Better peephole optimization for ARM. This is a commit of

http://codereview.chromium.org/2004006 for Subrato De.


git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@4662 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/arm/assembler-arm.cc

@@ -268,6 +268,20 @@ const Instr kBlxRegMask =
     15 * B24 | 15 * B20 | 15 * B16 | 15 * B12 | 15 * B8 | 15 * B4;
 const Instr kBlxRegPattern =
     B24 | B21 | 15 * B16 | 15 * B12 | 15 * B8 | 3 * B4;
+// A mask for the Rd register for push, pop, ldr, str instructions.
+const Instr kRdMask = 0x0000f000;
+static const Instr kLdrRegFpOffsetPattern =
+    al | B26 | L | Offset | fp.code() * B16;
+static const Instr kStrRegFpOffsetPattern =
+    al | B26 | Offset | fp.code() * B16;
+static const Instr kLdrRegFpNegOffsetPattern =
+    al | B26 | L | NegOffset | fp.code() * B16;
+static const Instr kStrRegFpNegOffsetPattern =
+    al | B26 | NegOffset | fp.code() * B16;
+static const Instr kLdrStrInstrTypeMask = 0xffff0000;
+static const Instr kLdrStrInstrArgumentMask = 0x0000ffff;
+static const Instr kLdrStrOffsetMask = 0x00000fff;
+static const int kRdShift = 12;
 
 // Spare buffer.
 static const int kMinimalBufferSize = 4*KB;
@@ -395,6 +409,43 @@ Instr Assembler::SetLdrRegisterImmediateOffset(Instr instr, int offset) {
 }
 
 
+Register Assembler::GetRd(Instr instr) {
+  Register reg;
+  reg.code_ = ((instr & kRdMask) >> kRdShift);
+  return reg;
+}
+
+
+bool Assembler::IsPush(Instr instr) {
+  return ((instr & ~kRdMask) == kPushRegPattern);
+}
+
+
+bool Assembler::IsPop(Instr instr) {
+  return ((instr & ~kRdMask) == kPopRegPattern);
+}
+
+
+bool Assembler::IsStrRegFpOffset(Instr instr) {
+  return ((instr & kLdrStrInstrTypeMask) == kStrRegFpOffsetPattern);
+}
+
+
+bool Assembler::IsLdrRegFpOffset(Instr instr) {
+  return ((instr & kLdrStrInstrTypeMask) == kLdrRegFpOffsetPattern);
+}
+
+
+bool Assembler::IsStrRegFpNegOffset(Instr instr) {
+  return ((instr & kLdrStrInstrTypeMask) == kStrRegFpNegOffsetPattern);
+}
+
+
+bool Assembler::IsLdrRegFpNegOffset(Instr instr) {
+  return ((instr & kLdrStrInstrTypeMask) == kLdrRegFpNegOffsetPattern);
+}
+
+
 // Labels refer to positions in the (to be) generated code.
 // There are bound, linked, and unused labels.
 //
@@ -1086,20 +1137,179 @@ void Assembler::ldr(Register dst, const MemOperand& src, Condition cond) {
   }
   addrmod2(cond | B26 | L, dst, src);
 
