video_core: Bloodborne stabilization pt1 (#543)

* shader_recompiler: Writelane elimination pass + null image fix

* spirv: Implement image derivatives

* texture_cache: Reduce page bit size

* clang format

* slot_vector: Back to debug assert

* vk_graphics_pipeline: Handle null tsharp

* spirv: Revert some change

* vk_instance: Support primitive restart on list topology

* page_manager: Adjust windows exception handler

* clang format

* Remove subres tracking

* Will be done separately

src/common/slot_vector.h

@@ -28,9 +28,13 @@ struct SlotId {
 
 template <class T>
 class SlotVector {
-    constexpr static std::size_t InitialCapacity = 1024;
+    constexpr static std::size_t InitialCapacity = 2048;
 
 public:
+    SlotVector() {
+        Reserve(InitialCapacity);
+    }
+
     ~SlotVector() noexcept {
         std::size_t index = 0;
         for (u64 bits : stored_bitset) {
@@ -67,19 +71,6 @@ public:
         return SlotId{index};
     }
 
-    template <typename... Args>
-    [[nodiscard]] SlotId swap_and_insert(SlotId existing_id, Args&&... args) noexcept {
-        const u32 index = FreeValueIndex();
-        T& existing_value = values[existing_id.index].object;
-
-        new (&values[index].object) T(std::move(existing_value));
-        existing_value.~T();
-        new (&values[existing_id.index].object) T(std::forward<Args>(args)...);
-        SetStorageBit(index);
-
-        return SlotId{index};
-    }
-
     void erase(SlotId id) noexcept {
         values[id.index].object.~T();
         free_list.push_back(id.index);
@@ -151,7 +142,8 @@ private:
 
         const std::size_t old_free_size = free_list.size();
         free_list.resize(old_free_size + (new_capacity - values_capacity));
-        std::iota(free_list.begin() + old_free_size, free_list.end(),
+        const std::size_t new_free_size = free_list.size();
+        std::iota(free_list.rbegin(), free_list.rbegin() + new_free_size - old_free_size,
                   static_cast<u32>(values_capacity));
 
         delete[] values;

src/core/libraries/kernel/thread_management.cpp

@@ -1123,7 +1123,6 @@ int PS4_SYSV_ABI posix_pthread_join(ScePthread thread, void** res) {
 }
 
 int PS4_SYSV_ABI scePthreadDetach(ScePthread thread) {
-    LOG_INFO(Kernel_Pthread, "thread create name = {}", thread->name);
     thread->is_detached = true;
     return ORBIS_OK;
 }

src/shader_recompiler/backend/spirv/emit_spirv_image.cpp

@@ -16,6 +16,12 @@ struct ImageOperands {
                                                    static_cast<u32>(new_mask));
         operands.push_back(value);
     }
+    void Add(spv::ImageOperandsMask new_mask, Id value1, Id value2) {
+        mask = static_cast<spv::ImageOperandsMask>(static_cast<u32>(mask) |
+                                                   static_cast<u32>(new_mask));
+        operands.push_back(value1);
+        operands.push_back(value2);
+    }
 
     void AddOffset(EmitContext& ctx, const IR::Value& offset,
                    bool can_use_runtime_offsets = false) {
@@ -53,6 +59,15 @@ struct ImageOperands {
         }
     }
 
+    void AddDerivatives(EmitContext& ctx, Id derivatives) {
+        if (!Sirit::ValidId(derivatives)) {
+            return;
+        }
+        const Id dx{ctx.OpVectorShuffle(ctx.F32[2], derivatives, derivatives, 0, 1)};
+        const Id dy{ctx.OpVectorShuffle(ctx.F32[2], derivatives, derivatives, 2, 3)};
+        Add(spv::ImageOperandsMask::Grad, dx, dy);
+    }
+
     spv::ImageOperandsMask mask{};
     boost::container::static_vector<Id, 4> operands;
 };
@@ -117,7 +132,7 @@ Id EmitImageGather(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords,
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     const u32 comp = inst->Flags<IR::TextureInstInfo>().gather_comp.Value();
     ImageOperands operands;
-    operands.AddOffset(ctx, offset);
+    operands.AddOffset(ctx, offset, true);
     return ctx.OpImageGather(ctx.F32[4], sampled_image, coords, ctx.ConstU32(comp), operands.mask,
                              operands.operands);
 }
@@ -129,7 +144,7 @@ Id EmitImageGatherDref(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords,
     const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
     const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
     ImageOperands operands;
-    operands.AddOffset(ctx, offset);
+    operands.AddOffset(ctx, offset, true);
     return ctx.OpImageDrefGather(ctx.F32[4], sampled_image, coords, dref, operands.mask,
                                  operands.operands);
 }
@@ -181,9 +196,17 @@ Id EmitImageQueryLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords) {
     return ctx.OpImageQueryLod(ctx.F32[2], sampled_image, coords);
 }
 
-Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
+Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id derivatives,
-                     Id derivatives, const IR::Value& offset, Id lod_clamp) {
+                     const IR::Value& offset, Id lod_clamp) {
-    UNREACHABLE_MSG("SPIR-V Instruction");
+    const auto& texture = ctx.images[handle & 0xFFFF];
+    const Id image = ctx.OpLoad(texture.image_type, texture.id);
+    const Id sampler = ctx.OpLoad(ctx.sampler_type, ctx.samplers[handle >> 16]);
+    const Id sampled_image = ctx.OpSampledImage(texture.sampled_type, image, sampler);
+    ImageOperands operands;
+    operands.AddDerivatives(ctx, derivatives);
+    operands.AddOffset(ctx, offset);
+    return ctx.OpImageSampleExplicitLod(ctx.F32[4], sampled_image, coords, operands.mask,
+                                        operands.operands);
 }
 
 Id EmitImageRead(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords) {

src/shader_recompiler/backend/spirv/emit_spirv_instructions.h

@@ -387,8 +387,8 @@ Id EmitImageFetch(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, const
                   Id lod, Id ms);
 Id EmitImageQueryDimensions(EmitContext& ctx, IR::Inst* inst, u32 handle, Id lod, bool skip_mips);
 Id EmitImageQueryLod(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords);
-Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords,
+Id EmitImageGradient(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id derivatives,
-                     Id derivatives, const IR::Value& offset, Id lod_clamp);
+                     const IR::Value& offset, Id lod_clamp);
 Id EmitImageRead(EmitContext& ctx, IR::Inst* inst, const IR::Value& index, Id coords);
 void EmitImageWrite(EmitContext& ctx, IR::Inst* inst, u32 handle, Id coords, Id color);
 
@@ -407,5 +407,8 @@ Id EmitImageAtomicExchange32(EmitContext& ctx, IR::Inst* inst, u32 handle, Id co
 Id EmitLaneId(EmitContext& ctx);
 Id EmitWarpId(EmitContext& ctx);
 Id EmitQuadShuffle(EmitContext& ctx, Id value, Id index);
+Id EmitReadFirstLane(EmitContext& ctx, Id value);
+Id EmitReadLane(EmitContext& ctx, Id value, u32 lane);
+Id EmitWriteLane(EmitContext& ctx, Id value, Id write_value, u32 lane);
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp

@@ -22,4 +22,16 @@ Id EmitQuadShuffle(EmitContext& ctx, Id value, Id index) {
     return ctx.OpGroupNonUniformQuadBroadcast(ctx.U32[1], SubgroupScope(ctx), value, index);
 }
 
+Id EmitReadFirstLane(EmitContext& ctx, Id value) {
+    UNREACHABLE();
+}
+
+Id EmitReadLane(EmitContext& ctx, Id value, u32 lane) {
+    UNREACHABLE();
+}
+
+Id EmitWriteLane(EmitContext& ctx, Id value, Id write_value, u32 lane) {
+    return ctx.u32_zero_value;
+}
+
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/frontend/translate/data_share.cpp

@@ -127,7 +127,6 @@ void Translator::DS_ADD_U32(const GcnInst& inst, bool rtn) {
     const IR::U32 data{GetSrc(inst.src[1])};
     const IR::U32 offset = ir.Imm32(u32(inst.control.ds.offset0));
     const IR::U32 addr_offset = ir.IAdd(addr, offset);
-    IR::VectorReg dst_reg{inst.dst[0].code};
     const IR::Value original_val = ir.SharedAtomicIAdd(addr_offset, data);
     if (rtn) {
         SetDst(inst.dst[0], IR::U32{original_val});
@@ -139,7 +138,6 @@ void Translator::DS_MIN_U32(const GcnInst& inst, bool rtn) {
     const IR::U32 data{GetSrc(inst.src[1])};
     const IR::U32 offset = ir.Imm32(u32(inst.control.ds.offset0));
     const IR::U32 addr_offset = ir.IAdd(addr, offset);
-    IR::VectorReg dst_reg{inst.dst[0].code};
     const IR::Value original_val = ir.SharedAtomicIMin(addr_offset, data, false);
     if (rtn) {
         SetDst(inst.dst[0], IR::U32{original_val});
@@ -151,7 +149,6 @@ void Translator::DS_MAX_U32(const GcnInst& inst, bool rtn) {
     const IR::U32 data{GetSrc(inst.src[1])};
     const IR::U32 offset = ir.Imm32(u32(inst.control.ds.offset0));
     const IR::U32 addr_offset = ir.IAdd(addr, offset);
-    IR::VectorReg dst_reg{inst.dst[0].code};
     const IR::Value original_val = ir.SharedAtomicIMax(addr_offset, data, false);
     if (rtn) {
         SetDst(inst.dst[0], IR::U32{original_val});
@@ -168,13 +165,18 @@ void Translator::V_READFIRSTLANE_B32(const GcnInst& inst) {
 }
 
