recompiler: added support for discard on export with masked EXEC

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -45,6 +45,7 @@ void EmitSetVccHi(EmitContext& ctx);
 void EmitPrologue(EmitContext& ctx);
 void EmitEpilogue(EmitContext& ctx);
 void EmitDiscard(EmitContext& ctx);
+void EmitDiscardCond(EmitContext& ctx, Id condition);
 void EmitBarrier(EmitContext& ctx);
 void EmitWorkgroupMemoryBarrier(EmitContext& ctx);
 void EmitDeviceMemoryBarrier(EmitContext& ctx);

src/shader_recompiler/backend/spirv/emit_spirv_special.cpp

@@ -14,6 +14,17 @@ void EmitDiscard(EmitContext& ctx) {
     ctx.OpDemoteToHelperInvocationEXT();
 }
 
+void EmitDiscardCond(EmitContext& ctx, Id condition) {
+    const Id kill_label{ctx.OpLabel()};
+    const Id merge_label{ctx.OpLabel()};
+    ctx.OpSelectionMerge(merge_label, spv::SelectionControlMask::MaskNone);
+    ctx.OpBranchConditional(condition, kill_label, merge_label);
+    ctx.AddLabel(kill_label);
+    ctx.OpDemoteToHelperInvocationEXT();
+    ctx.OpBranch(merge_label);
+    ctx.AddLabel(merge_label);
+}
+
 void EmitEmitVertex(EmitContext& ctx, const IR::Value& stream) {
     throw NotImplementedException("Geometry streams");
 }

src/shader_recompiler/frontend/control_flow_graph.cpp

@@ -121,7 +121,7 @@ void CFG::EmitBlocks() {
 
 void CFG::LinkBlocks() {
     const auto get_block = [this](u32 address) {
-        const auto it = blocks.find(address, Compare{});
+        auto it = blocks.find(address, Compare{});
         ASSERT_MSG(it != blocks.end() && it->begin == address);
         return &*it;
     };
@@ -131,7 +131,10 @@ void CFG::LinkBlocks() {
         // If the block doesn't end with a branch we simply
         // need to link with the next block.
         if (!end_inst.IsTerminateInstruction()) {
-            block.branch_true = get_block(block.end);
+            auto* next_block = get_block(block.end);
+            ++next_block->num_predecessors;
+
+            block.branch_true = next_block;
             block.end_class = EndClass::Branch;
             continue;
         }
@@ -141,11 +144,20 @@ void CFG::LinkBlocks() {
         const u32 branch_pc = block.end - end_inst.length;
         const u32 target_pc = end_inst.BranchTarget(branch_pc);
         if (end_inst.IsUnconditionalBranch()) {
-            block.branch_true = get_block(target_pc);
+            auto* target_block = get_block(target_pc);
+            ++target_block->num_predecessors;
+
+            block.branch_true = target_block;
             block.end_class = EndClass::Branch;
         } else if (end_inst.IsConditionalBranch()) {
-            block.branch_true = get_block(target_pc);
+            auto* target_block = get_block(target_pc);
-            block.branch_false = get_block(block.end);
+            ++target_block->num_predecessors;
+
+            auto* end_block = get_block(block.end);
+            ++end_block->num_predecessors;
+
+            block.branch_true = target_block;
+            block.branch_false = end_block;
             block.end_class = EndClass::Branch;
         } else if (end_inst.opcode == Opcode::S_ENDPGM) {
             const auto& prev_inst = inst_list[block.end_index - 1];

src/shader_recompiler/frontend/control_flow_graph.h

@@ -36,6 +36,7 @@ struct Block : Hook {
     u32 end;
     u32 begin_index;
     u32 end_index;
+    u32 num_predecessors{};
     IR::Condition cond{};
     GcnInst end_inst{};
     EndClass end_class{};

src/shader_recompiler/frontend/structured_control_flow.cpp

@@ -631,6 +631,7 @@ private:
             case StatementType::Code: {
                 ensure_block();
                 if (!stmt.block->is_dummy) {
+                    current_block->has_multiple_predecessors = stmt.block->num_predecessors > 1;
                     const u32 start = stmt.block->begin_index;
                     const u32 size = stmt.block->end_index - start + 1;
                     Translate(current_block, stmt.block->begin, inst_list.subspan(start, size),

src/shader_recompiler/frontend/translate/export.cpp

@@ -1,11 +1,17 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#include "common/logging/log.h"
 #include "shader_recompiler/frontend/translate/translate.h"
 
 namespace Shader::Gcn {
 
 void Translator::EXP(const GcnInst& inst) {
+    if (ir.block->has_multiple_predecessors) {
+        LOG_WARNING(Render_Recompiler, "An ambiguous export appeared in translation");
+        ir.Discard(ir.LogicalNot(ir.GetExec()));
+    }
+
     const auto& exp = inst.control.exp;
     const IR::Attribute attrib{exp.target};
     const std::array vsrc = {

src/shader_recompiler/ir/basic_block.h

@@ -149,6 +149,8 @@ public:
     std::array<Value, NumScalarRegs> ssa_sreg_values;
     std::array<Value, NumVectorRegs> ssa_vreg_values;
 
+    bool has_multiple_predecessors{false};
+
 private:
     /// Memory pool for instruction list
     ObjectPool<Inst>* inst_pool;

src/shader_recompiler/ir/ir_emitter.cpp

@@ -115,6 +115,10 @@ void IREmitter::Discard() {
     Inst(Opcode::Discard);
 }
 
+void IREmitter::Discard(const U1& cond) {
+    Inst(Opcode::DiscardCond, cond);
+}
+
 void IREmitter::Barrier() {
     Inst(Opcode::Barrier);
 }

src/shader_recompiler/ir/ir_emitter.h

@@ -42,6 +42,7 @@ public:
     void Prologue();
     void Epilogue();
     void Discard();
+    void Discard(const U1& cond);
 
     void Barrier();
     void WorkgroupMemoryBarrier();

src/shader_recompiler/ir/microinstruction.cpp

@@ -49,6 +49,7 @@ bool Inst::MayHaveSideEffects() const noexcept {
     case Opcode::Prologue:
     case Opcode::Epilogue:
     case Opcode::Discard:
+    case Opcode::DiscardCond:
     case Opcode::SetAttribute:
     case Opcode::StoreBufferF32:
     case Opcode::StoreBufferF32x2:

src/shader_recompiler/ir/opcodes.inc

@@ -13,6 +13,7 @@ OPCODE(PhiMove,                                             Void,           Opaq
 OPCODE(Prologue,                                            Void,                                                                                           )
 OPCODE(Epilogue,                                            Void,                                                                                           )
 OPCODE(Discard,                                             Void,                                                                                           )
+OPCODE(DiscardCond,                                         Void,           U1,                                                                             )
 
 // Constant memory operations
 OPCODE(ReadConst,                                           U32,            U32x2,          U32,                                                            )

src/shader_recompiler/ir/passes/shader_info_collection_pass.cpp

@@ -37,6 +37,7 @@ void Visit(Info& info, IR::Inst& inst) {
         info.uses_group_quad = true;
         break;
     case IR::Opcode::Discard:
+    case IR::Opcode::DiscardCond:
         info.has_discard = true;
         break;
     case IR::Opcode::ImageGather: