gpu: route direct shader register ranges through handlers

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -2853,6 +2853,20 @@ KYTY_CP_OP_PARSER(CpOpSetShaderReg) {
 	auto pfunc = g_hw_sh_func[cmd_offset];
 
 	if (pfunc == nullptr) {
+		auto num_values = KYTY_PM4_LEN(cmd_id) - 2u;
+		bool handled    = num_values != 0;
+		for (uint32_t i = 0; i < num_values; i++) {
+			if (cmd_offset + i >= Pm4::SH_NUM || g_hw_sh_indirect_func[cmd_offset + i] == nullptr) {
+				handled = false;
+				break;
+			}
+		}
+		if (handled) {
+			for (uint32_t i = 0; i < num_values; i++) {
+				g_hw_sh_indirect_func[cmd_offset + i](cp, cmd_offset + i, buffer[1 + i]);
+			}
+			return num_values + 1u;
+		}
 		EXIT("unknown shader register\n\t%05" PRIx32 ":\n\tcmd_id = %08" PRIx32
 		     "\n\tcmd_offset = %08" PRIx32 "\n",
 		     num_dw - dw, cmd_id, cmd_offset);

tests/ShaderRecompilerComputeTests.cpp

@@ -22084,6 +22084,25 @@ void CheckPm4StencilInfoValueLane(RenderContext &renderer) {
   std::printf("[host]    %-32s ok\n", "Pm4StencilInfoValueLane");
 }
 
+void CheckPm4DirectShaderRegisterFallback(RenderContext &renderer) {
+  GraphicsInitJmpTables();
+  CommandProcessor processor(renderer);
+  constexpr uint64_t shader_address = 0x0000123456789a00ull;
+  std::array<uint32_t, 4> packet{
+      KYTY_PM4(4, Pm4::IT_SET_SH_REG, Pm4::R_ZERO),
+      Pm4::SPI_SHADER_PGM_LO_ES,
+      static_cast<uint32_t>(shader_address >> 8u),
+      static_cast<uint32_t>(shader_address >> 40u),
+  };
+  Pm4Execution execution;
+  Require("Pm4DirectShaderRegisterFallback", "ES program base",
+          processor.Process(execution, packet.data(), packet.size()) ==
+                  Pm4ProcessResult::Complete &&
+              processor.GetShCtx().GetVs().es_regs.data_addr == shader_address,
+          "direct SET_SH_REG did not route its ES base pair through the register handlers");
+  std::printf("[host]    %-32s ok\n", "Pm4DirectShaderRegisterFallback");
+}
+
 void CheckPm4PolygonOffsetRegisters(RenderContext &renderer) {
   GraphicsInitJmpTables();
   CommandProcessor processor(renderer);
@@ -22676,6 +22695,7 @@ int main(int argc, char **argv) {
   CheckVulkan13FeatureRequirements();
   CheckPm4AcquireMemNoOp(vulkan.RuntimeRenderer());
   CheckPm4StencilInfoValueLane(vulkan.RuntimeRenderer());
+  CheckPm4DirectShaderRegisterFallback(vulkan.RuntimeRenderer());
   CheckPm4PolygonOffsetRegisters(vulkan.RuntimeRenderer());
   CheckPm4ContextStateOperations(vulkan.RuntimeRenderer());
   CheckPm4WaitResume(vulkan.RuntimeRenderer());