Remove unused classes

shader/other: Implement MEMBAR.CTS

.gitmodules

@@ -28,3 +28,6 @@
 [submodule "libzip"]
     path = externals/libzip/libzip
     url = https://github.com/nih-at/libzip.git
+[submodule "xbyak"]
+    path = externals/xbyak
+    url = https://github.com/herumi/xbyak.git

externals/CMakeLists.txt

@@ -75,3 +75,11 @@ if (ENABLE_WEB_SERVICE)
     target_compile_definitions(httplib INTERFACE -DCPPHTTPLIB_OPENSSL_SUPPORT)
     target_link_libraries(httplib INTERFACE OpenSSL::SSL OpenSSL::Crypto)
 endif()
+
+if (NOT TARGET xbyak)
+    if (ARCHITECTURE_x86 OR ARCHITECTURE_x86_64)
+        add_library(xbyak INTERFACE)
+        target_include_directories(xbyak SYSTEM INTERFACE ./xbyak/xbyak)
+        target_compile_definitions(xbyak INTERFACE XBYAK_NO_OP_NAMES)
+    endif()
+endif()

src/common/CMakeLists.txt

@@ -171,10 +171,12 @@ if(ARCHITECTURE_x86_64)
         PRIVATE
             x64/cpu_detect.cpp
             x64/cpu_detect.h
+            x64/xbyak_abi.h
+            x64/xbyak_util.h
     )
 endif()
 
 create_target_directory_groups(common)
 
 target_link_libraries(common PUBLIC Boost::boost fmt::fmt microprofile)
-target_link_libraries(common PRIVATE lz4::lz4 zstd::zstd)
+target_link_libraries(common PRIVATE lz4::lz4 zstd::zstd xbyak)

