Modify 'movq reg, reg' encoding in 64-bit code.

runtime/vm/assembler_x64.cc

Index: runtime/vm/assembler_x64.cc
===================================================================
--- runtime/vm/assembler_x64.cc	(revision 6416)
+++ runtime/vm/assembler_x64.cc	(working copy)
@@ -278,12 +278,15 @@
 }
 
 
+// Use 0x89 encoding (instead of 0x8B encoding), which is expected by gdb64
+// older than 7.3.1-gg5 when disassembling a function's prolog (movq rbp, rsp)
+// for proper unwinding of Dart frames (use --generate_gdb_symbols and -O0).
 void Assembler::movq(Register dst, Register src) {
   AssemblerBuffer::EnsureCapacity ensured(&buffer_);
-  Operand operand(src);
-  EmitOperandREX(dst, operand, REX_W);
-  EmitUint8(0x8B);
-  EmitOperand(dst & 7, operand);
+  Operand operand(dst);
+  EmitOperandREX(src, operand, REX_W);
+  EmitUint8(0x89);

[Lasse Reichstein Nielsen, 2012/04/12 09:57:38]

You could decide whether to use 0x89 or 0x8b depending on whether src or dst is
rsp (or r12), i.e., use the encoding that doesn't require an extra SIP byte if
possible. That would give you 0x89 for 'movq rbp, rsp' and also save a byte for
'mov rsp, rbp'.
(The same trick can be done for most operations that have both r/m,r and r,r/m
addressing modes).

[regis, 2012/04/12 16:46:16]

This is a good point for memory addressing modes, but I do not think we ever
need a SIB byte in movq reg, reg.

+  EmitOperand(src & 7, operand);
 }