This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.
Index Nav: | [Date Index] [Subject Index] [Author Index] [Thread Index] | |
---|---|---|
Message Nav: | [Date Prev] [Date Next] | [Thread Prev] [Thread Next] |
Other format: | [Raw text] |
Hi, This patch implements basic block SLP, i.e., vectorization of straight-line code sequences out-of-loops (as opposed to the similar capability we already have that exploits such opportunities within a loop iteration, in a loop-aware manner). This first version supports only simple cases of code sequences that end with a group of adjacent stores and contain only aligned and non-aliasing data-refs of same type. It does not include a cost model, since it does not introduce any overheads. Basic block SLP pass is invoked right after the unrolling pass which follows the loop-vectorizer (before auto-par pass), thereby letting the loop-aware SLP have a go first (within the loop-vectorizer), taking advantage of the loop-context if possible. The pass is enabled with -fslp-vectorize flag, which is turned on by default when -ftree-vectorize is set. Data-refs analysis is a reduced data-refs analysis that ignores any evolution and any loop data dependences if exist. Inside the vectorizer the implementation is pretty simple and mostly reuses the existing vectorizer code. Most of the patch is adding a basic block argument to the existing functions. The new parts are in tree-vect-slp.c I need a review for parts that are not in the vectorizer: data refs analysis and pass management. Bootstrapped with vectorization enabled and tested without regressions on powerpc64-suse-linux. On x86_64-suse-linux I got the following failures: 1. gcc.target/x86_64/abi/test_struct_returning.c fails during execution. The problem seems to be in alignment and is not SLP specific. I opened a PR 39907 for it with a test for loop-based vectorization. 2. there are 23 tests that fail with: "error: alignment of array elements is greater than element size" (e.g. gcc.dg/torture/stackalign/nested-2.c and g++.dg/torture/stackalign/eh-alloca-1.C). They all have __attribute__((aligned(64))). The failure occurs when we are trying to vectorize the basic block and get a vector type for the aligned type. I guess, otherwise, we just never get to this check. Hence, these failures are not really related to SLP. What is the best way to handle this? Thanks, Ira ChangeLog: * doc/passes.texi (Tree-SSA passes): Document SLP pass. * tree-pass.h (pass_slp_vectorize): New pass. * timevar.def (TV_TREE_SLP_VECTORIZATION): Define. * tree-vectorizer.c (timevar.h): Include.: (vect_location): Fix comment. (vect_slp_memsyms_to_rename): Likewise. (vect_set_verbosity_level): Update verbosity level only with higher values. (vect_set_dump_settings): Add an argument. Ignore user defined verbosity if dump flags require higher level of verbosity. Print to stderr only for loop vectorization. (vectorize_loops): Update call to vect_set_dump_settings.. (execute_vect_slp): New function. (gate_vect_slp): Likewise. (struct gimple_opt_pass pass_slp_vectorize): New. * tree-vectorizer.h (struct _bb_vec_info): Define along macros to access its members. (vec_info_for_bb): New function. (struct _stmt_vec_info): Add bb_vinfo and a macro for its access. (vect_slp_memsyms_to_rename): Declare. (VECTORIZATION_ENABLED): New macro. (SLP_ENABLED, SLP_DISABLED): Likewise. (vect_is_simple_use): Add bb_vec_info argument. (new_stmt_vec_info, vect_analyze_data_ref_dependences, vect_analyze_data_refs_alignment, vect_verify_datarefs_alignment, vect_analyze_data_ref_accesses, vect_analyze_data_refs, vect_schedule_slp, vect_analyze_slp): Likewise. (vect_analyze_stmt): Add slp_tree argument. (find_bb_location): Declare. (vect_slp_analyze_bb, vect_slp_transform_bb): Likewise. * tree-vect-loop.c (new_loop_vec_info): Adjust function calls. (vect_analyze_loop_operations, vect_analyze_loop, get_initial_def_for_induction, vect_create_epilog_for_reduction,, vect_finalize_reduction, vectorizable_reduction, vectorizable_live_operation, vect_transform_loop): Likewise. * tree-data-ref.c (dr_analyze_innermost): Update comment, skip evolution analysis if analyzing a basic block. (dr_analyze_indices): Likewise. (initialize_data_dependence_relation): Skip the test whether the, object is invariant for basic blocks.: (compute_all_dependences): Skip dependence analysis for data references in basic blocks. (find_data_references_in_stmt): Don't fail in case of invariant access in basic block. (find_data_references_in_bb): New function. (find_data_references_in_loop): Move code to find_data_references_in_bb and add a call to it. (compute_data_dependences_for_bb): New function. * tree-data-ref.h (compute_data_dependences_for_bb): Declare. * tree-vect-data-refs.c (vect_check_interleaving): Adjust to the case that STEP is 0. (vect_analyze_data_ref_dependence): Check for interleaving in case of unknown dependence in basic block and fail in case of dependence in basic block. (vect_analyze_data_ref_dependences): Add bb_vinfo argument, get data dependence