ARM64: updated literal pool implementation.

src/arm64/assembler-arm64.h

 // Copyright 2013 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #ifndef V8_ARM64_ASSEMBLER_ARM64_H_
 #define V8_ARM64_ASSEMBLER_ARM64_H_
 
 #include <list>
+#include <vector>
 #include <map>
 
 #include "src/cpu.h"
 #include "src/globals.h"
 #include "src/utils.h"
 #include "src/assembler.h"
 #include "src/serialize.h"
 #include "src/arm64/instructions-arm64.h"
 
 
@@ skipping 714 matching lines @@
   Register base_;
   Register regoffset_;
   ptrdiff_t offset_;
   AddrMode addrmode_;
   Shift shift_;
   Extend extend_;
   unsigned shift_amount_;
 };
 
 
+class ConstPool {
+ public:
+  explicit ConstPool(Assembler* assm)
+      : assm_(assm),
+        first_use_(-1),
+        shared_entries_count(0) {}
+  void RecordEntry(intptr_t data, RelocInfo::Mode mode);
+  int EntryCount() const {
+    return shared_entries_count + unique_entries_.size();
+  }
+  bool IsEmpty() const {
+    return shared_entries_.empty() && unique_entries_.empty();
+  }
+  // Distance in bytes between the current pc and the first instruction
+  // using the pool. If there are no pending entries return kMaxInt.
+  int DistanceToFirstUse();
+  // Offset after which instructions using the pool will be out of range.
+  int MaxPcOffset();
+  // Maximum size the constant pool can be with current entries. It always
+  // includes alignment padding and branch over.
+  int WorstCaseSize();
+  // Size in bytes of the literal pool *if* it is emitted at the current
+  // pc. The size will include the branch over the pool if it was requested.
+  int SizeIfEmittedAtCurrentPc(bool require_jump);
+  void Emit(bool require_jump);

[rmcilroy, 2014/06/17 13:09:57]

comments on Emit and Clear please

[Rodolph Perfetta (ARM), 2014/06/18 16:53:23]

Done.

+  void Clear();
+
+ private:
+  bool CanBeShared(RelocInfo::Mode mode);
+  void EmitMarker();
+  void EmitGuard();
+  void EmitEntries();
+
+  Assembler* assm_;
+  // Keep track of the first instruction requiring a constant pool entry
+  // since the previous constant pool was emitted.
+  int first_use_;
+  // values, pc offset(s) of entries which can be shared.
+  std::multimap<uint64_t, int> shared_entries_;
+  // Number of distinct literal in shared entries.
+  int shared_entries_count;
+  // values, pc offset of entries which cannot be shared.
+  std::vector<std::pair<uint64_t, int> > unique_entries_;
+};
+
+
 // -----------------------------------------------------------------------------
 // Assembler.
 
 class Assembler : public AssemblerBase {
  public:
   // Create an assembler. Instructions and relocation information are emitted
   // into a buffer, with the instructions starting from the beginning and the
   // relocation information starting from the end of the buffer. See CodeDesc
   // for a detailed comment on the layout (globals.h).
   //
   // If the provided buffer is NULL, the assembler allocates and grows its own
   // buffer, and buffer_size determines the initial buffer size. The buffer is
   // owned by the assembler and deallocated upon destruction of the assembler.
   //
   // If the provided buffer is not NULL, the assembler uses the provided buffer
   // for code generation and assumes its size to be buffer_size. If the buffer
   // is too small, a fatal error occurs. No deallocation of the buffer is done
   // upon destruction of the assembler.
   Assembler(Isolate* arg_isolate, void* buffer, int buffer_size);
 
   virtual ~Assembler();
 
   virtual void AbortedCodeGeneration() {
-    num_pending_reloc_info_ = 0;
+    constpool_.Clear();
   }
 
   // System functions ---------------------------------------------------------
   // Start generating code from the beginning of the buffer, discarding any code
   // and data that has already been emitted into the buffer.
   //
   // In order to avoid any accidental transfer of state, Reset ASSERTs that the
   // constant pool is not blocked.
   void Reset();
 
@@ skipping 128 matching lines @@
   void StartBlockConstPool();
 
   // Resume constant pool emission. Need to be called as many time as
   // StartBlockConstPool to have an effect.
   void EndBlockConstPool();
 
   bool is_const_pool_blocked() const;
   static bool IsConstantPoolAt(Instruction* instr);
   static int ConstantPoolSizeAt(Instruction* instr);
   // See Assembler::CheckConstPool for more info.
-  void ConstantPoolMarker(uint32_t size);
   void EmitPoolGuard();
-  void ConstantPoolGuard();
 