-  // Eliminate pattern: push(r), pop(r)
-  //   str(r, MemOperand(sp, 4, NegPreIndex), al)
-  //   ldr(r, MemOperand(sp, 4, PostIndex), al)
-  // Both instructions can be eliminated.
+  // Eliminate pattern: push(ry), pop(rx)
+  //   str(ry, MemOperand(sp, 4, NegPreIndex), al)
+  //   ldr(rx, MemOperand(sp, 4, PostIndex), al)
+  // Both instructions can be eliminated if ry = rx.
+  // If ry != rx, a register copy from ry to rx is inserted
+  // after eliminating the push and the pop instructions.
   int pattern_size = 2 * kInstrSize;
+  Instr push_instr = instr_at(pc_ - 2 * kInstrSize);
+  Instr pop_instr = instr_at(pc_ - 1 * kInstrSize);
+
   if (FLAG_push_pop_elimination &&
+     last_bound_pos_ <= (pc_offset() - pattern_size) &&
+     reloc_info_writer.last_pc() <= (pc_ - pattern_size) &&
+     IsPush(push_instr) &&
+     IsPop(pop_instr)) {
+    if ((pop_instr & kRdMask) != (push_instr & kRdMask)) {
+      // For consecutive push and pop on different registers,
+      // we delete both the push & pop and insert a register move.
+      // push ry, pop rx --> mov rx, ry
+      Register reg_pushed, reg_popped;
+      reg_pushed = GetRd(push_instr);
+      reg_popped = GetRd(pop_instr);
+      pc_ -= 2 * kInstrSize;
+      // Insert a mov instruction, which is better than a pair of push & pop
+      mov(reg_popped, reg_pushed);
+      if (FLAG_print_push_pop_elimination) {
+        PrintF("%x push/pop (diff reg) replaced by a reg move\n", pc_offset());
+      }
+    } else {
+      // For consecutive push and pop on the same register,
+      // both the push and the pop can be deleted.
+      pc_ -= 2 * kInstrSize;
+      if (FLAG_print_push_pop_elimination) {
+        PrintF("%x push/pop (same reg) eliminated\n", pc_offset());
+      }
+    }
+  }
+
+  pattern_size = 2 * kInstrSize;
+  if (FLAG_peephole_optimization &&
       last_bound_pos_ <= (pc_offset() - pattern_size) &&
-      reloc_info_writer.last_pc() <= (pc_ - pattern_size) &&
-      // Pattern.
-      instr_at(pc_ - 1 * kInstrSize) == (kPopRegPattern | dst.code() * B12) &&
-      instr_at(pc_ - 2 * kInstrSize) == (kPushRegPattern | dst.code() * B12)) {
-    pc_ -= 2 * kInstrSize;
-    if (FLAG_print_push_pop_elimination) {
-      PrintF("%x push/pop (same reg) eliminated\n", pc_offset());
+      reloc_info_writer.last_pc() <= (pc_ - pattern_size)) {
+    Instr str_instr = instr_at(pc_ - 2 * kInstrSize);
+    Instr ldr_instr = instr_at(pc_ - 1 * kInstrSize);
+
+    if ((IsStrRegFpOffset(str_instr) &&
+         IsLdrRegFpOffset(ldr_instr)) ||
+       (IsStrRegFpNegOffset(str_instr) &&
+         IsLdrRegFpNegOffset(ldr_instr))) {
+      if ((ldr_instr & kLdrStrInstrArgumentMask) ==
+            (str_instr & kLdrStrInstrArgumentMask)) {
+        // Pattern: Ldr/str same fp+offset, same register.
+        //
+        // The following:
+        // str rx, [fp, #-12]
+        // ldr rx, [fp, #-12]
+        //
+        // Becomes:
+        // str rx, [fp, #-12]
+
+        pc_ -= 1 * kInstrSize;
+        if (FLAG_print_peephole_optimization) {
+          PrintF("%x str/ldr (fp + same offset), same reg\n", pc_offset());
+        }
+      } else if ((ldr_instr & kLdrStrOffsetMask) ==
+                 (str_instr & kLdrStrOffsetMask)) {
+        // Pattern: Ldr/str same fp+offset, different register.
+        //
+        // The following:
+        // str rx, [fp, #-12]
+        // ldr ry, [fp, #-12]
+        //
+        // Becomes:
+        // str rx, [fp, #-12]
+        // mov ry, rx
+
+        Register reg_stored, reg_loaded;
+        reg_stored = GetRd(str_instr);
+        reg_loaded = GetRd(ldr_instr);
+        pc_ -= 1 * kInstrSize;
+        // Insert a mov instruction, which is better than ldr.
+        mov(reg_loaded, reg_stored);
+        if (FLAG_print_peephole_optimization) {
+          PrintF("%x str/ldr (fp + same offset), diff reg \n", pc_offset());
+        }
+      }
+    }
+  }
+
+  pattern_size = 3 * kInstrSize;
+  if (FLAG_push_pop_elimination &&
+    last_bound_pos_ <= (pc_offset() - pattern_size) &&
+    reloc_info_writer.last_pc() <= (pc_ - pattern_size)) {
+    Instr mem_write_instr = instr_at(pc_ - 3 * kInstrSize);
+    Instr ldr_instr = instr_at(pc_ - 2 * kInstrSize);
+    Instr mem_read_instr = instr_at(pc_ - 1 * kInstrSize);
+    if (IsPush(mem_write_instr) &&
+        IsPop(mem_read_instr)) {
+      if ((IsLdrRegFpOffset(ldr_instr) ||
+        IsLdrRegFpNegOffset(ldr_instr))) {
+        if ((mem_write_instr & kRdMask) ==
+              (mem_read_instr & kRdMask)) {
+          // Pattern: push & pop from/to same register,
+          // with a fp+offset ldr in between
+          //
+          // The following:
+          // str rx, [sp, #-4]!
+          // ldr rz, [fp, #-24]
+          // ldr rx, [sp], #+4
+          //
+          // Becomes:
+          // if(rx == rz)
+          //   delete all
+          // else
+          //   ldr rz, [fp, #-24]
+
+          if ((mem_write_instr & kRdMask) == (ldr_instr & kRdMask)) {
+            pc_ -= 3 * kInstrSize;
+          } else {
+            pc_ -= 3 * kInstrSize;
+            // Reinsert back the ldr rz.
+            emit(ldr_instr);
+          }
+          if (FLAG_print_push_pop_elimination) {
+            PrintF("%x push/pop -dead ldr fp+offset in middle\n", pc_offset());
+          }
+        } else {
+          // Pattern: push & pop from/to different registers
+          // with a fp+offset ldr in between
+          //
+          // The following:
+          // str rx, [sp, #-4]!
+          // ldr rz, [fp, #-24]
+          // ldr ry, [sp], #+4
+          //
+          // Becomes:
+          // if(ry == rz)
+          //   mov ry, rx;
+          // else if(rx != rz)
+          //   ldr rz, [fp, #-24]
+          //   mov ry, rx
+          // else if((ry != rz) || (rx == rz)) becomes:
+          //   mov ry, rx
+          //   ldr rz, [fp, #-24]
+
+          Register reg_pushed, reg_popped;
+          if ((mem_read_instr & kRdMask) == (ldr_instr & kRdMask)) {
+            reg_pushed = GetRd(mem_write_instr);
+            reg_popped = GetRd(mem_read_instr);
+            pc_ -= 3 * kInstrSize;
+            mov(reg_popped, reg_pushed);
+          } else if ((mem_write_instr & kRdMask)
+                                != (ldr_instr & kRdMask)) {
+            reg_pushed = GetRd(mem_write_instr);
+            reg_popped = GetRd(mem_read_instr);
+            pc_ -= 3 * kInstrSize;
+            emit(ldr_instr);
+            mov(reg_popped, reg_pushed);
+          } else if (((mem_read_instr & kRdMask)
+                                     != (ldr_instr & kRdMask)) ||
+                    ((mem_write_instr & kRdMask)
+                                     == (ldr_instr & kRdMask)) ) {
+            reg_pushed = GetRd(mem_write_instr);
+            reg_popped = GetRd(mem_read_instr);
+            pc_ -= 3 * kInstrSize;
+            mov(reg_popped, reg_pushed);
+            emit(ldr_instr);
+          }
+          if (FLAG_print_push_pop_elimination) {
+            PrintF("%x push/pop (ldr fp+off in middle)\n", pc_offset());
+          }
+        }
+      }
     }
   }
 }

