9 files changed, 84 insertions, 10 deletions

diff --git a/src/common/settings.h b/src/common/settings.h
index 5b34169a8..e61d9cd7f 100644
--- a/src/common/settings.h
+++ b/src/common/settings.h
@@ -606,6 +606,7 @@ struct Values {
     BasicSetting<bool> dump_exefs{false, "dump_exefs"};
     BasicSetting<bool> dump_nso{false, "dump_nso"};
     BasicSetting<bool> dump_shaders{false, "dump_shaders"};
+    BasicSetting<bool> dump_macros{false, "dump_macros"};
     BasicSetting<bool> enable_fs_access_log{false, "enable_fs_access_log"};
     BasicSetting<bool> reporting_services{false, "reporting_services"};
     BasicSetting<bool> quest_flag{false, "quest_flag"};
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index d4652b167..7d0cb8fce 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -173,6 +173,8 @@ void Maxwell3D::ProcessMethodCall(u32 method, u32 argument, u32 nonshadow_argume
     case MAXWELL3D_REG_INDEX(shadow_ram_control):
         shadow_state.shadow_ram_control = static_cast<Regs::ShadowRamControl>(nonshadow_argument);
         return;
+    case MAXWELL3D_REG_INDEX(macros.upload_address):
+        return macro_engine->ClearCode(regs.macros.upload_address);
     case MAXWELL3D_REG_INDEX(macros.data):
         return macro_engine->AddCode(regs.macros.upload_address, argument);
     case MAXWELL3D_REG_INDEX(macros.bind):
diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp
index 76e8bc656..a7302f7c1 100644
--- a/src/video_core/engines/maxwell_dma.cpp
+++ b/src/video_core/engines/maxwell_dma.cpp
@@ -134,7 +134,8 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
 
     // Deswizzle the input and copy it over.
     UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
-    const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
+    const u32 bytes_per_pixel =
+        regs.launch_dma.remap_enable ? regs.pitch_out / regs.line_length_in : 1;
     const Parameters& src_params = regs.src_params;
     const u32 width = src_params.width;
     const u32 height = src_params.height;
@@ -166,7 +167,8 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
     UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
 
     const auto& dst_params = regs.dst_params;
-    const u32 bytes_per_pixel = regs.pitch_in / regs.line_length_in;
+    const u32 bytes_per_pixel =
+        regs.launch_dma.remap_enable ? regs.pitch_in / regs.line_length_in : 1;
     const u32 width = dst_params.width;
     const u32 height = dst_params.height;
     const u32 depth = dst_params.depth;
@@ -210,7 +212,8 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
 }
 
 void MaxwellDMA::FastCopyBlockLinearToPitch() {
-    const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
+    const u32 bytes_per_pixel =
+        regs.launch_dma.remap_enable ? regs.pitch_out / regs.line_length_in : 1;
     const size_t src_size = GOB_SIZE;
     const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
     u32 pos_x = regs.src_params.origin.x;
diff --git a/src/video_core/macro/macro.cpp b/src/video_core/macro/macro.cpp
index a033d03be..e7279efcd 100644
--- a/src/video_core/macro/macro.cpp
+++ b/src/video_core/macro/macro.cpp
@@ -2,11 +2,15 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include <cstring>
+#include <fstream>
 #include <optional>
+#include <span>
 
 #include <boost/container_hash/hash.hpp>
 
 #include "common/assert.h"
+#include "common/fs/fs.h"
+#include "common/fs/path_util.h"
 #include "common/settings.h"
 #include "video_core/macro/macro.h"
 #include "video_core/macro/macro_hle.h"
@@ -15,6 +19,23 @@
 
 namespace Tegra {
 
+static void Dump(u64 hash, std::span<const u32> code) {
+    const auto base_dir{Common::FS::GetYuzuPath(Common::FS::YuzuPath::DumpDir)};
+    const auto macro_dir{base_dir / "macros"};
+    if (!Common::FS::CreateDir(base_dir) || !Common::FS::CreateDir(macro_dir)) {
+        LOG_ERROR(Common_Filesystem, "Failed to create macro dump directories");
+        return;
+    }
+    const auto name{macro_dir / fmt::format("{:016x}.macro", hash)};
+    std::fstream macro_file(name, std::ios::out | std::ios::binary);
+    if (!macro_file) {
+        LOG_ERROR(Common_Filesystem, "Unable to open or create file at {}",
+                  Common::FS::PathToUTF8String(name));
+        return;
+    }
+    macro_file.write(reinterpret_cast<const char*>(code.data()), code.size_bytes());
+}
+
 MacroEngine::MacroEngine(Engines::Maxwell3D& maxwell3d)
     : hle_macros{std::make_unique<Tegra::HLEMacro>(maxwell3d)} {}
 
