Hello, I have been working a bit more on the IPA-SRA reimplementation that I have previously posted here in:
https://gcc.gnu.org/ml/gcc-patches/2019-05/msg00472.html This reimplementation is a full IPA pass that can handle strictly connected components in the call-graph, can take advantage of LTO and does not weirdly switch functions in pass-pipeline like our current quasi-IPA SRA does. Unlike the current IPA-SRA it can also remove return values, even in SCCs. On the other hand, it is less powerful when it comes to structures passed by reference. By design it will not create references to bits of an aggregate because that turned out to be just obfuscation in practice. However, it also cannot usually split aggregates passed by reference that are just passed to another function (where splitting would be useful) because it cannot perform the same TBAA analysis like the current implementation which already knows what types it should look at because it has access to bodies of all functions attempts to modify (for an example see https://gcc.gnu.org/ml/gcc-patches/2019-05/msg00955.html). Since May I have switched the order of IPA-CP and IPA-SRA, so that IPA-CP is run before IPA-SRA and it can propagate constants in aggregates before these are split and tracking is lost. This has uncovered two small bugs, as did LTO-building Mozilla Firefox with the patch and another issue was discovered when bootstrapping on ppc64le. All of these are fixed. I have also added two dbgcounters, one for processing parameters and another for return values removal. The patch has successfully bootstrapped, LTO-bootstrapped and LTO-profilebootstrapped on x86_64-linux. Testing on x86_64 fixes 24 LTO tests in guality/pr68860-1.c but regresses 13 guality tests, 3 of which are not LTO. Ther patch also passed bootstrap and testing on aarch64-linux and ppc64le-linux. When building the big libxul library of Mozilla Firefox with both LTO and profile feedback, the patch has modified 21093 functions out of 259745 it has seen, so slightly over 8%. In the process, it has removed 6693 unused return values and 1538 unused parameters and split 7126 parameters into 8394 smaller components. Note that IPA-CP can also remove unused scalar parameters in simple situations, and since it now runs first, the 1538 are only those which IPA-CP left behind for one reason or another. The built Firefux could browse the Amazon web site for a while without any apparent problem. After the patches are approved, I would like to commit them in one big commit but in order to ease review I have split the big patch into four smaller (but still rather big) ones: 1. Removal of old IPA-SRA, 2. New parameter manipulation infrastructure, 3. New IPA-SRA implementation, and 4. Modifications to the testsuite. When applied one after another they should compile but I have only exhaustively tested the complete patch. If you want to experiment with new IPA-SRA or the parameter manipulation infrastructure it introduces, I suggest that you clone the jamborm/ipa-sra branch from our git "mirror" (https://gcc.gnu.org/git/?p=gcc.git;a=shortlog;h=refs/heads/jamborm/ipa-sra). Thanks in advance for any questions, comments and suggestions, Martin 2019-07-23 Martin Jambor <mjam...@suse.cz> * coretypes.h (cgraph_edge): Declare. * ipa-param-manipulation.c: Rewrite. * ipa-param-manipulation.h: Likewise. * Makefile.in (GTFILES): Added ipa-param-manipulation.h and ipa-sra.c. (OBJS): Added ipa-sra.o. * cgraph.h (ipa_replace_map): Removed fields old_tree, replace_p and ref_p, added fields param_adjustments and performed_splits. (struct cgraph_clone_info): Remove ags_to_skip and combined_args_to_skip, new field param_adjustments. (cgraph_node::create_clone): Changed parameters to use ipa_param_adjustments. (cgraph_node::create_virtual_clone): Likewise. (cgraph_node::create_virtual_clone_with_body): Likewise. (tree_function_versioning): Likewise. (cgraph_build_function_type_skip_args): Removed. * cgraph.c (cgraph_edge::redirect_call_stmt_to_callee): Convert to using ipa_param_adjustments. (clone_of_p): Likewise. * cgraphclones.c (cgraph_build_function_type_skip_args): Removed. (build_function_decl_skip_args): Likewise. (duplicate_thunk_for_node): Adjust parameters using ipa_param_body_adjustments, copy param_adjustments instead of args_to_skip. (cgraph_node::create_clone): Convert to using ipa_param_adjustments. (cgraph_node::create_virtual_clone): Likewise. (cgraph_node::create_version_clone_with_body): Likewise. (cgraph_materialize_clone): Likewise. (symbol_table::materialize_all_clones): Likewise. * ipa-fnsummary.c (ipa_fn_summary_t::duplicate): Simplify