RE: [PATCH] tcg: `reachable_code_pass()` remove empty else-branch

[Taylor Simpson]

> -----Original Message-----
> From: Anton Johansson <anjo@rev.ng>
> Sent: Wednesday, March 1, 2023 7:22 AM
> To: qemu-devel@nongnu.org
> Cc: ale@rev.ng; richard.henderson@linaro.org; Taylor Simpson
> <tsimpson@quicinc.com>
> Subject: [PATCH] tcg: `reachable_code_pass()` remove empty else-branch
> 
> This patch extends reachable_code_pass() to also deal with empty else-
> branches of the form
> 
>   br $L0
>   set_label $L1
>   set_label $L0
> 
> converting them to
> 
>   set_label $L1
> 
> when $L0 is only referenced by the br op.  This type of empty-else branch will
> be emitted by idef-parser in the Hexagon frontend once CANCEL statements
> have been ignored.
> 
> Signed-off-by: Anton Johansson <anjo@rev.ng>
> ---
>  tcg/tcg.c | 41 ++++++++++++++++++++++++++++++-----------
>  1 file changed, 30 insertions(+), 11 deletions(-)
> 
> diff --git a/tcg/tcg.c b/tcg/tcg.c
> index a4a3da6804..531bc74231 100644
> --- a/tcg/tcg.c
> +++ b/tcg/tcg.c
> @@ -2664,21 +2664,40 @@ static void reachable_code_pass(TCGContext *s)
>                  dead = false;
>                  remove = false;
> 
> -                /*
> -                 * Optimization can fold conditional branches to 
> unconditional.
> -                 * If we find a label with one reference which is preceded by
> -                 * an unconditional branch to it, remove both.  This needed 
> to
> -                 * wait until the dead code in between them was removed.
> -                 */
> -                if (label->refs == 1) {
> -                    TCGOp *op_prev = QTAILQ_PREV(op, link);

Can't we just insert a while loop here to move op_prev back across labels?

    while (op_next->opc == INDEX_op_set_label) {
        op_prev = QTAILQ_PREV(op, op_prev);
    }

> -                    if (op_prev->opc == INDEX_op_br &&
> -                        label == arg_label(op_prev->args[0])) {

Also, here is the patch that exposes the need for this optimization
https://lists.gnu.org/archive/html/qemu-devel/2023-01/msg07236.html

Thanks,
Taylor