MacroHLE: Add HLE replacement for base vertex and base instance.

12 files changed, 91 insertions, 6 deletions

diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp
index 73b67f0af..e4802bf9e 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp
@@ -339,6 +339,10 @@ Id EmitGetAttribute(EmitContext& ctx, IR::Attribute attr, Id vertex) {
             const Id base{ctx.OpLoad(ctx.U32[1], ctx.base_vertex)};
             return ctx.OpBitcast(ctx.F32[1], ctx.OpISub(ctx.U32[1], index, base));
         }
+    case IR::Attribute::BaseInstance:
+        return ctx.OpBitcast(ctx.F32[1], ctx.OpLoad(ctx.U32[1], ctx.base_instance));
+    case IR::Attribute::BaseVertex:
+        return ctx.OpBitcast(ctx.F32[1], ctx.OpLoad(ctx.U32[1], ctx.base_vertex));
     case IR::Attribute::FrontFace:
         return ctx.OpSelect(ctx.F32[1], ctx.OpLoad(ctx.U1, ctx.front_face),
                             ctx.OpBitcast(ctx.F32[1], ctx.Const(std::numeric_limits<u32>::max())),
@@ -380,6 +384,10 @@ Id EmitGetAttributeU32(EmitContext& ctx, IR::Attribute attr, Id) {
             const Id base{ctx.OpLoad(ctx.U32[1], ctx.base_vertex)};
             return ctx.OpISub(ctx.U32[1], index, base);
         }
+    case IR::Attribute::BaseInstance:
+        return ctx.OpLoad(ctx.U32[1], ctx.base_instance);
+    case IR::Attribute::BaseVertex:
+        return ctx.OpLoad(ctx.U32[1], ctx.base_vertex);
     default:
         throw NotImplementedException("Read U32 attribute {}", attr);
     }
diff --git a/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp b/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
index 41dc6d031..563a5fc49 100644
--- a/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
+++ b/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
@@ -1379,18 +1379,28 @@ void EmitContext::DefineInputs(const IR::Program& program) {
     if (loads[IR::Attribute::InstanceId]) {
         if (profile.support_vertex_instance_id) {
             instance_id = DefineInput(*this, U32[1], true, spv::BuiltIn::InstanceId);
+            if (loads[IR::Attribute::BaseInstance]) {
+                base_instance = DefineInput(*this, U32[1], true, spv::BuiltIn::BaseVertex);
+            }
         } else {
             instance_index = DefineInput(*this, U32[1], true, spv::BuiltIn::InstanceIndex);
             base_instance = DefineInput(*this, U32[1], true, spv::BuiltIn::BaseInstance);
         }
+    } else if (loads[IR::Attribute::BaseInstance]) {
+        base_instance = DefineInput(*this, U32[1], true, spv::BuiltIn::BaseInstance);
     }
     if (loads[IR::Attribute::VertexId]) {
         if (profile.support_vertex_instance_id) {
             vertex_id = DefineInput(*this, U32[1], true, spv::BuiltIn::VertexId);
+            if (loads[IR::Attribute::BaseVertex]) {
+                base_vertex = DefineInput(*this, U32[1], true, spv::BuiltIn::BaseVertex);
+            }
         } else {
             vertex_index = DefineInput(*this, U32[1], true, spv::BuiltIn::VertexIndex);
             base_vertex = DefineInput(*this, U32[1], true, spv::BuiltIn::BaseVertex);
         }
+    } else if (loads[IR::Attribute::BaseVertex]) {
+        base_vertex = DefineInput(*this, U32[1], true, spv::BuiltIn::BaseVertex);
     }
     if (loads[IR::Attribute::FrontFace]) {
         front_face = DefineInput(*this, U1, true, spv::BuiltIn::FrontFacing);
diff --git a/src/shader_recompiler/environment.h b/src/shader_recompiler/environment.h
index 402f2664f..b9b4455f6 100644
--- a/src/shader_recompiler/environment.h
+++ b/src/shader_recompiler/environment.h
@@ -34,6 +34,11 @@ public:
 
     [[nodiscard]] virtual std::array<u32, 3> WorkgroupSize() const = 0;
 
+    [[nodiscard]] virtual bool HasHLEMacroState() const = 0;
+
+    [[nodiscard]] virtual std::optional<ReplaceConstant> GetReplaceConstBuffer(
+        u32 bank, u32 offset) = 0;
+
     virtual void Dump(u64 hash) = 0;
 
