[turbofan] Fewer gap moves for tail calls

Restore frame pointer directly from stack rather than copying it and
restoring. Also restore return address register directly on platforms that
support it.

BUG=v8:4076
LOG=n

Review URL: https://codereview.chromium.org/1488553002

Cr-Commit-Position: refs/heads/master@{#32410}

src/base/build_config.h

@@ -183,6 +183,13 @@
 #error Unknown target architecture endianness
 #endif
 
+#if defined(V8_TARGET_ARCH_IA32) || defined(V8_TARGET_ARCH_X64) || \
+    defined(V8_TARGET_ARCH_X87)
+#define V8_TARGET_ARCH_STORES_RETURN_ADDRESS_ON_STACK 1
+#else
+#define V8_TARGET_ARCH_STORES_RETURN_ADDRESS_ON_STACK 0
+#endif
+
 // Number of bits to represent the page size for paged spaces. The value of 20
 // gives 1Mb bytes per page.
 #if V8_HOST_ARCH_PPC && V8_TARGET_ARCH_PPC && V8_OS_LINUX

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

@@ -359,13 +359,6 @@ void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   if (sp_slot_delta > 0) {
     __ add(sp, sp, Operand(sp_slot_delta * kPointerSize));
   }
-  if (frame()->needs_frame()) {
-    if (FLAG_enable_embedded_constant_pool) {
-      __ ldm(ia_w, sp, pp.bit() | fp.bit() | lr.bit());
-    } else {
-      __ ldm(ia_w, sp, fp.bit() | lr.bit());
-    }
-  }
   frame_access_state()->SetFrameAccessToDefault();
 }
 
@@ -376,6 +369,13 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ sub(sp, sp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
+  if (frame()->needs_frame()) {
+    if (FLAG_enable_embedded_constant_pool) {
+      __ ldr(cp, MemOperand(fp, StandardFrameConstants::kConstantPoolOffset));
+    }
+    __ ldr(lr, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
+    __ ldr(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
+  }
   frame_access_state()->SetFrameAccessToSP();
 }
 

src/compiler/arm64/code-generator-arm64.cc

@@ -464,9 +464,6 @@ void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   if (sp_slot_delta > 0) {
     __ Add(jssp, jssp, Operand(sp_slot_delta * kPointerSize));
   }
-  if (frame()->needs_frame()) {
-    __ Pop(fp, lr);
-  }
   frame_access_state()->SetFrameAccessToDefault();
 }
 
@@ -477,6 +474,10 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ Sub(jssp, jssp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
+  if (frame()->needs_frame()) {
+    __ Ldr(lr, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
+    __ Ldr(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
+  }
   frame_access_state()->SetFrameAccessToSP();
 }
 

src/compiler/code-generator.cc

@@ -675,12 +675,10 @@ int CodeGenerator::TailCallFrameStackSlotDelta(int stack_param_delta) {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
   int spill_slots = frame()->GetSpillSlotCount();
   bool has_frame = descriptor->IsJSFunctionCall() || spill_slots > 0;
-  // Leave the PC and saved frame pointer on the stack.
+  // Leave the PC on the stack on platforms that have that as part of their ABI
+  int pc_slots = V8_TARGET_ARCH_STORES_RETURN_ADDRESS_ON_STACK ? 1 : 0;
   int sp_slot_delta =
-      has_frame
-          ? (frame()->GetTotalFrameSlotCount() -
-             (StandardFrameConstants::kFixedFrameSizeFromFp / kPointerSize))
-          : 0;
+      has_frame ? (frame()->GetTotalFrameSlotCount() - pc_slots) : 0;
   // Discard only slots that won't be used by new parameters.
   sp_slot_delta += stack_param_delta;
   return sp_slot_delta;

src/compiler/ia32/code-generator-ia32.cc

@@ -337,9 +337,6 @@ void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   if (sp_slot_delta > 0) {
     __ add(esp, Immediate(sp_slot_delta * kPointerSize));
   }
-  if (frame()->needs_frame()) {
-    __ pop(ebp);
-  }
   frame_access_state()->SetFrameAccessToDefault();
 }
 
