MIPS: Use JIC/JIALC offset when possible

src/mips/macro-assembler-mips.cc

 // Copyright 2012 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.
 
 #include <limits.h>  // For LONG_MIN, LONG_MAX.
 
 #if V8_TARGET_ARCH_MIPS
 
 #include "src/base/bits.h"
 #include "src/base/division-by-constant.h"
@@ skipping 3582 matching lines @@
     DCHECK(offset == 0);
     if (IsMipsArchVariant(kMips32r6) && bdslot == PROTECT) {
       return BranchAndLinkShortHelperR6(0, L, cond, rs, rt);
     } else {
       return BranchAndLinkShortHelper(0, L, cond, rs, rt, bdslot);
     }
   }
   return false;
 }
 
+void MacroAssembler::Jump(Register target, int16_t offset, Condition cond,
+                          Register rs, const Operand& rt, BranchDelaySlot bd) {
+  BlockTrampolinePoolScope block_trampoline_pool(this);
+  DCHECK(is_int16(offset));

[dusan.simicic, 2017/04/25 11:24:10]

'offset' always fits in int16 in this method, DCHECK() is not necessary? 

+  if (IsMipsArchVariant(kMips32r6) && bd == PROTECT) {
+    if (cond == cc_always) {
+      jic(target, offset);
+    } else {
+      BRANCH_ARGS_CHECK(cond, rs, rt);
+      Branch(2, NegateCondition(cond), rs, rt);
+      jic(target, offset);
+    }
+  } else {
+    if (offset != 0) {
+      Addu(target, target, offset);
+    }
+    if (cond == cc_always) {
+      jr(target);
+    } else {
+      BRANCH_ARGS_CHECK(cond, rs, rt);
+      Branch(2, NegateCondition(cond), rs, rt);
+      jr(target);
+    }
+    // Emit a nop in the branch delay slot if required.
+    if (bd == PROTECT) nop();
+  }
+}
 
-void MacroAssembler::Jump(Register target,
-                          Condition cond,
-                          Register rs,
-                          const Operand& rt,
+void MacroAssembler::Jump(Register target, Register base, int16_t offset,
+                          Condition cond, Register rs, const Operand& rt,
                           BranchDelaySlot bd) {
+  DCHECK(is_int16(offset));

[dusan.simicic, 2017/04/25 11:24:10]

Same comment as above.

   BlockTrampolinePoolScope block_trampoline_pool(this);
   if (IsMipsArchVariant(kMips32r6) && bd == PROTECT) {
     if (cond == cc_always) {
-      jic(target, 0);
+      jic(base, offset);
     } else {
       BRANCH_ARGS_CHECK(cond, rs, rt);
       Branch(2, NegateCondition(cond), rs, rt);
-      jic(target, 0);
+      jic(base, offset);
     }
   } else {
+    if (offset != 0) {
+      Addu(target, base, offset);
+    } else {  // Call through target
+      if (!target.is(base)) mov(target, base);
+    }
     if (cond == cc_always) {
       jr(target);
     } else {
+      BRANCH_ARGS_CHECK(cond, rs, rt);
+      Branch(2, NegateCondition(cond), rs, rt);
+      jr(target);
+    }
+    // Emit a nop in the branch delay slot if required.
+    if (bd == PROTECT) nop();
+  }
+}
+
+void MacroAssembler::Jump(Register target, const Operand& offset,
+                          Condition cond, Register rs, const Operand& rt,
+                          BranchDelaySlot bd) {
+  BlockTrampolinePoolScope block_trampoline_pool(this);
+  if (IsMipsArchVariant(kMips32r6) && bd == PROTECT &&
+      !is_int16(offset.immediate())) {
+    uint32_t aui_offset, jic_offset;
+    Assembler::UnpackTargetAddressUnsigned(offset.immediate(), aui_offset,
+                                           jic_offset);
+    RecordRelocInfo(RelocInfo::EXTERNAL_REFERENCE, offset.immediate());
+    aui(target, target, aui_offset);
+    if (cond == cc_always) {
+      jic(target, jic_offset);
+    } else {
+      BRANCH_ARGS_CHECK(cond, rs, rt);
+      Branch(2, NegateCondition(cond), rs, rt);
+      jic(target, jic_offset);
+    }
+  } else {
+    if (offset.immediate() != 0) {
+      Addu(target, target, offset);
+    }
+    if (cond == cc_always) {
+      jr(target);
+    } else {
       BRANCH_ARGS_CHECK(cond, rs, rt);
       Branch(2, NegateCondition(cond), rs, rt);
       jr(target);
     }
     // Emit a nop in the branch delay slot if required.
     if (bd == PROTECT) nop();
   }
 }
 
