Hi, On Mon, Dec 12 2022, Jan Hubicka wrote: >> Hi, >> >> I'm re-posting patches which I have posted at the end of stage 1 but >> which have not passed review yet. >> >> 8<-------------------------------------------------------------------- >> >> Fortran descriptors are structures which are often constructed just >> for a particular argument of a particular call where it is passed by >> reference. When the called function is under compiler's control, it >> can be beneficial to split up the descriptor and pass it in individual >> parameters. Unfortunately, currently we allow IPA-SRA to replace a >> pointer with a set of replacements which are at most twice as big in >> total and for descriptors we'd need to bump that factor to seven. >> >> This patch looks for parameters which are ADDR_EXPRs of local >> variables which are written to and passed as arguments by reference >> but are never loaded from and marks them with a flag in the call edge >> summary. The IPA analysis phase then identifies formal parameters >> which are always fed such arguments and then is more lenient when it >> comoes to size. >> >> In order not to store to maximums per formal parameter, I calculate >> the more lenient one by multiplying the existing one with a new >> parameter. If it is preferable to keep the maximums independent, we >> can do that. Documentation for the new parameter is missing as I >> still need to re-base the patch on a version which has sphinx. I will >> write it before committing. >> >> I have disable IPA-SRA in pr48636-2.f90 in order to be able to keep >> using its dump-scan expressions. The new testcase is basically a copy >> of it with different options and IPA-SRA dump scans. >> >> Bootstrapped and tested individually when I originally posted it and >> now bootstrapped and LTO-bootstrapped and tested as part of the whole >> series. OK for master? >> >> >> gcc/ChangeLog: >> >> 2022-11-11 Martin Jambor <mjam...@suse.cz> >> >> * ipa-sra.c (isra_param_desc): New field not_specially_constructed. >> (struct isra_param_flow): New field constructed_for_calls. >> (isra_call_summary::dump): Dump the new flag. >> (loaded_decls): New variable. >> (dump_isra_param_descriptor): New parameter hints, dump >> not_specially_constructed if it is true. >> (dump_isra_param_descriptors): New parameter hints, pass it to >> dump_isra_param_descriptor. >> (ipa_sra_function_summaries::duplicate): Duplicate new flag. >> (create_parameter_descriptors): Adjust comment. >> (get_gensum_param_desc): Bail out when decl2desc is NULL. >> (scan_expr_access): Add loaded local variables to loaded_decls. >> (scan_function): Survive if final_bbs is NULL. >> (isra_analyze_call): Compute constructed_for_calls flag. >> (process_scan_results): Be optimistic about size limits. Do not dump >> computed param hints when dumpint IPA-SRA structures. >> (isra_write_edge_summary): Stream constructed_for_calls. >> (isra_read_edge_summary): Likewise. >> (ipa_sra_dump_all_summaries): New parameter hints, pass it to >> dump_isra_param_descriptor. >> (flip_all_hints_pessimistic): New function. >> (flip_all_param_hints_pessimistic): Likewise. >> (propagate_param_hints): Likewise. >> (disable_unavailable_parameters): Renamed to >> adjust_parameter_descriptions. Expand size limits for parameters >> which are specially contstructed by all callers. Check limits again.p >> (ipa_sra_analysis): Pass required hints to ipa_sra_dump_all_summaries. >> Add hint propagation. >> (ipa_sra_summarize_function): Initialize and destory loaded_decls, >> rearrange so that scan_function is called even when there are no >> candidates. >> * params.opt (ipa-sra-ptrwrap-growth-factor): New parameter. > > Hmm, this is quite specific heuristics, but I do not have much better > ideas, so it is OK :) >
Yeah, it kind of is. I was wondering whether it should really only target Fortran array descriptors (and have them marked some way) but eventually decided for this - but I do not expect the code to trigger too much for non-Fortran code. > Can this be useful also for inlining? IPA-SRA deallocates its summaries after the analysis phase so we'd need to postpone that for later. Otherwise it should be quite directly usable - perhaps after a check that the "hint propagation" bit of IPA-SRA has been run, the flag starts with the optimistic value. Martin