   // Prevent veneer pool emission until EndBlockVeneerPool is called.
   // Call to this function can be nested but must be followed by an equal
   // number of call to EndBlockConstpool.
   void StartBlockVeneerPool();
 
   // Resume constant pool emission. Need to be called as many time as
   // StartBlockVeneerPool to have an effect.
   void EndBlockVeneerPool();
 
@@ skipping 760 matching lines @@
 
   // Parameters are described in arm64/instructions-arm64.h.
   void debug(const char* message, uint32_t code, Instr params = BREAK);
 
   // Required by V8.
   void dd(uint32_t data) { dc32(data); }
   void db(uint8_t data) { dc8(data); }
 
   // Code generation helpers --------------------------------------------------
 
-  unsigned num_pending_reloc_info() const { return num_pending_reloc_info_; }
+  bool IsConstPoolEmpty() const { return constpool_.IsEmpty(); }
+
+  Instruction* pc() const { return Instruction::Cast(pc_); }
 
   Instruction* InstructionAt(int offset) const {
     return reinterpret_cast<Instruction*>(buffer_ + offset);
   }
 
   ptrdiff_t InstructionOffset(Instruction* instr) const {
     return reinterpret_cast<byte*>(instr) - buffer_;
   }
 
   // Register encoding.
@@ skipping 334 matching lines @@
   }
 
   void GrowBuffer();
   void CheckBufferSpace();
   void CheckBuffer();
 
   // Pc offset of the next constant pool check.
   int next_constant_pool_check_;
 
   // Constant pool generation
-  // Pools are emitted in the instruction stream, preferably after unconditional
-  // jumps or after returns from functions (in dead code locations).
-  // If a long code sequence does not contain unconditional jumps, it is
-  // necessary to emit the constant pool before the pool gets too far from the
-  // location it is accessed from. In this case, we emit a jump over the emitted
-  // constant pool.
+  // Pools are emitted in the instruction stream. They are emitted when:
+  //  * the distance to the first use is above a pre-defined distance or
+  //  * the numbers of entries in the pool is above a pre-defined size or
+  //  * code generation is finished

[rmcilroy, 2014/06/17 13:09:57]

Is it no longer the case that they are emitted in dead code locations (or was
the previous comment wrong)?  dead code locations are not necessarily only when
code generation is finished.

[Rodolph Perfetta (ARM), 2014/06/18 16:53:23]

No we no longer emit in dead code location. On ARM because of the small load
literal range it was a nice optimisation but on a64 it would only contribute to
ICache pollution.

On 2014/06/17 13:09:57, rmcilroy wrote:

[rmcilroy, 2014/06/19 09:58:16]

Interesting. Sounds good in that case.

+  // If a pool needs to be emitted before code generation is finished a branch
+  // over the emitted pool will be inserted.
+
   // Constants in the pool may be addresses of functions that gets relocated;
   // if so, a relocation info entry is associated to the constant pool entry.
 
   // Repeated checking whether the constant pool should be emitted is rather
   // expensive. By default we only check again once a number of instructions
   // has been generated. That also means that the sizing of the buffers is not
   // an exact science, and that we rely on some slop to not overrun buffers.
   static const int kCheckConstPoolIntervalInst = 128;
   static const int kCheckConstPoolInterval =
     kCheckConstPoolIntervalInst * kInstructionSize;
 
-  // Constants in pools are accessed via pc relative addressing, which can
-  // reach +/-4KB thereby defining a maximum distance between the instruction
-  // and the accessed constant.
-  static const int kMaxDistToConstPool = 4 * KB;
-  static const int kMaxNumPendingRelocInfo =
-    kMaxDistToConstPool / kInstructionSize;
+  // Distance to first use after a which a pool will be emitted. Pool entries
+  // are accessed with pc relative load therefore this cannot be more than
+  // 1 * MB. Since constant pool emission checks are interval based this value
+  // is an approximation.
+  static const int kApproximateDistToConstPool = 64 * KB;

[rmcilroy, 2014/06/17 13:09:57]

How about kApproxMaxDistToConstPool?

[Rodolph Perfetta (ARM), 2014/06/18 16:53:23]

Done.