 
 void MacroAssembler::Jump(intptr_t target,
                           RelocInfo::Mode rmode,
                           Condition cond,
                           Register rs,
                           const Operand& rt,
                           BranchDelaySlot bd) {
+  BlockTrampolinePoolScope block_trampoline_pool(this);
   Label skip;
   if (cond != cc_always) {
     Branch(USE_DELAY_SLOT, &skip, NegateCondition(cond), rs, rt);
   }
   // The first instruction of 'li' may be placed in the delay slot.
   // This is not an issue, t9 is expected to be clobbered anyway.

[dusan.simicic, 2017/04/25 11:24:10]

This comment is related to 'else' block now. Maybe is better to move it down.

-  li(t9, Operand(target, rmode));
-  Jump(t9, al, zero_reg, Operand(zero_reg), bd);
+  if (IsMipsArchVariant(kMips32r6) && bd == PROTECT) {
+    uint32_t lui_offset, jic_offset;
+    UnpackTargetAddressUnsigned(target, lui_offset, jic_offset);
+    DCHECK(MustUseReg(rmode));
+    RecordRelocInfo(rmode, target);
+    lui(t9, lui_offset);
+    Jump(t9, jic_offset, al, zero_reg, Operand(zero_reg), bd);
+  } else {
+    li(t9, Operand(target, rmode));
+    Jump(t9, 0, al, zero_reg, Operand(zero_reg), bd);
+  }
   bind(&skip);
 }
 
 
 void MacroAssembler::Jump(Address target,
                           RelocInfo::Mode rmode,
                           Condition cond,
                           Register rs,
                           const Operand& rt,
                           BranchDelaySlot bd) {
   DCHECK(!RelocInfo::IsCodeTarget(rmode));
   Jump(reinterpret_cast<intptr_t>(target), rmode, cond, rs, rt, bd);
 }
 
 
 void MacroAssembler::Jump(Handle<Code> code,
                           RelocInfo::Mode rmode,
                           Condition cond,
                           Register rs,
                           const Operand& rt,
                           BranchDelaySlot bd) {
   DCHECK(RelocInfo::IsCodeTarget(rmode));
   AllowDeferredHandleDereference embedding_raw_address;
   Jump(reinterpret_cast<intptr_t>(code.location()), rmode, cond, rs, rt, bd);
 }
 
-
-int MacroAssembler::CallSize(Register target,
-                             Condition cond,
-                             Register rs,
-                             const Operand& rt,
+int MacroAssembler::CallSize(Register target, int16_t offset, Condition cond,
+                             Register rs, const Operand& rt,
                              BranchDelaySlot bd) {
   int size = 0;
 
   if (cond == cc_always) {
     size += 1;
   } else {
     size += 3;
   }
 
   if (bd == PROTECT && !IsMipsArchVariant(kMips32r6)) size += 1;
 
+  if (!IsMipsArchVariant(kMips32r6) && offset != 0) {
+    size += 1;
+  }
+
   return size * kInstrSize;
 }
 
