From c67741f8c505a9c17ffaf60f1cc37458b7e1301c Mon Sep 17 00:00:00 2001 From: Richard Carlsson Date: Wed, 29 Apr 2015 14:25:07 +0200 Subject: Avoid repeated calls to system_info in hipe loader Make hipe_unified_loader not call system_info(hipe_architecture) repeatedly. Also clean up some ugly architecture-dependent case switches. --- lib/kernel/src/hipe_unified_loader.erl | 122 +++++++++++++++++++-------------- 1 file changed, 72 insertions(+), 50 deletions(-) (limited to 'lib/kernel/src') diff --git a/lib/kernel/src/hipe_unified_loader.erl b/lib/kernel/src/hipe_unified_loader.erl index 49d4a8fe54..949a3c9207 100644 --- a/lib/kernel/src/hipe_unified_loader.erl +++ b/lib/kernel/src/hipe_unified_loader.erl @@ -82,6 +82,13 @@ chunk_name(Architecture) -> %% HW32 %% HiPE, x86, Win32 end. +word_size(Architecture) -> + case Architecture of + amd64 -> 8; + ppc64 -> 8; + _ -> 4 + end. + %%======================================================================== -spec load_native_code(Mod, binary()) -> 'no_native' | {'module', Mod} @@ -212,18 +219,22 @@ load_common(Mod, Bin, Beam, OldReferencesToPatch) -> bad_crc; true -> put(closures_to_patch, []), + Architecture=erlang:system_info(hipe_architecture), + WordSize = word_size(Architecture), + WriteWord = write_word_fun(WordSize), %% Create data segment {ConstAddr,ConstMap2} = - create_data_segment(ConstAlign, ConstSize, ConstMap), + create_data_segment(ConstAlign, ConstSize, ConstMap, WriteWord), %% Find callees for which we may need trampolines. - CalleeMFAs = find_callee_mfas(Refs), + CalleeMFAs = find_callee_mfas(Refs, Architecture), %% Write the code to memory. {CodeAddress,Trampolines} = enter_code(CodeSize, CodeBinary, CalleeMFAs, Mod, Beam), %% Construct CalleeMFA-to-trampoline mapping. - TrampolineMap = mk_trampoline_map(CalleeMFAs, Trampolines), + TrampolineMap = mk_trampoline_map(CalleeMFAs, Trampolines, + Architecture), %% Patch references to code labels in data seg. - ok = patch_consts(LabelMap, ConstAddr, CodeAddress), + ok = patch_consts(LabelMap, ConstAddr, CodeAddress, WriteWord), %% Find out which functions are being loaded (and where). %% Note: Addresses are sorted descending. {MFAs,Addresses} = exports(ExportMap, CodeAddress), @@ -275,14 +286,26 @@ load_common(Mod, Bin, Beam, OldReferencesToPatch) -> %% Scan the list of patches and build a set (returned as a tuple) %% of the callees for which we may need trampolines. %% -find_callee_mfas(Patches) when is_list(Patches) -> - case erlang:system_info(hipe_architecture) of - amd64 -> []; - arm -> find_callee_mfas(Patches, gb_sets:empty(), false); - powerpc -> find_callee_mfas(Patches, gb_sets:empty(), true); - ppc64 -> find_callee_mfas(Patches, gb_sets:empty(), true); - ultrasparc -> []; - x86 -> [] +find_callee_mfas(Patches, Architecture) when is_list(Patches) -> + case needs_trampolines(Architecture) of + true -> find_callee_mfas(Patches, gb_sets:empty(), + no_erts_trampolines(Architecture)); + _ -> [] + end. + +needs_trampolines(Architecture) -> + case Architecture of + arm -> true; + powerpc -> true; + ppc64 -> true; + _ -> false + end. + +no_erts_trampolines(Architecture) -> + case Architecture of + powerpc -> true; + ppc64 -> true; + _ -> false end. find_callee_mfas([{Type,Data}|Patches], MFAs, SkipErtsSyms) -> @@ -318,14 +341,9 @@ add_callee_mfas([], MFAs, _SkipErtsSyms) -> MFAs. %%---------------------------------------------------------------- %% -mk_trampoline_map([], []) -> []; % archs not using trampolines -mk_trampoline_map(CalleeMFAs, Trampolines) -> - SizeofLong = - case erlang:system_info(hipe_architecture) of - amd64 -> 8; - ppc64 -> 8; - _ -> 4 - end, +mk_trampoline_map([], [], _) -> []; % archs not using trampolines +mk_trampoline_map(CalleeMFAs, Trampolines, Architecture) -> + SizeofLong = word_size(Architecture), mk_trampoline_map(tuple_size(CalleeMFAs), CalleeMFAs, Trampolines, SizeofLong, gb_trees:empty()). @@ -621,22 +639,24 @@ patch_load_mfa(CodeAddress, DestMFA, Addresses, RemoteOrLocal) -> %%---------------------------------------------------------------- %% Patch references to code labels in the data segment. %% -patch_consts(Labels, DataAddress, CodeAddress) -> +patch_consts(Labels, DataAddress, CodeAddress, WriteWord) -> lists:foreach(fun (L) -> - patch_label_or_labels(L, DataAddress, CodeAddress) + patch_label_or_labels(L, DataAddress, CodeAddress, + WriteWord) end, Labels). -patch_label_or_labels({Pos,Offset}, DataAddress, CodeAddress) -> +patch_label_or_labels({Pos,Offset}, DataAddress, CodeAddress, WriteWord) -> ?ASSERT(assert_local_patch(CodeAddress+Offset)), - write_word(DataAddress+Pos, CodeAddress+Offset); -patch_label_or_labels({sorted,Base,UnOrderdList}, DataAddress, CodeAddress) -> - sort_and_write(UnOrderdList, Base, DataAddress, CodeAddress). + WriteWord(DataAddress+Pos, CodeAddress+Offset); +patch_label_or_labels({sorted,Base,UnOrderdList}, DataAddress, CodeAddress, + WriteWord) -> + sort_and_write(UnOrderdList, Base, DataAddress, CodeAddress, WriteWord). -sort_and_write(UnOrderdList, Base, DataAddress, CodeAddress) -> +sort_and_write(UnOrderdList, Base, DataAddress, CodeAddress, WriteWord) -> WriteAndInc = fun ({_, Offset}, DataPos) -> ?ASSERT(assert_local_patch(CodeAddress+Offset)), - write_word(DataPos, CodeAddress+Offset) + WriteWord(DataPos, CodeAddress+Offset) end, lists:foldl(WriteAndInc, DataAddress+Base, sort_on_representation(UnOrderdList)). @@ -662,17 +682,18 @@ patch_instr(Address, Value, Type) -> %% XXX: It appears this is used for inserting both code addresses %% and other data. In HiPE, code addresses are still 32-bit on %% some 64-bit machines. -write_word(DataAddress, DataWord) -> - case erlang:system_info(hipe_architecture) of - amd64 -> - hipe_bifs:write_u64(DataAddress, DataWord), - DataAddress+8; - ppc64 -> - hipe_bifs:write_u64(DataAddress, DataWord), - DataAddress+8; - _ -> - hipe_bifs:write_u32(DataAddress, DataWord), - DataAddress+4 +write_word_fun(WordSize) -> + case WordSize of + 8 -> + fun (DataAddress, DataWord) -> + hipe_bifs:write_u64(DataAddress, DataWord), + DataAddress+8 + end; + 4 -> + fun (DataAddress, DataWord) -> + hipe_bifs:write_u32(DataAddress, DataWord), + DataAddress+4 + end end. %%-------------------------------------------------------------------- @@ -688,30 +709,31 @@ bif_address(Name) when is_atom(Name) -> %% memory, and produces a ConstMap2 mapping each constant's ConstNo to %% its runtime address, tagged if the constant is a term. %% -create_data_segment(DataAlign, DataSize, DataList) -> +create_data_segment(DataAlign, DataSize, DataList, WriteWord) -> %%io:format("create_data_segment: \nDataAlign: ~p\nDataSize: ~p\nDataList: ~p\n",[DataAlign,DataSize,DataList]), DataAddress = hipe_bifs:alloc_data(DataAlign, DataSize), - enter_data(DataList, [], DataAddress, DataSize). + enter_data(DataList, [], DataAddress, DataSize, WriteWord). -enter_data(List, ConstMap2, DataAddress, DataSize) -> +enter_data(List, ConstMap2, DataAddress, DataSize, WriteWord) -> case List of [ConstNo,Offset,Type,Data|Rest] when is_integer(Offset) -> %%?msg("Const ~w\n",[[ConstNo,Offset,Type,Data]]), ?ASSERT((Offset >= 0) and (Offset =< DataSize)), - Res = enter_datum(Type, Data, DataAddress+Offset), - enter_data(Rest, [{ConstNo,Res}|ConstMap2], DataAddress, DataSize); + Res = enter_datum(Type, Data, DataAddress+Offset, WriteWord), + enter_data(Rest, [{ConstNo,Res}|ConstMap2], DataAddress, DataSize, + WriteWord); [] -> {DataAddress, ConstMap2} end. -enter_datum(Type, Data, Address) -> +enter_datum(Type, Data, Address, WriteWord) -> case ?EXT2CONST_TYPE(Type) of term -> %% Address is unused for terms hipe_bifs:term_to_word(hipe_bifs:merge_term(Data)); sorted_block -> L = lists:sort([hipe_bifs:term_to_word(Term) || Term <- Data]), - write_words(L, Address), + write_words(L, Address, WriteWord), Address; block -> case Data of @@ -719,7 +741,7 @@ enter_datum(Type, Data, Address) -> write_bytes(Lbls, Address); {Lbls, SortOrder} -> SortedLbls = [Lbl || {_,Lbl} <- lists:sort(group(Lbls, SortOrder))], - write_words(SortedLbls, Address); + write_words(SortedLbls, Address, WriteWord); Lbls -> write_bytes(Lbls, Address) end, @@ -734,9 +756,9 @@ group([B1,B2,B3,B4|Ls], [O|Os]) -> bytes_to_32(B4,B3,B2,B1) -> (B4 bsl 24) bor (B3 bsl 16) bor (B2 bsl 8) bor B1. -write_words([W|Rest], Addr) -> - write_words(Rest, write_word(Addr, W)); -write_words([], Addr) when is_integer(Addr) -> true. +write_words([W|Rest], Addr, WriteWord) -> + write_words(Rest, WriteWord(Addr, W), WriteWord); +write_words([], Addr, _) when is_integer(Addr) -> true. write_bytes([B|Rest], Addr) -> hipe_bifs:write_u8(Addr, B), -- cgit v1.2.3