src/arm/assembler-arm.h

@@ -987,6 +987,13 @@ class Assembler : public Malloced {
   static bool IsLdrRegisterImmediate(Instr instr);
   static int GetLdrRegisterImmediateOffset(Instr instr);
   static Instr SetLdrRegisterImmediateOffset(Instr instr, int offset);
+  static Register GetRd(Instr instr);
+  static bool IsPush(Instr instr);
+  static bool IsPop(Instr instr);
+  static bool IsStrRegFpOffset(Instr instr);
+  static bool IsLdrRegFpOffset(Instr instr);
+  static bool IsStrRegFpNegOffset(Instr instr);
+  static bool IsLdrRegFpNegOffset(Instr instr);
 
 
  protected:

src/flag-definitions.h

@@ -104,6 +104,10 @@ DEFINE_bool(push_pop_elimination, true,
             "eliminate redundant push/pops in assembly code")
 DEFINE_bool(print_push_pop_elimination, false,
             "print elimination of redundant push/pops in assembly code")
+DEFINE_bool(peephole_optimization, true,
+            "perform peephole optimizations in assembly code")
+DEFINE_bool(print_peephole_optimization, false,
+            "print peephole optimizations in assembly code")
 DEFINE_bool(enable_sse2, true,
             "enable use of SSE2 instructions if available")
 DEFINE_bool(enable_sse3, true,