 
 // Note: To call gcc-compiled C code on mips, you must call thru t9.
-void MacroAssembler::Call(Register target,
-                          Condition cond,
-                          Register rs,
-                          const Operand& rt,
-                          BranchDelaySlot bd) {
+void MacroAssembler::Call(Register target, int16_t offset, Condition cond,
+                          Register rs, const Operand& rt, BranchDelaySlot bd) {
+  DCHECK(is_int16(offset));
 #ifdef DEBUG
   int size = IsPrevInstrCompactBranch() ? kInstrSize : 0;
 #endif
+
+  BlockTrampolinePoolScope block_trampoline_pool(this);
+  Label start;
+  bind(&start);
+  if (IsMipsArchVariant(kMips32r6) && bd == PROTECT) {
+    if (cond == cc_always) {
+      jialc(target, offset);
+    } else {
+      BRANCH_ARGS_CHECK(cond, rs, rt);
+      Branch(2, NegateCondition(cond), rs, rt);
+      jialc(target, offset);
+    }
+  } else {
+    if (offset != 0) {
+      Addu(target, target, offset);
+    }
+    if (cond == cc_always) {
+      jalr(target);
+    } else {
+      BRANCH_ARGS_CHECK(cond, rs, rt);
+      Branch(2, NegateCondition(cond), rs, rt);
+      jalr(target);
+    }
+    // Emit a nop in the branch delay slot if required.
+    if (bd == PROTECT) nop();
+  }
+
+#ifdef DEBUG
+  CHECK_EQ(size + CallSize(target, offset, cond, rs, rt, bd),
+           SizeOfCodeGeneratedSince(&start));
+#endif
+}
+
+// Note: To call gcc-compiled C code on mips, you must call thru t9.
+void MacroAssembler::Call(Register target, Register base, int16_t offset,
+                          Condition cond, Register rs, const Operand& rt,
+                          BranchDelaySlot bd) {
+  DCHECK(is_uint16(offset));
+#ifdef DEBUG
+  int size = IsPrevInstrCompactBranch() ? kInstrSize : 0;
+#endif
 
   BlockTrampolinePoolScope block_trampoline_pool(this);
   Label start;
   bind(&start);
   if (IsMipsArchVariant(kMips32r6) && bd == PROTECT) {
     if (cond == cc_always) {
-      jialc(target, 0);
+      jialc(base, offset);
     } else {
       BRANCH_ARGS_CHECK(cond, rs, rt);
       Branch(2, NegateCondition(cond), rs, rt);
-      jialc(target, 0);
+      jialc(base, offset);
     }
   } else {
+    if (offset != 0) {
+      Addu(target, base, offset);
+    } else {  // Call through target
+      if (!target.is(base)) mov(target, base);
+    }
     if (cond == cc_always) {
       jalr(target);
     } else {
       BRANCH_ARGS_CHECK(cond, rs, rt);
       Branch(2, NegateCondition(cond), rs, rt);
       jalr(target);
     }
     // Emit a nop in the branch delay slot if required.
     if (bd == PROTECT) nop();
   }
 
 #ifdef DEBUG
-  CHECK_EQ(size + CallSize(target, cond, rs, rt, bd),
+  CHECK_EQ(size + CallSize(target, offset, cond, rs, rt, bd),
            SizeOfCodeGeneratedSince(&start));
 #endif
 }
 
 
 int MacroAssembler::CallSize(Address target,
                              RelocInfo::Mode rmode,
                              Condition cond,
                              Register rs,
                              const Operand& rt,
                              BranchDelaySlot bd) {
-  int size = CallSize(t9, cond, rs, rt, bd);
-  return size + 2 * kInstrSize;
+  int size = CallSize(t9, 0, cond, rs, rt, bd);
+  if (IsMipsArchVariant(kMips32r6) && bd == PROTECT && cond == cc_always)
+    return size + 1 * kInstrSize;
+  else
+    return size + 2 * kInstrSize;
 }
 
 
 void MacroAssembler::Call(Address target,
                           RelocInfo::Mode rmode,
                           Condition cond,
                           Register rs,
                           const Operand& rt,
                           BranchDelaySlot bd) {
+  CheckBuffer();
   BlockTrampolinePoolScope block_trampoline_pool(this);
   Label start;
   bind(&start);
   int32_t target_int = reinterpret_cast<int32_t>(target);
-  li(t9, Operand(target_int, rmode), CONSTANT_SIZE);
-  Call(t9, cond, rs, rt, bd);
+  if (IsMipsArchVariant(kMips32r6) && bd == PROTECT && cond == cc_always) {
+    uint32_t lui_offset, jialc_offset;
+    UnpackTargetAddressUnsigned(target_int, lui_offset, jialc_offset);
+    if (MustUseReg(rmode)) {
+      RecordRelocInfo(rmode, target_int);
+    }
+    lui(t9, lui_offset);
+    Call(t9, jialc_offset, cond, rs, rt, bd);
+  } else {
+    li(t9, Operand(target_int, rmode), CONSTANT_SIZE);
+    Call(t9, 0, cond, rs, rt, bd);
+  }
   DCHECK_EQ(CallSize(target, rmode, cond, rs, rt, bd),
             SizeOfCodeGeneratedSince(&start));
 }
 
 
 int MacroAssembler::CallSize(Handle<Code> code,
                              RelocInfo::Mode rmode,
                              TypeFeedbackId ast_id,
                              Condition cond,
                              Register rs,
@@ skipping 24 matching lines @@
   Call(reinterpret_cast<Address>(code.location()), rmode, cond, rs, rt, bd);
   DCHECK_EQ(CallSize(code, rmode, ast_id, cond, rs, rt, bd),
             SizeOfCodeGeneratedSince(&start));
 }
 
 
 void MacroAssembler::Ret(Condition cond,
                          Register rs,
                          const Operand& rt,
                          BranchDelaySlot bd) {
-  Jump(ra, cond, rs, rt, bd);
+  Jump(ra, 0, cond, rs, rt, bd);
 }
 
 
 void MacroAssembler::BranchLong(Label* L, BranchDelaySlot bdslot) {
   if (IsMipsArchVariant(kMips32r6) && bdslot == PROTECT &&
       (!L->is_bound() || is_near_r6(L))) {
     BranchShortHelperR6(0, L);
   } else {
     BlockTrampolinePoolScope block_trampoline_pool(this);
     uint32_t imm32;
     imm32 = jump_address(L);
     if (IsMipsArchVariant(kMips32r6) && bdslot == PROTECT) {
       uint32_t lui_offset, jic_offset;
       UnpackTargetAddressUnsigned(imm32, lui_offset, jic_offset);
       {
         BlockGrowBufferScope block_buf_growth(this);
         // Buffer growth (and relocation) must be blocked for internal
         // references until associated instructions are emitted and
         // available to be patched.
         RecordRelocInfo(RelocInfo::INTERNAL_REFERENCE_ENCODED);
         lui(at, lui_offset);
         jic(at, jic_offset);
       }
       CheckBuffer();
     } else {
       {
         BlockGrowBufferScope block_buf_growth(this);
         // Buffer growth (and relocation) must be blocked for internal
-        // references
-        // until associated instructions are emitted and available to be
-        // patched.
+        // references until associated instructions are emitted and
+        // available to be patched.
         RecordRelocInfo(RelocInfo::INTERNAL_REFERENCE_ENCODED);
         lui(at, (imm32 & kHiMask) >> kLuiShift);
         ori(at, at, (imm32 & kImm16Mask));
       }
       CheckBuffer();
       jr(at);
       // Emit a nop in the branch delay slot if required.
       if (bdslot == PROTECT) nop();
     }
   }
 }
 
 
 void MacroAssembler::BranchAndLinkLong(Label* L, BranchDelaySlot bdslot) {
   if (IsMipsArchVariant(kMips32r6) && bdslot == PROTECT &&
       (!L->is_bound() || is_near_r6(L))) {
     BranchAndLinkShortHelperR6(0, L);
   } else {
     BlockTrampolinePoolScope block_trampoline_pool(this);
     uint32_t imm32;
     imm32 = jump_address(L);
     if (IsMipsArchVariant(kMips32r6) && bdslot == PROTECT) {
-      uint32_t lui_offset, jic_offset;
-      UnpackTargetAddressUnsigned(imm32, lui_offset, jic_offset);
+      uint32_t lui_offset, jialc_offset;
+      UnpackTargetAddressUnsigned(imm32, lui_offset, jialc_offset);
       {
         BlockGrowBufferScope block_buf_growth(this);
         // Buffer growth (and relocation) must be blocked for internal
         // references until associated instructions are emitted and
         // available to be patched.
         RecordRelocInfo(RelocInfo::INTERNAL_REFERENCE_ENCODED);
         lui(at, lui_offset);
-        jialc(at, jic_offset);
+        jialc(at, jialc_offset);
       }
       CheckBuffer();
     } else {
       {
         BlockGrowBufferScope block_buf_growth(this);
         // Buffer growth (and relocation) must be blocked for internal
-        // references
-        // until associated instructions are emitted and available to be
-        // patched.
+        // references until associated instructions are emitted and
+        // available to be patched.
         RecordRelocInfo(RelocInfo::INTERNAL_REFERENCE_ENCODED);
         lui(at, (imm32 & kHiMask) >> kLuiShift);
         ori(at, at, (imm32 & kImm16Mask));
       }
       CheckBuffer();
       jalr(at);
       // Emit a nop in the branch delay slot if required.
       if (bdslot == PROTECT) nop();
     }
   }
@@ skipping 2158 matching lines @@
 
