target-i386: fix cmpxchg instruction emulation

When the i386 cmpxchg instruction is executed with a memory operand
and the comparison result is "unequal", do the memory write before
changing the accumulator instead of the other way around, because
otherwise the new accumulator value will incorrectly be used in the
comparison when the instruction is restarted after a page fault.

This bug was originally reported on 2010-04-25 as
https://bugs.launchpad.net/qemu/+bug/569760Signed-off-by: NAndreas Gustafsson <gson@gson.org>

target-i386/translate.c

@@ -4870,20 +4870,23 @@ static target_ulong disas_insn(DisasContext *s, target_ulong pc_start)
             tcg_gen_sub_tl(t2, cpu_regs[R_EAX], t0);
             gen_extu(ot, t2);
             tcg_gen_brcondi_tl(TCG_COND_EQ, t2, 0, label1);
+            label2 = gen_new_label();
             if (mod == 3) {
-                label2 = gen_new_label();
                 gen_op_mov_reg_v(ot, R_EAX, t0);
                 tcg_gen_br(label2);
                 gen_set_label(label1);
                 gen_op_mov_reg_v(ot, rm, t1);
-                gen_set_label(label2);
             } else {
-                tcg_gen_mov_tl(t1, t0);
+                /* perform no-op store cycle like physical cpu; must be
+                   before changing accumulator to ensure idempotency if
+                   the store faults and the instruction is restarted */
+                gen_op_st_v(ot + s->mem_index, t0, a0);
                 gen_op_mov_reg_v(ot, R_EAX, t0);
+                tcg_gen_br(label2);
                 gen_set_label(label1);
-                /* always store */
                 gen_op_st_v(ot + s->mem_index, t1, a0);
             }
+            gen_set_label(label2);
             tcg_gen_mov_tl(cpu_cc_src, t0);
             tcg_gen_mov_tl(cpu_cc_dst, t2);
             s->cc_op = CC_OP_SUBB + ot;