12 files changed, 103 insertions, 41 deletions

diff --git a/src/common/telemetry.cpp b/src/common/telemetry.cpp
index 6241d08b3..98c82cd17 100644
--- a/src/common/telemetry.cpp
+++ b/src/common/telemetry.cpp
@@ -55,22 +55,50 @@ void AppendBuildInfo(FieldCollection& fc) {
 
 void AppendCPUInfo(FieldCollection& fc) {
 #ifdef ARCHITECTURE_x86_64
-    fc.AddField(FieldType::UserSystem, "CPU_Model", Common::GetCPUCaps().cpu_string);
-    fc.AddField(FieldType::UserSystem, "CPU_BrandString", Common::GetCPUCaps().brand_string);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_AES", Common::GetCPUCaps().aes);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_AVX", Common::GetCPUCaps().avx);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_AVX2", Common::GetCPUCaps().avx2);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_AVX512", Common::GetCPUCaps().avx512);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_BMI1", Common::GetCPUCaps().bmi1);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_BMI2", Common::GetCPUCaps().bmi2);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_FMA", Common::GetCPUCaps().fma);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_FMA4", Common::GetCPUCaps().fma4);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_SSE", Common::GetCPUCaps().sse);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_SSE2", Common::GetCPUCaps().sse2);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_SSE3", Common::GetCPUCaps().sse3);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_SSSE3", Common::GetCPUCaps().ssse3);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_SSE41", Common::GetCPUCaps().sse4_1);
-    fc.AddField(FieldType::UserSystem, "CPU_Extension_x64_SSE42", Common::GetCPUCaps().sse4_2);
+
+    const auto& caps = Common::GetCPUCaps();
+    const auto add_field = [&fc](std::string_view field_name, const auto& field_value) {
+        fc.AddField(FieldType::UserSystem, field_name, field_value);
+    };
+    add_field("CPU_Model", caps.cpu_string);
+    add_field("CPU_BrandString", caps.brand_string);
+
+    add_field("CPU_Extension_x64_SSE", caps.sse);
+    add_field("CPU_Extension_x64_SSE2", caps.sse2);
+    add_field("CPU_Extension_x64_SSE3", caps.sse3);
+    add_field("CPU_Extension_x64_SSSE3", caps.ssse3);
+    add_field("CPU_Extension_x64_SSE41", caps.sse4_1);
+    add_field("CPU_Extension_x64_SSE42", caps.sse4_2);
+
+    add_field("CPU_Extension_x64_AVX", caps.avx);
+    add_field("CPU_Extension_x64_AVX_VNNI", caps.avx_vnni);
+    add_field("CPU_Extension_x64_AVX2", caps.avx2);
+
+    // Skylake-X/SP level AVX512, for compatibility with the previous telemetry field
+    add_field("CPU_Extension_x64_AVX512",
+              caps.avx512f && caps.avx512cd && caps.avx512vl && caps.avx512dq && caps.avx512bw);
+
+    add_field("CPU_Extension_x64_AVX512F", caps.avx512f);
+    add_field("CPU_Extension_x64_AVX512CD", caps.avx512cd);
+    add_field("CPU_Extension_x64_AVX512VL", caps.avx512vl);
+    add_field("CPU_Extension_x64_AVX512DQ", caps.avx512dq);
+    add_field("CPU_Extension_x64_AVX512BW", caps.avx512bw);
+    add_field("CPU_Extension_x64_AVX512BITALG", caps.avx512bitalg);
+    add_field("CPU_Extension_x64_AVX512VBMI", caps.avx512vbmi);
+
+    add_field("CPU_Extension_x64_AES", caps.aes);
+    add_field("CPU_Extension_x64_BMI1", caps.bmi1);
+    add_field("CPU_Extension_x64_BMI2", caps.bmi2);
+    add_field("CPU_Extension_x64_F16C", caps.f16c);
+    add_field("CPU_Extension_x64_FMA", caps.fma);
+    add_field("CPU_Extension_x64_FMA4", caps.fma4);
+    add_field("CPU_Extension_x64_GFNI", caps.gfni);
+    add_field("CPU_Extension_x64_INVARIANT_TSC", caps.invariant_tsc);
+    add_field("CPU_Extension_x64_LZCNT", caps.lzcnt);
+    add_field("CPU_Extension_x64_MOVBE", caps.movbe);
+    add_field("CPU_Extension_x64_PCLMULQDQ", caps.pclmulqdq);
+    add_field("CPU_Extension_x64_POPCNT", caps.popcnt);
+    add_field("CPU_Extension_x64_SHA", caps.sha);
 #else
     fc.AddField(FieldType::UserSystem, "CPU_Model", "Other");
 #endif
diff --git a/src/common/telemetry.h b/src/common/telemetry.h
index 4d632f7eb..3524c857e 100644
--- a/src/common/telemetry.h
+++ b/src/common/telemetry.h
@@ -8,6 +8,7 @@
 #include <map>
 #include <memory>
 #include <string>
+#include <string_view>
 #include "common/common_funcs.h"
 #include "common/common_types.h"
 
