ARM64: optimize call immediate

BUG=
R=ulan@chromium.org

Review URL: https://codereview.chromium.org/209923002

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@20207 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/arm64/assembler-arm64-inl.h

@@ -589,20 +589,19 @@ Address Assembler::return_address_from_call_start(Address pc) {
   // sequences:
   //
   // Without relocation:
-  //  movz  ip0, #(target & 0x000000000000ffff)
-  //  movk  ip0, #(target & 0x00000000ffff0000)
-  //  movk  ip0, #(target & 0x0000ffff00000000)
-  //  movk  ip0, #(target & 0xffff000000000000)
-  //  blr   ip0
+  //  movz  temp, #(target & 0x000000000000ffff)
+  //  movk  temp, #(target & 0x00000000ffff0000)
+  //  movk  temp, #(target & 0x0000ffff00000000)
+  //  blr   temp
   //
   // With relocation:
-  //  ldr   ip0, =target
-  //  blr   ip0
+  //  ldr   temp, =target
+  //  blr   temp
   //
   // The return address is immediately after the blr instruction in both cases,
   // so it can be found by adding the call size to the address at the start of
   // the call sequence.
-  STATIC_ASSERT(Assembler::kCallSizeWithoutRelocation == 5 * kInstructionSize);
+  STATIC_ASSERT(Assembler::kCallSizeWithoutRelocation == 4 * kInstructionSize);
   STATIC_ASSERT(Assembler::kCallSizeWithRelocation == 2 * kInstructionSize);
 
   Instruction* instr = reinterpret_cast<Instruction*>(pc);
@@ -610,8 +609,7 @@ Address Assembler::return_address_from_call_start(Address pc) {
     // Verify the instruction sequence.
     ASSERT(instr->following(1)->IsMovk());
     ASSERT(instr->following(2)->IsMovk());
-    ASSERT(instr->following(3)->IsMovk());
-    ASSERT(instr->following(4)->IsBranchAndLinkToRegister());
+    ASSERT(instr->following(3)->IsBranchAndLinkToRegister());
     return pc + Assembler::kCallSizeWithoutRelocation;
   } else {
     // Verify the instruction sequence.

src/arm64/assembler-arm64.h

@@ -818,16 +818,15 @@ class Assembler : public AssemblerBase {
   // as it will choose the correct value for a given relocation mode.
   //
   // Without relocation:
-  //  movz  ip0, #(target & 0x000000000000ffff)
-  //  movk  ip0, #(target & 0x00000000ffff0000)
-  //  movk  ip0, #(target & 0x0000ffff00000000)
-  //  movk  ip0, #(target & 0xffff000000000000)
-  //  blr   ip0
+  //  movz  temp, #(target & 0x000000000000ffff)
+  //  movk  temp, #(target & 0x00000000ffff0000)
+  //  movk  temp, #(target & 0x0000ffff00000000)
+  //  blr   temp
   //
   // With relocation:
-  //  ldr   ip0, =target
-  //  blr   ip0
-  static const int kCallSizeWithoutRelocation = 5 * kInstructionSize;
+  //  ldr   temp, =target
+  //  blr   temp
+  static const int kCallSizeWithoutRelocation = 4 * kInstructionSize;
   static const int kCallSizeWithRelocation = 2 * kInstructionSize;
 
   // Size of the generated code in bytes

src/arm64/macro-assembler-arm64.cc

@@ -2024,11 +2024,13 @@ void MacroAssembler::Call(Address target, RelocInfo::Mode rmode) {
   Register temp = temps.AcquireX();
 
   if (rmode == RelocInfo::NONE64) {
+    // Addresses are 48 bits so we never need to load the upper 16 bits.
     uint64_t imm = reinterpret_cast<uint64_t>(target);
+    // If we don't use ARM tagged addresses, the 16 higher bits must be 0.
+    ASSERT(((imm >> 48) & 0xffff) == 0);
     movz(temp, (imm >> 0) & 0xffff, 0);
     movk(temp, (imm >> 16) & 0xffff, 16);
     movk(temp, (imm >> 32) & 0xffff, 32);
-    movk(temp, (imm >> 48) & 0xffff, 48);
   } else {
     LoadRelocated(temp, Operand(reinterpret_cast<intptr_t>(target), rmode));
   }