-		{
-			IO.error("RegisterEndpointNotificationCallback() failed: %s (0x%08x)", std::system_category().message(hr), static_cast<u32>(hr));
-			m_device_enumerator->Release();
-		}
-	}
-	else
-	{
-		IO.error("Device enumerator invalid");
-	}
-#endif
-}
-
-audio_device_listener::~audio_device_listener()
-{
-#ifdef _WIN32
-	if (m_device_enumerator != nullptr)
-	{
-		m_device_enumerator->UnregisterEndpointNotificationCallback(this);
-		m_device_enumerator->Release();
-	}
-
-	if (m_com_init_success)
-	{
-		CoUninitialize();
-	}
-#endif
-}
-
-bool audio_device_listener::input_device_changed()
-{
-	return m_input_device_changed.test_and_reset();
-}
-
-bool audio_device_listener::output_device_changed()
-{
-	return m_output_device_changed.test_and_reset();
-}
-
-#ifdef _WIN32
-template <>
-void fmt_class_string<ERole>::format(std::string& out, u64 arg)
-{
-	format_enum(out, arg, [](auto value)
-	{
-		switch (value)
-		{
-		case eConsole: return "eConsole";
-		case eMultimedia: return "eMultimedia";
-		case eCommunications: return "eCommunications";
-		}
-
-		return unknown;
-	});
-}
-
-template <>
-void fmt_class_string<EDataFlow>::format(std::string& out, u64 arg)
-{
-	format_enum(out, arg, [](auto value)
-	{
-		switch (value)
-		{
-		case eRender: return "eRender";
-		case eCapture: return "eCapture";
-		case eAll: return "eAll";
-		}
-
-		return unknown;
-	});
-}
-
-HRESULT audio_device_listener::OnDefaultDeviceChanged(EDataFlow flow, ERole role, LPCWSTR new_default_device_id)
-{
-	IO.notice("OnDefaultDeviceChanged(flow=%s, role=%s, new_default_device_id=0x%x)", flow, role, new_default_device_id);
-
-	if (!new_default_device_id)
-	{
-		IO.notice("OnDefaultDeviceChanged(): new_default_device_id empty");
-		return S_OK;
-	}
-
-	// Listen only for one device role, otherwise we're going to receive more than one
-	// notification for flow type
-	if (role != eConsole)
-	{
-		IO.notice("OnDefaultDeviceChanged(): we don't care about this device");
-		return S_OK;
-	}
-
-	const std::wstring tmp(new_default_device_id);
-	const std::string new_device_id = wchar_to_utf8(tmp.c_str());
-
-	if (flow == eRender || flow == eAll)
-	{
-		if (output_device_id != new_device_id)
-		{
-			output_device_id = new_device_id;
-
-			IO.warning("Default output device changed: new device = '%s'", output_device_id);
-
-			m_output_device_changed = true;
-		}
-	}
-
-	if (flow == eCapture || flow == eAll)
-	{
-		if (input_device_id != new_device_id)
-		{
-			input_device_id = new_device_id;
-
-			IO.warning("Default input device changed: new device = '%s'", input_device_id);
-
-			m_input_device_changed = true;
-		}
-	}
-
-	return S_OK;
-}
-#endif
@@ -1,42 +0,0 @@
-#pragma once
-
-#ifdef _WIN32
-#include <MMDeviceAPI.h>
-#endif
-
-#include "util/atomic.hpp"
-
-#ifdef _WIN32
-class audio_device_listener : public IMMNotificationClient
-#else
-class audio_device_listener
-#endif
-{
-public:
-	audio_device_listener();
-	~audio_device_listener();
-
-	bool input_device_changed();
-	bool output_device_changed();
-
-private:
-#ifdef _WIN32
-	std::string input_device_id;
-	std::string output_device_id;
-
-	IFACEMETHODIMP_(ULONG) AddRef() override { return 1; };
-	IFACEMETHODIMP_(ULONG) Release() override { return 1; };
-	IFACEMETHODIMP QueryInterface(REFIID /*iid*/, void** /*object*/) override { return S_OK; };
-	IFACEMETHODIMP OnPropertyValueChanged(LPCWSTR /*device_id*/, const PROPERTYKEY /*key*/) override { return S_OK; };
-	IFACEMETHODIMP OnDeviceAdded(LPCWSTR /*device_id*/) override { return S_OK; };
-	IFACEMETHODIMP OnDeviceRemoved(LPCWSTR /*device_id*/) override { return S_OK; };
-	IFACEMETHODIMP OnDeviceStateChanged(LPCWSTR /*device_id*/, DWORD /*new_state*/) override { return S_OK; };
-	IFACEMETHODIMP OnDefaultDeviceChanged(EDataFlow flow, ERole role, LPCWSTR new_default_device_id) override;
-
-	IMMDeviceEnumerator* m_device_enumerator = nullptr;
-	bool m_com_init_success = false;
-#endif
-
-	atomic_t<bool> m_input_device_changed = false;
-	atomic_t<bool> m_output_device_changed = false;
-};
@@ -2,6 +2,7 @@ add_library(rpcs3_emu
    cache_utils.cpp
    IdManager.cpp
    localized_string.cpp
+    savestate_utils.cpp
    System.cpp
    system_config.cpp
    system_config_types.cpp
@@ -36,7 +37,6 @@ target_include_directories(rpcs3_emu
    PUBLIC
        ${RPCS3_SRC_DIR})

