Jit: Use RangeSet for physical_addresses

This makes JitBaseBlockCache::ErasePhysicalRange around 50% faster and
PPCAnalyzer::Analyze around 40% faster. Rogue Squadron 2's notoriously
laggy action of switching to and from cockpit view is made something
like 20-30% faster by this, though this is a very rough measurement.

Source/Core/Common/RangeSet.h

@@ -24,6 +24,8 @@ public:
 
     const T& to() const { return It->second; }
 
+    std::pair<T, T> operator*() { return {from(), to()}; }
+
     const_iterator& operator++()
     {
       ++It;
@@ -243,6 +245,18 @@ public:
     return get_from(it) <= value && value < get_to(it);
   }
 
+  bool overlaps(T from, T to) const
+  {
+    if (from >= to)
+      return false;
+
+    auto it = Map.lower_bound(to);
+    if (it == Map.begin())
+      return false;
+    --it;
+    return get_from(it) < to && from < get_to(it);
+  }
+
   std::size_t size() const { return Map.size(); }
 
   bool empty() const { return Map.empty(); }

Source/Core/Common/RangeSizeSet.h

@@ -69,6 +69,8 @@ public:
 
     const T& to() const { return It->second.To; }
 
+    std::pair<T, T> operator*() { return {from(), to()}; }
+
     const_iterator& operator++()
     {
       ++It;
@@ -351,6 +353,18 @@ public:
     return get_from(it) <= value && value < get_to(it);
   }
 
+  bool overlaps(T from, T to) const
+  {
+    if (from >= to)
+      return false;
+
+    auto it = Map.lower_bound(to);
+    if (it == Map.begin())
+      return false;
+    --it;
+    return get_from(it) < to && from < get_to(it);
+  }
+
   std::size_t size() const { return Map.size(); }
 
   bool empty() const { return Map.empty(); }

Source/Core/Core/PowerPC/JitArm64/Jit.h

@@ -7,9 +7,8 @@
 #include <map>
 #include <optional>
 
-#include <rangeset/rangesizeset.h>
-
 #include "Common/Arm64Emitter.h"
+#include "Common/RangeSizeSet.h"
 
 #include "Core/PowerPC/JitArm64/JitArm64Cache.h"
 #include "Core/PowerPC/JitArm64/JitArm64_RegCache.h"
@@ -431,10 +430,10 @@ protected:
   u8* m_near_code_end = nullptr;
   bool m_near_code_write_failed = false;
 
-  HyoutaUtilities::RangeSizeSet<u8*> m_free_ranges_near_0;
+  Common::RangeSizeSet<u8*> m_free_ranges_near_0;
-  HyoutaUtilities::RangeSizeSet<u8*> m_free_ranges_near_1;
+  Common::RangeSizeSet<u8*> m_free_ranges_near_1;
-  HyoutaUtilities::RangeSizeSet<u8*> m_free_ranges_far_0;
+  Common::RangeSizeSet<u8*> m_free_ranges_far_0;
-  HyoutaUtilities::RangeSizeSet<u8*> m_free_ranges_far_1;
+  Common::RangeSizeSet<u8*> m_free_ranges_far_1;
 
   std::unique_ptr<HostDisassembler> m_disassembler;
 };

Source/Core/Core/PowerPC/JitCommon/JitCache.cpp

@@ -31,8 +31,7 @@ using namespace Gen;
 
 bool JitBlock::OverlapsPhysicalRange(u32 address, u32 length) const
 {
-  return physical_addresses.lower_bound(address) !=
+  return physical_addresses.overlaps(address, address + length);
-         physical_addresses.lower_bound(address + length);
 }
 
 void JitBlock::ProfileData::BeginProfiling(ProfileData* data)
@@ -171,10 +170,13 @@ void JitBaseBlockCache::FinalizeBlock(JitBlock& block, bool block_link,
                                  original_buffer_transform_view.end());
   }
 
-  for (u32 addr : block.physical_addresses)
+  for (auto [range_start, range_end] : block.physical_addresses)
   {
-    valid_block.Set(addr / 32);
+    for (u32 i = range_start & ~31; i < range_end; i += 32)
-    block_range_map[addr & BLOCK_RANGE_MAP_MASK].insert(&block);
+      valid_block.Set(i / 32);
+
+    for (u32 i = range_start & BLOCK_RANGE_MAP_MASK; i < range_end; i += BLOCK_RANGE_SIZE)
+      block_range_map[i].insert(&block);
   }
 
   if (block_link)
@@ -362,9 +364,15 @@ void JitBaseBlockCache::ErasePhysicalRange(u32 address, u32 length)
       {
         // If the block overlaps, also remove all other occupied slots in the other macro blocks.
         // This will leak empty macro blocks, but they may be reused or cleared later on.
-        for (u32 addr : block->physical_addresses)
+        for (auto [range_start, range_end] : block->physical_addresses)
-          if ((addr & BLOCK_RANGE_MAP_MASK) != start->first)
+        {
-            block_range_map[addr & BLOCK_RANGE_MAP_MASK].erase(block);
+          DEBUG_ASSERT(range_start != range_end);
+          for (u32 i = range_start & BLOCK_RANGE_MAP_MASK; i < range_end; i += BLOCK_RANGE_SIZE)
+          {
+            if (i != start->first)
+              block_range_map[i].erase(block);
+          }
+        }
 