 void Translator::V_READLANE_B32(const GcnInst& inst) {
-    ASSERT(info.stage != Stage::Compute);
+    const IR::ScalarReg dst{inst.dst[0].code};
-    SetDst(inst.dst[0], GetSrc(inst.src[0]));
+    const IR::U32 value{GetSrc(inst.src[0])};
+    const IR::U32 lane{GetSrc(inst.src[1])};
+    ir.SetScalarReg(dst, ir.ReadLane(value, lane));
 }
 
 void Translator::V_WRITELANE_B32(const GcnInst& inst) {
-    ASSERT(info.stage != Stage::Compute);
+    const IR::VectorReg dst{inst.dst[0].code};
-    SetDst(inst.dst[0], GetSrc(inst.src[0]));
+    const IR::U32 value{GetSrc(inst.src[0])};
+    const IR::U32 lane{GetSrc(inst.src[1])};
+    const IR::U32 old_value{GetSrc(inst.dst[0])};
+    ir.SetVectorReg(dst, ir.WriteLane(old_value, value, lane));
 }
 
 } // namespace Shader::Gcn

src/shader_recompiler/frontend/translate/scalar_alu.cpp

@@ -440,13 +440,16 @@ void Translator::S_SUB_U32(const GcnInst& inst) {
 void Translator::S_GETPC_B64(u32 pc, const GcnInst& inst) {
     // This only really exists to let resource tracking pass know
     // there is an inline cbuf.
-    SetDst(inst.dst[0], ir.Imm32(pc));
+    const IR::ScalarReg dst{inst.dst[0].code};
+    ir.SetScalarReg(dst, ir.Imm32(pc));
+    ir.SetScalarReg(dst + 1, ir.Imm32(0));
 }
 
 void Translator::S_ADDC_U32(const GcnInst& inst) {
     const IR::U32 src0{GetSrc(inst.src[0])};
     const IR::U32 src1{GetSrc(inst.src[1])};
-    SetDst(inst.dst[0], ir.IAdd(ir.IAdd(src0, src1), ir.GetSccLo()));
+    const IR::U32 carry{ir.Select(ir.GetScc(), ir.Imm32(1U), ir.Imm32(0U))};
+    SetDst(inst.dst[0], ir.IAdd(ir.IAdd(src0, src1), carry));
 }
 
 void Translator::S_MAX_U32(const GcnInst& inst) {

src/shader_recompiler/frontend/translate/vector_memory.cpp

@@ -17,6 +17,7 @@ void Translator::EmitVectorMemory(const GcnInst& inst) {
     case Opcode::IMAGE_SAMPLE_C_O:
     case Opcode::IMAGE_SAMPLE_B:
     case Opcode::IMAGE_SAMPLE_C_LZ_O:
+    case Opcode::IMAGE_SAMPLE_D:
         return IMAGE_SAMPLE(inst);
     case Opcode::IMAGE_GATHER4_C:
     case Opcode::IMAGE_GATHER4_LZ:
@@ -162,12 +163,15 @@ void Translator::IMAGE_SAMPLE(const GcnInst& inst) {
         flags.test(MimgModifier::LodBias) ? ir.GetVectorReg<IR::F32>(addr_reg++) : IR::F32{};
     const IR::F32 dref =
         flags.test(MimgModifier::Pcf) ? ir.GetVectorReg<IR::F32>(addr_reg++) : IR::F32{};
-
+    const IR::Value derivatives = [&] -> IR::Value {
-    // Derivatives are tricky because their number depends on the texture type which is located in
+        if (!flags.test(MimgModifier::Derivative)) {
-    // T#. We don't have access to T# though until resource tracking pass. For now assume no
+            return {};
-    // derivatives are present, otherwise we don't know where coordinates are placed in the address
+        }
-    // stream.
+        addr_reg = addr_reg + 4;
-    ASSERT_MSG(!flags.test(MimgModifier::Derivative), "Derivative image instruction");
+        return ir.CompositeConstruct(
+            ir.GetVectorReg<IR::F32>(addr_reg - 4), ir.GetVectorReg<IR::F32>(addr_reg - 3),
+            ir.GetVectorReg<IR::F32>(addr_reg - 2), ir.GetVectorReg<IR::F32>(addr_reg - 1));
+    }();
 
