MIPS[64]: Port '[Deopt] Remove jump table in prologue of deopt entries.'

Port commit 4ab96a9a81

Original message:
> Remove the use of a jump table in the prologue of the deopt entries
> and instead pass the bailout id explicitly in a register when calling
> the deopt entry routine from optimized code. This unifies the logic
> with the way the Arm64 code works. It saves the following amount of
> memory in code stubs:
>
>  - arm:  384KB
>  - ia32: 480KB
>  - x64:  240KB
>
> This could be offset by a slight increase in the size of optimized code
> for loading the immediate, however this impact should be minimal and
> will scale with the maximum number of bailout ids (e.g., the size of
> code will increase by one instruction per bailout id on Arm, therefore
> ~98,000 bailouts will be needed before the overhead is greater than
> the current fixed table size).
>
> Change-Id: I838604b48fa04cbd45320c7b9dac0de08fd8eb25
> Reviewed-on: https://chromium-review.googlesource.com/c/1398224
> Commit-Queue: Ross McIlroy <rmcilroy@chromium.org>
> Reviewed-by: Jaroslav Sevcik <jarin@chromium.org>
> Cr-Commit-Position: refs/heads/master@{#58636}

Change-Id: I4d070b90ebd4f9d4e82eaa74fe6d41c3a39d93e8
Reviewed-on: https://chromium-review.googlesource.com/c/1400848
Reviewed-by: Sreten Kovacevic <skovacevic@wavecomp.com>
Commit-Queue: Sreten Kovacevic <skovacevic@wavecomp.com>
Cr-Commit-Position: refs/heads/master@{#58655}

src/mips/assembler-mips.cc

@@ -202,11 +202,6 @@ bool RelocInfo::IsInConstantPool() {
   return false;
 }
 
-int RelocInfo::GetDeoptimizationId(Isolate* isolate, DeoptimizeKind kind) {
-  DCHECK(IsRuntimeEntry(rmode_));
-  return Deoptimizer::GetDeoptimizationId(isolate, target_address(), kind);
-}
-
 uint32_t RelocInfo::wasm_call_tag() const {
   DCHECK(rmode_ == WASM_CALL || rmode_ == WASM_STUB_CALL);
   return static_cast<uint32_t>(

src/mips/deoptimizer-mips.cc

@@ -15,10 +15,9 @@ namespace internal {
 // This code tries to be close to ia32 code so that any changes can be
 // easily ported.
 void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
-                                                Isolate* isolate, int count,
+                                                Isolate* isolate,
                                                 DeoptimizeKind deopt_kind) {
   NoRootArrayScope no_root_array(masm);
-  GenerateDeoptimizationEntriesPrologue(masm, count);
 
   // Unlike on ARM we don't save all the registers, just the useful ones.
   // For the rest, there are gaps on the stack, so the offsets remain the same.
@@ -64,16 +63,14 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
   const int kSavedRegistersAreaSize =
       (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize + kFloatRegsSize;
 
-  // Get the bailout id from the stack.
+  // Get the bailout id is passed as kRootRegister by the caller.
-  __ lw(a2, MemOperand(sp, kSavedRegistersAreaSize));
+  __ mov(a2, kRootRegister);
 
   // Get the address of the location in the code object (a3) (return
   // address for lazy deoptimization) and compute the fp-to-sp delta in
   // register t0.
   __ mov(a3, ra);
-  // Correct one word for bailout id.
+  __ Addu(t0, sp, Operand(kSavedRegistersAreaSize));
-  __ Addu(t0, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
-
   __ Subu(t0, fp, t0);
 
   // Allocate a new deoptimizer object.
@@ -139,8 +136,8 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
     __ swc1(f0, MemOperand(a1, dst_offset));
   }
 
-  // Remove the bailout id and the saved registers from the stack.
+  // Remove the saved registers from the stack.
-  __ Addu(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
+  __ Addu(sp, sp, Operand(kSavedRegistersAreaSize));
 
   // Compute a pointer to the unwinding limit in register a2; that is
   // the first stack slot not part of the input frame.
@@ -240,73 +237,6 @@ const int Deoptimizer::table_entry_size_ = 2 * kInstrSize;
 const int Deoptimizer::table_entry_size_ = 3 * kInstrSize;
 #endif
 
-void Deoptimizer::GenerateDeoptimizationEntriesPrologue(MacroAssembler* masm,
-                                                        int count) {
-  Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm);
-
-  // Create a sequence of deoptimization entries.
-  // Note that registers are still live when jumping to an entry.
-  Label table_start, done, trampoline_jump;
-  __ bind(&table_start);
-
-#ifdef _MIPS_ARCH_MIPS32R6
-  int kMaxEntriesBranchReach =
-      (1 << (kImm26Bits - 2)) / (table_entry_size_ / kInstrSize);
-#else
-  int kMaxEntriesBranchReach =
-      (1 << (kImm16Bits - 2)) / (table_entry_size_ / kInstrSize);
-#endif
-
-  if (count <= kMaxEntriesBranchReach) {
-    // Common case.
-    for (int i = 0; i < count; i++) {
-      Label start;
-      __ bind(&start);
-      DCHECK(is_int16(i));
-      if (IsMipsArchVariant(kMips32r6)) {
-        __ li(kScratchReg, i);
-        __ BranchShort(PROTECT, &done);
-      } else {
-        __ BranchShort(USE_DELAY_SLOT, &done);  // Expose delay slot.
-        __ li(kScratchReg, i);                  // In the delay slot.
-        __ nop();
-      }
-      DCHECK_EQ(table_entry_size_, masm->SizeOfCodeGeneratedSince(&start));
-    }
-
-    DCHECK_EQ(masm->SizeOfCodeGeneratedSince(&table_start),
-              count * table_entry_size_);
-    __ bind(&done);
-    __ Push(kScratchReg);
-  } else {
-    DCHECK(!IsMipsArchVariant(kMips32r6));
-    // Uncommon case, the branch cannot reach.
-    // Create mini trampoline to reach the end of the table
-    for (int i = 0, j = 0; i < count; i++, j++) {
-      Label start;
-      __ bind(&start);
-      DCHECK(is_int16(i));
-      if (j >= kMaxEntriesBranchReach) {
-        j = 0;
-        __ li(kScratchReg, i);
-        __ bind(&trampoline_jump);
-        trampoline_jump = Label();
-        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);
-        __ nop();
-      } else {
-        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);  // Expose delay slot.
-        __ li(kScratchReg, i);                             // In the delay slot.
-        __ nop();
-      }
-      DCHECK_EQ(table_entry_size_, masm->SizeOfCodeGeneratedSince(&start));
-    }
-
-    DCHECK_EQ(masm->SizeOfCodeGeneratedSince(&table_start),
-              count * table_entry_size_);
-    __ bind(&trampoline_jump);
-    __ Push(kScratchReg);
-  }
-}
 
 bool Deoptimizer::PadTopOfStackRegister() { return false; }
 

src/mips/macro-assembler-mips.cc

@@ -5519,6 +5519,17 @@ void TurboAssembler::ResetSpeculationPoisonRegister() {
   li(kSpeculationPoisonRegister, -1);
 }
 
+void TurboAssembler::CallForDeoptimization(Address target, int deopt_id) {
+  NoRootArrayScope no_root_array(this);
+
+  // Save the deipt id in kRootRegister (we don't need the roots array from now
+  // on).
+  DCHECK_LE(deopt_id, 0xFFFF);
+  li(kRootRegister, deopt_id);
+
+  Call(target, RelocInfo::RUNTIME_ENTRY);
+}
+
 }  // namespace internal
 }  // namespace v8
 

src/mips/macro-assembler-mips.h

@@ -252,11 +252,7 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
   // The return address on the stack is used by frame iteration.
   void StoreReturnAddressAndCall(Register target);
 
-  void CallForDeoptimization(Address target, int deopt_id,
+  void CallForDeoptimization(Address target, int deopt_id);
-                             RelocInfo::Mode rmode) {
-    USE(deopt_id);
-    Call(target, rmode);
-  }
 
   void Ret(COND_ARGS);
   inline void Ret(BranchDelaySlot bd, Condition cond = al,

src/mips64/assembler-mips64.cc

@@ -179,11 +179,6 @@ bool RelocInfo::IsInConstantPool() {
   return false;
 }
 
-int RelocInfo::GetDeoptimizationId(Isolate* isolate, DeoptimizeKind kind) {
-  DCHECK(IsRuntimeEntry(rmode_));
-  return Deoptimizer::GetDeoptimizationId(isolate, target_address(), kind);
-}
-
 uint32_t RelocInfo::wasm_call_tag() const {
   DCHECK(rmode_ == WASM_CALL || rmode_ == WASM_STUB_CALL);
   return static_cast<uint32_t>(

src/mips64/deoptimizer-mips64.cc

@@ -15,10 +15,9 @@ namespace internal {
 // This code tries to be close to ia32 code so that any changes can be
 // easily ported.
 void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
-                                                Isolate* isolate, int count,
+                                                Isolate* isolate,
                                                 DeoptimizeKind deopt_kind) {
   NoRootArrayScope no_root_array(masm);
-  GenerateDeoptimizationEntriesPrologue(masm, count);
 
   // Unlike on ARM we don't save all the registers, just the useful ones.
   // For the rest, there are gaps on the stack, so the offsets remain the same.
@@ -65,15 +64,14 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
   const int kSavedRegistersAreaSize =
       (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize + kFloatRegsSize;
 
-  // Get the bailout id from the stack.
+  // Get the bailout is passed as kRootRegister by the caller.
-  __ Ld(a2, MemOperand(sp, kSavedRegistersAreaSize));
+  __ mov(a2, kRootRegister);
 
   // Get the address of the location in the code object (a3) (return
   // address for lazy deoptimization) and compute the fp-to-sp delta in
   // register a4.
   __ mov(a3, ra);
-  // Correct one word for bailout id.
+  __ Daddu(a4, sp, Operand(kSavedRegistersAreaSize));
-  __ Daddu(a4, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
 