-
 # Utilities
 target_sources(rpcs3_emu PRIVATE
    ../util/atomic.cpp
@@ -58,6 +58,7 @@ target_sources(rpcs3_emu PRIVATE
    ../../Utilities/rXml.cpp
    ../../Utilities/sema.cpp
    ../../Utilities/simple_ringbuf.cpp
+    ../../Utilities/stack_trace.cpp
    ../../Utilities/StrFmt.cpp
    ../../Utilities/Thread.cpp
    ../../Utilities/version.cpp
@@ -115,20 +116,22 @@ target_sources(rpcs3_emu PRIVATE
    ../Loader/TRP.cpp
 )

-
 # Audio
 target_sources(rpcs3_emu PRIVATE
-    Audio/audio_device_listener.cpp
    Audio/audio_resampler.cpp
    Audio/AudioDumper.cpp
    Audio/AudioBackend.cpp
    Audio/Cubeb/CubebBackend.cpp
+    Audio/Cubeb/cubeb_enumerator.cpp
 )

 if(USE_FAUDIO)
    find_package(SDL2)
    if(SDL2_FOUND AND NOT SDL2_VERSION VERSION_LESS 2.0.9)
-        target_sources(rpcs3_emu PRIVATE Audio/FAudio/FAudioBackend.cpp)
+        target_sources(rpcs3_emu PRIVATE
+            Audio/FAudio/FAudioBackend.cpp
+            Audio/FAudio/faudio_enumerator.cpp
+        )
        target_link_libraries(rpcs3_emu PUBLIC 3rdparty::faudio)
    endif()
 endif()
@@ -139,6 +142,7 @@ if(WIN32)
    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} /DELAYLOAD:xaudio2_9redist.dll")
    target_sources(rpcs3_emu PRIVATE
        Audio/XAudio2/XAudio2Backend.cpp
+        Audio/XAudio2/xaudio2_enumerator.cpp
    )
 endif()

@@ -182,6 +186,7 @@ target_sources(rpcs3_emu PRIVATE
    Cell/lv2/sys_event.cpp
    Cell/lv2/sys_event_flag.cpp
    Cell/lv2/sys_fs.cpp
+    Cell/lv2/sys_game.cpp
    Cell/lv2/sys_gamepad.cpp
    Cell/lv2/sys_gpio.cpp
    Cell/lv2/sys_hid.cpp
@@ -259,6 +264,7 @@ target_sources(rpcs3_emu PRIVATE
    Cell/Modules/cellMsgDialog.cpp
    Cell/Modules/cellMusic.cpp
    Cell/Modules/cellMusicDecode.cpp
+    Cell/Modules/cellMusicSelectionContext.cpp
    Cell/Modules/cellMusicExport.cpp
    Cell/Modules/cellNetAoi.cpp
    Cell/Modules/cellNetCtl.cpp
@@ -332,7 +338,7 @@ target_sources(rpcs3_emu PRIVATE
    Cell/Modules/sceNpUtil.cpp
    Cell/Modules/StaticHLE.cpp
    Cell/Modules/sys_crashdump.cpp
-    Cell/Modules/sys_game.cpp
+    Cell/Modules/sys_game_.cpp
    Cell/Modules/sys_heap.cpp
    Cell/Modules/sys_io_.cpp
    Cell/Modules/sys_libc_.cpp
@@ -408,7 +414,6 @@ target_sources(rpcs3_emu PRIVATE
    Memory/vm.cpp
 )

