backend/arm64: Add verbose debugging output

9 files changed, 216 insertions, 0 deletions

diff --git a/src/dynarmic/CMakeLists.txt b/src/dynarmic/CMakeLists.txt
index 6e39253c..5b084d9a 100644
--- a/src/dynarmic/CMakeLists.txt
+++ b/src/dynarmic/CMakeLists.txt
@@ -373,6 +373,8 @@ if ("arm64" IN_LIST ARCHITECTURE)
         backend/arm64/reg_alloc.cpp
         backend/arm64/reg_alloc.h
         backend/arm64/stack_layout.h
+        backend/arm64/verbose_debugging_output.cpp
+        backend/arm64/verbose_debugging_output.h
         common/spin_lock_arm64.cpp
         common/spin_lock_arm64.h
     )
diff --git a/src/dynarmic/backend/arm64/a32_address_space.cpp b/src/dynarmic/backend/arm64/a32_address_space.cpp
index 8a630c93..8723326b 100644
--- a/src/dynarmic/backend/arm64/a32_address_space.cpp
+++ b/src/dynarmic/backend/arm64/a32_address_space.cpp
@@ -174,6 +174,7 @@ IR::Block A32AddressSpace::GenerateIR(IR::LocationDescriptor descriptor) const {
         Optimization::ConstantPropagation(ir_block);
         Optimization::DeadCodeElimination(ir_block);
     }
+    Optimization::IdentityRemovalPass(ir_block);
     Optimization::VerificationPass(ir_block);
 
     return ir_block;
@@ -385,6 +386,8 @@ EmitConfig A32AddressSpace::GetEmitConfig() {
         .state_exclusive_state_offset = offsetof(A32JitState, exclusive_state),
 
         .coprocessors = conf.coprocessors,
+
+        .very_verbose_debugging_output = conf.very_verbose_debugging_output,
     };
 }
 
diff --git a/src/dynarmic/backend/arm64/a64_address_space.cpp b/src/dynarmic/backend/arm64/a64_address_space.cpp
index 06966695..28016384 100644
--- a/src/dynarmic/backend/arm64/a64_address_space.cpp
+++ b/src/dynarmic/backend/arm64/a64_address_space.cpp
@@ -562,6 +562,8 @@ EmitConfig A64AddressSpace::GetEmitConfig() {
         .state_exclusive_state_offset = offsetof(A64JitState, exclusive_state),
 
         .coprocessors{},
+
+        .very_verbose_debugging_output = conf.very_verbose_debugging_output,
     };
 }
 
diff --git a/src/dynarmic/backend/arm64/emit_arm64.cpp b/src/dynarmic/backend/arm64/emit_arm64.cpp
index e5ad165c..05358731 100644
--- a/src/dynarmic/backend/arm64/emit_arm64.cpp
+++ b/src/dynarmic/backend/arm64/emit_arm64.cpp
@@ -11,6 +11,7 @@
 #include "dynarmic/backend/arm64/emit_context.h"
 #include "dynarmic/backend/arm64/fpsr_manager.h"
 #include "dynarmic/backend/arm64/reg_alloc.h"
+#include "dynarmic/backend/arm64/verbose_debugging_output.h"
 #include "dynarmic/ir/basic_block.h"
 #include "dynarmic/ir/microinstruction.h"
 #include "dynarmic/ir/opcodes.h"
@@ -175,6 +176,10 @@ static void EmitAddCycles(oaknut::CodeGenerator& code, EmitContext& ctx, size_t
 }
 
 EmittedBlockInfo EmitArm64(oaknut::CodeGenerator& code, IR::Block block, const EmitConfig& conf, FastmemManager& fastmem_manager) {
+    if (conf.very_verbose_debugging_output) {
+        std::puts(IR::DumpBlock(block).c_str());
+    }
+
     EmittedBlockInfo ebi;
 
     FpsrManager fpsr_manager{code, conf.state_fpsr_offset};
@@ -223,6 +228,10 @@ EmittedBlockInfo EmitArm64(oaknut::CodeGenerator& code, IR::Block block, const E
 
         reg_alloc.UpdateAllUses();
         reg_alloc.AssertAllUnlocked();
+
+        if (conf.very_verbose_debugging_output) {
+            EmitVerboseDebuggingOutput(code, ctx);
+        }
     }
 
     fpsr_manager.Spill();
diff --git a/src/dynarmic/backend/arm64/emit_arm64.h b/src/dynarmic/backend/arm64/emit_arm64.h
index c15d9266..6d13e7e3 100644
--- a/src/dynarmic/backend/arm64/emit_arm64.h
+++ b/src/dynarmic/backend/arm64/emit_arm64.h
@@ -155,6 +155,9 @@ struct EmitConfig {
 
     // A32 specific
     std::array<std::shared_ptr<A32::Coprocessor>, 16> coprocessors{};
+
+    // Debugging
+    bool very_verbose_debugging_output;
 };
 
