Run Windows games on Apple Silicon — free, open source, with an open compatibility database that tells you what actually works.
Swift apple-silicond3dmetaldxmtdxvkfexgame-porting-toolkitgamingmac-gamingmacosmetalmoltenvksteamswiftuivkd3d-protonwhisky-alternativewindows-gameswinewine-wrapper
Wine 11 engines r20 and r21 carry patch 0021 (engine-wine-11.0-20261006T161716Z-0021-es2-compat): macOS lists GL_ARB_ES2_compatibility only in 4.1 core contexts, so a program in a legacy 2.1 context got nothing for glClearDepthf, glDepthRangef, glGetShaderPrecisionFormat, glReleaseShaderCompiler and glShaderBinary and called the empty pointer. Paperback: The Game (Loom engine) crashed at address 0 on every engine; on an M4, r21 installed from this manifest into an empty home, an environment booted and both of the game's builds reached the title menu (highball-db#349). Both revisions are unreleased, so they change in place
Gauthier Piarrette2d02138687f8
diff --git a/Tests/HighballKitTests/BundledEngineTests.swift b/Tests/HighballKitTests/BundledEngineTests.swiftindex c41a0bb..7ee0d2f 100644--- a/Tests/HighballKitTests/BundledEngineTests.swift+++ b/Tests/HighballKitTests/BundledEngineTests.swift@@ -67,7 +67,8 @@ final class BundledEngineTests: XCTestCase {}/// r21 is r19 and r20 is r18 with the Wine rebuilt for patch 0020 (the Mac driver stops doubling- /// HORZRES and VERTRES in retina mode, highball#261) and the refreshed 0012, numbered the same+ /// HORZRES and VERTRES in retina mode, highball#261), 0021 (GL_ARB_ES2_compatibility's calls in+ /// legacy OpenGL contexts, highball-db#349) and the refreshed 0012, numbered the same/// way: the D3DMetal 4 one below, so every older plain pin is offered the D3DMetal 3 revision/// and Forza Horizon 6's pin on r18 is offered r20.func testR21AndR20AreR19AndR18WithTheRetinaFixedWine() throws {diff --git a/spike/engines/x64-crossover26.3-r20.json b/spike/engines/x64-crossover26.3-r20.jsonindex 257236d..c4e1bd4 100644--- a/spike/engines/x64-crossover26.3-r20.json+++ b/spike/engines/x64-crossover26.3-r20.json@@ -72,10 +72,10 @@"kind": "engine","license": "LGPL-2.1+","note": "Built from CodeWeavers' crossover-sources-26.3.0.tar.gz (sha pinned in highball-engine/inputs.json) by highball-engine's build workflow with patches 0001-0003, 0005-0012, 0014, 0015 and 0016, from the audio-trial branch (main without the held 0013). 0009 swaps %gs:0x20 between Tib.FiberData (what PE code reads as GetCurrentFiber()) and the pthread QoS class (what macOS's libdispatch and pthread read) at every crossing between PE and unix code, so a fiber-based game no longer corrupts the scheduling slot: with 0008 alone Metaphor: ReFantazio opened a black window and hung in libdispatch. Stripped; Wine Mono 10.4.1 and Gecko 2.47.4 included.",- "sha256": "2992e120e52295f0f1b9983954545b4b7b1bdab86934daae67d364b35b33df9c",- "size": 334865553,- "url": "https://github.com/gauthierpiarrette/highball-engine/releases/download/engine-wine-11.0-20261006T120237Z-0020-retina/engine-wine-11.0-20261006T120237Z-0020-retina.tar.gz",- "version": "wine-11.0 (CrossOver 26.3 tree) Highball build 2026-10-06, patches 0001-0020 (0020: the Mac driver no longer doubles HORZRES and VERTRES in retina mode; 0012 refreshed to athei's 335ebe1c46)"+ "sha256": "51355c786168c5c84eec51e3bc364c8a1acd87437e1418b27887a51d42a5657b",+ "size": 334864515,+ "url": "https://github.com/gauthierpiarrette/highball-engine/releases/download/engine-wine-11.0-20261006T161716Z-0021-es2-compat/engine-wine-11.0-20261006T161716Z-0021-es2-compat.tar.gz",+ "version": "wine-11.0 (CrossOver 26.3 tree) Highball build 2026-10-06, patches 0001-0021 (0020: the Mac driver no longer doubles HORZRES and VERTRES in retina mode; 0021: GL_ARB_ES2_compatibility's calls answer in legacy OpenGL contexts; 0012 refreshed to athei's 335ebe1c46)"},"winetricks": {"kind": "tool",@@ -134,7 +134,7 @@},"minMacOS": "27.0","notes": [- "r20 (2026-10-06): r18 (D3DMetal 4, macOS 27) rebuilt with patch 0020 and the refreshed 0012. 0020 (athei/wine fa43d6db2): in retina mode the Mac driver reported HORZRES and VERTRES twice the screen size that GetSystemMetrics, EnumDisplaySettings and DESKTOPHORZRES report (7680x4320 against 3840x2160 on an M4's 1920x1080 display, measured on r17), so a game sizing its window or fullscreen picture from them got it twice too large; Red Dead Redemption 2's 1080p fullscreen came out cropped for a player with Retina on (highball#261). With retina off nothing changes. 