On 05/12/2015 05:53 AM, Richard Henderson wrote:
On 05/07/2015 06:12 PM, Alexander Graf wrote:
+static ExitStatus op_laa(DisasContext *s, DisasOps *o)
+{
+ TCGv_i64 m2 = tcg_temp_new_i64();
+
+ /* XXX should be atomic */
+ tcg_gen_qemu_ld32s(m2, o->in2, get_mem_index(s));
+
+ /* Set r1 to the unmodified contents of m2 */
+ tcg_gen_mov_i64(o->out, m2);
+
+ /* m2 = r3 + m2 */
+ tcg_gen_add_i64(m2, o->in1, m2);
+ tcg_gen_qemu_st32(m2, o->in2, get_mem_index(s));
+
+ /* Set in2 to the input operand for cc calculation */
+ tcg_gen_mov_i64(o->in2, o->out);
+
+ tcg_temp_free_i64(m2);
+ return NO_EXIT;
+}
I don't believe that this computes the right cc.
You need to place m2 into in2, not out.
That said, maybe
tcg_temp_free_i64(o->in2);
o->in2 = m2;
is the right way to perform that copy. Since we
know that in2 was in2_a2, allocating a temp for an
address, we also know we can both (1) free it and
(2) put another temp in there that will be freed.
Well, alternatively I could try to make out be the memory value at the
end of the op and invent new load/store helpers for atomic instructions.
That way the op bit should shrink significantly and we get the correct
operand in s->out automatically.
Don't worry about the atomic-ness of the operation
until some of the many patch sets going round on
that very subject are resolved. This will be good
enough for system mode for now.
I think maybe the whole target ought to be cleaned
up for the new TCGMemOps. For this, you could share
the code for LAA and LAAG with the TCGMemOp in insn->data.
That's an interesting idea. Let me see what I can pull off.
Alex