src/common/x64/xbyak_abi.h

@@ -0,0 +1,266 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <bitset>
+#include <initializer_list>
+#include <xbyak.h>
+#include "common/assert.h"
+
+namespace Common::X64 {
+
+inline int RegToIndex(const Xbyak::Reg& reg) {
+    using Kind = Xbyak::Reg::Kind;
+    ASSERT_MSG((reg.getKind() & (Kind::REG | Kind::XMM)) != 0,
+               "RegSet only support GPRs and XMM registers.");
+    ASSERT_MSG(reg.getIdx() < 16, "RegSet only supports XXM0-15.");
+    return reg.getIdx() + (reg.getKind() == Kind::REG ? 0 : 16);
+}
+
+inline Xbyak::Reg64 IndexToReg64(int reg_index) {
+    ASSERT(reg_index < 16);
+    return Xbyak::Reg64(reg_index);
+}
+
+inline Xbyak::Xmm IndexToXmm(int reg_index) {
+    ASSERT(reg_index >= 16 && reg_index < 32);
+    return Xbyak::Xmm(reg_index - 16);
+}
+
+inline Xbyak::Reg IndexToReg(int reg_index) {
+    if (reg_index < 16) {
+        return IndexToReg64(reg_index);
+    } else {
+        return IndexToXmm(reg_index);
+    }
+}
+
+inline std::bitset<32> BuildRegSet(std::initializer_list<Xbyak::Reg> regs) {
+    std::bitset<32> bits;
+    for (const Xbyak::Reg& reg : regs) {
+        bits[RegToIndex(reg)] = true;
+    }
+    return bits;
+}
+
+const std::bitset<32> ABI_ALL_GPRS(0x0000FFFF);
+const std::bitset<32> ABI_ALL_XMMS(0xFFFF0000);
+
+#ifdef _WIN32
+
+// Microsoft x64 ABI
+const Xbyak::Reg ABI_RETURN = Xbyak::util::rax;
+const Xbyak::Reg ABI_PARAM1 = Xbyak::util::rcx;
+const Xbyak::Reg ABI_PARAM2 = Xbyak::util::rdx;
+const Xbyak::Reg ABI_PARAM3 = Xbyak::util::r8;
+const Xbyak::Reg ABI_PARAM4 = Xbyak::util::r9;
+
+const std::bitset<32> ABI_ALL_CALLER_SAVED = BuildRegSet({
+    // GPRs
+    Xbyak::util::rcx,
+    Xbyak::util::rdx,
+    Xbyak::util::r8,
+    Xbyak::util::r9,
+    Xbyak::util::r10,
+    Xbyak::util::r11,
+    // XMMs
+    Xbyak::util::xmm0,
+    Xbyak::util::xmm1,
+    Xbyak::util::xmm2,
+    Xbyak::util::xmm3,
+    Xbyak::util::xmm4,
+    Xbyak::util::xmm5,
+});
+
+const std::bitset<32> ABI_ALL_CALLEE_SAVED = BuildRegSet({
+    // GPRs
+    Xbyak::util::rbx,
+    Xbyak::util::rsi,
+    Xbyak::util::rdi,
+    Xbyak::util::rbp,
+    Xbyak::util::r12,
+    Xbyak::util::r13,
+    Xbyak::util::r14,
+    Xbyak::util::r15,
+    // XMMs
+    Xbyak::util::xmm6,
+    Xbyak::util::xmm7,
+    Xbyak::util::xmm8,
+    Xbyak::util::xmm9,
+    Xbyak::util::xmm10,
+    Xbyak::util::xmm11,
+    Xbyak::util::xmm12,
+    Xbyak::util::xmm13,
+    Xbyak::util::xmm14,
+    Xbyak::util::xmm15,
+});
+
+constexpr size_t ABI_SHADOW_SPACE = 0x20;
+
+#else
+
+// System V x86-64 ABI
+const Xbyak::Reg ABI_RETURN = Xbyak::util::rax;
+const Xbyak::Reg ABI_PARAM1 = Xbyak::util::rdi;
+const Xbyak::Reg ABI_PARAM2 = Xbyak::util::rsi;
+const Xbyak::Reg ABI_PARAM3 = Xbyak::util::rdx;
+const Xbyak::Reg ABI_PARAM4 = Xbyak::util::rcx;
+
+const std::bitset<32> ABI_ALL_CALLER_SAVED = BuildRegSet({
+    // GPRs
+    Xbyak::util::rcx,
+    Xbyak::util::rdx,
+    Xbyak::util::rdi,
+    Xbyak::util::rsi,
+    Xbyak::util::r8,
+    Xbyak::util::r9,
+    Xbyak::util::r10,
+    Xbyak::util::r11,
+    // XMMs
+    Xbyak::util::xmm0,
+    Xbyak::util::xmm1,
+    Xbyak::util::xmm2,
+    Xbyak::util::xmm3,
+    Xbyak::util::xmm4,
+    Xbyak::util::xmm5,
+    Xbyak::util::xmm6,
+    Xbyak::util::xmm7,
+    Xbyak::util::xmm8,
+    Xbyak::util::xmm9,
+    Xbyak::util::xmm10,
+    Xbyak::util::xmm11,
+    Xbyak::util::xmm12,
+    Xbyak::util::xmm13,
+    Xbyak::util::xmm14,
+    Xbyak::util::xmm15,
+});
+
+const std::bitset<32> ABI_ALL_CALLEE_SAVED = BuildRegSet({
+    // GPRs
+    Xbyak::util::rbx,
+    Xbyak::util::rbp,
+    Xbyak::util::r12,
+    Xbyak::util::r13,
+    Xbyak::util::r14,
+    Xbyak::util::r15,
+});
+
+constexpr size_t ABI_SHADOW_SPACE = 0;
+
+#endif
+
+inline void ABI_CalculateFrameSize(std::bitset<32> regs, size_t rsp_alignment,
+                                   size_t needed_frame_size, s32* out_subtraction,
+                                   s32* out_xmm_offset) {
+    const auto count = (regs & ABI_ALL_GPRS).count();
+    rsp_alignment -= count * 8;
+    size_t subtraction = 0;
+    const auto xmm_count = (regs & ABI_ALL_XMMS).count();
+    if (xmm_count) {
+        // If we have any XMMs to save, we must align the stack here.
+        subtraction = rsp_alignment & 0xF;
+    }
+    subtraction += 0x10 * xmm_count;
+    size_t xmm_base_subtraction = subtraction;
+    subtraction += needed_frame_size;
+    subtraction += ABI_SHADOW_SPACE;
+    // Final alignment.
+    rsp_alignment -= subtraction;
+    subtraction += rsp_alignment & 0xF;
+
+    *out_subtraction = (s32)subtraction;
+    *out_xmm_offset = (s32)(subtraction - xmm_base_subtraction);
+}
+
+inline size_t ABI_PushRegistersAndAdjustStack(Xbyak::CodeGenerator& code, std::bitset<32> regs,
+                                              size_t rsp_alignment, size_t needed_frame_size = 0) {
+    s32 subtraction, xmm_offset;
+    ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size, &subtraction, &xmm_offset);
+    for (std::size_t i = 0; i < regs.size(); ++i) {
+        if (regs[i] && ABI_ALL_GPRS[i]) {
+            code.push(IndexToReg64(static_cast<int>(i)));
+        }
+    }
+    if (subtraction != 0) {
+        code.sub(code.rsp, subtraction);
+    }
+
+    for (int i = 0; i < regs.count(); i++) {
+        if (regs.test(i) & ABI_ALL_GPRS.test(i)) {
+            code.push(IndexToReg64(i));
+        }
+    }
+
+    for (std::size_t i = 0; i < regs.size(); ++i) {
+        if (regs[i] && ABI_ALL_XMMS[i]) {
+            code.movaps(code.xword[code.rsp + xmm_offset], IndexToXmm(static_cast<int>(i)));
+            xmm_offset += 0x10;
+        }
+    }
+
+    return ABI_SHADOW_SPACE;
+}
+
+inline void ABI_PopRegistersAndAdjustStack(Xbyak::CodeGenerator& code, std::bitset<32> regs,
+                                           size_t rsp_alignment, size_t needed_frame_size = 0) {
+    s32 subtraction, xmm_offset;
+    ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size, &subtraction, &xmm_offset);
+
+    for (std::size_t i = 0; i < regs.size(); ++i) {
+        if (regs[i] && ABI_ALL_XMMS[i]) {
+            code.movaps(IndexToXmm(static_cast<int>(i)), code.xword[code.rsp + xmm_offset]);
+            xmm_offset += 0x10;
+        }
+    }
+
+    if (subtraction != 0) {
+        code.add(code.rsp, subtraction);
+    }
+
+    // GPRs need to be popped in reverse order
+    for (int i = 15; i >= 0; i--) {
+        if (regs[i]) {
+            code.pop(IndexToReg64(i));
+        }
+    }
+}
+
+inline size_t ABI_PushRegistersAndAdjustStackGPS(Xbyak::CodeGenerator& code, std::bitset<32> regs,
+                                                 size_t rsp_alignment,
+                                                 size_t needed_frame_size = 0) {
+    s32 subtraction, xmm_offset;
+    ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size, &subtraction, &xmm_offset);
+
+    for (std::size_t i = 0; i < regs.size(); ++i) {
+        if (regs[i] && ABI_ALL_GPRS[i]) {
+            code.push(IndexToReg64(static_cast<int>(i)));
+        }
+    }
+
+    if (subtraction != 0) {
+        code.sub(code.rsp, subtraction);
+    }
+
+    return ABI_SHADOW_SPACE;
+}
+
+inline void ABI_PopRegistersAndAdjustStackGPS(Xbyak::CodeGenerator& code, std::bitset<32> regs,
+                                              size_t rsp_alignment, size_t needed_frame_size = 0) {
+    s32 subtraction, xmm_offset;
+    ABI_CalculateFrameSize(regs, rsp_alignment, needed_frame_size, &subtraction, &xmm_offset);
+
+    if (subtraction != 0) {
+        code.add(code.rsp, subtraction);
+    }
+
+    // GPRs need to be popped in reverse order
+    for (int i = 15; i >= 0; i--) {
+        if (regs[i]) {
+            code.pop(IndexToReg64(i));
+        }
+    }
+}
+
+} // namespace Common::X64