 EmittedBlockInfo EmitArm64(oaknut::CodeGenerator& code, IR::Block block, const EmitConfig& emit_conf, FastmemManager& fastmem_manager);
diff --git a/src/dynarmic/backend/arm64/reg_alloc.cpp b/src/dynarmic/backend/arm64/reg_alloc.cpp
index 83accac2..a8fd2ec7 100644
--- a/src/dynarmic/backend/arm64/reg_alloc.cpp
+++ b/src/dynarmic/backend/arm64/reg_alloc.cpp
@@ -7,14 +7,18 @@
 
 #include <algorithm>
 #include <array>
+#include <iterator>
 
 #include <mcl/assert.hpp>
 #include <mcl/bit/bit_field.hpp>
+#include <mcl/bit_cast.hpp>
 #include <mcl/mp/metavalue/lift_value.hpp>
 #include <mcl/stdint.hpp>
 
 #include "dynarmic/backend/arm64/abi.h"
+#include "dynarmic/backend/arm64/emit_context.h"
 #include "dynarmic/backend/arm64/fpsr_manager.h"
+#include "dynarmic/backend/arm64/verbose_debugging_output.h"
 #include "dynarmic/common/always_false.h"
 
 namespace Dynarmic::Backend::Arm64 {
@@ -234,6 +238,34 @@ void RegAlloc::AssertNoMoreUses() const {
     ASSERT(std::all_of(spills.begin(), spills.end(), is_empty));
 }
 
+void RegAlloc::EmitVerboseDebuggingOutput(EmitContext& ctx) {
+    code.MOV(X19, mcl::bit_cast<u64>(&PrintVerboseDebuggingOutputLine));  // Non-volatile register
+
+    const auto do_location = [&](HostLocInfo& info, HostLocType type, size_t index) {
+        using namespace oaknut::util;
+        for (const IR::Inst* value : info.values) {
+            const auto inst_offset = std::distance(ctx.block.begin(), IR::Block::iterator(const_cast<IR::Inst*>(value)));
+            code.MOV(X0, SP);
+            code.MOV(X1, static_cast<u64>(type));
+            code.MOV(X2, index);
+            code.MOV(X3, mcl::bit_cast<u64>(inst_offset));
+            code.MOV(X4, static_cast<u64>(value->GetType()));
+            code.BLR(X19);
+        }
+    };
+
+    for (size_t i = 0; i < gprs.size(); i++) {
+        do_location(gprs[i], HostLocType::X, i);
+    }
+    for (size_t i = 0; i < fprs.size(); i++) {
+        do_location(fprs[i], HostLocType::Q, i);
+    }
+    do_location(flags, HostLocType::Nzcv, 0);
+    for (size_t i = 0; i < spills.size(); i++) {
+        do_location(spills[i], HostLocType::Spill, i);
+    }
+}
+
 template<HostLoc::Kind kind>
 int RegAlloc::GenerateImmediate(const IR::Value& value) {
     ASSERT(value.GetType() != IR::Type::U1);
diff --git a/src/dynarmic/backend/arm64/reg_alloc.h b/src/dynarmic/backend/arm64/reg_alloc.h
index a101ee6d..6c97d4af 100644
--- a/src/dynarmic/backend/arm64/reg_alloc.h
+++ b/src/dynarmic/backend/arm64/reg_alloc.h
@@ -23,6 +23,7 @@
 
 namespace Dynarmic::Backend::Arm64 {
 
+struct EmitContext;
 class FpsrManager;
 class RegAlloc;
 
@@ -290,6 +291,8 @@ public:
     void AssertAllUnlocked() const;
     void AssertNoMoreUses() const;
 