 void MacroAssembler::PrepareCallCFunction(int num_reg_arguments,
                                           Register scratch) {
   PrepareCallCFunction(num_reg_arguments, 0, scratch);
 }
 
 
 void MacroAssembler::CallCFunction(ExternalReference function,
                                    int num_reg_arguments,
                                    int num_double_arguments) {
-  li(t8, Operand(function));
-  CallCFunctionHelper(t8, num_reg_arguments, num_double_arguments);
+  if (IsMipsArchVariant(kMips32r6)) {
+    uint32_t lui_offset, jialc_offset;
+    UnpackTargetAddressUnsigned(Operand(function).immediate(), lui_offset,
+                                jialc_offset);
+    if (MustUseReg(Operand(function).rmode())) {
+      RecordRelocInfo(Operand(function).rmode(), Operand(function).immediate());
+    }
+    lui(t9, lui_offset);
+    CallCFunctionHelper(t9, jialc_offset, num_reg_arguments,
+                        num_double_arguments);
+  } else {
+    li(t9, Operand(function));
+    CallCFunctionHelper(t9, 0, num_reg_arguments, num_double_arguments);
+  }
 }
 
 
 void MacroAssembler::CallCFunction(Register function,
                                    int num_reg_arguments,
                                    int num_double_arguments) {
-  CallCFunctionHelper(function, num_reg_arguments, num_double_arguments);
+  CallCFunctionHelper(function, 0, num_reg_arguments, num_double_arguments);
 }
 
 
 void MacroAssembler::CallCFunction(ExternalReference function,
                                    int num_arguments) {
   CallCFunction(function, num_arguments, 0);
 }
 
 
 void MacroAssembler::CallCFunction(Register function,
                                    int num_arguments) {
   CallCFunction(function, num_arguments, 0);
 }
 
-
-void MacroAssembler::CallCFunctionHelper(Register function,
+void MacroAssembler::CallCFunctionHelper(Register function_base,
+                                         int16_t function_offset,
                                          int num_reg_arguments,
                                          int num_double_arguments) {
   DCHECK(has_frame());
   // Make sure that the stack is aligned before calling a C function unless
   // running in the simulator. The simulator has its own alignment check which
   // provides more information.
   // The argument stots are presumed to have been set up by
   // PrepareCallCFunction. The C function must be called via t9, for mips ABI.
 
 #if V8_HOST_ARCH_MIPS
@@ skipping 10 matching lines @@
       stop("Unexpected alignment in CallCFunction");
       bind(&alignment_as_expected);
     }
   }
 #endif  // V8_HOST_ARCH_MIPS
 
   // Just call directly. The function called cannot cause a GC, or
   // allow preemption, so the return address in the link register
   // stays correct.
 
-  if (!function.is(t9)) {
-    mov(t9, function);
-    function = t9;
+  if (!function_base.is(t9)) {
+    mov(t9, function_base);
+    function_base = t9;
   }
 
-  Call(function);
+  Call(function_base, function_offset);
 
   int stack_passed_arguments = CalculateStackPassedWords(
       num_reg_arguments, num_double_arguments);
 
   if (base::OS::ActivationFrameAlignment() > kPointerSize) {
     lw(sp, MemOperand(sp, stack_passed_arguments * kPointerSize));
   } else {
     Addu(sp, sp, Operand(stack_passed_arguments * kPointerSize));
   }
 }
@@ skipping 314 matching lines @@
 }
 
 
 CodePatcher::~CodePatcher() {
   // Indicate that code has changed.
   if (flush_cache_ == FLUSH) {
     Assembler::FlushICache(masm_.isolate(), address_, size_);
   }
 
   // Check that the code was patched as expected.
+
   DCHECK(masm_.pc_ == address_ + size_);
   DCHECK(masm_.reloc_info_writer.pos() == address_ + size_ + Assembler::kGap);
 }
 
 
 void CodePatcher::Emit(Instr instr) {
   masm()->emit(instr);
 }
 
 
@@ skipping 29 matching lines @@
   if (mag.shift > 0) sra(result, result, mag.shift);
   srl(at, dividend, 31);
   Addu(result, result, Operand(at));
 }
 
 
 }  // namespace internal
 }  // namespace v8
 
 #endif  // V8_TARGET_ARCH_MIPS