   __ Dsubu(a4, fp, a4);
 
@@ -140,8 +138,8 @@ void Deoptimizer::GenerateDeoptimizationEntries(MacroAssembler* masm,
     __ Swc1(f0, MemOperand(a1, dst_offset));
   }
 
-  // Remove the bailout id and the saved registers from the stack.
+  // Remove the saved registers from the stack.
-  __ Daddu(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
+  __ Daddu(sp, sp, Operand(kSavedRegistersAreaSize));
 
   // Compute a pointer to the unwinding limit in register a2; that is
   // the first stack slot not part of the input frame.
@@ -239,74 +237,6 @@ const int Deoptimizer::table_entry_size_ = 2 * kInstrSize;
 const int Deoptimizer::table_entry_size_ = 3 * kInstrSize;
 #endif
 
-void Deoptimizer::GenerateDeoptimizationEntriesPrologue(MacroAssembler* masm,
-                                                        int count) {
-  Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm);
-
-  // Create a sequence of deoptimization entries.
-  // Note that registers are still live when jumping to an entry.
-  Label table_start, done, trampoline_jump;
-  __ bind(&table_start);
-#ifdef _MIPS_ARCH_MIPS64R6
-  int kMaxEntriesBranchReach =
-      (1 << (kImm26Bits - 2)) / (table_entry_size_ / kInstrSize);
-#else
-  int kMaxEntriesBranchReach =
-      (1 << (kImm16Bits - 2)) / (table_entry_size_ / kInstrSize);
-#endif
-
-  if (count <= kMaxEntriesBranchReach) {
-    // Common case.
-    for (int i = 0; i < count; i++) {
-      Label start;
-      __ bind(&start);
-      DCHECK(is_int16(i));
-      if (kArchVariant == kMips64r6) {
-        __ li(kScratchReg, i);
-        __ BranchShort(PROTECT, &done);
-      } else {
-        __ BranchShort(USE_DELAY_SLOT, &done);  // Expose delay slot.
-        __ li(kScratchReg, i);                  // In the delay slot.
-        __ nop();
-      }
-
-      DCHECK_EQ(table_entry_size_, masm->SizeOfCodeGeneratedSince(&start));
-    }
-
-    DCHECK_EQ(masm->SizeOfCodeGeneratedSince(&table_start),
-              count * table_entry_size_);
-    __ bind(&done);
-    __ Push(kScratchReg);
-  } else {
-    DCHECK_NE(kArchVariant, kMips64r6);
-    // Uncommon case, the branch cannot reach.
-    // Create mini trampoline to reach the end of the table
-    for (int i = 0, j = 0; i < count; i++, j++) {
-      Label start;
-      __ bind(&start);
-      DCHECK(is_int16(i));
-      if (j >= kMaxEntriesBranchReach) {
-        j = 0;
-        __ li(kScratchReg, i);
-        __ bind(&trampoline_jump);
-        trampoline_jump = Label();
-        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);
-        __ nop();
-      } else {
-        __ BranchShort(USE_DELAY_SLOT, &trampoline_jump);  // Expose delay slot.
-        __ li(kScratchReg, i);                             // In the delay slot.
-        __ nop();
-      }
-      DCHECK_EQ(table_entry_size_, masm->SizeOfCodeGeneratedSince(&start));
-    }
-
-    DCHECK_EQ(masm->SizeOfCodeGeneratedSince(&table_start),
-              count * table_entry_size_);
-    __ bind(&trampoline_jump);
-    __ Push(kScratchReg);
-  }
-}
-
 bool Deoptimizer::PadTopOfStackRegister() { return false; }
 
 void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {

src/mips64/macro-assembler-mips64.cc

@@ -5886,6 +5886,16 @@ void TurboAssembler::ResetSpeculationPoisonRegister() {
   li(kSpeculationPoisonRegister, -1);
 }
 
+void TurboAssembler::CallForDeoptimization(Address target, int deopt_id) {
+  NoRootArrayScope no_root_array(this);
+
+  // Save the deopt id in kRootRegister (we don't need the roots array from now
+  // on).
+  DCHECK_LE(deopt_id, 0xFFFF);
+  li(kRootRegister, deopt_id);
+  Call(target, RelocInfo::RUNTIME_ENTRY);
+}
+
 }  // namespace internal
 }  // namespace v8
 

src/mips64/macro-assembler-mips64.h

@@ -277,11 +277,7 @@ class V8_EXPORT_PRIVATE TurboAssembler : public TurboAssemblerBase {
   // The return address on the stack is used by frame iteration.
   void StoreReturnAddressAndCall(Register target);
 
-  void CallForDeoptimization(Address target, int deopt_id,
+  void CallForDeoptimization(Address target, int deopt_id);
-                             RelocInfo::Mode rmode) {
-    USE(deopt_id);
-    Call(target, rmode);
-  }
 
   void Ret(COND_ARGS);
   inline void Ret(BranchDelaySlot bd, Condition cond = al,