+    void EmitVerboseDebuggingOutput(EmitContext& ctx);
+
 private:
     friend struct Argument;
     template<typename>
diff --git a/src/dynarmic/backend/arm64/verbose_debugging_output.cpp b/src/dynarmic/backend/arm64/verbose_debugging_output.cpp
new file mode 100644
index 00000000..fbb26bc4
--- /dev/null
+++ b/src/dynarmic/backend/arm64/verbose_debugging_output.cpp
@@ -0,0 +1,104 @@
+/* This file is part of the dynarmic project.
+ * Copyright (c) 2023 MerryMage
+ * SPDX-License-Identifier: 0BSD
+ */
+
+#include "dynarmic/backend/arm64/verbose_debugging_output.h"
+
+#include <fmt/format.h>
+#include <oaknut/oaknut.hpp>
+
+#include "dynarmic/backend/arm64/emit_context.h"
+#include "dynarmic/ir/type.h"
+
+namespace Dynarmic::Backend::Arm64 {
+
+using namespace oaknut::util;
+
+void EmitVerboseDebuggingOutput(oaknut::CodeGenerator& code, EmitContext& ctx) {
+    code.SUB(SP, SP, sizeof(RegisterData));
+    for (int i = 0; i < 30; i++) {
+        if (i == 18) {
+            continue;  // Platform register
+        }
+        code.STR(oaknut::XReg{i}, SP, offsetof(RegisterData, x) + i * sizeof(u64));
+    }
+    for (int i = 0; i < 32; i++) {
+        code.STR(oaknut::QReg{i}, SP, offsetof(RegisterData, q) + i * sizeof(Vector));
+    }
+    code.MRS(X0, oaknut::SystemReg::NZCV);
+    code.STR(X0, SP, offsetof(RegisterData, nzcv));
+    code.ADD(X0, SP, sizeof(RegisterData) + offsetof(StackLayout, spill));
+    code.STR(X0, SP, offsetof(RegisterData, spill));
+
+    ctx.reg_alloc.EmitVerboseDebuggingOutput(ctx);
+
+    code.LDR(X0, SP, offsetof(RegisterData, nzcv));
+    code.MSR(oaknut::SystemReg::NZCV, X0);
+    for (int i = 0; i < 32; i++) {
+        code.LDR(oaknut::QReg{i}, SP, offsetof(RegisterData, q) + i * sizeof(Vector));
+    }
+    for (int i = 0; i < 30; i++) {
+        if (i == 18) {
+            continue;  // Platform register
+        }
+        code.LDR(oaknut::XReg{i}, SP, offsetof(RegisterData, x) + i * sizeof(u64));
+    }
+    code.ADD(SP, SP, sizeof(RegisterData));
+}
+
+void PrintVerboseDebuggingOutputLine(RegisterData& reg_data, HostLocType reg_type, size_t reg_index, size_t inst_index, IR::Type inst_type) {
+    fmt::print("dynarmic debug: %{:05} = ", inst_index);
+
+    Vector value = [&]() -> Vector {
+        switch (reg_type) {
+        case HostLocType::X:
+            return {reg_data.x[reg_index], 0};
+        case HostLocType::Q:
+            return reg_data.q[reg_index];
+        case HostLocType::Nzcv:
+            return {reg_data.nzcv, 0};
+        case HostLocType::Spill:
+            return (*reg_data.spill)[reg_index];
+        }
+        fmt::print("invalid reg_type! ");
+        return {0, 0};
+    }();
+
+    switch (inst_type) {
+    case IR::Type::U1:
+    case IR::Type::U8:
+        fmt::print("{:02x}", value[0] & 0xff);
+        break;
+    case IR::Type::U16:
+        fmt::print("{:04x}", value[0] & 0xffff);
+        break;
+    case IR::Type::U32:
+    case IR::Type::NZCVFlags:
+        fmt::print("{:08x}", value[0] & 0xffffffff);
+        break;
+    case IR::Type::U64:
+        fmt::print("{:016x}", value[0]);
+        break;
+    case IR::Type::U128:
+        fmt::print("{:016x}{:016x}", value[1], value[0]);
+        break;
+    case IR::Type::A32Reg:
+    case IR::Type::A32ExtReg:
+    case IR::Type::A64Reg:
+    case IR::Type::A64Vec:
+    case IR::Type::CoprocInfo:
+    case IR::Type::Cond:
+    case IR::Type::Void:
+    case IR::Type::Table:
+    case IR::Type::AccType:
+    case IR::Type::Opaque:
+    default:
+        fmt::print("invalid inst_type!");
+        break;
+    }
+
+    fmt::print("\n");
+}
+
+}  // namespace Dynarmic::Backend::Arm64
\ No newline at end of file
diff --git a/src/dynarmic/backend/arm64/verbose_debugging_output.h b/src/dynarmic/backend/arm64/verbose_debugging_output.h
new file mode 100644
index 00000000..1099edc4
--- /dev/null
+++ b/src/dynarmic/backend/arm64/verbose_debugging_output.h
@@ -0,0 +1,58 @@
+/* This file is part of the dynarmic project.
+ * Copyright (c) 2023 MerryMage
+ * SPDX-License-Identifier: 0BSD
+ */
+
+#pragma once
+
+#include <array>
+
+#include <mcl/stdint.hpp>
+
+#include "dynarmic/backend/arm64/stack_layout.h"
+
+namespace oaknut {
+struct PointerCodeGeneratorPolicy;
+template<typename>
+class BasicCodeGenerator;
+using CodeGenerator = BasicCodeGenerator<PointerCodeGeneratorPolicy>;
+struct Label;
+}  // namespace oaknut
+
+namespace Dynarmic::IR {
+enum class Type;
+}  // namespace Dynarmic::IR
+
+namespace Dynarmic::Backend::Arm64 {
+
+struct EmitContext;
+
+using Vector = std::array<u64, 2>;
+
+#ifdef _MSC_VER
+#    pragma warning(push)
+#    pragma warning(disable : 4324)  // Structure was padded due to alignment specifier
+#endif
+
+enum class HostLocType {
+    X,
+    Q,
+    Nzcv,
+    Spill,
+};
+
+struct alignas(16) RegisterData {
+    std::array<u64, 30> x;
+    std::array<Vector, 32> q;
+    u32 nzcv;
+    decltype(StackLayout::spill)* spill;
+};
+
+#ifdef _MSC_VER
+#    pragma warning(pop)
+#endif
+
+void EmitVerboseDebuggingOutput(oaknut::CodeGenerator& code, EmitContext& ctx);
+void PrintVerboseDebuggingOutputLine(RegisterData& reg_data, HostLocType reg_type, size_t reg_index, size_t inst_index, IR::Type inst_type);
+
+}  // namespace Dynarmic::Backend::Arm64
\ No newline at end of file