     [[nodiscard]] const ProgramHeader& SPH() const noexcept {
diff --git a/src/shader_recompiler/frontend/ir/attribute.cpp b/src/shader_recompiler/frontend/ir/attribute.cpp
index 7d3d882e4..73e189a89 100644
--- a/src/shader_recompiler/frontend/ir/attribute.cpp
+++ b/src/shader_recompiler/frontend/ir/attribute.cpp
@@ -446,6 +446,10 @@ std::string NameOf(Attribute attribute) {
         return "ViewportMask";
     case Attribute::FrontFace:
         return "FrontFace";
+    case Attribute::BaseInstance:
+        return "BaseInstance";
+    case Attribute::BaseVertex:
+        return "BaseVertex";
     }
     return fmt::format("<reserved attribute {}>", static_cast<int>(attribute));
 }
diff --git a/src/shader_recompiler/frontend/ir/attribute.h b/src/shader_recompiler/frontend/ir/attribute.h
index 6ee3947b1..364d8a912 100644
--- a/src/shader_recompiler/frontend/ir/attribute.h
+++ b/src/shader_recompiler/frontend/ir/attribute.h
@@ -219,6 +219,10 @@ enum class Attribute : u64 {
     FixedFncTexture9Q = 231,
     ViewportMask = 232,
     FrontFace = 255,
+
+    // Implementation attributes
+    BaseInstance = 256,
+    BaseVertex = 257,
 };
 
 constexpr size_t NUM_GENERICS = 32;
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.cpp b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
index 0cdac0eff..eb2e49a68 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.cpp
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.cpp
@@ -294,6 +294,14 @@ F32 IREmitter::GetAttribute(IR::Attribute attribute, const U32& vertex) {
     return Inst<F32>(Opcode::GetAttribute, attribute, vertex);
 }
 
+U32 IREmitter::GetAttributeU32(IR::Attribute attribute) {
+    return GetAttributeU32(attribute, Imm32(0));
+}
+
+U32 IREmitter::GetAttributeU32(IR::Attribute attribute, const U32& vertex) {
+    return Inst<U32>(Opcode::GetAttributeU32, attribute, vertex);
+}
+
 void IREmitter::SetAttribute(IR::Attribute attribute, const F32& value, const U32& vertex) {
     Inst(Opcode::SetAttribute, attribute, value, vertex);
 }
diff --git a/src/shader_recompiler/frontend/ir/ir_emitter.h b/src/shader_recompiler/frontend/ir/ir_emitter.h
index 2df992feb..7aaaa4ab0 100644
--- a/src/shader_recompiler/frontend/ir/ir_emitter.h
+++ b/src/shader_recompiler/frontend/ir/ir_emitter.h
@@ -74,6 +74,8 @@ public:
 
     [[nodiscard]] F32 GetAttribute(IR::Attribute attribute);
     [[nodiscard]] F32 GetAttribute(IR::Attribute attribute, const U32& vertex);
+    [[nodiscard]] U32 GetAttributeU32(IR::Attribute attribute);
+    [[nodiscard]] U32 GetAttributeU32(IR::Attribute attribute, const U32& vertex);
     void SetAttribute(IR::Attribute attribute, const F32& value, const U32& vertex);
 
     [[nodiscard]] F32 GetAttributeIndexed(const U32& phys_address);
diff --git a/src/shader_recompiler/frontend/maxwell/translate_program.cpp b/src/shader_recompiler/frontend/maxwell/translate_program.cpp
index 3adbd2b16..ac159d24b 100644
--- a/src/shader_recompiler/frontend/maxwell/translate_program.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate_program.cpp
@@ -219,7 +219,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     }
     Optimization::SsaRewritePass(program);
 
-    Optimization::ConstantPropagationPass(program);
+    Optimization::ConstantPropagationPass(env, program);
 
     Optimization::PositionPass(env, program);
 
diff --git a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
index 826f9a54a..ac10405f3 100644
--- a/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
+++ b/src/shader_recompiler/ir_opt/constant_propagation_pass.cpp
@@ -7,6 +7,7 @@
 #include <type_traits>
 
 #include "common/bit_cast.h"
+#include "shader_recompiler/environment.h"
 #include "shader_recompiler/exception.h"
 #include "shader_recompiler/frontend/ir/ir_emitter.h"
 #include "shader_recompiler/frontend/ir/value.h"
@@ -515,6 +516,8 @@ void FoldBitCast(IR::Inst& inst, IR::Opcode reverse) {
             case IR::Attribute::PrimitiveId:
             case IR::Attribute::InstanceId:
             case IR::Attribute::VertexId:
+            case IR::Attribute::BaseVertex:
+            case IR::Attribute::BaseInstance:
                 break;
             default:
                 return;
@@ -644,7 +647,37 @@ void FoldFSwizzleAdd(IR::Block& block, IR::Inst& inst) {
     }
 }
 
-void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
+void FoldConstBuffer(Environment& env, IR::Block& block, IR::Inst& inst) {
+    const IR::Value bank{inst.Arg(0)};
+    const IR::Value offset{inst.Arg(1)};
+    if (!bank.IsImmediate() || !offset.IsImmediate()) {
+        return;
+    }
+    const auto bank_value = bank.U32();
+    const auto offset_value = offset.U32();
+    auto replacement = env.GetReplaceConstBuffer(bank_value, offset_value);
+    if (!replacement) {
+        return;
+    }
+    const auto new_attribute = [replacement]() {
+        switch (*replacement) {
+        case ReplaceConstant::BaseInstance:
+            return IR::Attribute::BaseInstance;
+        case ReplaceConstant::BaseVertex:
+            return IR::Attribute::BaseVertex;
+        default:
+            throw NotImplementedException("Not implemented replacement variable {}", *replacement);
+        }
+    }();
+    IR::IREmitter ir{block, IR::Block::InstructionList::s_iterator_to(inst)};
+    if (inst.GetOpcode() == IR::Opcode::GetCbufU32) {
+        inst.ReplaceUsesWith(ir.GetAttributeU32(new_attribute));
+    } else {
+        inst.ReplaceUsesWith(ir.GetAttribute(new_attribute));
+    }
+}
+
+void ConstantPropagation(Environment& env, IR::Block& block, IR::Inst& inst) {
     switch (inst.GetOpcode()) {
     case IR::Opcode::GetRegister:
         return FoldGetRegister(inst);
@@ -789,18 +822,24 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
                                     IR::Opcode::CompositeInsertF16x4);
     case IR::Opcode::FSwizzleAdd:
         return FoldFSwizzleAdd(block, inst);
+    case IR::Opcode::GetCbufF32:
+    case IR::Opcode::GetCbufU32:
+        if (env.HasHLEMacroState()) {
+            return FoldConstBuffer(env, block, inst);
+        }
+        break;
     default:
         break;
     }
 }
 } // Anonymous namespace
 
-void ConstantPropagationPass(IR::Program& program) {
+void ConstantPropagationPass(Environment& env, IR::Program& program) {
     const auto end{program.post_order_blocks.rend()};
     for (auto it = program.post_order_blocks.rbegin(); it != end; ++it) {
         IR::Block* const block{*it};
         for (IR::Inst& inst : block->Instructions()) {
-            ConstantPropagation(*block, inst);
+            ConstantPropagation(env, *block, inst);
         }
     }
 }
diff --git a/src/shader_recompiler/ir_opt/passes.h b/src/shader_recompiler/ir_opt/passes.h
index 11bfe801a..1f8f2ba95 100644
--- a/src/shader_recompiler/ir_opt/passes.h
+++ b/src/shader_recompiler/ir_opt/passes.h
@@ -13,7 +13,7 @@ struct HostTranslateInfo;
 namespace Shader::Optimization {
 
 void CollectShaderInfoPass(Environment& env, IR::Program& program);
-void ConstantPropagationPass(IR::Program& program);
+void ConstantPropagationPass(Environment& env, IR::Program& program);
 void DeadCodeEliminationPass(IR::Program& program);
 void GlobalMemoryToStorageBufferPass(IR::Program& program);
 void IdentityRemovalPass(IR::Program& program);
diff --git a/src/shader_recompiler/shader_info.h b/src/shader_recompiler/shader_info.h
index d9c6e92db..ea0f48344 100644
--- a/src/shader_recompiler/shader_info.h
+++ b/src/shader_recompiler/shader_info.h
@@ -16,6 +16,11 @@
 
 namespace Shader {
 
+enum class ReplaceConstant : u32 {
+    BaseInstance,
+    BaseVertex,
+};
+
 enum class TextureType : u32 {
     Color1D,
     ColorArray1D,
diff --git a/src/shader_recompiler/varying_state.h b/src/shader_recompiler/varying_state.h
index 7b28a285f..18a9aaf50 100644
--- a/src/shader_recompiler/varying_state.h
+++ b/src/shader_recompiler/varying_state.h
@@ -11,7 +11,7 @@
 namespace Shader {
 
 struct VaryingState {
-    std::bitset<256> mask{};
+    std::bitset<512> mask{};
 
     void Set(IR::Attribute attribute, bool state = true) {
         mask[static_cast<size_t>(attribute)] = state;