     // Now we can load body components as noted in Table 8.9 Image Opcodes with Sampler
     // Since these are at most 4 dwords, we load them into a single uvec4 and place them
@@ -177,6 +181,10 @@ void Translator::IMAGE_SAMPLE(const GcnInst& inst) {
         ir.GetVectorReg<IR::F32>(addr_reg), ir.GetVectorReg<IR::F32>(addr_reg + 1),
         ir.GetVectorReg<IR::F32>(addr_reg + 2), ir.GetVectorReg<IR::F32>(addr_reg + 3));
 
+    // Derivatives are tricky because their number depends on the texture type which is located in
+    // T#. We don't have access to T# though until resource tracking pass. For now assume if
+    // derivatives are present, that a 2D image is bound.
+    const bool has_derivatives = flags.test(MimgModifier::Derivative);
     const bool explicit_lod = flags.any(MimgModifier::Level0, MimgModifier::Lod);
 
     IR::TextureInstInfo info{};
@@ -186,9 +194,13 @@ void Translator::IMAGE_SAMPLE(const GcnInst& inst) {
     info.force_level0.Assign(flags.test(MimgModifier::Level0));
     info.has_offset.Assign(flags.test(MimgModifier::Offset));
     info.explicit_lod.Assign(explicit_lod);
+    info.has_derivatives.Assign(has_derivatives);
 
     // Issue IR instruction, leaving unknown fields blank to patch later.
     const IR::Value texel = [&]() -> IR::Value {
+        if (has_derivatives) {
+            return ir.ImageGradient(handle, body, derivatives, offset, {}, info);
+        }
         if (!flags.test(MimgModifier::Pcf)) {
             if (explicit_lod) {
                 return ir.ImageSampleExplicitLod(handle, body, offset, info);

src/shader_recompiler/ir/ir_emitter.cpp

@@ -209,10 +209,6 @@ U1 IREmitter::GetVcc() {
     return Inst<U1>(Opcode::GetVcc);
 }
 
-U32 IREmitter::GetSccLo() {
-    return Inst<U32>(Opcode::GetSccLo);
-}
-
 U32 IREmitter::GetVccLo() {
     return Inst<U32>(Opcode::GetVccLo);
 }
@@ -445,6 +441,18 @@ U32 IREmitter::QuadShuffle(const U32& value, const U32& index) {
     return Inst<U32>(Opcode::QuadShuffle, value, index);
 }
 
+U32 IREmitter::ReadFirstLane(const U32& value) {
+    return Inst<U32>(Opcode::ReadFirstLane, value);
+}
+
+U32 IREmitter::ReadLane(const U32& value, const U32& lane) {
+    return Inst<U32>(Opcode::ReadLane, value, lane);
+}
+
+U32 IREmitter::WriteLane(const U32& value, const U32& write_value, const U32& lane) {
+    return Inst<U32>(Opcode::WriteLane, value, write_value, lane);
+}
+
 F32F64 IREmitter::FPAdd(const F32F64& a, const F32F64& b) {
     if (a.Type() != b.Type()) {
         UNREACHABLE_MSG("Mismatching types {} and {}", a.Type(), b.Type());

src/shader_recompiler/ir/ir_emitter.h

@@ -65,7 +65,6 @@ public:
     [[nodiscard]] U1 GetScc();
     [[nodiscard]] U1 GetExec();
     [[nodiscard]] U1 GetVcc();
-    [[nodiscard]] U32 GetSccLo();
     [[nodiscard]] U32 GetVccLo();
     [[nodiscard]] U32 GetVccHi();
     void SetScc(const U1& value);
@@ -122,6 +121,9 @@ public:
     [[nodiscard]] U32 LaneId();
     [[nodiscard]] U32 WarpId();
     [[nodiscard]] U32 QuadShuffle(const U32& value, const U32& index);
+    [[nodiscard]] U32 ReadFirstLane(const U32& value);
+    [[nodiscard]] U32 ReadLane(const U32& value, const U32& lane);
+    [[nodiscard]] U32 WriteLane(const U32& value, const U32& write_value, const U32& lane);
 
     [[nodiscard]] Value CompositeConstruct(const Value& e1, const Value& e2);
     [[nodiscard]] Value CompositeConstruct(const Value& e1, const Value& e2, const Value& e3);

src/shader_recompiler/ir/opcodes.inc

@@ -58,7 +58,6 @@ OPCODE(SetAttribute,                                        Void,           Attr
 OPCODE(GetScc,                                             U1,             Void,                                                                            )
 OPCODE(GetExec,                                            U1,             Void,                                                                            )
 OPCODE(GetVcc,                                             U1,             Void,                                                                            )
-OPCODE(GetSccLo,                                           U32,            Void,                                                                            )
 OPCODE(GetVccLo,                                           U32,            Void,                                                                            )
 OPCODE(GetVccHi,                                           U32,            Void,                                                                            )
 OPCODE(SetScc,                                             Void,           U1,                                                                              )
@@ -330,19 +329,22 @@ OPCODE(ImageRead,                                           U32x4,          Opaq
 OPCODE(ImageWrite,                                          Void,           Opaque,         Opaque,         U32x4,                                          )
 
