emit_x64: Do not clear fast_dispatch_table unnecessarily

Reduces invalidation overhead

src/backend/x64/a32_emit_x64.cpp

@@ -158,12 +158,11 @@ void A32EmitX64::ClearCache() {
 
 void A32EmitX64::InvalidateCacheRanges(const boost::icl::interval_set<u32>& ranges) {
     InvalidateBasicBlocks(block_ranges.InvalidateRanges(ranges));
-    ClearFastDispatchTable();
 }
 
 void A32EmitX64::ClearFastDispatchTable() {
     if (config.enable_fast_dispatch) {
-        fast_dispatch_table.fill({0xFFFFFFFFFFFFFFFFull, nullptr});
+        fast_dispatch_table.fill({});
     }
 }
 
@@ -278,6 +277,16 @@ void A32EmitX64::GenTerminalHandlers() {
         code.mov(ptr[rbp + offsetof(FastDispatchEntry, code_ptr)], rax);
         code.jmp(rax);
         PerfMapRegister(terminal_handler_fast_dispatch_hint, code.getCurr(), "a32_terminal_handler_fast_dispatch_hint");
+
+        code.align();
+        fast_dispatch_table_lookup = code.getCurr<FastDispatchEntry&(*)(u64)>();
+        code.mov(code.ABI_PARAM2, reinterpret_cast<u64>(fast_dispatch_table.data()));
+        if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE42)) {
+            code.crc32(code.ABI_PARAM1.cvt32(), code.ABI_PARAM2.cvt32());
+        }
+        code.and_(code.ABI_PARAM1.cvt32(), fast_dispatch_table_mask);
+        code.lea(code.ABI_RETURN, code.ptr[code.ABI_PARAM1 + code.ABI_PARAM2]);
+        code.ret();
     }
 }
 
@@ -1511,4 +1520,11 @@ void A32EmitX64::EmitPatchMovRcx(CodePtr target_code_ptr) {
     code.EnsurePatchLocationSize(patch_location, 10);
 }
 
+void A32EmitX64::Unpatch(const IR::LocationDescriptor& location) {
+    EmitX64::Unpatch(location);
+    if (config.enable_fast_dispatch) {
+        (*fast_dispatch_table_lookup)(location.Value()) = {};
+    }
+}
+
 } // namespace Dynarmic::Backend::X64

src/backend/x64/a32_emit_x64.h

@@ -52,8 +52,8 @@ protected:
     BlockRangeInformation<u32> block_ranges;
 
     struct FastDispatchEntry {
-        u64 location_descriptor;
+        u64 location_descriptor = 0xFFFF'FFFF'FFFF'FFFFull;
-        const void* code_ptr;
+        const void* code_ptr = nullptr;
     };
     static_assert(sizeof(FastDispatchEntry) == 0x10);
     static constexpr u64 fast_dispatch_table_mask = 0xFFFF0;
@@ -67,6 +67,7 @@ protected:
 
     const void* terminal_handler_pop_rsb_hint;
     const void* terminal_handler_fast_dispatch_hint = nullptr;
+    FastDispatchEntry& (*fast_dispatch_table_lookup)(u64) = nullptr;
     void GenTerminalHandlers();
 
     // Microinstruction emitters
@@ -112,6 +113,7 @@ protected:
     void EmitTerminalImpl(IR::Term::CheckHalt terminal, IR::LocationDescriptor initial_location) override;
 
     // Patching
+    void Unpatch(const IR::LocationDescriptor& target_desc) override;
     void EmitPatchJg(const IR::LocationDescriptor& target_desc, CodePtr target_code_ptr = nullptr) override;
     void EmitPatchJmp(const IR::LocationDescriptor& target_desc, CodePtr target_code_ptr = nullptr) override;
     void EmitPatchMovRcx(CodePtr target_code_ptr = nullptr) override;

src/backend/x64/a64_emit_x64.cpp

@@ -126,12 +126,11 @@ void A64EmitX64::ClearCache() {
 
 void A64EmitX64::InvalidateCacheRanges(const boost::icl::interval_set<u64>& ranges) {
     InvalidateBasicBlocks(block_ranges.InvalidateRanges(ranges));
-    ClearFastDispatchTable();
 }
 
 void A64EmitX64::ClearFastDispatchTable() {
     if (conf.enable_fast_dispatch) {
-        fast_dispatch_table.fill({0xFFFFFFFFFFFFFFFFull, nullptr});
+        fast_dispatch_table.fill({});
     }
 }
 