-
 # RSX
 target_sources(rpcs3_emu PRIVATE
    RSX/gcm_enums.cpp
@@ -430,6 +435,7 @@ target_sources(rpcs3_emu PRIVATE
    RSX/Null/NullGSRender.cpp
    RSX/Overlays/overlay_animation.cpp
    RSX/Overlays/overlay_controls.cpp
+    RSX/Overlays/overlay_cursor.cpp
    RSX/Overlays/overlay_edit_text.cpp
    RSX/Overlays/overlay_fonts.cpp
    RSX/Overlays/overlay_list_view.cpp
@@ -457,10 +463,15 @@ target_sources(rpcs3_emu PRIVATE
    RSX/Program/VertexProgramDecompiler.cpp
    RSX/Capture/rsx_capture.cpp
    RSX/Capture/rsx_replay.cpp
+    RSX/GL/glutils/blitter.cpp
    RSX/GL/glutils/buffer_object.cpp
+    RSX/GL/glutils/capabilities.cpp
    RSX/GL/glutils/common.cpp
+    RSX/GL/glutils/fbo.cpp
    RSX/GL/glutils/image.cpp
+    RSX/GL/glutils/program.cpp
    RSX/GL/glutils/ring_buffer.cpp
+    RSX/GL/glutils/sampler.cpp
    RSX/GL/GLCommonDecompiler.cpp
    RSX/GL/GLCompute.cpp
    RSX/GL/GLDraw.cpp
@@ -538,7 +549,7 @@ target_link_libraries(rpcs3_emu
 if(APPLE)
    check_function_exists(clock_gettime HAVE_CLOCK_GETTIME)
    if (${HAVE_CLOCK_GETTIME})
-	    target_compile_definitions(rpcs3_emu PUBLIC -DHAVE_CLOCK_GETTIME)
+        target_compile_definitions(rpcs3_emu PUBLIC -DHAVE_CLOCK_GETTIME)
    endif()
 endif()

@@ -5,6 +5,7 @@
 #include "Emu/System.h"
 #include "Emu/system_config.h"
 #include "Emu/Memory/vm_locking.h"
+#include "Emu/Memory/vm_reservation.h"
 #include "Emu/IdManager.h"
 #include "Emu/GDB.h"
 #include "Emu/Cell/PPUThread.h"
@@ -49,10 +50,14 @@ void fmt_class_string<cpu_flag>::format(std::string& out, u64 arg)
 		case cpu_flag::pause: return "p";
 		case cpu_flag::suspend: return "s";
 		case cpu_flag::ret: return "ret";
+		case cpu_flag::again: return "a";
 		case cpu_flag::signal: return "sig";
 		case cpu_flag::memory: return "mem";
 		case cpu_flag::pending: return "pend";
 		case cpu_flag::pending_recheck: return "pend-re";
+		case cpu_flag::notify: return "ntf";
+		case cpu_flag::yield: return "y";
+		case cpu_flag::preempt: return "PREEMPT";
 		case cpu_flag::dbg_global_pause: return "G-PAUSE";
 		case cpu_flag::dbg_pause: return "PAUSE";
 		case cpu_flag::dbg_step: return "STEP";
@@ -83,15 +88,18 @@ struct cpu_prof
 		// Total number of samples
 		u64 samples = 0, idle = 0;

-		// Avoid printing replicas
-		bool printed = false;
+		// Avoid printing replicas or when not much changed
+		u64 new_samples = 0;
+
+		static constexpr u64 min_print_samples = 500;
+		static constexpr u64 min_print_all_samples = min_print_samples * 20;

 		void reset()
 		{
 			freq.clear();
 			samples = 0;
 			idle = 0;
-			printed = false;
+			new_samples = 0;
 		}

 		static std::string format(const std::multimap<u64, u64, std::greater<u64>>& chart, u64 samples, u64 idle, bool extended_print = false)
@@ -127,8 +135,13 @@ struct cpu_prof
 		// Print info
 		void print(const std::shared_ptr<cpu_thread>& ptr)
 		{
-			if (printed || samples == idle)
+			if (new_samples < min_print_samples || samples == idle)
 			{
+				if (cpu_flag::exit - ptr->state)
+				{
+					profiler.notice("Thread \"%s\" [0x%08x]: %u samples, %u new (%.4f%% idle): Not enough new samples have been collected since the last print.", ptr->get_name(), ptr->id, samples, new_samples, 100. * idle / samples);
+				}
+
 				return;
 			}

@@ -142,13 +155,22 @@ struct cpu_prof

 			// Print results
 			const std::string results = format(chart, samples, idle);
-			profiler.notice("Thread \"%s\" [0x%08x]: %u samples (%.4f%% idle):%s", ptr->get_name(), ptr->id, samples, 100. * idle / samples, results);
+			profiler.notice("Thread \"%s\" [0x%08x]: %u samples, %u new (%.4f%% idle):%s", ptr->get_name(), ptr->id, samples, new_samples, 100. * idle / samples, results);

-			printed = true;
+			new_samples = 0;
 		}

-		static void print_all(const std::unordered_map<std::shared_ptr<cpu_thread>, sample_info>& threads)
+		static void print_all(std::unordered_map<std::shared_ptr<cpu_thread>, sample_info>& threads)
 		{
+			u64 new_samples = 0;
+
+			// Print all results and cleanup
+			for (auto& [ptr, info] : threads)
+			{
+				new_samples += info.new_samples;
+				info.print(ptr);
+			}
+
 			std::multimap<u64, u64, std::greater<u64>> chart;

 			std::unordered_map<u64, u64, value_hash<u64>> freq;
@@ -157,7 +179,7 @@ struct cpu_prof

 			for (auto& [_, info] : threads)
 			{
-				// This function collects thread information regardless of 'printed' member state
+				// This function collects thread information regardless of 'new_samples' member state
 				for (auto& [name, count] : info.freq)
 				{
 					freq[name] += count;
@@ -167,18 +189,24 @@ struct cpu_prof
 				idle += info.idle;
 			}

-			for (auto& [name, count] : freq)
-			{
-				chart.emplace(count, name);
-			}
-
 			if (samples == idle)
 			{
 				return;
 			}

+			if (new_samples < min_print_all_samples && thread_ctrl::state() != thread_state::aborting)
+			{
+				profiler.notice("All Threads: %u samples, %u new (%.4f%% idle): Not enough new samples have been collected since the last print.", samples, new_samples, 100. * idle / samples);
+				return;
+			}
+
+			for (auto& [name, count] : freq)
+			{
+				chart.emplace(count, name);
+			}
+
 			const std::string results = format(chart, samples, idle, true);
-			profiler.notice("All Threads: %u samples (%.4f%% idle):%s", samples, 100. * idle / samples, results);
+			profiler.notice("All Threads: %u samples, %u new (%.4f%% idle):%s", samples, new_samples, 100. * idle / samples, results);
 		}
 	};

@@ -239,7 +267,7 @@ struct cpu_prof
 			// Sample active threads
 			for (auto& [ptr, info] : threads)
 			{
-				if (cpu_flag::exit - ptr->state)
+				if (auto state = +ptr->state; cpu_flag::exit - state)
 				{
 					// Get short function hash
 					const u64 name = atomic_storage<u64>::load(ptr->block_hash);
@@ -247,9 +275,10 @@ struct cpu_prof
 					// Append occurrence
 					info.samples++;

-					if (auto state = +ptr->state; !::is_paused(state) && !::is_stopped(state) && cpu_flag::wait - state)
+					if (cpu_flag::wait - state)
 					{
 						info.freq[name]++;
+						info.new_samples++;

 						// Append verification time to fixed common name 0000000...chunk-0x3fffc
 						if (name >> 16 && (name & 0xffff) == 0)
@@ -257,6 +286,12 @@ struct cpu_prof
 					}
 					else
 					{
+						if (state & (cpu_flag::dbg_pause + cpu_flag::dbg_global_pause))
+						{
+							// Idle state caused by emulation pause is not accounted for
+							continue;
+						}
+
 						info.idle++;
 					}
 				}
@@ -270,11 +305,13 @@ struct cpu_prof
 			{
 				profiler.success("Flushing profiling results...");

-				// Print all results and cleanup
-				for (auto& [ptr, info] : threads)
-				{
-					info.print(ptr);
-				}
+				sample_info::print_all(threads);
+			}
+
+			if (Emu.IsPaused())
+			{
+				thread_ctrl::wait_for(5000);
+				continue;
 			}

 			// Wait, roughly for 20µs
@@ -282,11 +319,6 @@ struct cpu_prof
 		}

 		// Print all remaining results
-		for (auto& [ptr, info] : threads)
-		{
-			info.print(ptr);
-		}
-
 		sample_info::print_all(threads);
 	}

@@ -573,6 +605,7 @@ void cpu_thread::operator()()
 		if (!(state0 & cpu_flag::stop))
 		{
 			cpu_task();
+			state += cpu_flag::wait;

 			if (state & cpu_flag::ret && state.test_and_reset(cpu_flag::ret))
 			{
@@ -582,7 +615,7 @@ void cpu_thread::operator()()
 			continue;
 		}

-		thread_ctrl::wait_on(state, state0);
+		state.wait(state0);

 		if (state & cpu_flag::ret && state.test_and_reset(cpu_flag::ret))
 		{
@@ -634,20 +667,22 @@ cpu_thread::cpu_thread(u32 id)

 void cpu_thread::cpu_wait(bs_t<cpu_flag> old)
 {
-	thread_ctrl::wait_on(state, old);
+	state.wait(old);
 }

+static atomic_t<u32> s_dummy_atomic = 0;
+
 bool cpu_thread::check_state() noexcept
 {
 	bool cpu_sleep_called = false;
 	bool cpu_can_stop = true;
-	bool escape, retval;
+	bool escape{}, retval{};

 	while (true)
 	{
 		// Process all flags in a single atomic op
 		bs_t<cpu_flag> state1;
-		const auto state0 = state.fetch_op([&](bs_t<cpu_flag>& flags)
+		auto state0 = state.fetch_op([&](bs_t<cpu_flag>& flags)
 		{
 			bool store = false;

@@ -687,7 +722,6 @@ bool cpu_thread::check_state() noexcept
 			{
 				// Sticky flag, indicates check_state() is not allowed to return true
 				flags -= cpu_flag::temp;
-				flags -= cpu_flag::wait;
 				cpu_can_stop = false;
 				store = true;
 			}
@@ -699,6 +733,12 @@ bool cpu_thread::check_state() noexcept
 				store = true;
 			}

+			if (flags & cpu_flag::notify)
+			{
+				flags -= cpu_flag::notify;
+				store = true;
+			}
+
 			// Can't process dbg_step if we only paused temporarily
 			if (cpu_can_stop && flags & cpu_flag::dbg_step)
 			{
@@ -720,10 +760,10 @@ bool cpu_thread::check_state() noexcept
 			}

 			// Atomically clean wait flag and escape
-			if (!(flags & (cpu_flag::exit + cpu_flag::ret + cpu_flag::stop)))
+			if (!is_stopped(flags) && flags.none_of(cpu_flag::ret))
 			{
 				// Check pause flags which hold thread inside check_state (ignore suspend/debug flags on cpu_flag::temp)
-				if (flags & (cpu_flag::pause + cpu_flag::memory) || (cpu_can_stop && flags & (cpu_flag::dbg_global_pause + cpu_flag::dbg_pause + cpu_flag::suspend)))
+				if (flags & (cpu_flag::pause + cpu_flag::memory) || (cpu_can_stop && flags & (cpu_flag::dbg_global_pause + cpu_flag::dbg_pause + cpu_flag::suspend + cpu_flag::yield + cpu_flag::preempt)))
 				{
 					if (!(flags & cpu_flag::wait))
 					{
@@ -731,6 +771,12 @@ bool cpu_thread::check_state() noexcept
 						store = true;
 					}

