PPC: perf enhancement: omit redundant compare with 0

R=mbrandy@us.ibm.com, michael_dawson@ca.ibm.com
BUG=

Review URL: https://codereview.chromium.org/1222133004

Cr-Commit-Position: refs/heads/master@{#29459}

src/ppc/assembler-ppc.cc

@@ -219,6 +219,7 @@ Assembler::Assembler(Isolate* isolate, void* buffer, int buffer_size)
                                                kMaxBlockTrampolineSectionSize;
   internal_trampoline_exception_ = false;
   last_bound_pos_ = 0;
+  optimizable_cmpi_pos_ = -1;
   trampoline_emitted_ = FLAG_force_long_branches;
   unbound_labels_count_ = 0;
   ClearRecordedAstId();
@@ -1030,9 +1031,17 @@ void Assembler::cmpl(Register src1, Register src2, CRegister cr) {
 void Assembler::cmpwi(Register src1, const Operand& src2, CRegister cr) {
   intptr_t imm16 = src2.imm_;
   int L = 0;
+  int pos = pc_offset();
   DCHECK(is_int16(imm16));
   DCHECK(cr.code() >= 0 && cr.code() <= 7);
   imm16 &= kImm16Mask;
+
+  // For cmpwi against 0, save postition and cr for later examination
+  // of potential optimization.
+  if (imm16 == 0 && pos > 0 && last_bound_pos_ != pos) {
+    optimizable_cmpi_pos_ = pos;
+    cmpi_cr_ = cr;
+  }
   emit(CMPI | cr.code() * B23 | L * B21 | src1.code() * B16 | imm16);
 }
 

src/ppc/assembler-ppc.h

@@ -755,11 +755,49 @@ class Assembler : public AssemblerBase {
   // Convenience branch instructions using labels
   void b(Label* L, LKBit lk = LeaveLK) { b(branch_offset(L, false), lk); }
 
+  inline CRegister cmpi_optimization(CRegister cr) {
+    // Check whether the branch is preceeded by an optimizable cmpi against 0.
+    // The cmpi can be deleted if it is also preceeded by an instruction that
+    // sets the register used by the compare and supports a dot form.
+    unsigned int sradi_mask = kOpcodeMask | kExt2OpcodeVariant2Mask;
+    unsigned int srawi_mask = kOpcodeMask | kExt2OpcodeMask;
+    int pos = pc_offset();
+    int cmpi_pos = pc_offset() - kInstrSize;
+
+    if (cmpi_pos > 0 && optimizable_cmpi_pos_ == cmpi_pos &&
+        cmpi_cr_.code() == cr.code() && last_bound_pos_ != pos) {
+      int xpos = cmpi_pos - kInstrSize;
+      int xinstr = instr_at(xpos);
+      int cmpi_ra = (instr_at(cmpi_pos) & 0x1f0000) >> 16;
+      // ra is at the same bit position for the three cases below.
+      int ra = (xinstr & 0x1f0000) >> 16;
+      if (cmpi_ra == ra) {
+        if ((xinstr & sradi_mask) == (EXT2 | SRADIX)) {
+          cr = cr0;
+          instr_at_put(xpos, xinstr | SetRC);
+          pc_ -= kInstrSize;
+        } else if ((xinstr & srawi_mask) == (EXT2 | SRAWIX)) {
+          cr = cr0;
+          instr_at_put(xpos, xinstr | SetRC);
+          pc_ -= kInstrSize;
+        } else if ((xinstr & kOpcodeMask) == ANDIx) {
+          cr = cr0;
+          pc_ -= kInstrSize;
+          // nothing to do here since andi. records.
+        }
+        // didn't match one of the above, must keep cmpwi.
+      }
+    }
+    return cr;
+  }
+
   void bc_short(Condition cond, Label* L, CRegister cr = cr7,
                 LKBit lk = LeaveLK) {
     DCHECK(cond != al);
     DCHECK(cr.code() >= 0 && cr.code() <= 7);
 
+    cr = cmpi_optimization(cr);
+
     int b_offset = branch_offset(L, false);
 
     switch (cond) {
@@ -804,6 +842,8 @@ class Assembler : public AssemblerBase {
     DCHECK(cond != al);
     DCHECK(cr.code() >= 0 && cr.code() <= 7);
 
+    cr = cmpi_optimization(cr);
+
     switch (cond) {
       case eq:
         isel(rt, ra, rb, encode_crbit(cr, CR_EQ));
@@ -1452,6 +1492,9 @@ class Assembler : public AssemblerBase {
 
   // The bound position, before this we cannot do instruction elimination.
   int last_bound_pos_;
+  // Optimizable cmpi information.
+  int optimizable_cmpi_pos_;
+  CRegister cmpi_cr_;
 
   ConstantPoolBuilder constant_pool_builder_;
 

src/ppc/constants-ppc.h

@@ -336,7 +336,8 @@ enum {
   kBOfieldMask = 0x1f << 21,
   kOpcodeMask = 0x3f << 26,
   kExt1OpcodeMask = 0x3ff << 1,
-  kExt2OpcodeMask = 0x1f << 1,
+  kExt2OpcodeMask = 0x3ff << 1,
+  kExt2OpcodeVariant2Mask = 0x1ff << 2,
   kExt5OpcodeMask = 0x3 << 2,
   kBOMask = 0x1f << 21,
   kBIMask = 0x1F << 16,