 
-
-  // Average distance beetween a constant pool and the first instruction
-  // accessing the constant pool. Longer distance should result in less I-cache
-  // pollution.
-  // In practice the distance will be smaller since constant pool emission is
-  // forced after function return and sometimes after unconditional branches.
-  static const int kAvgDistToConstPool =
-    kMaxDistToConstPool - kCheckConstPoolInterval;
+  // Number of pool entries after which a pool will be emitted. Since constant
+  // pool emission checks are interval based this value is an approximation.
+  static const int kApproximatePoolEntryCount = 512;

[rmcilroy, 2014/06/17 13:09:57]

kApproxMaxPoolEntryCount?

[Rodolph Perfetta (ARM), 2014/06/18 16:53:23]

Done.

 
   // Emission of the constant pool may be blocked in some code sequences.
   int const_pool_blocked_nesting_;  // Block emission if this is not zero.
   int no_const_pool_before_;  // Block emission before this pc offset.
 
-  // Keep track of the first instruction requiring a constant pool entry
-  // since the previous constant pool was emitted.
-  int first_const_pool_use_;
-
   // Emission of the veneer pools may be blocked in some code sequences.
   int veneer_pool_blocked_nesting_;  // Block emission if this is not zero.
 
   // Relocation info generation
   // Each relocation is encoded as a variable size value
   static const int kMaxRelocSize = RelocInfoWriter::kMaxSize;
   RelocInfoWriter reloc_info_writer;
 
   // Relocation info records are also used during code generation as temporary
   // containers for constants and code target addresses until they are emitted
   // to the constant pool. These pending relocation info records are temporarily
   // stored in a separate buffer until a constant pool is emitted.
   // If every instruction in a long sequence is accessing the pool, we need one
   // pending relocation entry per instruction.
 
-  // the buffer of pending relocation info
-  RelocInfo pending_reloc_info_[kMaxNumPendingRelocInfo];
-  // number of pending reloc info entries in the buffer
-  int num_pending_reloc_info_;
+  // The pending constant pool.
+  ConstPool constpool_;
 
   // Relocation for a type-recording IC has the AST id added to it.  This
   // member variable is a way to pass the information from the call site to
   // the relocation info.
   TypeFeedbackId recorded_ast_id_;
 
   inline TypeFeedbackId RecordedAstId();
   inline void ClearRecordedAstId();
 
  protected:
@@ skipping 66 matching lines @@
   // This function deletes the information related to the label by traversing
   // the label chain, and for each PC-relative instruction in the chain checking
   // if pending unresolved information exists. Its complexity is proportional to
   // the length of the label chain.
   void DeleteUnresolvedBranchInfoForLabelTraverse(Label* label);
 
  private:
   PositionsRecorder positions_recorder_;
   friend class PositionsRecorder;
   friend class EnsureSpace;
+  friend class ConstPool;
 };
 
 class PatchingAssembler : public Assembler {
  public:
   // Create an Assembler with a buffer starting at 'start'.
   // The buffer size is
   //   size of instructions to patch + kGap
   // Where kGap is the distance from which the Assembler tries to grow the
   // buffer.
   // If more or fewer instructions than expected are generated or if some
@@ skipping 12 matching lines @@
     StartBlockPools();
   }
 
   ~PatchingAssembler() {
     // Const pool should still be blocked.
     ASSERT(is_const_pool_blocked());
     EndBlockPools();
     // Verify we have generated the number of instruction we expected.
     ASSERT((pc_offset() + kGap) == buffer_size_);
     // Verify no relocation information has been emitted.
-    ASSERT(num_pending_reloc_info() == 0);
+    ASSERT(IsConstPoolEmpty());
     // Flush the Instruction cache.
     size_t length = buffer_size_ - kGap;
     CPU::FlushICache(buffer_, length);
   }
 
   static const int kMovInt64NInstrs = 4;
   void MovInt64(const Register& rd, int64_t imm);
 
   // See definition of PatchAdrFar() for details.
   static const int kAdrFarPatchableNNops = kMovInt64NInstrs - 1;
   static const int kAdrFarPatchableNInstrs = kAdrFarPatchableNNops + 3;
   void PatchAdrFar(Instruction* target);
 };
 
 
 class EnsureSpace BASE_EMBEDDED {
  public:
   explicit EnsureSpace(Assembler* assembler) {
     assembler->CheckBufferSpace();
   }
 };
 
 } }  // namespace v8::internal
 
 #endif  // V8_ARM64_ASSEMBLER_ARM64_H_