instances from either loop or basic block vectorization info. (vect_compute_data_ref_alignment): Check if it is loop vectorization before calling nested_in_vect_loop_p. (vect_compute_data_refs_alignment): Add bb_vinfo argument, get data dependence instances from either loop or basic block vectorization info. (vect_verify_datarefs_alignment): Likewise. (vect_enhance_data_refs_alignment): Adjust function calls. (vect_analyze_data_refs_alignment): Likewise. (vect_analyze_group_access): Fix printing. Skip different checks if DR_STEP is 0. Keep strided stores either in loop or basic block vectorization data structure. (vect_analyze_data_ref_access): Fix comments, allow zero step in basic blocks. (vect_analyze_data_ref_accesses): Add bb_vinfo argument, get data dependence instances from either loop or basic block vectorization info. (vect_analyze_data_refs): Update comment. Call compute_data_dependences_for_bb to analyze basic blocks. (vect_create_addr_base_for_vector_ref): Check for outer loop only in case of loop vectorization. In case of basic block vectorization use data-ref itself as a base. (vect_create_data_ref_ptr): In case of basic block vectorization: don't advance the pointer, add new statements before the current statement. Adjust function calls. (vect_supportable_dr_alignment): Support only aligned accesses in basic block vectorization. * common.opt (fslp-vectorize): New flag. * tree-vect-patterns.c (widened_name_p): Adjust function calls. (vect_pattern_recog_1): Likewise. * tree-vect-stmts.c (process_use): Likewise. (vect_init_vector): Add new statements in the beginning of the basic block in case of basic block SLP. (vect_get_vec_def_for_operand): Adjust function calls. (vect_finish_stmt_generation): Likewise. (vectorizable_call): Add assert that it is loop vectorization, adjust function calls. (vectorizable_conversion, vectorizable_assignment): Likewise., (vectorizable_operation): In case of basic block SLP, take vectorization factor from statement's type and skip the relevance check. Adjust function calls. (vectorizable_type_demotion): Add assert that it is loop vectorization, adjust function calls. (vectorizable_type_promotion): Likewise. (vectorizable_store): Check for outer loop only in case of loop vectorization. Adjust function calls. For basic blocks, skip the relevance check and don't advance pointers. (vectorizable_load): Likewise. (vectorizable_condition): Add assert that it is loop vectorization, adjust function calls. (vect_analyze_stmt): Add argument. In case of basic block SLP, check that it is not reduction, get vector type, call only supported functions, skip loop specific parts. (vect_transform_stmt): Check for outer loop only in case of loop vectorization. (new_stmt_vec_info): Add new argument and initialize bb_vinfo. (vect_is_simple_use): Fix comment, add new argument, fix conditions for external definition.: * passes.c (pass_slp_vectorize): New pass. * tree-vect-slp.c (find_bb_location): New function., (vect_get_and_check_slp_defs): Add argument, adjust function calls, check for patterns only in loops. (vect_build_slp_tree): Add argument, adjust function calls, fail in case of multiple types in basic block SLP. (vect_mark_slp_stmts_relevant): New function. (vect_supported_load_permutation_p): Fix comment. (vect_analyze_slp_instance): Add argument. In case of basic block SLP, take vectorization factor from statement's type, check that unrolling factor is 1. Adjust function call. Save SLP instance in either loop or basic block vectorization structure. Return FALSE, if SLP failed. (vect_analyze_slp): Add argument. Get strided stores groups from either loop or basic block vectorization structure. Return FALSE if basic block SLP failed. (new_bb_vec_info): New function. (destroy_bb_vec_info, vect_slp_analyze_node_operations, vect_slp_analyze_operations, vect_slp_analyze_bb): Likewise. (vect_schedule_slp): Add argument. Get SLP instances from either loop or basic block vectorization structure. Set vectorization factor to be 1 for basic block SLP. (vect_slp_transform_bb): New function. testsuite/ChangeLog: * gcc.dg/vect/bb-slp-1.c: New test. * gcc.dg/vect/bb-slp-2.c, gcc.dg/vect/bb-slp-3.c, gcc.dg/vect/bb-slp-4.c, gcc.dg/vect/bb-slp-5.c, gcc.dg/vect/bb-slp-6.c, gcc.dg/vect/bb-slp-7.c, gcc.dg/vect/bb-slp-8.c, gcc.dg/vect/bb-slp-9.c, gcc.dg/vect/bb-slp-10.c, gcc.dg/vect/bb-slp-11.c, gcc.dg/vect/no-tree-reassoc-bb-slp-12.c, gcc.dg/vect/bb-slp-13.c, gcc.dg/vect/bb-slp-14.c, gcc.dg/vect/bb-slp-15.c, gcc.dg/vect/bb-slp-16.c, gcc.dg/vect/bb-slp-17.c, gcc.dg/vect/bb-slp-18.c, gcc.dg/vect/bb-slp-19.c, gcc.dg/vect/bb-slp-20.c, gcc.dg/vect/bb-slp-21.c, gcc.dg/vect/bb-slp-22.c: Likewise. * gcc.dg/vect/vect.exp: Run basic block SLP tests. Patch:. (See attached file: slp-tests.txt)(See attached file: slp.txt)
Attachment:
slp-tests.txt
Description: Text document
Attachment:
slp.txt
Description: Text document
Index Nav: | [Date Index] [Subject Index] [Author Index] [Thread Index] | |
---|---|---|
Message Nav: | [Date Prev] [Date Next] | [Thread Prev] [Thread Next] |