         // And remove the block.
         DestroyBlock(*block);
@@ -404,8 +412,11 @@ void JitBaseBlockCache::EraseSingleBlock(const JitBlock& block)
 
   JitBlock& mutable_block = block_map_iter->second;
 
-  for (const u32 addr : mutable_block.physical_addresses)
+  for (auto [range_start, range_end] : mutable_block.physical_addresses)
-    block_range_map[addr & BLOCK_RANGE_MAP_MASK].erase(&mutable_block);
+  {
+    for (u32 i = range_start & BLOCK_RANGE_MAP_MASK; i < range_end; i += BLOCK_RANGE_SIZE)
+      block_range_map[i].erase(&mutable_block);
+  }
 
   DestroyBlock(mutable_block);
   block_map.erase(block_map_iter);  // The original JitBlock reference is now dangling.

Source/Core/Core/PowerPC/JitCommon/JitCache.h

@@ -10,13 +10,13 @@
 #include <functional>
 #include <map>
 #include <memory>
-#include <set>
 #include <type_traits>
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
 
 #include "Common/CommonTypes.h"
+#include "Common/RangeSet.h"
 #include "Core/HW/Memmap.h"
 #include "Core/PowerPC/Gekko.h"
 #include "Core/PowerPC/PPCAnalyst.h"
@@ -101,7 +101,7 @@ struct JitBlock : public JitBlockData
   std::vector<LinkData> linkData;
 
   // This set stores all physical addresses of all occupied instructions.
-  std::set<u32> physical_addresses;
+  Common::RangeSet<u32> physical_addresses;
 
   // This is only available when debugging is enabled. It is a trimmed-down copy of the
   // PPCAnalyst::CodeBuffer used to recompile this block, including repeat instructions.
@@ -218,7 +218,8 @@ private:
   // Range of overlapping code indexed by a masked physical address.
   // This is used for invalidation of memory regions. The range is grouped
   // in macro blocks of each 0x100 bytes.
-  static constexpr u32 BLOCK_RANGE_MAP_MASK = ~(0x100 - 1);
+  static constexpr u32 BLOCK_RANGE_SIZE = 0x100;
+  static constexpr u32 BLOCK_RANGE_MAP_MASK = ~(BLOCK_RANGE_SIZE - 1);
   std::map<u32, std::unordered_set<JitBlock*>> block_range_map;
 
   // This bitsets shows which cachelines overlap with any blocks.

Source/Core/Core/PowerPC/PPCAnalyst.cpp

@@ -849,7 +849,8 @@ u32 PPCAnalyzer::Analyze(u32 address, CodeBlock* block, CodeBuffer* buffer,
     code[i].inst = inst;
     code[i].skip = false;
     block->m_stats->numCycles += opinfo->num_cycles;
-    block->m_physical_addresses.insert(result.physical_address);
+    block->m_physical_addresses.insert(result.physical_address,
+                                       result.physical_address + sizeof(UGeckoInstruction));
 
     SetInstructionStats(block, &code[i], opinfo);
 

Source/Core/Core/PowerPC/PPCAnalyst.h

@@ -5,11 +5,11 @@
 
 #include <algorithm>
 #include <cstddef>
-#include <set>
 #include <vector>
 
 #include "Common/BitSet.h"
 #include "Common/CommonTypes.h"
+#include "Common/RangeSet.h"
 #include "Core/PowerPC/PPCTables.h"
 
 class PPCSymbolDB;
@@ -129,7 +129,7 @@ struct CodeBlock
   BitSet32 m_gpr_inputs;
 
   // Which memory locations are occupied by this block.
-  std::set<u32> m_physical_addresses;
+  Common::RangeSet<u32> m_physical_addresses;
 };
 
 class PPCAnalyzer

Source/Core/DolphinQt/Debugger/JitBlockTableModel.cpp

@@ -236,7 +236,8 @@ QVariant JitBlockTableModel::DisplayRoleData(const QModelIndex& index) const
   case Column::CodeBufferSize:
     return QString::number(jit_block.originalSize * sizeof(UGeckoInstruction));
   case Column::RepeatInstructions:
-    return QString::number(jit_block.originalSize - jit_block.physical_addresses.size());
+    return QString::number(jit_block.originalSize - jit_block.physical_addresses.get_stats().first /
+                                                        sizeof(UGeckoInstruction));
   case Column::HostNearCodeSize:
     return QString::number(jit_block.near_end - jit_block.near_begin);
   case Column::HostFarCodeSize:
@@ -329,7 +330,9 @@ QVariant JitBlockTableModel::SortRoleData(const QModelIndex& index) const
   case Column::CodeBufferSize:
     return static_cast<qulonglong>(jit_block.originalSize);
   case Column::RepeatInstructions:
-    return static_cast<qulonglong>(jit_block.originalSize - jit_block.physical_addresses.size());
+    return static_cast<qulonglong>(jit_block.originalSize -
+                                   jit_block.physical_addresses.get_stats().first /
+                                       sizeof(UGeckoInstruction));
   case Column::HostNearCodeSize:
     return static_cast<qulonglong>(jit_block.near_end - jit_block.near_begin);
   case Column::HostFarCodeSize: