diff options
Diffstat (limited to 'lib/compiler/src/beam_block.erl')
| -rw-r--r-- | lib/compiler/src/beam_block.erl | 97 |
1 files changed, 59 insertions, 38 deletions
diff --git a/lib/compiler/src/beam_block.erl b/lib/compiler/src/beam_block.erl index fe1ce6f60b..39ae8d5347 100644 --- a/lib/compiler/src/beam_block.erl +++ b/lib/compiler/src/beam_block.erl @@ -23,25 +23,32 @@ -module(beam_block). -export([module/2]). --import(lists, [reverse/1,reverse/2,foldl/3,member/2]). +-import(lists, [reverse/1,reverse/2,member/2]). -spec module(beam_utils:module_code(), [compile:option()]) -> {'ok',beam_utils:module_code()}. -module({Mod,Exp,Attr,Fs0,Lc}, _Opt) -> - Fs = [function(F) || F <- Fs0], +module({Mod,Exp,Attr,Fs0,Lc}, Opts) -> + Blockify = not member(no_blockify, Opts), + Fs = [function(F, Blockify) || F <- Fs0], {ok,{Mod,Exp,Attr,Fs,Lc}}. -function({function,Name,Arity,CLabel,Is0}) -> +function({function,Name,Arity,CLabel,Is0}, Blockify) -> try %% Collect basic blocks and optimize them. - Is1 = blockify(Is0), - Is2 = embed_lines(Is1), + Is2 = case Blockify of + true -> + Is1 = blockify(Is0), + embed_lines(Is1); + false -> + Is0 + end, Is3 = beam_utils:anno_defs(Is2), Is4 = move_allocates(Is3), Is5 = beam_utils:live_opt(Is4), Is6 = opt_blocks(Is5), - Is = beam_utils:delete_live_annos(Is6), + Is7 = beam_utils:delete_annos(Is6), + Is = opt_allocs(Is7), %% Done. {function,Name,Arity,CLabel,Is} @@ -136,17 +143,16 @@ embed_lines([], Acc) -> Acc. opt_blocks([{block,Bl0}|Is]) -> %% The live annotation at the beginning is not useful. - [{'%live',_,_}|Bl] = Bl0, + [{'%anno',_}|Bl] = Bl0, [{block,opt_block(Bl)}|opt_blocks(Is)]; opt_blocks([I|Is]) -> [I|opt_blocks(Is)]; opt_blocks([]) -> []. opt_block(Is0) -> - Is = find_fixpoint(fun(Is) -> - opt_tuple_element(opt(Is)) - end, Is0), - opt_alloc(Is). + find_fixpoint(fun(Is) -> + opt_tuple_element(opt(Is)) + end, Is0). find_fixpoint(OptFun, Is0) -> case OptFun(Is0) of @@ -196,7 +202,7 @@ move_allocates([I|Is]) -> [I|move_allocates(Is)]; move_allocates([]) -> []. -move_allocates_1([{'%def',_}|Is], Acc) -> +move_allocates_1([{'%anno',_}|Is], Acc) -> move_allocates_1(Is, Acc); move_allocates_1([I|Is], [{set,[],[],{alloc,Live0,Info}}|Acc]=Acc0) -> case alloc_may_pass(I) of @@ -240,10 +246,14 @@ opt([{set,_,_,{line,_}}=Line1, {set,[D2],[{integer,Idx2},Reg],{bif,element,{f,0}}}=I2|Is]) when Idx1 < Idx2, D1 =/= D2, D1 =/= Reg, D2 =/= Reg -> opt([Line2,I2,Line1,I1|Is]); +opt([{set,[D1],[{integer,Idx1},Reg],{bif,element,{f,L}}}=I1, + {set,[D2],[{integer,Idx2},Reg],{bif,element,{f,L}}}=I2|Is]) + when Idx1 < Idx2, D1 =/= D2, D1 =/= Reg, D2 =/= Reg -> + opt([I2,I1|Is]); opt([{set,Ds0,Ss,Op}|Is0]) -> {Ds,Is} = opt_moves(Ds0, Is0), [{set,Ds,Ss,Op}|opt(Is)]; -opt([{'%live',_,_}=I|Is]) -> +opt([{'%anno',_}=I|Is]) -> [I|opt(Is)]; opt([]) -> []. @@ -411,31 +421,47 @@ eliminate_use_of_from_reg([I]=Is, From, _To, Acc) -> no end. +%% opt_allocs(Instructions) -> Instructions. Optimize allocate +%% instructions inside blocks. If safe, replace an allocate_zero +%% instruction with the slightly cheaper allocate instruction. + +opt_allocs(Is) -> + D = beam_utils:index_labels(Is), + opt_allocs_1(Is, D). + +opt_allocs_1([{block,Bl0}|Is], D) -> + Bl = opt_alloc(Bl0, {D,Is}), + [{block,Bl}|opt_allocs_1(Is, D)]; +opt_allocs_1([I|Is], D) -> + [I|opt_allocs_1(Is, D)]; +opt_allocs_1([], _) -> []. + %% opt_alloc(Instructions) -> Instructions' %% Optimises all allocate instructions. opt_alloc([{set,[],[],{alloc,Live0,Info0}}, - {set,[],[],{alloc,Live,Info}}|Is]) -> + {set,[],[],{alloc,Live,Info}}|Is], D) -> Live = Live0, %Assertion. Alloc = combine_alloc(Info0, Info), I = {set,[],[],{alloc,Live,Alloc}}, - opt_alloc([I|Is]); -opt_alloc([{set,[],[],{alloc,R,{_,Ns,Nh,[]}}}|Is]) -> - [{set,[],[],opt_alloc(Is, Ns, Nh, R)}|Is]; -opt_alloc([I|Is]) -> [I|opt_alloc(Is)]; -opt_alloc([]) -> []. + opt_alloc([I|Is], D); +opt_alloc([{set,[],[],{alloc,R,{_,Ns,Nh,[]}}}|Is], D) -> + [{set,[],[],opt_alloc(Is, D, Ns, Nh, R)}|Is]; +opt_alloc([I|Is], D) -> [I|opt_alloc(Is, D)]; +opt_alloc([], _) -> []. combine_alloc({_,Ns,Nh1,Init}, {_,nostack,Nh2,[]}) -> {zero,Ns,beam_utils:combine_heap_needs(Nh1, Nh2),Init}. - + %% opt_alloc(Instructions, FrameSize, HeapNeed, LivingRegs) -> [Instr] %% Generates the optimal sequence of instructions for %% allocating and initalizing the stack frame and needed heap. -opt_alloc(_Is, nostack, Nh, LivingRegs) -> +opt_alloc(_Is, _D, nostack, Nh, LivingRegs) -> {alloc,LivingRegs,{nozero,nostack,Nh,[]}}; -opt_alloc(Is, Ns, Nh, LivingRegs) -> - InitRegs = init_yreg(Is, 0), +opt_alloc(Bl, {D,OuterIs}, Ns, Nh, LivingRegs) -> + Is = [{block,Bl}|OuterIs], + InitRegs = init_yregs(Ns, Is, D), case count_ones(InitRegs) of N when N*2 > Ns -> {alloc,LivingRegs,{nozero,Ns,Nh,gen_init(Ns, InitRegs)}}; @@ -451,19 +477,14 @@ gen_init(Fs, Regs, Y, Acc) when Regs band 1 =:= 0 -> gen_init(Fs, Regs, Y, Acc) -> gen_init(Fs, Regs bsr 1, Y+1, Acc). -%% init_yreg(Instructions, RegSet) -> RegSetInitialized -%% Calculate the set of initialized y registers. - -init_yreg([{set,_,_,{bif,_,_}}|_], Reg) -> Reg; -init_yreg([{set,_,_,{alloc,_,{gc_bif,_,_}}}|_], Reg) -> Reg; -init_yreg([{set,_,_,{alloc,_,{put_map,_,_}}}|_], Reg) -> Reg; -init_yreg([{set,Ds,_,_}|Is], Reg) -> init_yreg(Is, add_yregs(Ds, Reg)); -init_yreg(_Is, Reg) -> Reg. - -add_yregs(Ys, Reg) -> foldl(fun(Y, R0) -> add_yreg(Y, R0) end, Reg, Ys). - -add_yreg({y,Y}, Reg) -> Reg bor (1 bsl Y); -add_yreg(_, Reg) -> Reg. +init_yregs(Y, Is, D) when Y >= 0 -> + case beam_utils:is_killed({y,Y}, Is, D) of + true -> + (1 bsl Y) bor init_yregs(Y-1, Is, D); + false -> + init_yregs(Y-1, Is, D) + end; +init_yregs(_, _, _) -> 0. count_ones(Bits) -> count_ones(Bits, 0). count_ones(0, Acc) -> Acc; @@ -514,7 +535,7 @@ x_live([], Regs) -> Regs. %% Given a reversed instruction stream, determine the %% the registers that are defined. -defined_regs([{'%def',Def}|_], Regs) -> +defined_regs([{'%anno',{def,Def}}|_], Regs) -> Def bor Regs; defined_regs([{set,Ds,_,{alloc,Live,_}}|_], Regs) -> x_live(Ds, Regs bor ((1 bsl Live) - 1)); |