s390: optimize TF to use tmll and fix tmll sim

R=joransiu@ca.ibm.com, bjaideep@ca.ibm.com
BUG=

Review-Url: https://codereview.chromium.org/2622073006
Cr-Commit-Position: refs/heads/master@{#42297}

src/compiler/s390/code-generator-s390.cc

@@ -1675,21 +1675,30 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       break;
     case kS390_Tst32:
       if (HasRegisterInput(instr, 1)) {
-        __ AndP(r0, i.InputRegister(0), i.InputRegister(1));
+        __ lr(r0, i.InputRegister(0));
+        __ nr(r0, i.InputRegister(1));
       } else {
-        __ AndP(r0, i.InputRegister(0), i.InputImmediate(1));
+        Operand opnd = i.InputImmediate(1);
+        if (is_uint16(opnd.immediate())) {
+          __ tmll(i.InputRegister(0), opnd);
+        } else {
+          __ lr(r0, i.InputRegister(0));
+          __ nilf(r0, opnd);
+        }
       }
-      __ LoadAndTestP_ExtendSrc(r0, r0);
       break;
-#if V8_TARGET_ARCH_S390X
     case kS390_Tst64:
       if (HasRegisterInput(instr, 1)) {
         __ AndP(r0, i.InputRegister(0), i.InputRegister(1));
       } else {
-        __ AndP(r0, i.InputRegister(0), i.InputImmediate(1));
+        Operand opnd = i.InputImmediate(1);
+        if (is_uint16(opnd.immediate())) {
+          __ tmll(i.InputRegister(0), opnd);
+        } else {
+          __ AndP(r0, i.InputRegister(0), opnd);
+        }
       }
       break;
-#endif
     case kS390_Float64SilenceNaN: {
       DoubleRegister value = i.InputDoubleRegister(0);
       DoubleRegister result = i.OutputDoubleRegister();

src/compiler/s390/instruction-selector-s390.cc

@@ -1625,9 +1625,27 @@ void VisitWordCompareZero(InstructionSelector* selector, Node* user,
 
   if (selector->CanCover(user, value)) {
     switch (value->opcode()) {
-      case IrOpcode::kWord32Equal:
+      case IrOpcode::kWord32Equal: {
         cont->OverwriteAndNegateIfEqual(kEqual);
+        Int32BinopMatcher m(value);
+        if (m.right().Is(0)) {
+          // Try to combine the branch with a comparison.
+          Node* const user = m.node();
+          Node* const value = m.left().node();
+          if (selector->CanCover(user, value)) {
+            switch (value->opcode()) {
+              case IrOpcode::kInt32Sub:
+                return VisitWord32Compare(selector, value, cont);
+              case IrOpcode::kWord32And:
+                return VisitWordCompare(selector, value, kS390_Tst64, cont,
+                                        true, kUint32Imm);
+              default:
+                break;
+            }
+          }
+        }
         return VisitWord32Compare(selector, value, cont);
+      }
       case IrOpcode::kInt32LessThan:
         cont->OverwriteAndNegateIfEqual(kSignedLessThan);
         return VisitWord32Compare(selector, value, cont);
@@ -1641,9 +1659,27 @@ void VisitWordCompareZero(InstructionSelector* selector, Node* user,
         cont->OverwriteAndNegateIfEqual(kUnsignedLessThanOrEqual);
         return VisitWord32Compare(selector, value, cont);
 #if V8_TARGET_ARCH_S390X
-      case IrOpcode::kWord64Equal:
+      case IrOpcode::kWord64Equal: {
         cont->OverwriteAndNegateIfEqual(kEqual);
+        Int64BinopMatcher m(value);
+        if (m.right().Is(0)) {
+          // Try to combine the branch with a comparison.
+          Node* const user = m.node();
+          Node* const value = m.left().node();
+          if (selector->CanCover(user, value)) {
+            switch (value->opcode()) {
+              case IrOpcode::kInt64Sub:
+                return VisitWord64Compare(selector, value, cont);
+              case IrOpcode::kWord64And:
+                return VisitWordCompare(selector, value, kS390_Tst64, cont,
+                                        true, kUint32Imm);
+              default:
+                break;
+            }
+          }
+        }
         return VisitWord64Compare(selector, value, cont);
+      }
       case IrOpcode::kInt64LessThan:
         cont->OverwriteAndNegateIfEqual(kSignedLessThan);
         return VisitWord64Compare(selector, value, cont);

src/s390/simulator-s390.cc

@@ -7779,47 +7779,38 @@ EVALUATE(TMLH) {
 EVALUATE(TMLL) {
   DCHECK_OPCODE(TMLL);
   DECODE_RI_A_INSTRUCTION(instr, r1, i2);
-  int mask = i2 & 0x0000FFFF;
-  if (mask == 0) {
-    condition_reg_ = 0x0;
-    return length;
-  }
+  uint32_t mask = i2 & 0x0000FFFF;
   uint32_t r1_val = get_low_register<uint32_t>(r1);
   r1_val = r1_val & 0x0000FFFF;  // uses only the last 16bits
 
-  // Test if all selected bits are Zero
-  bool allSelectedBitsAreZeros = true;
-  for (int i = 0; i < 15; i++) {
-    if (mask & (1 << i)) {
-      if (r1_val & (1 << i)) {
-        allSelectedBitsAreZeros = false;
-        break;
-      }
-    }
-  }
-  if (allSelectedBitsAreZeros) {
+  // Test if all selected bits are zeros or mask is zero
+  if (0 == (mask & r1_val)) {
     condition_reg_ = 0x8;
     return length;  // Done!
   }
 
+  DCHECK(mask != 0);
   // Test if all selected bits are one
-  bool allSelectedBitsAreOnes = true;
-  for (int i = 0; i < 15; i++) {
-    if (mask & (1 << i)) {
-      if (!(r1_val & (1 << i))) {
-        allSelectedBitsAreOnes = false;
-        break;
-      }
-    }
-  }
-  if (allSelectedBitsAreOnes) {
+  if (mask == (mask & r1_val)) {
     condition_reg_ = 0x1;
     return length;  // Done!
   }
 
   // Now we know selected bits mixed zeros and ones
   // Test if the leftmost bit is zero or one
-  for (int i = 14; i >= 0; i--) {
+#if defined(__GNUC__)
+  int leadingZeros = __builtin_clz(mask);
+  mask = 0x80000000u >> leadingZeros;
+  if (mask & r1_val) {
+    // leftmost bit is one
+    condition_reg_ = 0x4;
+  } else {
+    // leftmost bit is zero
+    condition_reg_ = 0x2;
+  }
+  return length;  // Done!
+#else
+  for (int i = 15; i >= 0; i--) {
     if (mask & (1 << i)) {
       if (r1_val & (1 << i)) {
         // leftmost bit is one
@@ -7831,6 +7822,8 @@ EVALUATE(TMLL) {
       return length;  // Done!
     }
   }
+#endif
+  UNREACHABLE();
   return length;
 }