On 3/19/24 23:56, Richard Henderson wrote:
On 3/19/24 03:32, Pierrick Bouvier wrote:
static void plugin_gen_inject(struct qemu_plugin_tb *plugin_tb)
{
- TCGOp *op;
+ TCGOp *op, *next;
int insn_idx = -1;
pr_ops();
- QTAILQ_FOREACH(op, &tcg_ctx->ops, link) {
+ /*
+ * While injecting code, we cannot afford to reuse any ebb temps
+ * that might be live within the existing opcode stream.
+ * The simplest solution is to release them all and create new.
+ */
+ memset(tcg_ctx->free_temps, 0, sizeof(tcg_ctx->free_temps));
+
Not an expert at this, but wouldn't that break an existing TB that already has
some ops on
those temps?
No, this only affects allocation of new temps -- if free_temps is empty, a new
temp will
be allocated from tcg_ctx->nb_temps++.
Zeroing free_temps here ensures that we *do not* reuse a temp that might
already be live
across any plugin insertion point. Between insertion points, we will free
plugin temps
and only reuse those.
Thanks, by looking at tcg_temp_new_internal fn, and with your
explaination, it makes more sense.
r~
