x86 GenLongRegOrMemOp should be aware about xmm to core reg conversion

GenLongRegOrMemOp function expects arithmetic on core regs but does not
ensure that operand in core reg.

The patch adds the conversion if it is required.

Change-Id: Ibb6928b8cc2c63ede1a20d6ee45d9a64884231b6
Signed-off-by: Serguei Katkov <serguei.i.katkov@intel.com>

diff --git a/compiler/dex/quick/x86/int_x86.cc b/compiler/dex/quick/x86/int_x86.cc
index dcbaad9..1fe0af9 100644 (file)
--- a/compiler/dex/quick/x86/int_x86.cc
+++ b/compiler/dex/quick/x86/int_x86.cc
@@ -1091,6 +1091,8 @@ void X86Mir2Lir::GenLongRegOrMemOp(RegLocation rl_dest, RegLocation rl_src,
   X86OpCode x86op = GetOpcode(op, rl_dest, rl_src, false);
   if (rl_src.location == kLocPhysReg) {
     // Both operands are in registers.
+    // But we must ensure that rl_src is in pair
+    rl_src = EvalLocWide(rl_src, kCoreReg, true);
     if (rl_dest.reg.GetReg() == rl_src.reg.GetHighReg()) {
       // The registers are the same, so we would clobber it before the use.
       int temp_reg = AllocTemp();