0012 follows athei's 335ebe1c46: the x87 sidecar handshake checks the reply's id and size and says when the sidecar did not attach; it stays dormant until a sidecar is configured. Built by highball-engine's workflow on main (both patches built green on a branch first). Every other component is unchanged.",+ "r20 (2026-10-06): r18 (D3DMetal 4, macOS 27) rebuilt with patch 0020 and the refreshed 0012. 0020 (athei/wine fa43d6db2): in retina mode the Mac driver reported HORZRES and VERTRES twice the screen size that GetSystemMetrics, EnumDisplaySettings and DESKTOPHORZRES report (7680x4320 against 3840x2160 on an M4's 1920x1080 display, measured on r17), so a game sizing its window or fullscreen picture from them got it twice too large; Red Dead Redemption 2's 1080p fullscreen came out cropped for a player with Retina on (highball#261). With retina off nothing changes. 0012 follows athei's 335ebe1c46: the x87 sidecar handshake checks the reply's id and size and says when the sidecar did not attach; it stays dormant until a sidecar is configured. Built by highball-engine's workflow on main (both patches built green on a branch first). Every other component is unchanged. The same build carries 0021: macOS lists GL_ARB_ES2_compatibility only in 4.1 core contexts, so in a legacy 2.1 context opengl32 returned nothing for glClearDepthf, glDepthRangef, glGetShaderPrecisionFormat, glReleaseShaderCompiler and glShaderBinary, and a program written against OpenGL ES 2 called the empty pointer; they now answer through glClearDepth and glDepthRange or do nothing. Paperback: The Game (Loom engine) crashed at address 0 on every engine and reaches its title menu on this one, both builds (M4, macOS 27.0, highball-db#349).","r18 (2026-10-05): r17 with a rebuilt Wine with patches 0017 to 0019 and a new D3DMetal shim. 0017 adds kernelbase's GetIntegratedDisplaySize (upstream Wine c68ed23672), which Forza Horizon 6 calls once its window is up and died on. 0018 reports a process's commit charge (PagefileUsage, PrivateUsage) from the Mach task's footprint instead of its address space: a 9 MB program read 41,987 MB on an M4, and Forza Horizon 6 warned about low memory at 87 GB of 32. 0019 stops CredEnumerate failing when the macOS Keychain has no match for the filter, which threw away matching registry credentials: the Xbox Live sign-in library looks its device key and tokens up that way and asked for a new sign-in on every launch. The shim (d3dmetal-tsshim-20261005) answers a DirectX 12 shader model query D3DMetal does not know with the highest model it supports, and drops ResolveSubresource from a multisampled depth texture into a colour texture, which aborts D3DMetal 3.0 and 4.0b2 alike. It also carries Apple's D3DMetal 4.0 beta 2 (the Wine 10 line's r20 component), on macOS 27 only with the Metal 4 backend off (D3DM_MTL4=0), for Forza Horizon 6: its copy protection opens its own files through \\\\.\\GLOBALROOT paths that only the Wine 11 tree handles, and it refuses D3DMetal 3.0 as an unsupported graphics card (FH205) because only 4.0b2 reports enhanced barriers. The Metal 4 backend measured the same frame rate with Forza (50 against 49 fps in one scene on an M4). It is numbered below r19 on purpose, like r20 under r21 on the Wine 10 line: a recipe that pins an earlier Wine 11 revision is offered the newest revision carrying its pin, and that must stay a D3DMetal 3 one (r19), so Helldivers 2, Guardians of the Galaxy, Persona 5 Royal and Planet Coaster 2 do not change D3DMetal under their verified rows.","r17 (2026-10-03): r16 plus patch 0013 (athei's 539aa62). Wine turned data execution prevention off for the whole process as soon as any loaded module lacked IMAGE_DLLCHARACTERISTICS_NX_COMPAT, which most 32-bit games from before 2010 do. With it off every readable mapping gets PROT_EXEC, and under Rosetta a writable and executable page costs a Mach round trip on first touch, so a Direct3D 9 game streaming vertices through DXVK paid for every buffer write (highball#165). Now DEP follows the main executable and stays permanently on under Rosetta, as Windows's AlwaysOn policy does: ProcessExecuteFlags reports it permanent and a request to turn it off gets STATUS_ACCESS_DENIED; WINE_DISABLE_NX_COMPAT=0 in a game's variables restores the old behaviour for a program that writes code into data pages. Measured on an M4, macOS 27.0, with athei's d3d9stall (noover 2816 1500, DXVK d3d9, five runs each on r16): r16 spent 31 to 49 s in Lock+write per 1500 frames with two or three stalls over a second, this build 0.9 to 1.4 s with none. Every other component is byte-identical to r16.","r16 (2026-10-03): r15 with the DXMT build the default engine has shipped since 0.10.1 (dxmt-highball-20260930T161516Z-a444716) in place of the 4 September build. In the older build the encoding context's Metal device field was declared after the command helpers whose constructors build compute pipelines through it, so a game could die at start or minutes in with newComputePipelineStateWithDescriptor sent to a garbage object (NSInvalidArgumentException, unrecognised selector on an NSIndirectTaggedPointerString or whatever the memory held). Where Winds Meet crashed that way five times on r12 (highball#240), and Among Us's 64-bit build did in 8 of 12 launches on the default engine's r13 before the same fix. Every other component, the Wine build included, is byte-identical to r15, so only the 19 MB DXMT archive is new.",diff --git a/spike/engines/x64-crossover26.3-r21.json b/spike/engines/x64-crossover26.3-r21.jsonindex c899a83..d54207e 100644--- a/spike/engines/x64-crossover26.3-r21.json+++ b/spike/engines/x64-crossover26.3-r21.json@@ -71,10 +71,10 @@"kind": "engine","license": "LGPL-2.1+","note": "Built from CodeWeavers' crossover-sources-26.3.0.tar.gz (sha pinned in highball-engine/inputs.json) by highball-engine's build workflow with patches 0001-0003, 0005-0012, 0014, 0015 and 0016, from the audio-trial branch (main without the held 0013). 0009 swaps %gs:0x20 between Tib.FiberData (what PE code reads as GetCurrentFiber()) and the pthread QoS class (what macOS's libdispatch and pthread read) at every crossing between PE and unix code, so a fiber-based game no longer corrupts the scheduling slot: with 0008 alone Metaphor: ReFantazio opened a black window and hung in libdispatch. Stripped; Wine Mono 10.4.1 and Gecko 2.47.4 included.",- "sha256": "2992e120e52295f0f1b9983954545b4b7b1bdab86934daae67d364b35b33df9c",- "size": 334865553,- "url": "https://github.com/gauthierpiarrette/highball-engine/releases/download/engine-wine-11.0-20261006T120237Z-0020-retina/engine-wine-11.0-20261006T120237Z-0020-retina.tar.gz",- "version": "wine-11.0 (CrossOver 26.3 tree) Highball build 2026-10-06, patches 0001-0020 (0020: the Mac driver no longer doubles HORZRES and VERTRES in retina mode; 0012 refreshed to athei's 335ebe1c46)"+ "sha256": "51355c786168c5c84eec51e3bc364c8a1acd87437e1418b27887a51d42a5657b",+ "size": 334864515,+ "url": "https://github.com/gauthierpiarrette/highball-engine/releases/download/engine-wine-11.0-20261006T161716Z-0021-es2-compat/engine-wine-11.0-20261006T161716Z-0021-es2-compat.tar.gz",+ "version": "wine-11.0 (CrossOver 26.3 tree) Highball build 2026-10-06, patches 0001-0021 (0020: the Mac driver no longer doubles HORZRES and VERTRES in retina mode; 0021: GL_ARB_ES2_compatibility's calls answer in legacy OpenGL contexts; 0012 refreshed to athei's 335ebe1c46)"},"winetricks": {"kind": "tool",@@ -112,7 +112,7 @@},"minMacOS": "14.0","notes": [- "r21 (2026-10-06): r19 rebuilt with patch 0020 and the refreshed 0012. 0020 (athei/wine fa43d6db2): in retina mode the Mac driver reported HORZRES and VERTRES twice the screen size that GetSystemMetrics, EnumDisplaySettings and DESKTOPHORZRES report (7680x4320 against 3840x2160 on an M4's 1920x1080 display, measured on r17), so a game sizing its window or fullscreen picture from them got it twice too large; Red Dead Redemption 2's 1080p fullscreen came out cropped for a player with Retina on (highball#261). With retina off nothing changes. 0012 follows athei's 335ebe1c46: the x87 sidecar handshake checks the reply's id and size and says when the sidecar did not attach; it stays dormant until a sidecar is configured. Built by highball-engine's workflow on main (both patches built green on a branch first). Every other component is unchanged.",+ "r21 (2026-10-06): r19 rebuilt with patch 0020 and the refreshed 0012. 