A64/x64: Create a global_offset optimization for the page table.

Instead of looking up the page table like:
  table[addr >> 12][addr & 0xFFF]
We can use a global offset on the table to query the memory like:
  table[addr >> 12][addr]

This saves two instructions on *every* memory access within the recompiler.

Thanks at skmp for the idea.

include/dynarmic/A64/config.h

@@ -163,6 +163,13 @@ struct UserConfig {
     /// relevant memory callback.
     /// This is only used if page_table is not nullptr.
     bool silently_mirror_page_table = true;
+    /// Determines if the pointer in the page_table shall be offseted locally or globally.
+    /// 'false' will access page_table[addr >> bits][addr & mask]
+    /// 'true'  will access page_table[addr >> bits][addr]
+    /// Note: page_table[addr >> bits] will still be checked to verify active pages.
+    ///       So there might be wrongly faulted pages which maps to nullptr.
+    ///       This can be avoided by carefully allocating the memory region.
+    bool absolute_offset_page_table = false;
 
     /// This option relates to translation. Generally when we run into an unpredictable
     /// instruction the ExceptionRaised callback is called. If this is true, we define

src/backend/x64/a64_emit_x64.cpp

@@ -737,6 +737,9 @@ static Xbyak::RegExp EmitVAddrLookup(BlockOfCode& code, A64EmitContext& ctx, Xby
     code.mov(page_table, qword[page_table + tmp * sizeof(void*)]);
     code.test(page_table, page_table);
     code.jz(abort, code.T_NEAR);
+    if (ctx.conf.absolute_offset_page_table) {
+        return page_table + vaddr;
+    }
     code.mov(tmp, vaddr);
     code.and_(tmp, static_cast<u32>(page_size - 1));
     return page_table + tmp;