 // Image atomic operations
-OPCODE(ImageAtomicIAdd32,                                   U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicIAdd32,                                   U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicSMin32,                                   U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicSMin32,                                   U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicUMin32,                                   U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicUMin32,                                   U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicSMax32,                                   U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicSMax32,                                   U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicUMax32,                                   U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicUMax32,                                   U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicInc32,                                    U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicInc32,                                    U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicDec32,                                    U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicDec32,                                    U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicAnd32,                                    U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicAnd32,                                    U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicOr32,                                     U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicOr32,                                     U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicXor32,                                    U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicXor32,                                    U32,            Opaque,         Opaque,         U32,                                            )
-OPCODE(ImageAtomicExchange32,                               U32,            Opaque,            Opaque,            U32,                                      )
+OPCODE(ImageAtomicExchange32,                               U32,            Opaque,         Opaque,         U32,                                            )
 
 // Warp operations
 OPCODE(LaneId,                                              U32,                                                                                            )
 OPCODE(WarpId,                                              U32,                                                                                            )
 OPCODE(QuadShuffle,                                         U32,            U32,            U32                                                             )
+OPCODE(ReadFirstLane,                                       U32,            U32,            U32                                                             )
+OPCODE(ReadLane,                                            U32,            U32,            U32                                                             )
+OPCODE(WriteLane,                                           U32,            U32,            U32,            U32                                             )

src/shader_recompiler/ir/passes/constant_propogation_pass.cpp

@@ -250,6 +250,18 @@ void FoldCmpClass(IR::Inst& inst) {
     }
 }
 
+void FoldReadLane(IR::Inst& inst) {
+    const u32 lane = inst.Arg(1).U32();
+    IR::Inst* prod = inst.Arg(0).InstRecursive();
+    while (prod->GetOpcode() == IR::Opcode::WriteLane) {
+        if (prod->Arg(2).U32() == lane) {
+            inst.ReplaceUsesWith(prod->Arg(1));
+            return;
+        }
+        prod = prod->Arg(0).InstRecursive();
+    }
+}
+
 void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     switch (inst.GetOpcode()) {
     case IR::Opcode::IAdd32:
@@ -289,6 +301,8 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
     case IR::Opcode::SelectF32:
     case IR::Opcode::SelectF64:
         return FoldSelect(inst);
+    case IR::Opcode::ReadLane:
+        return FoldReadLane(inst);
     case IR::Opcode::FPNeg32:
         FoldWhenAllImmediates(inst, [](f32 a) { return -a; });
         return;

src/shader_recompiler/ir/passes/resource_tracking_pass.cpp

@@ -345,6 +345,7 @@ SharpLocation TrackSharp(const IR::Inst* inst) {
 
     // Retrieve SGPR pair that holds sbase
     const auto pred1 = [](const IR::Inst* inst) -> std::optional<IR::ScalarReg> {
+        ASSERT(inst->GetOpcode() != IR::Opcode::ReadConst);
         if (inst->GetOpcode() == IR::Opcode::GetUserData) {
             return inst->Arg(0).ScalarReg();
         }
@@ -402,24 +403,13 @@ s32 TryHandleInlineCbuf(IR::Inst& inst, Info& info, Descriptors& descriptors,
     // is used to define an inline constant buffer
 
     IR::Inst* handle = inst.Arg(0).InstRecursive();
-    IR::Inst* p0 = handle->Arg(0).InstRecursive();
+    if (!handle->AreAllArgsImmediates()) {
-    if (p0->GetOpcode() != IR::Opcode::IAdd32 || !p0->Arg(0).IsImmediate() ||
-        !p0->Arg(1).IsImmediate()) {
-        return -1;
-    }
-    IR::Inst* p1 = handle->Arg(1).InstRecursive();
-    if (p1->GetOpcode() != IR::Opcode::IAdd32) {
-        return -1;
-    }
-    if (!handle->Arg(3).IsImmediate() || !handle->Arg(2).IsImmediate()) {
         return -1;
     }
     // We have found this pattern. Build the sharp.
-    std::array<u32, 4> buffer;
+    std::array<u64, 2> buffer;
-    buffer[0] = info.pgm_base + p0->Arg(0).U32() + p0->Arg(1).U32();
+    buffer[0] = info.pgm_base + (handle->Arg(0).U32() | u64(handle->Arg(1).U32()) << 32);
-    buffer[1] = 0;
+    buffer[1] = handle->Arg(2).U32() | u64(handle->Arg(3).U32()) << 32;
-    buffer[2] = handle->Arg(2).U32();
-    buffer[3] = handle->Arg(3).U32();
     cbuf = std::bit_cast<AmdGpu::Buffer>(buffer);
     // Assign a binding to this sharp.
     return descriptors.Add(BufferResource{
@@ -617,7 +607,11 @@ void PatchImageInstruction(IR::Block& block, IR::Inst& inst, Info& info, Descrip
         const IR::Value arg = inst.Arg(arg_pos);
         ASSERT_MSG(arg.Type() == IR::Type::U32, "Unexpected offset type");
 
