[dynarmic] inlined pool in block + slab-like for each block

Signed-off-by: lizzie <lizzie@eden-emu.dev>

src/dynarmic/src/dynarmic/ir/basic_block.cpp

@@ -37,7 +37,21 @@ Block::Block(const LocationDescriptor& location)
 /// @param args            A sequence of Value instances used as arguments for the instruction.
 /// @returns Iterator to the newly created instruction.
 Block::iterator Block::PrependNewInst(iterator insertion_point, Opcode opcode, std::initializer_list<Value> args) noexcept {
-    IR::Inst* inst = new IR::Inst(opcode);
+    // First try using the "inline" buffer, otherwise fallback to a slower slab-like allocation scheme
+    // purpouse is to avoid many calls to new/delete which invoke malloc which invokes mmap
+    // just pool it!!! - reason why there is an inline buffer is because many small blocks are created
+    // with few instructions due to subpar optimisations on other passes... plus branch-heavy code will
+    // hugely benefit from the coherency of faster allocations...
+    IR::Inst* inst;
+    if (inlined_inst.size() < inlined_inst.max_size()) {
+        inst = &inlined_inst[inlined_inst.size()];
+        inlined_inst.emplace_back(opcode);
+    } else {
+        if (pooled_inst.empty() || pooled_inst.back().size() == pooled_inst.back().max_size())
+            pooled_inst.emplace_back();
+        inst = &pooled_inst.back()[pooled_inst.back().size()];
+        pooled_inst.back().emplace_back(opcode);
+    }
     DEBUG_ASSERT(args.size() == inst->NumArgs());
     std::for_each(args.begin(), args.end(), [&inst, index = size_t(0)](const auto& arg) mutable {
         inst->SetArg(index, arg);

src/dynarmic/src/dynarmic/ir/basic_block.h

@@ -13,6 +13,9 @@
 #include <optional>
 #include <string>
 
+#include <boost/container/container_fwd.hpp>
+#include <boost/container/static_vector.hpp>
+#include <boost/container/stable_vector.hpp>
 #include <mcl/container/intrusive_list.hpp>
 #include "dynarmic/common/common_types.h"
 
@@ -163,8 +166,12 @@ public:
         return cycle_count;
     }
 private:
+    /// "Hot cache" for small blocks so we don't call global allocator
+    boost::container::static_vector<Inst, 14> inlined_inst;
     /// List of instructions in this block.
     instruction_list_type instructions;
+    /// "Long/far" memory pool
+    boost::container::stable_vector<boost::container::static_vector<Inst, 32>> pooled_inst;
     /// Block to execute next if `cond` did not pass.
     std::optional<LocationDescriptor> cond_failed = {};
     /// Description of the starting location of this block
@@ -180,6 +187,7 @@ private:
     /// Number of cycles this block takes to execute.
     size_t cycle_count = 0;
 };
+static_assert(sizeof(Block) == 2048);
 
 /// Returns a string representation of the contents of block. Intended for debugging.
 std::string DumpBlock(const IR::Block& block) noexcept;