src/common/x64/xbyak_util.h

@@ -0,0 +1,47 @@
+// Copyright 2016 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <type_traits>
+#include <xbyak.h>
+#include "common/x64/xbyak_abi.h"
+
+namespace Common::X64 {
+
+// Constants for use with cmpps/cmpss
+enum {
+    CMP_EQ = 0,
+    CMP_LT = 1,
+    CMP_LE = 2,
+    CMP_UNORD = 3,
+    CMP_NEQ = 4,
+    CMP_NLT = 5,
+    CMP_NLE = 6,
+    CMP_ORD = 7,
+};
+
+constexpr bool IsWithin2G(uintptr_t ref, uintptr_t target) {
+    const u64 distance = target - (ref + 5);
+    return !(distance >= 0x8000'0000ULL && distance <= ~0x8000'0000ULL);
+}
+
+inline bool IsWithin2G(const Xbyak::CodeGenerator& code, uintptr_t target) {
+    return IsWithin2G(reinterpret_cast<uintptr_t>(code.getCurr()), target);
+}
+
+template <typename T>
+inline void CallFarFunction(Xbyak::CodeGenerator& code, const T f) {
+    static_assert(std::is_pointer_v<T>, "Argument must be a (function) pointer.");
+    size_t addr = reinterpret_cast<size_t>(f);
+    if (IsWithin2G(code, addr)) {
+        code.call(f);
+    } else {
+        // ABI_RETURN is a safe temp register to use before a call
+        code.mov(ABI_RETURN, addr);
+        code.call(ABI_RETURN);
+    }
+}
+
+} // namespace Common::X64

src/core/hle/service/bcat/bcat.cpp

@@ -14,6 +14,8 @@ BCAT::BCAT(Core::System& system, std::shared_ptr<Module> module,
         {0, &BCAT::CreateBcatService, "CreateBcatService"},
         {1, &BCAT::CreateDeliveryCacheStorageService, "CreateDeliveryCacheStorageService"},
         {2, &BCAT::CreateDeliveryCacheStorageServiceWithApplicationId, "CreateDeliveryCacheStorageServiceWithApplicationId"},
+        {3, nullptr, "CreateDeliveryCacheProgressService"},
+        {4, nullptr, "CreateDeliveryCacheProgressServiceWithApplicationId"},
     };
     // clang-format on
     RegisterHandlers(functions);

src/core/hle/service/bcat/module.cpp

@@ -143,10 +143,13 @@ public:
             {20401, nullptr, "UnregisterSystemApplicationDeliveryTask"},
             {20410, nullptr, "SetSystemApplicationDeliveryTaskTimer"},
             {30100, &IBcatService::SetPassphrase, "SetPassphrase"},
+            {30101, nullptr, "Unknown"},
+            {30102, nullptr, "Unknown2"},
             {30200, nullptr, "RegisterBackgroundDeliveryTask"},
             {30201, nullptr, "UnregisterBackgroundDeliveryTask"},
             {30202, nullptr, "BlockDeliveryTask"},
             {30203, nullptr, "UnblockDeliveryTask"},
+            {30210, nullptr, "SetDeliveryTaskTimer"},
             {30300, nullptr, "RegisterSystemApplicationDeliveryTasks"},
             {90100, nullptr, "EnumerateBackgroundDeliveryTask"},
             {90200, nullptr, "GetDeliveryList"},

src/video_core/CMakeLists.txt

@@ -229,7 +229,7 @@ endif()
 create_target_directory_groups(video_core)
 
 target_link_libraries(video_core PUBLIC common core)
-target_link_libraries(video_core PRIVATE glad)
+target_link_libraries(video_core PRIVATE glad xbyak)
 
 if (ENABLE_VULKAN)
     target_include_directories(video_core PRIVATE sirit ../../externals/Vulkan-Headers/include)

src/video_core/engines/maxwell_3d.cpp

@@ -457,8 +457,9 @@ void Maxwell3D::StampQueryResult(u64 payload, bool long_query) {
 
 void Maxwell3D::ProcessQueryGet() {
     // TODO(Subv): Support the other query units.
-    ASSERT_MSG(regs.query.query_get.unit == Regs::QueryUnit::Crop,
-               "Units other than CROP are unimplemented");
+    if (regs.query.query_get.unit != Regs::QueryUnit::Crop) {
+        LOG_DEBUG(HW_GPU, "Units other than CROP are unimplemented");
+    }
 