-        const auto read = [&](u32 offset) -> auto {
+        const auto read = [&](u32 offset) -> IR::U32 {
+            if (arg.IsImmediate()) {
+                const u16 comp = (arg.U32() >> offset) & 0x3F;
+                return ir.Imm32(s32(comp << 26) >> 26);
+            }
             return ir.BitFieldExtract(IR::U32{arg}, ir.Imm32(offset), ir.Imm32(6), true);
         };
 
@@ -637,7 +631,10 @@ void PatchImageInstruction(IR::Block& block, IR::Inst& inst, Info& info, Descrip
             UNREACHABLE();
         }
     }
-
+    if (inst_info.has_derivatives) {
+        ASSERT_MSG(image.GetType() == AmdGpu::ImageType::Color2D,
+                   "User derivatives only supported for 2D images");
+    }
     if (inst_info.has_lod_clamp) {
         const u32 arg_pos = [&]() -> u32 {
             switch (inst.GetOpcode()) {

src/shader_recompiler/ir/passes/ssa_rewrite_pass.cpp

@@ -32,7 +32,6 @@ struct SccFlagTag : FlagTag {};
 struct ExecFlagTag : FlagTag {};
 struct VccFlagTag : FlagTag {};
 struct VccLoTag : FlagTag {};
-struct SccLoTag : FlagTag {};
 struct VccHiTag : FlagTag {};
 
 struct GotoVariable : FlagTag {
@@ -45,7 +44,7 @@ struct GotoVariable : FlagTag {
 };
 
 using Variant = std::variant<IR::ScalarReg, IR::VectorReg, GotoVariable, SccFlagTag, ExecFlagTag,
-                             VccFlagTag, SccLoTag, VccLoTag, VccHiTag>;
+                             VccFlagTag, VccLoTag, VccHiTag>;
 using ValueMap = std::unordered_map<IR::Block*, IR::Value>;
 
 struct DefTable {
@@ -84,13 +83,6 @@ struct DefTable {
         exec_flag.insert_or_assign(block, value);
     }
 
-    const IR::Value& Def(IR::Block* block, SccLoTag) {
-        return scc_lo_flag[block];
-    }
-    void SetDef(IR::Block* block, SccLoTag, const IR::Value& value) {
-        scc_lo_flag.insert_or_assign(block, value);
-    }
-
     const IR::Value& Def(IR::Block* block, VccLoTag) {
         return vcc_lo_flag[block];
     }
@@ -133,10 +125,6 @@ IR::Opcode UndefOpcode(const VccLoTag) noexcept {
     return IR::Opcode::UndefU32;
 }
 
-IR::Opcode UndefOpcode(const SccLoTag) noexcept {
-    return IR::Opcode::UndefU32;
-}
-
 IR::Opcode UndefOpcode(const VccHiTag) noexcept {
     return IR::Opcode::UndefU32;
 }
@@ -336,9 +324,6 @@ void VisitInst(Pass& pass, IR::Block* block, IR::Inst& inst) {
     case IR::Opcode::SetVcc:
         pass.WriteVariable(VccFlagTag{}, block, inst.Arg(0));
         break;
-    case IR::Opcode::SetSccLo:
-        pass.WriteVariable(SccLoTag{}, block, inst.Arg(0));
-        break;
     case IR::Opcode::SetVccLo:
         pass.WriteVariable(VccLoTag{}, block, inst.Arg(0));
         break;
@@ -371,9 +356,6 @@ void VisitInst(Pass& pass, IR::Block* block, IR::Inst& inst) {
     case IR::Opcode::GetVcc:
         inst.ReplaceUsesWith(pass.ReadVariable(VccFlagTag{}, block));
         break;
-    case IR::Opcode::GetSccLo:
-        inst.ReplaceUsesWith(pass.ReadVariable(SccLoTag{}, block));
-        break;
     case IR::Opcode::GetVccLo:
         inst.ReplaceUsesWith(pass.ReadVariable(VccLoTag{}, block));
         break;

src/shader_recompiler/ir/reg.h

@@ -58,6 +58,7 @@ union TextureInstInfo {
     BitField<4, 1, u32> explicit_lod;
     BitField<5, 1, u32> has_offset;
     BitField<6, 2, u32> gather_comp;
+    BitField<8, 1, u32> has_derivatives;
 };
 
 union BufferInstInfo {

src/shader_recompiler/recompiler.cpp

@@ -56,11 +56,11 @@ IR::Program TranslateProgram(Common::ObjectPool<IR::Inst>& inst_pool,
 