0020 (athei/wine fa43d6db2): in retina mode the Mac driver reported HORZRES and VERTRES twice the screen size that GetSystemMetrics, EnumDisplaySettings and DESKTOPHORZRES report (7680x4320 against 3840x2160 on an M4's 1920x1080 display, measured on r17), so a game sizing its window or fullscreen picture from them got it twice too large; Red Dead Redemption 2's 1080p fullscreen came out cropped for a player with Retina on (highball#261). With retina off nothing changes. 0012 follows athei's 335ebe1c46: the x87 sidecar handshake checks the reply's id and size and says when the sidecar did not attach; it stays dormant until a sidecar is configured. Built by highball-engine's workflow on main (both patches built green on a branch first). Every other component is unchanged. The same build carries 0021: macOS lists GL_ARB_ES2_compatibility only in 4.1 core contexts, so in a legacy 2.1 context opengl32 returned nothing for glClearDepthf, glDepthRangef, glGetShaderPrecisionFormat, glReleaseShaderCompiler and glShaderBinary, and a program written against OpenGL ES 2 called the empty pointer; they now answer through glClearDepth and glDepthRange or do nothing. Paperback: The Game (Loom engine) crashed at address 0 on every engine and reaches its title menu on this one, both builds (M4, macOS 27.0, highball-db#349).","r19 (2026-10-05): r17 with a rebuilt Wine with patches 0017 to 0019 and a new D3DMetal shim. 0017 adds kernelbase's GetIntegratedDisplaySize (upstream Wine c68ed23672), which Forza Horizon 6 calls once its window is up and died on. 0018 reports a process's commit charge (PagefileUsage, PrivateUsage) from the Mach task's footprint instead of its address space: a 9 MB program read 41,987 MB on an M4, and Forza Horizon 6 warned about low memory at 87 GB of 32. 0019 stops CredEnumerate failing when the macOS Keychain has no match for the filter, which threw away matching registry credentials: the Xbox Live sign-in library looks its device key and tokens up that way and asked for a new sign-in on every launch. The shim (d3dmetal-tsshim-20261005) answers a DirectX 12 shader model query D3DMetal does not know with the highest model it supports, and drops ResolveSubresource from a multisampled depth texture into a colour texture, which aborts D3DMetal 3.0 and 4.0b2 alike. Every file is byte-identical to r18 outside renderers/d3dmetal; r18 is the D3DMetal 4 variant for macOS 27, numbered below so recipes pinned to earlier Wine 11 revisions keep D3DMetal 3.","r17 (2026-10-03): r16 plus patch 0013 (athei's 539aa62). Wine turned data execution prevention off for the whole process as soon as any loaded module lacked IMAGE_DLLCHARACTERISTICS_NX_COMPAT, which most 32-bit games from before 2010 do. With it off every readable mapping gets PROT_EXEC, and under Rosetta a writable and executable page costs a Mach round trip on first touch, so a Direct3D 9 game streaming vertices through DXVK paid for every buffer write (highball#165). Now DEP follows the main executable and stays permanently on under Rosetta, as Windows's AlwaysOn policy does: ProcessExecuteFlags reports it permanent and a request to turn it off gets STATUS_ACCESS_DENIED; WINE_DISABLE_NX_COMPAT=0 in a game's variables restores the old behaviour for a program that writes code into data pages. Measured on an M4, macOS 27.0, with athei's d3d9stall (noover 2816 1500, DXVK d3d9, five runs each on r16): r16 spent 31 to 49 s in Lock+write per 1500 frames with two or three stalls over a second, this build 0.9 to 1.4 s with none. Every other component is byte-identical to r16.","r16 (2026-10-03): r15 with the DXMT build the default engine has shipped since 0.10.1 (dxmt-highball-20260930T161516Z-a444716) in place of the 4 September build. In the older build the encoding context's Metal device field was declared after the command helpers whose constructors build compute pipelines through it, so a game could die at start or minutes in with newComputePipelineStateWithDescriptor sent to a garbage object (NSInvalidArgumentException, unrecognised selector on an NSIndirectTaggedPointerString or whatever the memory held). Where Winds Meet crashed that way five times on r12 (highball#240), and Among Us's 64-bit build did in 8 of 12 launches on the default engine's r13 before the same fix. Every other component, the Wine build included, is byte-identical to r15, so only the 19 MB DXMT archive is new.",