@@ -55,8 +56,8 @@ class Field : public FieldInterface {
 public:
     YUZU_NON_COPYABLE(Field);
 
-    Field(FieldType type_, std::string name_, T value_)
-        : name(std::move(name_)), type(type_), value(std::move(value_)) {}
+    Field(FieldType type_, std::string_view name_, T value_)
+        : name(name_), type(type_), value(std::move(value_)) {}
 
     ~Field() override = default;
 
@@ -123,7 +124,7 @@ public:
      * @param value Value for the field to add.
      */
     template <typename T>
-    void AddField(FieldType type, const char* name, T value) {
+    void AddField(FieldType type, std::string_view name, T value) {
         return AddField(std::make_unique<Field<T>>(type, name, std::move(value)));
     }
 
diff --git a/src/common/x64/cpu_detect.cpp b/src/common/x64/cpu_detect.cpp
index 99d87f586..d81edb140 100644
--- a/src/common/x64/cpu_detect.cpp
+++ b/src/common/x64/cpu_detect.cpp
@@ -93,10 +93,14 @@ static CPUCaps Detect() {
         caps.sse = Common::Bit<25>(cpu_id[3]);
         caps.sse2 = Common::Bit<26>(cpu_id[3]);
         caps.sse3 = Common::Bit<0>(cpu_id[2]);
+        caps.pclmulqdq = Common::Bit<1>(cpu_id[2]);
         caps.ssse3 = Common::Bit<9>(cpu_id[2]);
         caps.sse4_1 = Common::Bit<19>(cpu_id[2]);
         caps.sse4_2 = Common::Bit<20>(cpu_id[2]);
+        caps.movbe = Common::Bit<22>(cpu_id[2]);
+        caps.popcnt = Common::Bit<23>(cpu_id[2]);
         caps.aes = Common::Bit<25>(cpu_id[2]);
+        caps.f16c = Common::Bit<29>(cpu_id[2]);
 
         // AVX support requires 3 separate checks:
         //  - Is the AVX bit set in CPUID?
@@ -112,16 +116,26 @@ static CPUCaps Detect() {
 
         if (max_std_fn >= 7) {
             __cpuidex(cpu_id, 0x00000007, 0x00000000);
-            // Can't enable AVX2 unless the XSAVE/XGETBV checks above passed
-            caps.avx2 = caps.avx && Common::Bit<5>(cpu_id[1]);
+            // Can't enable AVX{2,512} unless the XSAVE/XGETBV checks above passed
+            if (caps.avx) {
+                caps.avx2 = Common::Bit<5>(cpu_id[1]);
+                caps.avx512f = Common::Bit<16>(cpu_id[1]);
+                caps.avx512dq = Common::Bit<17>(cpu_id[1]);
+                caps.avx512cd = Common::Bit<28>(cpu_id[1]);
+                caps.avx512bw = Common::Bit<30>(cpu_id[1]);
+                caps.avx512vl = Common::Bit<31>(cpu_id[1]);
+                caps.avx512vbmi = Common::Bit<1>(cpu_id[2]);
+                caps.avx512bitalg = Common::Bit<12>(cpu_id[2]);
+            }
+
             caps.bmi1 = Common::Bit<3>(cpu_id[1]);
             caps.bmi2 = Common::Bit<8>(cpu_id[1]);
-            // Checks for AVX512F, AVX512CD, AVX512VL, AVX512DQ, AVX512BW (Intel Skylake-X/SP)
-            if (Common::Bit<16>(cpu_id[1]) && Common::Bit<28>(cpu_id[1]) &&
-                Common::Bit<31>(cpu_id[1]) && Common::Bit<17>(cpu_id[1]) &&
-                Common::Bit<30>(cpu_id[1])) {
-                caps.avx512 = caps.avx2;
-            }
+            caps.sha = Common::Bit<29>(cpu_id[1]);
+
+            caps.gfni = Common::Bit<8>(cpu_id[2]);
+
+            __cpuidex(cpu_id, 0x00000007, 0x00000001);
+            caps.avx_vnni = caps.avx && Common::Bit<4>(cpu_id[0]);
         }
     }
 
diff --git a/src/common/x64/cpu_detect.h b/src/common/x64/cpu_detect.h
index 3e6d808f3..40c48b132 100644
--- a/src/common/x64/cpu_detect.h
+++ b/src/common/x64/cpu_detect.h
@@ -35,16 +35,31 @@ struct CPUCaps {
     bool ssse3 : 1;
     bool sse4_1 : 1;
     bool sse4_2 : 1;
-    bool lzcnt : 1;
+
     bool avx : 1;
+    bool avx_vnni : 1;
     bool avx2 : 1;
-    bool avx512 : 1;
+    bool avx512f : 1;
+    bool avx512dq : 1;
+    bool avx512cd : 1;
+    bool avx512bw : 1;
+    bool avx512vl : 1;
+    bool avx512vbmi : 1;
+    bool avx512bitalg : 1;
+
+    bool aes : 1;
     bool bmi1 : 1;
     bool bmi2 : 1;
+    bool f16c : 1;
     bool fma : 1;
     bool fma4 : 1;
-    bool aes : 1;
+    bool gfni : 1;
     bool invariant_tsc : 1;
+    bool lzcnt : 1;
+    bool movbe : 1;
+    bool pclmulqdq : 1;
+    bool popcnt : 1;
+    bool sha : 1;
 };
 
 /**
diff --git a/src/core/arm/dynarmic/arm_dynarmic_32.cpp b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
index 286976623..c1c843b8f 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_32.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_32.cpp
@@ -148,8 +148,8 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
     config.wall_clock_cntpct = uses_wall_clock;
 
     // Code cache size
-    config.code_cache_size = 512_MiB;
-    config.far_code_offset = 400_MiB;
+    config.code_cache_size = 128_MiB;
+    config.far_code_offset = 100_MiB;
 
     // Safe optimizations
     if (Settings::values.cpu_debug_mode) {
diff --git a/src/core/arm/dynarmic/arm_dynarmic_64.cpp b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
index d96226c41..aa74fce4d 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_64.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
@@ -208,8 +208,8 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
     config.wall_clock_cntpct = uses_wall_clock;
 
     // Code cache size
-    config.code_cache_size = 512_MiB;
-    config.far_code_offset = 400_MiB;
+    config.code_cache_size = 128_MiB;
+    config.far_code_offset = 100_MiB;
 
     // Safe optimizations
     if (Settings::values.cpu_debug_mode) {
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv.h b/src/shader_recompiler/backend/spirv/emit_spirv.h
index b412957c7..2b360e073 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv.h
+++ b/src/shader_recompiler/backend/spirv/emit_spirv.h
@@ -22,7 +22,7 @@ constexpr u32 NUM_TEXTURE_AND_IMAGE_SCALING_WORDS =
 struct RescalingLayout {
     alignas(16) std::array<u32, NUM_TEXTURE_SCALING_WORDS> rescaling_textures;
     alignas(16) std::array<u32, NUM_IMAGE_SCALING_WORDS> rescaling_images;
-    alignas(16) u32 down_factor;
+    u32 down_factor;
 };
 constexpr u32 RESCALING_LAYOUT_WORDS_OFFSET = offsetof(RescalingLayout, rescaling_textures);
 constexpr u32 RESCALING_LAYOUT_DOWN_FACTOR_OFFSET = offsetof(RescalingLayout, down_factor);
diff --git a/src/shader_recompiler/frontend/maxwell/translate_program.cpp b/src/shader_recompiler/frontend/maxwell/translate_program.cpp
index 248ad3ced..b22725584 100644
--- a/src/shader_recompiler/frontend/maxwell/translate_program.cpp
+++ b/src/shader_recompiler/frontend/maxwell/translate_program.cpp
@@ -212,11 +212,11 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     }
     Optimization::SsaRewritePass(program);
 
