Mali: Turn off any depth writes if depth test is set to NEVER.

GPU/Common/FragmentShaderGenerator.cpp

@@ -155,7 +155,9 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 		shading = doFlatShading ? "flat" : "";
 	}
 
-	bool useDiscardStencilBugWorkaround = id.Bit(FS_BIT_NO_DEPTH_CANNOT_DISCARD_STENCIL);
+	bool forceDepthWritesOff = id.Bit(FS_BIT_DEPTH_TEST_NEVER);
+
+	bool useDiscardStencilBugWorkaround = id.Bit(FS_BIT_NO_DEPTH_CANNOT_DISCARD_STENCIL) && !forceDepthWritesOff;
 
 	GEBlendSrcFactor replaceBlendFuncA = (GEBlendSrcFactor)id.Bits(FS_BIT_BLENDFUNC_A, 4);
 	GEBlendDstFactor replaceBlendFuncB = (GEBlendDstFactor)id.Bits(FS_BIT_BLENDFUNC_B, 4);
@@ -177,7 +179,7 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 	}
 
 	bool needFragCoord = readFramebufferTex || gstate_c.Use(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT);
-	bool writeDepth = gstate_c.Use(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT);
+	bool writeDepth = gstate_c.Use(GPU_ROUND_FRAGMENT_DEPTH_TO_16BIT) && !forceDepthWritesOff;
 
 	// TODO: We could have a separate mechanism to support more ops using the shader blending mechanism,
 // on hardware that can do proper bit math in fragment shaders.

GPU/Common/ShaderId.cpp

@@ -196,6 +196,7 @@ std::string FragmentShaderDesc(const FShaderID &id) {
 	if (id.Bit(FS_BIT_FLATSHADE)) desc << "Flat ";
 	if (id.Bit(FS_BIT_BGRA_TEXTURE)) desc << "BGRA ";
 	if (id.Bit(FS_BIT_UBERSHADER)) desc << "FragUber ";
+	if (id.Bit(FS_BIT_DEPTH_TEST_NEVER)) desc << "DepthNever ";
 	switch ((ShaderDepalMode)id.Bits(FS_BIT_SHADER_DEPAL_MODE, 2)) {
 	case ShaderDepalMode::OFF: break;
 	case ShaderDepalMode::NORMAL: desc << "Depal ";  break;
@@ -397,6 +398,14 @@ void ComputeFragmentShaderID(FShaderID *id_out, const ComputedPipelineState &pip
 			}
 		}
 
+		// Forcibly disable NEVER + depth-write on Mali.
+		// TODO: Take this from computed depth test instead of directly from the gstate.
+		// That will take more refactoring though.
+		if (bugs.Has(Draw::Bugs::NO_DEPTH_CANNOT_DISCARD_STENCIL_MALI) &&
+			gstate.getDepthTestFunction() == GE_COMP_NEVER && gstate.isDepthTestEnabled()) {
+			id.SetBit(FS_BIT_DEPTH_TEST_NEVER);
+		}
+
 		// In case the USE flag changes (for example, in multisampling we might disable input attachments),
 		// we don't want to accidentally use the wrong cached shader here. So moved it to a bit.
 		if (FragmentIdNeedsFramebufferRead(id)) {

GPU/Common/ShaderId.h

@@ -102,6 +102,7 @@ enum FShaderBit : uint8_t {
 	FS_BIT_STEREO = 58,
 	FS_BIT_USE_FRAMEBUFFER_FETCH = 59,
 	FS_BIT_UBERSHADER = 60,
+	FS_BIT_DEPTH_TEST_NEVER = 61,  // Only used on Mali. Set when depth == NEVER. We forcibly avoid writing to depth in this case, since it crashes the driver.
 };
 
 static inline FShaderBit operator +(FShaderBit bit, int i) {

GPU/GPUCommonHW.cpp

@@ -124,7 +124,7 @@ const CommonCommandTableEntry commonCommandTable[] = {
 	{ GE_CMD_BLENDFIXEDB, FLAG_FLUSHBEFOREONCHANGE, DIRTY_BLEND_STATE | DIRTY_FRAGMENTSHADER_STATE },
 	{ GE_CMD_MASKRGB, FLAG_FLUSHBEFOREONCHANGE, DIRTY_BLEND_STATE | DIRTY_FRAGMENTSHADER_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_COLORWRITEMASK },
 	{ GE_CMD_MASKALPHA, FLAG_FLUSHBEFOREONCHANGE, DIRTY_BLEND_STATE | DIRTY_FRAGMENTSHADER_STATE | DIRTY_DEPTHSTENCIL_STATE | DIRTY_COLORWRITEMASK },
-	{ GE_CMD_ZTEST, FLAG_FLUSHBEFOREONCHANGE, DIRTY_DEPTHSTENCIL_STATE },
+	{ GE_CMD_ZTEST, FLAG_FLUSHBEFOREONCHANGE, DIRTY_DEPTHSTENCIL_STATE | DIRTY_FRAGMENTSHADER_STATE },
 	{ GE_CMD_ZTESTENABLE, FLAG_FLUSHBEFOREONCHANGE, DIRTY_DEPTHSTENCIL_STATE | DIRTY_FRAGMENTSHADER_STATE },
 	{ GE_CMD_ZWRITEDISABLE, FLAG_FLUSHBEFOREONCHANGE, DIRTY_DEPTHSTENCIL_STATE | DIRTY_FRAGMENTSHADER_STATE },
 	{ GE_CMD_LOGICOP, FLAG_FLUSHBEFOREONCHANGE, DIRTY_BLEND_STATE | DIRTY_FRAGMENTSHADER_STATE },