@@ -24,6 +45,11 @@ void MacroEngine::AddCode(u32 method, u32 data) {
     uploaded_macro_code[method].push_back(data);
 }
 
+void MacroEngine::ClearCode(u32 method) {
+    macro_cache.erase(method);
+    uploaded_macro_code.erase(method);
+}
+
 void MacroEngine::Execute(u32 method, const std::vector<u32>& parameters) {
     auto compiled_macro = macro_cache.find(method);
     if (compiled_macro != macro_cache.end()) {
@@ -54,6 +80,9 @@ void MacroEngine::Execute(u32 method, const std::vector<u32>& parameters) {
         if (!mid_method.has_value()) {
             cache_info.lle_program = Compile(macro_code->second);
             cache_info.hash = boost::hash_value(macro_code->second);
+            if (Settings::values.dump_macros) {
+                Dump(cache_info.hash, macro_code->second);
+            }
         } else {
             const auto& macro_cached = uploaded_macro_code[mid_method.value()];
             const auto rebased_method = method - mid_method.value();
@@ -63,6 +92,9 @@ void MacroEngine::Execute(u32 method, const std::vector<u32>& parameters) {
                         code.size() * sizeof(u32));
             cache_info.hash = boost::hash_value(code);
             cache_info.lle_program = Compile(code);
+            if (Settings::values.dump_macros) {
+                Dump(cache_info.hash, code);
+            }
         }
 
         if (auto hle_program = hle_macros->GetHLEProgram(cache_info.hash)) {
diff --git a/src/video_core/macro/macro.h b/src/video_core/macro/macro.h
index 7e12c16dc..07d97ba39 100644
--- a/src/video_core/macro/macro.h
+++ b/src/video_core/macro/macro.h
@@ -117,6 +117,9 @@ public:
     // Store the uploaded macro code to compile them when they're called.
     void AddCode(u32 method, u32 data);
 
+    // Clear the code associated with a method.
+    void ClearCode(u32 method);
+
     // Compiles the macro if its not in the cache, and executes the compiled macro
     void Execute(u32 method, const std::vector<u32>& parameters);
 
diff --git a/src/video_core/macro/macro_jit_x64.cpp b/src/video_core/macro/macro_jit_x64.cpp
index dc2b490d4..dc5376501 100644
--- a/src/video_core/macro/macro_jit_x64.cpp
+++ b/src/video_core/macro/macro_jit_x64.cpp
@@ -23,7 +23,8 @@ MICROPROFILE_DEFINE(MacroJitExecute, "GPU", "Execute macro JIT", MP_RGB(255, 255
 namespace Tegra {
 namespace {
 constexpr Xbyak::Reg64 STATE = Xbyak::util::rbx;
-constexpr Xbyak::Reg32 RESULT = Xbyak::util::ebp;
+constexpr Xbyak::Reg32 RESULT = Xbyak::util::r10d;
+constexpr Xbyak::Reg64 MAX_PARAMETER = Xbyak::util::r11;
 constexpr Xbyak::Reg64 PARAMETERS = Xbyak::util::r12;
 constexpr Xbyak::Reg32 METHOD_ADDRESS = Xbyak::util::r14d;
 constexpr Xbyak::Reg64 BRANCH_HOLDER = Xbyak::util::r15;
@@ -31,6 +32,7 @@ constexpr Xbyak::Reg64 BRANCH_HOLDER = Xbyak::util::r15;
 constexpr std::bitset<32> PERSISTENT_REGISTERS = Common::X64::BuildRegSet({
     STATE,
     RESULT,
+    MAX_PARAMETER,
     PARAMETERS,
     METHOD_ADDRESS,
     BRANCH_HOLDER,
@@ -80,7 +82,7 @@ private:
         u32 carry_flag{};
     };
     static_assert(offsetof(JITState, maxwell3d) == 0, "Maxwell3D is not at 0x0");
-    using ProgramType = void (*)(JITState*, const u32*);
+    using ProgramType = void (*)(JITState*, const u32*, const u32*);
 
     struct OptimizerState {
         bool can_skip_carry{};
@@ -112,7 +114,7 @@ void MacroJITx64Impl::Execute(const std::vector<u32>& parameters, u32 method) {
     JITState state{};
     state.maxwell3d = &maxwell3d;
     state.registers = {};
-    program(&state, parameters.data());
+    program(&state, parameters.data(), parameters.data() + parameters.size());
 }
 
 void MacroJITx64Impl::Compile_ALU(Macro::Opcode opcode) {
@@ -488,6 +490,7 @@ void MacroJITx64Impl::Compile() {
     // JIT state
     mov(STATE, Common::X64::ABI_PARAM1);
     mov(PARAMETERS, Common::X64::ABI_PARAM2);
+    mov(MAX_PARAMETER, Common::X64::ABI_PARAM3);
     xor_(RESULT, RESULT);
     xor_(METHOD_ADDRESS, METHOD_ADDRESS);
     xor_(BRANCH_HOLDER, BRANCH_HOLDER);
@@ -598,7 +601,22 @@ bool MacroJITx64Impl::Compile_NextInstruction() {
     return true;
 }
 