+    Optimization::ConstantPropagationPass(program);
+
     Optimization::GlobalMemoryToStorageBufferPass(program);
     Optimization::TexturePass(env, program);
 
-    Optimization::ConstantPropagationPass(program);
-
     if (Settings::values.resolution_info.active) {
         Optimization::RescalingPass(program);
     }
diff --git a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp
index 38592afd0..ddf497e32 100644
--- a/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp
+++ b/src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp
@@ -334,7 +334,8 @@ std::optional<LowAddrInfo> TrackLowAddress(IR::Inst* inst) {
 /// Tries to track the storage buffer address used by a global memory instruction
 std::optional<StorageBufferAddr> Track(const IR::Value& value, const Bias* bias) {
     const auto pred{[bias](const IR::Inst* inst) -> std::optional<StorageBufferAddr> {
-        if (inst->GetOpcode() != IR::Opcode::GetCbufU32) {
+        if (inst->GetOpcode() != IR::Opcode::GetCbufU32 &&
+            inst->GetOpcode() != IR::Opcode::GetCbufU32x2) {
             return std::nullopt;
         }
         const IR::Value index{inst->Arg(0)};
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.cpp b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
index 3e96c0f60..4d73427b4 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pass.cpp
+++ b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
@@ -2,6 +2,7 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <array>
 #include <cstring>
 #include <memory>
 #include <optional>
@@ -292,7 +293,7 @@ std::pair<VkBuffer, VkDeviceSize> QuadIndexedPass::Assemble(
             .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
             .dstAccessMask = VK_ACCESS_INDEX_READ_BIT,
         };
-        const std::array push_constants{base_vertex, index_shift};
+        const std::array<u32, 2> push_constants{base_vertex, index_shift};
         const VkDescriptorSet set = descriptor_allocator.Commit();
         device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
         cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
diff --git a/src/yuzu/configuration/config.cpp b/src/yuzu/configuration/config.cpp
index c2b66ff14..4b943c6ba 100644
--- a/src/yuzu/configuration/config.cpp
+++ b/src/yuzu/configuration/config.cpp
@@ -1155,6 +1155,8 @@ void Config::SaveCpuValues() {
         WriteBasicSetting(Settings::values.cpuopt_misc_ir);
         WriteBasicSetting(Settings::values.cpuopt_reduce_misalign_checks);
         WriteBasicSetting(Settings::values.cpuopt_fastmem);
+        WriteBasicSetting(Settings::values.cpuopt_fastmem_exclusives);
+        WriteBasicSetting(Settings::values.cpuopt_recompile_exclusives);
     }
 
     qt_config->endGroup();
diff --git a/src/yuzu/main.cpp b/src/yuzu/main.cpp
index b3a8da0ea..1d459bdb3 100644
--- a/src/yuzu/main.cpp
+++ b/src/yuzu/main.cpp
@@ -249,9 +249,9 @@ GMainWindow::GMainWindow()
 #ifdef ARCHITECTURE_x86_64
     const auto& caps = Common::GetCPUCaps();
     std::string cpu_string = caps.cpu_string;
-    if (caps.avx || caps.avx2 || caps.avx512) {
+    if (caps.avx || caps.avx2 || caps.avx512f) {
         cpu_string += " | AVX";
-        if (caps.avx512) {
+        if (caps.avx512f) {
             cpu_string += "512";
         } else if (caps.avx2) {
             cpu_string += '2';