+					if (flags & (cpu_flag::yield + cpu_flag::preempt) && cpu_can_stop)
+					{
+						flags -= (cpu_flag::yield + cpu_flag::preempt);
+						store = true;
+					}
+
 					escape = false;
 					state1 = flags;
 					return store;
@@ -746,9 +792,9 @@ bool cpu_thread::check_state() noexcept
 			}
 			else
 			{
-				if (cpu_can_stop && !(flags & cpu_flag::wait))
+				if (flags & cpu_flag::wait)
 				{
-					flags += cpu_flag::wait;
+					flags -= cpu_flag::wait;
 					store = true;
 				}

@@ -760,6 +806,34 @@ bool cpu_thread::check_state() noexcept
 			return store;
 		}).first;

+		if (state0 & cpu_flag::preempt && cpu_can_stop)
+		{
+			if (cpu_flag::wait - state0)
+			{
+				if (!escape || !retval)
+				{
+					// Yield itself
+					state0 += cpu_flag::yield;
+					escape = false;
+				}
+			}
+
+			if (const u128 bits = s_cpu_bits)
+			{
+				reader_lock lock(s_cpu_lock);
+
+				cpu_counter::for_all_cpu(bits & s_cpu_bits, [](cpu_thread* cpu)
+				{
+					if (cpu->state.none_of(cpu_flag::wait + cpu_flag::yield))
+					{
+						cpu->state += cpu_flag::yield;
+					}
+
+					return true;
+				});
+			}
+		}
+
 		if (escape)
 		{
 			if (s_tls_thread_slot == umax && !retval)
@@ -768,16 +842,16 @@ bool cpu_thread::check_state() noexcept
 				cpu_counter::add(this);
 			}

-			constexpr auto pending_and_temp = (cpu_flag::pending + cpu_flag::temp);
-
-			if ((state0 & pending_and_temp) == cpu_flag::pending)
+			if (cpu_can_stop && state0 & cpu_flag::pending)
 			{
 				// Execute pending work
 				cpu_work();

-				if ((state1 ^ state) - pending_and_temp)
+				if ((state1 ^ state) - cpu_flag::pending)
 				{
 					// Work could have changed flags
+					// Reset internal flags as if check_state() has just been called
+					cpu_sleep_called = false;
 					continue;
 				}
 			}
@@ -848,6 +922,32 @@ bool cpu_thread::check_state() noexcept
 						break;
 					}
 				}
+
+				continue;
+			}
+
+			if (state0 & cpu_flag::yield && cpu_flag::wait - state0 && cpu_can_stop)
+			{
+				if (auto spu = try_get<spu_thread>())
+				{
+					if (spu->raddr && spu->rtime == vm::reservation_acquire(spu->raddr) && spu->getllar_spin_count < 10)
+					{
+						// Reservation operation is a critical section (but this may result in false positives)
+						continue;
+					}
+				}
+				else if (auto ppu = try_get<ppu_thread>())
+				{
+					if (ppu->raddr && ppu->rtime == vm::reservation_acquire(ppu->raddr))
+					{
+						// Same
+						continue;
+					}
+				}
+
+				// Short sleep when yield flag is present alone (makes no sense when other methods which can stop thread execution have been done)
+				// Pass a mask of a single bit which is often unused to avoid notifications
+				s_dummy_atomic.wait(0, 1u << 30, atomic_wait_timeout{80'000});
 			}
 		}
 	}
@@ -874,8 +974,27 @@ void cpu_thread::notify()

 cpu_thread& cpu_thread::operator=(thread_state)
 {
-	state += cpu_flag::exit;
-	state.notify_one(cpu_flag::exit);
+	if (state & cpu_flag::exit)
+	{
+		// Must be notified elsewhere or self-raised
+		return *this;
+	}
+
+	const auto old = state.fetch_add(cpu_flag::exit);
+
+	if (old & cpu_flag::wait && old.none_of(cpu_flag::again + cpu_flag::exit))
+	{
+		state.notify_one(cpu_flag::exit);
+
+		if (auto thread = try_get<spu_thread>())
+		{
+			if (u32 resv = atomic_storage<u32>::load(thread->raddr))
+			{
+				vm::reservation_notifier(resv).notify_one();
+			}
+		}
+	}
+
 	return *this;
 }

