shader: fix crash from packed ancillary used as DPP old operand (#1769)

shader: emit plain EXEC selects for DPP writes that reach every active lane (Street Fighter 6)

src/graphics/shader/recompiler/backend/spirv/spirvEmitterFlow.cpp

@@ -132,7 +132,7 @@ uint32_t EmitDppWriteCondition(ValueEmitContext& ctx, const IR::DppMoveFlags& fl
 	enable(flags.bank_mask, 2u);
 	enable(flags.row_mask, 4u);
 	// Only row shifts (0x101-0x10f, 0x111-0x11f) can vacate a lane; other controls always read one.
-	const bool row_shift = !flags.dpp8 && flags.control > 0x100u && flags.control < 0x120u &&
+	const bool row_shift = flags.control > 0x100u && flags.control < 0x120u &&
 	                       flags.control != 0x110u;
 	if (!flags.bound_control && row_shift) {
 		writable = Binary(state, spv::OpLogicalAnd, TypeBool(state), writable,

src/graphics/shader/recompiler/frontend/translate/Translate.cpp

@@ -319,7 +319,10 @@ void Translator::WriteRawU32(const Decoder::Operand& operand, IR::U32 value) {
 		case Decoder::OperandKind::Vgpr: {
 			const auto reg = static_cast<IR::VectorReg>(operand.reg);
 			const auto old = ir.GetVectorReg(reg);
-			if (operand.dpp) {
+			// DPP skips EXEC lanes only through row/bank masks, or for invalid sources without
+			// BOUND_CTRL (dpp8 reads zero); otherwise it is a plain masked write.
+			if (operand.dpp && (operand.dpp_row_mask != 0xfu || operand.dpp_bank_mask != 0xfu ||
+			                    (!operand.dpp_bound_ctrl && !operand.dpp8))) {
 				value = IR::U32(ir.Emit(IR::ValueOpcode::DppUpdateU32, {value, old, ir.GetExec()},
 				                        DppFlags(operand)));
 			} else {

tests/shaderCfgTests.cpp

@@ -7045,6 +7045,17 @@ void TestPixelAncillaryLayerInput() {
     CheckSpirvBinaryValidates(wqm_result.spirv);
   }
 
+  uint32_t dpp[] = {
+      EncodeVop3Word0(0x148, 6), EncodeVop3Word1(5 + 256, 128 + 16, 128 + 11),
+      EncodeVop1(0x01, 5, 250), EncodeVop1Dpp(6) | (1u << 19u), // Bound-zero DPP overwrite.
+      EncodeExp0(0x00, 0x1), EncodeExp1(5, 0, 0, 0), EncodeSopp(0x01),
+  };
+  auto dpp_result = RecompileForTest(dpp, options);
+  Check(ProgramHasInput(dpp_result.program, StageInputKind::Layer) &&
+            !ProgramHasInput(dpp_result.program, StageInputKind::PackedAncillary),
+        "DPP write retained the old ancillary register value");
+  CheckSpirvBinaryValidates(dpp_result.spirv);
+
   const uint32_t loop[] = {
       EncodeSop1(0x04, 8, 126),  // Save the original EXEC mask.
       EncodeSop1(0x0a, 126, 8),   // Widen for helper lanes.
@@ -7098,6 +7109,9 @@ void TestPixelAncillaryLayerInput() {
   shader[5] = EncodeExp1(6, 7, 5, 0);
   ExpectFatal([&] { (void)RecompileForTest(shader, options); },
               "direct raw ancillary export was silently replaced after field lowering");
+  dpp[3] = EncodeVop1Dpp(6); // Without BOUND_CTRL, lanes with inactive sources keep the value.
+  ExpectFatal([&] { (void)RecompileForTest(dpp, options); },
+              "DPP write without BOUND_CTRL silently replaced the ancillary value");
 #endif
 }