+static void WarnInvalidParameter(uintptr_t parameter, uintptr_t max_parameter) {
+    LOG_CRITICAL(HW_GPU,
+                 "Macro JIT: invalid parameter access 0x{:x} (0x{:x} is the last parameter)",
+                 parameter, max_parameter - sizeof(u32));
+}
+
 Xbyak::Reg32 MacroJITx64Impl::Compile_FetchParameter() {
+    Xbyak::Label parameter_ok{};
+    cmp(PARAMETERS, MAX_PARAMETER);
+    jb(parameter_ok, T_NEAR);
+    Common::X64::ABI_PushRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
+    mov(Common::X64::ABI_PARAM1, PARAMETERS);
+    mov(Common::X64::ABI_PARAM2, MAX_PARAMETER);
+    Common::X64::CallFarFunction(*this, &WarnInvalidParameter);
+    Common::X64::ABI_PopRegistersAndAdjustStack(*this, PersistentCallerSavedRegs(), 0);
+    L(parameter_ok);
     mov(eax, dword[PARAMETERS]);
     add(PARAMETERS, sizeof(u32));
     return eax;
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 8ef79753f..159b71161 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -563,12 +563,11 @@ void RasterizerOpenGL::SyncViewport() {
         flags[Dirty::FrontFace] = false;
 
         GLenum mode = MaxwellToGL::FrontFace(regs.front_face);
-        bool flip_faces = false;
-        if (regs.screen_y_control.triangle_rast_flip != 0 &&
-            regs.viewport_transform[0].scale_y < 0.0f) {
+        bool flip_faces = true;
+        if (regs.screen_y_control.triangle_rast_flip != 0) {
             flip_faces = !flip_faces;
         }
-        if (regs.viewport_transform[0].scale_z < 0.0f) {
+        if (regs.viewport_transform[0].scale_y < 0.0f) {
             flip_faces = !flip_faces;
         }
         if (flip_faces) {
diff --git a/src/yuzu/configuration/configure_debug.cpp b/src/yuzu/configuration/configure_debug.cpp
index bd50f7a68..d6e8b5ead 100644
--- a/src/yuzu/configuration/configure_debug.cpp
+++ b/src/yuzu/configuration/configure_debug.cpp
@@ -53,6 +53,8 @@ void ConfigureDebug::SetConfiguration() {
     ui->enable_nsight_aftermath->setChecked(Settings::values.enable_nsight_aftermath.GetValue());
     ui->dump_shaders->setEnabled(runtime_lock);
     ui->dump_shaders->setChecked(Settings::values.dump_shaders.GetValue());
+    ui->dump_macros->setEnabled(runtime_lock);
+    ui->dump_macros->setChecked(Settings::values.dump_macros.GetValue());
     ui->disable_macro_jit->setEnabled(runtime_lock);
     ui->disable_macro_jit->setChecked(Settings::values.disable_macro_jit.GetValue());
     ui->disable_loop_safety_checks->setEnabled(runtime_lock);
@@ -83,6 +85,7 @@ void ConfigureDebug::ApplyConfiguration() {
     Settings::values.cpu_debug_mode = ui->enable_cpu_debugging->isChecked();
     Settings::values.enable_nsight_aftermath = ui->enable_nsight_aftermath->isChecked();
     Settings::values.dump_shaders = ui->dump_shaders->isChecked();
+    Settings::values.dump_macros = ui->dump_macros->isChecked();
     Settings::values.disable_shader_loop_safety_checks =
         ui->disable_loop_safety_checks->isChecked();
     Settings::values.disable_macro_jit = ui->disable_macro_jit->isChecked();
diff --git a/src/yuzu/configuration/configure_debug.ui b/src/yuzu/configuration/configure_debug.ui
index c1d90d588..863a3fd57 100644
--- a/src/yuzu/configuration/configure_debug.ui
+++ b/src/yuzu/configuration/configure_debug.ui
@@ -118,6 +118,19 @@
         </property>
        </widget>
       </item>
+      <item row="0" column="2">
+       <widget class="QCheckBox" name="dump_macros">
+        <property name="enabled">
+         <bool>true</bool>
+        </property>
+        <property name="toolTip">
+         <string>When checked, it will dump all the macro programs of the GPU</string>
+        </property>
+        <property name="text">
+         <string>Dump Maxwell Macros</string>
+        </property>
+       </widget>
+      </item>
       <item row="0" column="1">
        <widget class="QCheckBox" name="disable_macro_jit">
         <property name="enabled">