@@ -945,6 +1064,24 @@ u32* cpu_thread::get_pc2()
 	return nullptr;
 }

+cpu_thread* cpu_thread::get_next_cpu()
+{
+	switch (id_type())
+	{
+	case 1:
+	{
+		return static_cast<ppu_thread*>(this)->next_cpu;
+	}
+	case 2:
+	{
+		return static_cast<spu_thread*>(this)->next_cpu;
+	}
+	default: break;
+	}
+
+	return nullptr;
+}
+
 std::shared_ptr<CPUDisAsm> make_disasm(const cpu_thread* cpu);

 void cpu_thread::dump_all(std::string& ret) const
@@ -1195,3 +1332,78 @@ u32 CPUDisAsm::DisAsmBranchTarget(s32 /*imm*/)
 	// Unused
 	return 0;
 }
+
+extern bool try_lock_spu_threads_in_a_state_compatible_with_savestates(bool revert_lock)
+{
+	const u64 start = get_system_time();
+
+	// Attempt to lock for half a second, if somehow takes longer abort it
+	do
+	{
+		if (revert_lock)
+		{
+			// Revert the operation of this function
+			break;
+		}
+
+		if (cpu_thread::suspend_all(nullptr, {}, []()
+		{
+			return idm::select<named_thread<spu_thread>>([](u32, spu_thread& spu)
+			{
+				if (!spu.unsavable)
+				{
+					if (Emu.IsPaused())
+					{
+						// If emulation is paused, we can only hope it's already in a state compatible with savestates
+						return !!(spu.state & (cpu_flag::dbg_global_pause + cpu_flag::dbg_pause));
+					}
+					else
+					{
+						ensure(!spu.state.test_and_set(cpu_flag::dbg_global_pause));
+					}
+
+					return false;
+				}
+
+				return true;
+			}).ret;
+		}))
+		{
+			if (Emu.IsPaused())
+			{
+				return false;
+			}
+
+			// It's faster to lock once
+			reader_lock lock(id_manager::g_mutex);
+
+			idm::select<named_thread<spu_thread>>([](u32, spu_thread& spu)
+			{
+				spu.state -= cpu_flag::dbg_global_pause;
+			}, idm::unlocked);
+
+			// For faster signalling, first remove state flags then batch notifications
+			idm::select<named_thread<spu_thread>>([](u32, spu_thread& spu)
+			{
+				if (spu.state & cpu_flag::wait)
+				{
+					spu.state.notify_one();
+				}
+			}, idm::unlocked);
+
+			continue;
+		}
+
+		return true;
+	} while (std::this_thread::yield(), get_system_time() - start <= 500'000);
+
+	idm::select<named_thread<spu_thread>>([&](u32, named_thread<spu_thread>& spu)
+	{
+		if (spu.state.test_and_reset(cpu_flag::dbg_global_pause))
+		{
+			spu.state.notify_one();
+		}
+	});
+
+	return false;
+}
@@ -19,10 +19,14 @@ enum class cpu_flag : u32
 	pause, // Thread suspended by suspend_all technique
 	suspend, // Thread suspended
 	ret, // Callback return requested
+	again, // Thread must complete the syscall after deserialization
 	signal, // Thread received a signal (HLE)
 	memory, // Thread must unlock memory mutex
 	pending, // Thread has postponed work
 	pending_recheck, // Thread needs to recheck if there is pending work before ::pending removal
+	notify, // Flag meant solely to allow atomic notification on state without changing other flags
+	yield, // Thread is being requested to yield its execution time if it's running
+	preempt, // Thread is being requested to preempt the execution of all CPU threads

 	dbg_global_pause, // Emulation paused
 	dbg_pause, // Thread paused
@@ -34,7 +38,7 @@ enum class cpu_flag : u32
 // Test stopped state
 constexpr bool is_stopped(bs_t<cpu_flag> state)
 {
-	return !!(state & (cpu_flag::stop + cpu_flag::exit));
+	return !!(state & (cpu_flag::stop + cpu_flag::exit + cpu_flag::again));
 }

 // Test paused state
@@ -139,6 +143,7 @@ public:

 	u32 get_pc() const;
 	u32* get_pc2(); // Last PC before stepping for the debugger (may be null)
+	cpu_thread* get_next_cpu(); // Access next_cpu member if the is one

 	void notify();
 	cpu_thread& operator=(thread_state);
@@ -172,7 +177,7 @@ public:
 	virtual void cpu_sleep() {}