ipa_replace_map check. * ipa-cp.c (get_replacement_map): Do not initialize removed fields. (initialize_node_lattices): Make aware that some parameters might have already been removed. (want_remove_some_param_p): New function. (create_specialized_node): Convert to using ipa_param_adjustments and deal with possibly pre-existing adjustments. * lto-cgraph.c (output_cgraph_opt_summary_p): Likewise. (output_node_opt_summary): Do not stream removed fields. Stream parameter adjustments instead of argumetns to skip. (input_node_opt_summary): Likewise. (input_node_opt_summary): Likewise. * lto-section-in.c (lto_section_name): Added ipa-sra section. * lto-streamer.h (lto_section_type): Likewise. * tree-inline.h (copy_body_data): New fields killed_new_ssa_names and param_body_adjs. (copy_decl_to_var): Declare. * tree-inline.c (update_clone_info): Do not remap old_tree. (remap_gimple_stmt): Use ipa_param_body_adjustments to modify gimple statements, walk all extra generated statements and remap their operands. (redirect_all_calls): Add killed SSA names to a hash set. (remap_ssa_name): Do not remap killed SSA names. (copy_arguments_for_versioning): Renames to copy_arguments_nochange, half of functionality moved to ipa_param_body_adjustments. (copy_decl_to_var): Make exported. (copy_body): Destroy killed_new_ssa_names hash set. (expand_call_inline): Remap performed splits. (update_clone_info): Likewise. (tree_function_versioning): Simplify tree_map processing. Updated to accept ipa_param_adjustments and use ipa_param_body_adjustments. * omp-simd-clone.c (simd_clone_vector_of_formal_parm_types): Adjust for the new interface. (simd_clone_clauses_extract): Likewise, make args an auto_vec. (simd_clone_compute_base_data_type): Likewise. (simd_clone_init_simd_arrays): Adjust for the new interface. (simd_clone_adjust_argument_types): Likewise. (struct modify_stmt_info): Likewise. (ipa_simd_modify_stmt_ops): Likewise. (ipa_simd_modify_function_body): Likewise. (simd_clone_adjust): Likewise. * tree-sra.c: Removed IPA-SRA. Include tree-sra.h. (type_internals_preclude_sra_p): Make public. * tree-sra.h: New file. * ipa-inline-transform.c (save_inline_function_body): Update to refelct new tree_function_versioning signature. * ipa-prop.c (adjust_agg_replacement_values): Use a helper from ipa_param_adjustments to get current parameter indices. (ipcp_modif_dom_walker::before_dom_children): Likewise. (ipcp_update_bits): Likewise. (ipcp_update_vr): Likewise. * ipa-split.c (split_function): Convert to using ipa_param_adjustments. * ipa-sra.c: New file. * multiple_target.c (create_target_clone): Update to reflet new type of create_version_clone_with_body. * trans-mem.c (ipa_tm_create_version): Update to reflect new type of tree_function_versioning. (modify_function): Update to reflect new type of tree_function_versioning. * params.def (PARAM_IPA_SRA_MAX_REPLACEMENTS): New. * passes.def: Remove old IPA-SRA and add new one. * tree-pass.h (make_pass_early_ipa_sra): Remove declaration. (make_pass_ipa_sra): Declare. * dbgcnt.def: Remove eipa_sra. Added ipa_sra_params and ipa_sra_retvalues. * doc/invoke.texi (ipa-sra-max-replacements): New. testsuite/ * g++.dg/ipa/pr81248.C: Adjust dg-options and dump-scan. * gcc.dg/ipa/ipa-sra-1.c: Likewise. * gcc.dg/ipa/ipa-sra-10.c: Likewise. * gcc.dg/ipa/ipa-sra-11.c: Likewise. * gcc.dg/ipa/ipa-sra-3.c: Likewise. * gcc.dg/ipa/ipa-sra-4.c: Likewise. * gcc.dg/ipa/ipa-sra-5.c: Likewise. * gcc.dg/ipa/ipacost-2.c: Disable ipa-sra. * gcc.dg/ipa/ipcp-agg-9.c: Likewise. * gcc.dg/ipa/pr78121.c: Adjust scan pattern. * gcc.dg/ipa/vrp1.c: Likewise. * gcc.dg/ipa/vrp2.c: Likewise. * gcc.dg/ipa/vrp3.c: Likewise. * gcc.dg/ipa/vrp7.c: Likewise. * gcc.dg/ipa/vrp8.c: Likewise. * gcc.dg/noreorder.c: use noipa attribute instead of noinline. * gcc.dg/ipa/20040703-wpa.c: New test. * gcc.dg/ipa/ipa-sra-12.c: New test. * gcc.dg/ipa/ipa-sra-13.c: Likewise. * gcc.dg/ipa/ipa-sra-14.c: Likewise. * gcc.dg/ipa/ipa-sra-15.c: Likewise. * gcc.dg/ipa/ipa-sra-16.c: Likewise. * gcc.dg/ipa/ipa-sra-17.c: Likewise. * gcc.dg/ipa/ipa-sra-18.c: Likewise. * gcc.dg/ipa/ipa-sra-19.c: Likewise. * gcc.dg/ipa/ipa-sra-20.c: Likewise. * gcc.dg/ipa/ipa-sra-21.c: Likewise. * gcc.dg/ipa/ipa-sra-22.c: Likewise. * gcc.dg/sso/ipa-sra-1.c: Likewise. * g++.dg/ipa/ipa-sra-2.C: Likewise. * g++.dg/ipa/ipa-sra-3.C: Likewise. * gcc.dg/tree-ssa/ipa-cp-1.c: Make return value used. * g++.dg/ipa/devirt-19.C: Add missing return, add -fipa-cp-clone option. * g++.dg/lto/devirt-19_0.C: Add -fipa-cp-clone option. * gcc.dg/ipa/ipa-sra-2.c: Mark to be removed. * gcc.dg/ipa/ipa-sra-6.c: Likewise.
