A64: Add support for --optimize-for-size in PushMultipleTimes

src/a64/macro-assembler-a64.cc

Index: src/a64/macro-assembler-a64.cc
diff --git a/src/a64/macro-assembler-a64.cc b/src/a64/macro-assembler-a64.cc
index 3d2d24e9e9277c2be15b3099de9e7cdca94e53df..9b600bc195be0d9d8505565db5ba36ad02003afb 100644
--- a/src/a64/macro-assembler-a64.cc
+++ b/src/a64/macro-assembler-a64.cc
@@ -672,10 +672,18 @@ void MacroAssembler::PopCPURegList(CPURegList registers) {
 void MacroAssembler::PushMultipleTimes(int count, Register src) {
   int size = src.SizeInBytes();
 
-  // TODO(all): Use a loop when optimizing for size.
-  TODO_UNIMPLEMENTED("PushMultipleTimes: Support --optimize-for-size.");
-
   PrepareForPush(count, size);
+
+  if (FLAG_optimize_for_size && count > 4) {

[Rodolph Perfetta (ARM), 2014/02/06 15:14:36]

The code below is 4 instructions long, PushMultipleTimes can handle 8 push in 4
instructions (using stp) so the break even point should be 8.

+    Label loop;
+    __ Mov(Tmp0(), count);
+    __ Bind(&loop);
+    PushHelper(1, size, src, NoReg, NoReg, NoReg);

[ulan, 2014/02/06 14:53:12]

Optimization: push 4 values at once (count / 4) times, and then set count =
(count % 4) and fall through to the old code.

[Rodolph Perfetta (ARM), 2014/02/06 15:14:36]

if count is even you can push two at a time with no extra memory cost.

[Rodolph Perfetta (ARM), 2014/02/06 15:14:36]

4 at a time will be faster but bigger, is that what we want?

On 2014/02/06 14:53:12, ulan wrote:

+    __ Subs(Tmp0(), Tmp0(), 1);
+    __ B(ne, &loop);
+    return;
+  }
+
   // Push up to four registers at a time if possible because if the current
   // stack pointer is csp and the register size is 32, registers must be pushed
   // in blocks of four in order to maintain the 16-byte alignment for csp.