 	// Callback for cpu_flag::pending
-	virtual void cpu_work() {}
+	virtual void cpu_work() { state -= cpu_flag::pending + cpu_flag::pending_recheck; }

 	// Callback for cpu_flag::ret
 	virtual void cpu_return() {}
@@ -32,7 +32,21 @@ cpu_translator::cpu_translator(llvm::Module* _module, bool is_be)

 		if (m_use_ssse3)
 		{
+#if defined(ARCH_X64)
 			return m_ir->CreateCall(get_intrinsic(llvm::Intrinsic::x86_ssse3_pshuf_b_128), {data0, index});
+#elif defined(ARCH_ARM64)
+			// Modified from sse2neon
+			// movi    v2.16b, #143
+			// and     v1.16b, v1.16b, v2.16b
+			// tbl     v0.16b, { v0.16b }, v1.16b
+			auto mask = llvm::ConstantInt::get(get_type<u8[16]>(), 0x8F);
+			auto and_mask = llvm::ConstantInt::get(get_type<bool[16]>(), true);
+			auto vec_len = llvm::ConstantInt::get(get_type<u32>(), 16);
+			auto index_masked = m_ir->CreateCall(get_intrinsic<u8[16]>(llvm::Intrinsic::vp_and), {index, mask, and_mask, vec_len});
+			return m_ir->CreateCall(get_intrinsic<u8[16]>(llvm::Intrinsic::aarch64_neon_tbl1), {data0, index_masked});
+#else
+#error "Unimplemented"
+#endif
 		}
 		else
 		{
@@ -140,6 +154,14 @@ void cpu_translator::initialize(llvm::LLVMContext& context, llvm::ExecutionEngin
 		m_use_avx512_icl = true;
 		m_use_vnni = true;
 	}
+
+	// Aarch64 CPUs
+	if (cpu == "cyclone" || cpu.contains("cortex"))
+	{
+		m_use_fma = true;
+		// AVX does not use intrinsics so far
+		m_use_avx = true;
+	}
 }

 llvm::Value* cpu_translator::bitcast(llvm::Value* val, llvm::Type* type) const
@@ -30,6 +30,7 @@
 #endif

 #include "util/types.hpp"
+#include "util/sysinfo.hpp"
 #include "Utilities/StrFmt.h"
 #include "Utilities/BitField.h"
 #include "Utilities/JIT.h"
@@ -2896,12 +2897,8 @@ protected:
 	bool m_is_be;

 	// Allow PSHUFB intrinsic
-#ifdef ARCH_X64
 	bool m_use_ssse3 = true;
-#else
-	// TODO: fix the pshufb arm64 native impl using TBL instruction
-	bool m_use_ssse3 = false;
-#endif
+
 	// Allow FMA
 	bool m_use_fma = false;

@@ -2979,6 +2976,7 @@ public:
 		m_engine->updateGlobalMapping({lame.data(), lame.size()}, reinterpret_cast<uptr>(_func));

 		const auto inst = m_ir->CreateCall(func, {args...});
+		inst->setTailCallKind(llvm::CallInst::TCK_NoTail);
 #ifdef _WIN32
 		inst->setCallingConv(llvm::CallingConv::Win64);
 #endif
@@ -3445,6 +3443,11 @@ public:
 	template <typename T1, typename T2, typename T3>
 	value_t<u8[16]> vperm2b(T1 a, T2 b, T3 c)
 	{
+		if (!utils::has_fast_vperm2b())
+		{
+			return vperm2b256to128(a, b, c);
+		}
+
 		value_t<u8[16]> result;

 		const auto data0 = a.eval(m_ir);
@@ -39,6 +39,8 @@ public:
 	{
 		return value;
 	}
+
+	ENABLE_BITWISE_SERIALIZATION;
 };

 enum CellNotAnError : s32
@@ -88,6 +88,8 @@ enum : u32

 struct alignas(16) spu_mfc_cmd
 {
+	ENABLE_BITWISE_SERIALIZATION;
+
 	MFC cmd;
 	u8 tag;
 	u16 size;
--- a/Show More
+++ b/Show More