     // Run optimization passes
     Shader::Optimization::SsaRewritePass(program.post_order_blocks);
-    Shader::Optimization::ResourceTrackingPass(program);
     Shader::Optimization::ConstantPropagationPass(program.post_order_blocks);
     if (program.info.stage != Stage::Compute) {
         Shader::Optimization::LowerSharedMemToRegisters(program);
     }
+    Shader::Optimization::ResourceTrackingPass(program);
     Shader::Optimization::IdentityRemovalPass(program.blocks);
     Shader::Optimization::DeadCodeEliminationPass(program);
     Shader::Optimization::CollectShaderInfoPass(program);

src/video_core/amdgpu/resource.h

@@ -179,6 +179,10 @@ struct Image {
         return base_address << 8;
     }
 
+    operator bool() const noexcept {
+        return base_address != 0;
+    }
+
     u32 DstSelect() const {
         return dst_sel_x | (dst_sel_y << 3) | (dst_sel_z << 6) | (dst_sel_w << 9);
     }

src/video_core/page_manager.cpp

@@ -51,7 +51,8 @@ struct PageManager::Impl {
         if (ec == EXCEPTION_ACCESS_VIOLATION) {
             const auto info = pExp->ExceptionRecord->ExceptionInformation;
             if (info[0] == 1) { // Write violation
-                rasterizer->InvalidateMemory(info[1], sizeof(u64));
+                const VAddr addr_aligned = Common::AlignDown(info[1], PAGESIZE);
+                rasterizer->InvalidateMemory(addr_aligned, PAGESIZE);
                 return EXCEPTION_CONTINUE_EXECUTION;
             } /* else {
                 UNREACHABLE();
@@ -199,7 +200,8 @@ struct PageManager::Impl {
         const greg_t err = ctx->uc_mcontext.gregs[REG_ERR];
 #endif
         if (err & 0x2) {
-            rasterizer->InvalidateMemory(address, sizeof(u64));
+            const VAddr addr_aligned = Common::AlignDown(address, PAGESIZE);
+            rasterizer->InvalidateMemory(addr_aligned, PAGESIZE);
         } else {
             // Read not supported!
             UNREACHABLE();

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -396,13 +396,18 @@ void GraphicsPipeline::BindResources(const Liverpool::Regs& regs,
 
         boost::container::static_vector<AmdGpu::Image, 16> tsharps;
         for (const auto& image_desc : stage->images) {
-            const auto& tsharp = tsharps.emplace_back(
+            const auto tsharp =
-                stage->ReadUd<AmdGpu::Image>(image_desc.sgpr_base, image_desc.dword_offset));
+                stage->ReadUd<AmdGpu::Image>(image_desc.sgpr_base, image_desc.dword_offset);
-            VideoCore::ImageInfo image_info{tsharp};
+            if (tsharp) {
-            VideoCore::ImageViewInfo view_info{tsharp, image_desc.is_storage};
+                tsharps.emplace_back(tsharp);
-            const auto& image_view = texture_cache.FindTexture(image_info, view_info);
+                VideoCore::ImageInfo image_info{tsharp};
-            const auto& image = texture_cache.GetImage(image_view.image_id);
+                VideoCore::ImageViewInfo view_info{tsharp, image_desc.is_storage};
-            image_infos.emplace_back(VK_NULL_HANDLE, *image_view.image_view, image.layout);
+                const auto& image_view = texture_cache.FindTexture(image_info, view_info);
+                const auto& image = texture_cache.GetImage(image_view.image_id);
+                image_infos.emplace_back(VK_NULL_HANDLE, *image_view.image_view, image.layout);
+            } else {
+                image_infos.emplace_back(VK_NULL_HANDLE, VK_NULL_HANDLE, vk::ImageLayout::eGeneral);
+            }
             set_writes.push_back({
                 .dstSet = VK_NULL_HANDLE,
                 .dstBinding = binding++,

src/video_core/renderer_vulkan/vk_instance.cpp

@@ -210,6 +210,8 @@ bool Instance::CreateDevice() {
     color_write_en &= add_extension(VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
     const bool calibrated_timestamps = add_extension(VK_EXT_CALIBRATED_TIMESTAMPS_EXTENSION_NAME);
     const bool robustness = add_extension(VK_EXT_ROBUSTNESS_2_EXTENSION_NAME);
+    const bool topology_restart =
+        add_extension(VK_EXT_PRIMITIVE_TOPOLOGY_LIST_RESTART_EXTENSION_NAME);
 