@@ -325,7 +324,7 @@ void A64EmitX64::GenTerminalHandlers() {
         code.L(rsb_cache_miss);
         code.mov(r12, reinterpret_cast<u64>(fast_dispatch_table.data()));
         if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE42)) {
-            code.crc32(rbp, r12d);
+            code.crc32(rbx, r12d);
         }
         code.and_(ebp, fast_dispatch_table_mask);
         code.lea(rbp, ptr[r12 + rbp]);
@@ -338,6 +337,16 @@ void A64EmitX64::GenTerminalHandlers() {
         code.mov(ptr[rbp + offsetof(FastDispatchEntry, code_ptr)], rax);
         code.jmp(rax);
         PerfMapRegister(terminal_handler_fast_dispatch_hint, code.getCurr(), "a64_terminal_handler_fast_dispatch_hint");
+
+        code.align();
+        fast_dispatch_table_lookup = code.getCurr<FastDispatchEntry&(*)(u64)>();
+        code.mov(code.ABI_PARAM2, reinterpret_cast<u64>(fast_dispatch_table.data()));
+        if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE42)) {
+            code.crc32(code.ABI_PARAM1, code.ABI_PARAM2);
+        }
+        code.and_(code.ABI_PARAM1.cvt32(), fast_dispatch_table_mask);
+        code.lea(code.ABI_RETURN, code.ptr[code.ABI_PARAM1 + code.ABI_PARAM2]);
+        code.ret();
     }
 }
 
@@ -1244,4 +1253,11 @@ void A64EmitX64::EmitPatchMovRcx(CodePtr target_code_ptr) {
     code.EnsurePatchLocationSize(patch_location, 10);
 }
 
+void A64EmitX64::Unpatch(const IR::LocationDescriptor& location) {
+    EmitX64::Unpatch(location);
+    if (conf.enable_fast_dispatch) {
+        (*fast_dispatch_table_lookup)(location.Value()) = {};
+    }
+}
+
 } // namespace Dynarmic::Backend::X64

src/backend/x64/a64_emit_x64.h

@@ -52,8 +52,8 @@ protected:
     BlockRangeInformation<u64> block_ranges;
 
     struct FastDispatchEntry {
-        u64 location_descriptor;
+        u64 location_descriptor = 0xFFFF'FFFF'FFFF'FFFFull;
-        const void* code_ptr;
+        const void* code_ptr = nullptr;
     };
     static_assert(sizeof(FastDispatchEntry) == 0x10);
     static constexpr u64 fast_dispatch_table_mask = 0xFFFFF0;
@@ -71,6 +71,7 @@ protected:
 
     const void* terminal_handler_pop_rsb_hint;
     const void* terminal_handler_fast_dispatch_hint = nullptr;
+    FastDispatchEntry& (*fast_dispatch_table_lookup)(u64) = nullptr;
     void GenTerminalHandlers();
 
     void EmitDirectPageTableMemoryRead(A64EmitContext& ctx, IR::Inst* inst, size_t bitsize);
@@ -101,6 +102,7 @@ protected:
     void EmitTerminalImpl(IR::Term::CheckHalt terminal, IR::LocationDescriptor initial_location) override;
 
     // Patching
+    void Unpatch(const IR::LocationDescriptor& target_desc) override;
     void EmitPatchJg(const IR::LocationDescriptor& target_desc, CodePtr target_code_ptr = nullptr) override;
     void EmitPatchJmp(const IR::LocationDescriptor& target_desc, CodePtr target_code_ptr = nullptr) override;
     void EmitPatchMovRcx(CodePtr target_code_ptr = nullptr) override;

src/backend/x64/emit_x64.h

@@ -109,7 +109,7 @@ protected:
         std::vector<CodePtr> mov_rcx;
     };
     void Patch(const IR::LocationDescriptor& target_desc, CodePtr target_code_ptr);
-    void Unpatch(const IR::LocationDescriptor& target_desc);
+    virtual void Unpatch(const IR::LocationDescriptor& target_desc);
     virtual void EmitPatchJg(const IR::LocationDescriptor& target_desc, CodePtr target_code_ptr = nullptr) = 0;
     virtual void EmitPatchJmp(const IR::LocationDescriptor& target_desc, CodePtr target_code_ptr = nullptr) = 0;
     virtual void EmitPatchMovRcx(CodePtr target_code_ptr = nullptr) = 0;