     switch (regs.query.query_get.operation) {
     case Regs::QueryOperation::Release:
@@ -534,8 +535,8 @@ void Maxwell3D::ProcessCounterReset() {
         rasterizer.ResetCounter(QueryType::SamplesPassed);
         break;
     default:
-        LOG_WARNING(Render_OpenGL, "Unimplemented counter reset={}",
-                    static_cast<int>(regs.counter_reset));
+        LOG_DEBUG(Render_OpenGL, "Unimplemented counter reset={}",
+                  static_cast<int>(regs.counter_reset));
         break;
     }
 }
@@ -592,8 +593,8 @@ std::optional<u64> Maxwell3D::GetQueryResult() {
                          system.GPU().GetTicks());
         return {};
     default:
-        UNIMPLEMENTED_MSG("Unimplemented query select type {}",
-                          static_cast<u32>(regs.query.query_get.select.Value()));
+        LOG_DEBUG(HW_GPU, "Unimplemented query select type {}",
+                  static_cast<u32>(regs.query.query_get.select.Value()));
         return 1;
     }
 }

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -2344,7 +2344,12 @@ private:
         return {};
     }
 
-    Expression MemoryBarrierGL(Operation) {
+    Expression MemoryBarrierGroup(Operation) {
+        code.AddLine("groupMemoryBarrier();");
+        return {};
+    }
+
+    Expression MemoryBarrierGlobal(Operation) {
         code.AddLine("memoryBarrier();");
         return {};
     }
@@ -2591,7 +2596,8 @@ private:
         &GLSLDecompiler::ShuffleIndexed,
 
         &GLSLDecompiler::Barrier,
-        &GLSLDecompiler::MemoryBarrierGL,
+        &GLSLDecompiler::MemoryBarrierGroup,
+        &GLSLDecompiler::MemoryBarrierGlobal,
     };
     static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));
 

src/video_core/renderer_vulkan/vk_shader_decompiler.cpp

@@ -2215,8 +2215,8 @@ private:
         return {};
     }
 
-    Expression MemoryBarrierGL(Operation) {
-        const auto scope = spv::Scope::Device;
+    template <spv::Scope scope>
+    Expression MemoryBarrier(Operation) {
         const auto semantics =
             spv::MemorySemanticsMask::AcquireRelease | spv::MemorySemanticsMask::UniformMemory |
             spv::MemorySemanticsMask::WorkgroupMemory |
@@ -2681,7 +2681,8 @@ private:
         &SPIRVDecompiler::ShuffleIndexed,
 
         &SPIRVDecompiler::Barrier,
-        &SPIRVDecompiler::MemoryBarrierGL,
+        &SPIRVDecompiler::MemoryBarrier<spv::Scope::Workgroup>,
+        &SPIRVDecompiler::MemoryBarrier<spv::Scope::Device>,
     };
     static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));
 

src/video_core/shader/decode/other.cpp

@@ -299,9 +299,19 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         break;
     }
     case OpCode::Id::MEMBAR: {
-        UNIMPLEMENTED_IF(instr.membar.type != Tegra::Shader::MembarType::GL);
         UNIMPLEMENTED_IF(instr.membar.unknown != Tegra::Shader::MembarUnknown::Default);
-        bb.push_back(Operation(OperationCode::MemoryBarrierGL));
+        const OperationCode type = [instr] {
+            switch (instr.membar.type) {
+            case Tegra::Shader::MembarType::CTA:
+                return OperationCode::MemoryBarrierGroup;
+            case Tegra::Shader::MembarType::GL:
+                return OperationCode::MemoryBarrierGlobal;
+            default:
+                UNIMPLEMENTED_MSG("MEMBAR type={}", static_cast<int>(instr.membar.type.Value()));
+                return OperationCode::MemoryBarrierGlobal;
+            }
+        }();
+        bb.push_back(Operation(type));
         break;
     }
     case OpCode::Id::DEPBAR: {

src/video_core/shader/node.h

@@ -233,8 +233,9 @@ enum class OperationCode {
     ThreadLtMask,   /// () -> uint
     ShuffleIndexed, /// (uint value, uint index) -> uint
 
-    Barrier,         /// () -> void
-    MemoryBarrierGL, /// () -> void
+    Barrier,             /// () -> void
+    MemoryBarrierGroup,  /// () -> void
+    MemoryBarrierGlobal, /// () -> void
 
     Amount,
 };