[ABTOPEMOHT]

Fixes #566 (Mewgenics: about half of the textures draw as black rectangles under Zink).

Cause

Metal on A-series iPhones and iPads has no BC texture sampling (supportsBCTextureCompression = NO, as madeira-d3d12/src/pe/mad_bcn.cpp already notes for the A16). So MoltenVK and KosmicKrisp report no BC formats, and Zink's is_format_supported(PIPE_FORMAT_DXT5_RGBA) is false.

Mesa can cover for this: with allow_compressed_fallback it still exposes GL_EXT_texture_compression_s3tc, _s3tc_srgb, ARB_texture_compression_rgtc, EXT_texture_compression_latc and ARB_texture_compression_bptc, and decodes those uploads to RGBA8 on the CPU (st_compressed_format_fallback, st_UnmapTextureImage). The DRI frontends read that option from driconf, where it defaults to true (DRI_CONF_ALLOW_COMPRESSED_FALLBACK(true) in driinfo_gallium.h). OSMesa never reads driconf: OSMesaCreateContextAttribs memsets attribs, so the option stayed false and the extensions were missing.

A game that uploads DXT data without checking for the extension (every desktop GPU has it) then gets GL_INVALID_ENUM from glCompressedTexImage2D. The texture stays incomplete and samples as opaque black. That is why switching gl-vulkan to KosmicKrisp did not help: both drivers run on the same Metal GPU without BC.

Confirmation

On the device from #566 (iPhone 12 Pro Max, A14, Madeira 0.2.0, Zink on MoltenVK), this line in Mewgenics' own config makes every texture draw correctly:

env.MESA_EXTENSION_OVERRIDE = +GL_EXT_texture_compression_s3tc +GL_EXT_texture_compression_s3tc_srgb +GL_ARB_texture_compression_rgtc +GL_ARB_texture_compression_bptc

It turns on the same extensions that this patch exposes, and the same CPU decode then runs.

Change

• build/mesa-ios/patches/0009-osmesa-allow-compressed-fallback.patch: OSMesa sets attribs.options.allow_compressed_fallback to true, the same default the DRI frontends use. MADEIRA_GL_COMPRESSED_FALLBACK=0 turns it off. A GPU that has BC keeps the native formats, because the fallback applies only when st->has_s3tc (or has_rgtc, has_bptc, ...) is false. This covers the MoltenVK and KosmicKrisp backends alike, because both go through OSMesa + Zink.
• build/mesa-ios/patches/README.md: entry for 0009.
• tests/x64/gls3tc-x64.c: uploads a solid-green 4x4 block as DXT1 and as DXT5, draws it, and reads a pixel back. Expected on an A14 to A16 device: fails before the patch (extension missing, pixel blue) and passes after it.

Testing

• The patch applies cleanly to Mesa 25.0.7 (patch -p1 --forward). It touches only src/gallium/frontends/osmesa/osmesa.c, which no other patch in the set changes.
• gls3tc-x64.c builds with clang (x86_64) and passes on a desktop GPU (NVIDIA, native S3TC), so the test logic itself is right.
• Not done: I could not build the iOS Mesa (build/mesa-ios/build.sh needs macOS and Xcode) or run the patched build on the device. The fix is confirmed only through the MESA_EXTENSION_OVERRIDE workaround above.

🤖 Generated with Claude Code