@@ -350,6 +347,9 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ sub(esp, Immediate(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
+  if (frame()->needs_frame()) {
+    __ mov(ebp, MemOperand(ebp, 0));
+  }
   frame_access_state()->SetFrameAccessToSP();
 }
 

src/compiler/instruction-selector.cc

@@ -507,12 +507,11 @@ void InstructionSelector::InitializeCallBuffer(Node* call, CallBuffer* buffer,
   }
   DCHECK_EQ(input_count, buffer->instruction_args.size() + pushed_count -
                              buffer->frame_state_value_count());
-  if (call_tail && stack_param_delta != 0) {
-    // For tail calls that change the size of their parameter list, move the
-    // saved caller return address, parent frame pointer and parent constant
-    // pool pointer to just above the parameters.
-
-    // Return address
+  if (V8_TARGET_ARCH_STORES_RETURN_ADDRESS_ON_STACK && call_tail &&
+      stack_param_delta != 0) {
+    // For tail calls that change the size of their parameter list and keep
+    // their return address on the stack, move the return address to just above
+    // the parameters.
     LinkageLocation saved_return_location =
         LinkageLocation::ForSavedCallerReturnAddress();
     InstructionOperand return_address =
@@ -520,26 +519,6 @@ void InstructionSelector::InitializeCallBuffer(Node* call, CallBuffer* buffer,
                                  saved_return_location, stack_param_delta),
                              saved_return_location);
     buffer->instruction_args.push_back(return_address);
-
-    // Parent frame pointer
-    LinkageLocation saved_frame_location =
-        LinkageLocation::ForSavedCallerFramePtr();
-    InstructionOperand saved_frame =
-        g.UsePointerLocation(LinkageLocation::ConvertToTailCallerLocation(
-                                 saved_frame_location, stack_param_delta),
-                             saved_frame_location);
-    buffer->instruction_args.push_back(saved_frame);
-
-    if (V8_EMBEDDED_CONSTANT_POOL) {
-      // Constant pool pointer
-      LinkageLocation saved_cp_location =
-          LinkageLocation::ForSavedCallerConstantPool();
-      InstructionOperand saved_cp =
-          g.UsePointerLocation(LinkageLocation::ConvertToTailCallerLocation(
-                                   saved_cp_location, stack_param_delta),
-                               saved_cp_location);
-      buffer->instruction_args.push_back(saved_cp);
-    }
   }
 }
 

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

@@ -467,9 +467,6 @@ void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   if (sp_slot_delta > 0) {
     __ addiu(sp, sp, sp_slot_delta * kPointerSize);
   }
-  if (frame()->needs_frame()) {
-    __ Pop(ra, fp);
-  }
   frame_access_state()->SetFrameAccessToDefault();
 }
 
@@ -480,6 +477,10 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ Subu(sp, sp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
+  if (frame()->needs_frame()) {
+    __ lw(ra, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
+    __ lw(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
+  }
   frame_access_state()->SetFrameAccessToSP();
 }
 

src/compiler/mips64/code-generator-mips64.cc

@@ -466,9 +466,6 @@ void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   if (sp_slot_delta > 0) {
     __ daddiu(sp, sp, sp_slot_delta * kPointerSize);
   }
-  if (frame()->needs_frame()) {
-    __ Pop(ra, fp);
-  }
   frame_access_state()->SetFrameAccessToDefault();
 }
 
@@ -479,6 +476,10 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ Dsubu(sp, sp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
+  if (frame()->needs_frame()) {
+    __ ld(ra, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
+    __ ld(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
+  }
   frame_access_state()->SetFrameAccessToSP();
 }
 

src/compiler/x64/code-generator-x64.cc

@@ -578,9 +578,6 @@ void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   if (sp_slot_delta > 0) {
     __ addq(rsp, Immediate(sp_slot_delta * kPointerSize));
   }
-  if (frame()->needs_frame()) {
-    __ popq(rbp);
-  }
   frame_access_state()->SetFrameAccessToDefault();
 }
 
@@ -591,6 +588,9 @@ void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
     __ subq(rsp, Immediate(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
+  if (frame()->needs_frame()) {
+    __ movq(rbp, MemOperand(rbp, 0));
+  }
   frame_access_state()->SetFrameAccessToSP();
 }