> This patch disables propagation of ipcp information into lto partitions
> where all instances of the node are marked to be inlined.
>
> Motivation:
> Incremental LTO needs stable values between compilations to be
> effective. This requirement fails with following example:
>
> void heavily_used_debug_function(int);
> ...
> heavily_used_debug_function(__LINE__);
>
> Ipcp creates long list of all __LINE__ arguments, and then
> propagates it with every function clone, even though for inlined
> functions this information is not useful.
>
> gcc/ChangeLog:
>
> * ipa-prop.cc (write_ipcp_transformation_info): Disable
> uneeded value propagation.
> * lto-cgraph.cc (lto_symtab_encoder_encode): Default values.
> (lto_symtab_encoder_always_inlined_p): New.
> (lto_set_symtab_encoder_not_always_inlined): New.
> (add_node_to): Set always inlined.
> * lto-streamer.h (struct lto_encoder_entry): New field.
> (lto_symtab_encoder_always_inlined_p): New.
OK
> ---
> gcc/ipa-prop.cc | 12 +++++++++---
> gcc/lto-cgraph.cc | 41 +++++++++++++++++++++++++++++++++++++----
> gcc/lto-streamer.h | 4 ++++
> 3 files changed, 50 insertions(+), 7 deletions(-)
>
> diff --git a/gcc/ipa-prop.cc b/gcc/ipa-prop.cc
> index 032358fde22..ef83ce3edb6 100644
> --- a/gcc/ipa-prop.cc
> +++ b/gcc/ipa-prop.cc
> @@ -5404,9 +5404,15 @@ write_ipcp_transformation_info (output_block *ob,
> cgraph_node *node,
> streamer_write_bitpack (&bp);
> }
>
> - streamer_write_uhwi (ob, vec_safe_length (ts->m_vr));
> - for (const ipa_vr &parm_vr : ts->m_vr)
> - parm_vr.streamer_write (ob);
> + /* If all instances of this node are inlined, ipcp info is not useful. */
> + if (!lto_symtab_encoder_always_inlined_p (encoder, node))
> + {
> + streamer_write_uhwi (ob, vec_safe_length (ts->m_vr));
> + for (const ipa_vr &parm_vr : ts->m_vr)
> + parm_vr.streamer_write (ob);
> + }
> + else
> + streamer_write_uhwi (ob, 0);
> }
>
> if (!encoder->map)
> {
> - lto_encoder_entry entry = {node, false, false, false};
> + lto_encoder_entry entry = {node, false, false, true, false};
>
> ref = encoder->nodes.length ();
> encoder->nodes.safe_push (entry);
> @@ -124,7 +124,7 @@ lto_symtab_encoder_encode (lto_symtab_encoder_t encoder,
> size_t *slot = encoder->map->get (node);
> if (!slot || !*slot)
> {
> - lto_encoder_entry entry = {node, false, false, false};
> + lto_encoder_entry entry = {node, false, false, true, false};
I wonder if we should enjoy C++ and add constructor for the entry :)
> +bool
> +lto_symtab_encoder_always_inlined_p (lto_symtab_encoder_t encoder,
> + struct cgraph_node *node)
always_inlined kind of triggers connection to always_inline attribute.
Perhaps the flag can be called only_for_inlining
to avoid confussion...
> +{
> + int index = lto_symtab_encoder_lookup (encoder, node);
> + return encoder->nodes[index].always_inlined;
> +}
> +
> +/* Specify that the NODE and its clones are not always inlined. */
> +
> +static void
> +lto_set_symtab_encoder_not_always_inlined (lto_symtab_encoder_t encoder,
> + struct cgraph_node *node)
> +{
> + int index = lto_symtab_encoder_encode (encoder, node);
> + gcc_checking_assert (encoder->nodes[index].node == node);
> + encoder->nodes[index].always_inlined = false;
> +}
> +
> /* Return TRUE if we should encode initializer of NODE (if any). */
>
> bool
> @@ -799,15 +820,27 @@ output_refs (lto_symtab_encoder_t encoder)
>
> static void
> add_node_to (lto_symtab_encoder_t encoder, struct cgraph_node *node,
> - bool include_body)
> + bool include_body, bool not_inlined)
> {
> if (node->clone_of)
> - add_node_to (encoder, node->clone_of, include_body);
> + add_node_to (encoder, node->clone_of, include_body, not_inlined);
> if (include_body)
> lto_set_symtab_encoder_encode_body (encoder, node);
> + if (not_inlined)
> + lto_set_symtab_encoder_not_always_inlined (encoder, node);
> lto_symtab_encoder_encode (encoder, node);
This is probably originaly done to me, but this does unnecesary multiple
hash lookups. Perhaps the flags (INLUDE_BODY, NOT_INLINED) can be
passed lto_symtab_encoder_encode?
OK whith those changes. The patch is good idea regardless caching,
since lto summaries can trigger extra unnecesary streaming.
Thanks,
Honza
