Codegen tweaks

Minor codegen cleanup.  Most significant part of change is
fixing dalvik register use counting to correctly record
cost of high word of register pair.  Significant boost to
Reversi benchmark; modest gain for Caffeinemark.

Change-Id: I41819e6d7be93e62d259240269339a94a934f312

4 files changed, 33 insertions, 19 deletions

diff --git a/src/compiler/codegen/arm/int_arm.cc b/src/compiler/codegen/arm/int_arm.cc
index 2736215ca6..fbc48d45b3 100644
--- a/src/compiler/codegen/arm/int_arm.cc
+++ b/src/compiler/codegen/arm/int_arm.cc
@@ -137,10 +137,24 @@ void ArmCodegen::GenFusedLongCmpImmBranch(CompilationUnit* cu, BasicBlock* bb, R
 
   switch(ccode) {
     case kCondEq:
-      OpCmpImmBranch(cu, kCondNe, high_reg, val_hi, not_taken);
-      break;
     case kCondNe:
-      OpCmpImmBranch(cu, kCondNe, high_reg, val_hi, taken);
+      LIR* target;
+      ConditionCode condition;
+      if (ccode == kCondEq) {
+        target = not_taken;
+        condition = kCondEq;
+      } else {
+        target = taken;
+        condition = kCondNe;
+      }
+      if (val == 0) {
+        int t_reg = AllocTemp(cu);
+        NewLIR4(cu, kThumb2OrrRRRs, t_reg, low_reg, high_reg, 0);
+        FreeTemp(cu, t_reg);
+        OpCondBranch(cu, condition, taken);
+        return;
+      }
+      OpCmpImmBranch(cu, kCondNe, high_reg, val_hi, target);
       break;
     case kCondLt:
       OpCmpImmBranch(cu, kCondLt, high_reg, val_hi, taken);
diff --git a/src/compiler/codegen/arm/utility_arm.cc b/src/compiler/codegen/arm/utility_arm.cc
index a670199580..d6ef6e5d04 100644
--- a/src/compiler/codegen/arm/utility_arm.cc
+++ b/src/compiler/codegen/arm/utility_arm.cc
@@ -500,6 +500,10 @@ LIR* ArmCodegen::OpRegRegImm(CompilationUnit* cu, OpKind op, int r_dest, int r_s
         alt_opcode = kThumb2AddRRR;
       }
       break;
+    case kOpRsub:
+      opcode = kThumb2RsubRRI8;
+      alt_opcode = kThumb2RsubRRR;
+      break;
     case kOpAdc:
       opcode = kThumb2AdcRRI8;
       alt_opcode = kThumb2AdcRRR;
diff --git a/src/compiler/codegen/gen_common.cc b/src/compiler/codegen/gen_common.cc
index 0a46593159..2eaa6b0343 100644
--- a/src/compiler/codegen/gen_common.cc
+++ b/src/compiler/codegen/gen_common.cc
@@ -1419,13 +1419,14 @@ void Codegen::GenArithOpIntLit(CompilationUnit* cu, Instruction::Code opcode,
   switch (opcode) {
     case Instruction::RSUB_INT_LIT8:
     case Instruction::RSUB_INT: {
-      int t_reg;
-      //TUNING: add support for use of Arm rsub op
       rl_src = LoadValue(cu, rl_src, kCoreReg);
-      t_reg = AllocTemp(cu);
-      LoadConstant(cu, t_reg, lit);
       rl_result = EvalLoc(cu, rl_dest, kCoreReg, true);
-      OpRegRegReg(cu, kOpSub, rl_result.low_reg, t_reg, rl_src.low_reg);
+      if (cu->instruction_set == kThumb2) {
+        OpRegRegImm(cu, kOpRsub, rl_result.low_reg, rl_src.low_reg, lit);
+      } else {
+        OpRegReg(cu, kOpNeg, rl_result.low_reg, rl_src.low_reg);
+        OpRegImm(cu, kOpAdd, rl_result.low_reg, lit);
+      }
       StoreValue(cu, rl_dest, rl_result);
       return;
     }
diff --git a/src/compiler/codegen/ralloc_util.cc b/src/compiler/codegen/ralloc_util.cc
index 1d5f3aca86..3a3aeba332 100644
--- a/src/compiler/codegen/ralloc_util.cc
+++ b/src/compiler/codegen/ralloc_util.cc
@@ -1082,27 +1082,22 @@ RegLocation GetSrcWide(CompilationUnit* cu, MIR* mir,
 static void CountRefs(CompilationUnit *cu, BasicBlock* bb, RefCounts* core_counts,
                       RefCounts* fp_counts)
 {
+  // TUNING: this routine could use some tweaking.
   if ((cu->disable_opt & (1 << kPromoteRegs)) ||
     !((bb->block_type == kEntryBlock) || (bb->block_type == kExitBlock) ||
       (bb->block_type == kDalvikByteCode))) {
     return;
   }
-  for (int i = 0; i < cu->num_ssa_regs;) {
+  for (int i = 0; i < cu->num_ssa_regs; i++) {
     RegLocation loc = cu->reg_location[i];
     RefCounts* counts = loc.fp ? fp_counts : core_counts;
     int p_map_idx = SRegToPMap(cu, loc.s_reg_low);
     //Don't count easily regenerated immediates
-    if (loc.fp || loc.wide || !IsInexpensiveConstant(cu, loc)) {
-      counts[p_map_idx].count += cu->use_counts.elem_list[i];
+    if (loc.fp || !IsInexpensiveConstant(cu, loc)) {
+      counts[p_map_idx].count += cu->raw_use_counts.elem_list[i];
     }
-    if (loc.wide) {
-      if (loc.fp) {
-        counts[p_map_idx].double_start = true;
-        counts[p_map_idx+1].count += cu->use_counts.elem_list[i+1];
-      }
-      i += 2;
-    } else {
-      i++;
+    if (loc.wide && loc.fp && !loc.high_word) {
+      counts[p_map_idx].double_start = true;
     }
   }
 }