     // These extensions are promoted by Vulkan 1.3, but for greater compatibility we use Vulkan 1.2
     // with extensions.
@@ -330,6 +332,9 @@ bool Instance::CreateDevice() {
         vk::PhysicalDeviceVertexInputDynamicStateFeaturesEXT{
             .vertexInputDynamicState = true,
         },
+        vk::PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT{
+            .primitiveTopologyListRestart = true,
+        },
 #ifdef __APPLE__
         feature_chain.get<vk::PhysicalDevicePortabilitySubsetFeaturesKHR>(),
 #endif
@@ -351,6 +356,9 @@ bool Instance::CreateDevice() {
     if (!workgroup_memory_explicit_layout) {
         device_chain.unlink<vk::PhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR>();
     }
+    if (!topology_restart) {
+        device_chain.unlink<vk::PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT>();
+    }
     if (robustness) {
         device_chain.get<vk::PhysicalDeviceRobustness2FeaturesEXT>().nullDescriptor =
             feature_chain.get<vk::PhysicalDeviceRobustness2FeaturesEXT>().nullDescriptor;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -280,9 +280,6 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline() {
             DumpShader(code, hash, stage, "bin");
         }
 
-        block_pool.ReleaseContents();
-        inst_pool.ReleaseContents();
-
         if (stage != Shader::Stage::Fragment && stage != Shader::Stage::Vertex) {
             LOG_ERROR(Render_Vulkan, "Unsupported shader stage {}. PL creation skipped.", stage);
             return {};

src/video_core/texture_cache/image_info.cpp

@@ -219,7 +219,12 @@ ImageInfo::ImageInfo(const AmdGpu::Image& image) noexcept {
     guest_address = image.Address();
 
     mips_layout.reserve(resources.levels);
+    tiling_idx = image.tiling_index;
+    UpdateSize();
+}
 
+void ImageInfo::UpdateSize() {
+    mips_layout.clear();
     MipInfo mip_info{};
     guest_size_bytes = 0;
     for (auto mip = 0u; mip < resources.levels; ++mip) {
@@ -265,7 +270,7 @@ ImageInfo::ImageInfo(const AmdGpu::Image& image) noexcept {
             ASSERT(!props.is_block);
             ASSERT(num_samples == 1);
             std::tie(mip_info.pitch, mip_info.size) =
-                ImageSizeMacroTiled(mip_w, mip_h, bpp, num_samples, image.tiling_index);
+                ImageSizeMacroTiled(mip_w, mip_h, bpp, num_samples, tiling_idx);
             break;
         }
         default: {

src/video_core/texture_cache/image_info.h

@@ -29,6 +29,8 @@ struct ImageInfo {
     bool IsPacked() const;
     bool IsDepthStencil() const;
 
+    void UpdateSize();
+
     struct {
         VAddr cmask_addr;
         VAddr fmask_addr;
@@ -69,6 +71,7 @@ struct ImageInfo {
     boost::container::small_vector<MipInfo, 14> mips_layout;
     VAddr guest_address{0};
     u32 guest_size_bytes{0};
+    u32 tiling_idx{0}; // TODO: merge with existing!
 };
 
 } // namespace VideoCore

src/video_core/texture_cache/texture_cache.cpp

@@ -18,11 +18,15 @@ TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler&
                            BufferCache& buffer_cache_, PageManager& tracker_)
     : instance{instance_}, scheduler{scheduler_}, buffer_cache{buffer_cache_}, tracker{tracker_},
       tile_manager{instance, scheduler} {
-    ImageInfo info;
+    ImageInfo info{};
     info.pixel_format = vk::Format::eR8G8B8A8Unorm;
     info.type = vk::ImageType::e2D;
+    info.tiling_idx = u32(AmdGpu::TilingMode::Texture_MicroTiled);
+    info.num_bits = 32;
+    info.UpdateSize();
     const ImageId null_id = slot_images.insert(instance, scheduler, info);
     ASSERT(null_id.index == 0);
+    slot_images[null_id].flags = ImageFlagBits{};
 
     ImageViewInfo view_info;
     void(slot_image_views.insert(instance, view_info, slot_images[null_id], null_id));

src/video_core/texture_cache/texture_cache.h

@@ -28,7 +28,7 @@ class TextureCache {
         using Entry = boost::container::small_vector<ImageId, 16>;
         static constexpr size_t AddressSpaceBits = 39;
         static constexpr size_t FirstLevelBits = 9;
-        static constexpr size_t PageBits = 22;
+        static constexpr size_t PageBits = 20;
     };
     using PageTable = MultiLevelPageTable<Traits>;