diff options
Diffstat (limited to 'lib/compiler/src')
-rw-r--r-- | lib/compiler/src/beam_asm.erl | 2 | ||||
-rw-r--r-- | lib/compiler/src/beam_dead.erl | 20 | ||||
-rw-r--r-- | lib/compiler/src/core_lint.erl | 6 | ||||
-rw-r--r-- | lib/compiler/src/sys_core_fold.erl | 10 | ||||
-rw-r--r-- | lib/compiler/src/v3_codegen.erl | 460 | ||||
-rw-r--r-- | lib/compiler/src/v3_core.erl | 4 | ||||
-rw-r--r-- | lib/compiler/src/v3_kernel.erl | 17 | ||||
-rw-r--r-- | lib/compiler/src/v3_kernel.hrl | 2 |
8 files changed, 333 insertions, 188 deletions
diff --git a/lib/compiler/src/beam_asm.erl b/lib/compiler/src/beam_asm.erl index 3dff51d7f6..453e00fce3 100644 --- a/lib/compiler/src/beam_asm.erl +++ b/lib/compiler/src/beam_asm.erl @@ -24,7 +24,7 @@ -export([module/4]). -export([encode/2]). --export_type([fail/0,label/0,reg/0,src/0,module_code/0,function_name/0]). +-export_type([fail/0,label/0,reg/0,reg_num/0,src/0,module_code/0,function_name/0]). -import(lists, [map/2,member/2,keymember/3,duplicate/2,splitwith/2]). -include("beam_opcodes.hrl"). diff --git a/lib/compiler/src/beam_dead.erl b/lib/compiler/src/beam_dead.erl index d379fdc4eb..1b152a2d6f 100644 --- a/lib/compiler/src/beam_dead.erl +++ b/lib/compiler/src/beam_dead.erl @@ -272,7 +272,8 @@ backward([{jump,{f,To0}},{move,Src,Reg}=Move|Is], D, Acc) -> end; backward([{jump,{f,To}}=J|[{bif,Op,{f,BifFail},Ops,Reg}|Is]=Is0], D, Acc) -> try replace_comp_op(To, Reg, Op, Ops, D) of - I -> backward(Is, D, I++Acc) + {Test,Jump} -> + backward([Jump,Test|Is], D, Acc) catch throw:not_possible -> case To =:= BifFail of @@ -446,7 +447,7 @@ prune_redundant([], _) -> []. replace_comp_op(To, Reg, Op, Ops, D) -> False = comp_op_find_shortcut(To, Reg, {atom,false}, D), True = comp_op_find_shortcut(To, Reg, {atom,true}, D), - [bif_to_test(Op, Ops, False),{jump,{f,True}}]. + {bif_to_test(Op, Ops, False),{jump,{f,True}}}. comp_op_find_shortcut(To0, Reg, Val, D) -> case shortcut_select_label(To0, Reg, Val, D) of @@ -483,15 +484,22 @@ not_possible() -> throw(not_possible). %% F1: is_eq_exact F2 Reg Lit2 F1: is_eq_exact F2 Reg Lit2 %% L2: .... L2: %% -combine_eqs(To, [Reg,{Type,_}=Lit1]=Ops, D, [{label,L1}|_]) - when Type =:= atom; Type =:= integer -> +combine_eqs(To, [Reg,{Type,_}=Lit1]=Ops, D, Acc) + when Type =:= atom; Type =:= integer -> + Next = case Acc of + [{label,Lbl}|_] -> Lbl; + [{jump,{f,Lbl}}|_] -> Lbl + end, case beam_utils:code_at(To, D) of [{test,is_eq_exact,{f,F2},[Reg,{Type,_}=Lit2]}, {label,L2}|_] when Lit1 =/= Lit2 -> - {select,select_val,Reg,{f,F2},[Lit1,{f,L1},Lit2,{f,L2}]}; + {select,select_val,Reg,{f,F2},[Lit1,{f,Next},Lit2,{f,L2}]}; + [{test,is_eq_exact,{f,F2},[Reg,{Type,_}=Lit2]}, + {jump,{f,L2}}|_] when Lit1 =/= Lit2 -> + {select,select_val,Reg,{f,F2},[Lit1,{f,Next},Lit2,{f,L2}]}; [{select,select_val,Reg,{f,F2},[{Type,_}|_]=List0}|_] -> List = remove_from_list(Lit1, List0), - {select,select_val,Reg,{f,F2},[Lit1,{f,L1}|List]}; + {select,select_val,Reg,{f,F2},[Lit1,{f,Next}|List]}; _Is -> {test,is_eq_exact,{f,To},Ops} end; diff --git a/lib/compiler/src/core_lint.erl b/lib/compiler/src/core_lint.erl index 7d3513c0ba..6e2114be56 100644 --- a/lib/compiler/src/core_lint.erl +++ b/lib/compiler/src/core_lint.erl @@ -353,12 +353,6 @@ expr(#c_case{arg=Arg,clauses=Cs}, Def, Rt, St0) -> Pc = case_patcount(Cs), St1 = body(Arg, Def, Pc, St0), clauses(Cs, Def, Pc, Rt, St1); -expr(#c_receive{clauses=Cs,timeout=#c_literal{val=infinity}, - action=#c_literal{}}, - Def, Rt, St) -> - %% If the timeout is 'infinity', the after code can never - %% be reached. We don't care if the return count is wrong. - clauses(Cs, Def, 1, Rt, St); expr(#c_receive{clauses=Cs,timeout=T,action=A}, Def, Rt, St0) -> St1 = expr(T, Def, 1, St0), St2 = body(A, Def, Rt, St1), diff --git a/lib/compiler/src/sys_core_fold.erl b/lib/compiler/src/sys_core_fold.erl index 6da68f1f4e..f8a467d6a9 100644 --- a/lib/compiler/src/sys_core_fold.erl +++ b/lib/compiler/src/sys_core_fold.erl @@ -2624,9 +2624,13 @@ delay_build_expr_1(#c_receive{clauses=Cs0, timeout=Timeout, action=A0}=Rec, TypeSig) -> Cs = delay_build_cs(Cs0, TypeSig), - A = case Timeout of - #c_literal{val=infinity} -> A0; - _ -> delay_build_expr(A0, TypeSig) + A = case {Timeout,A0} of + {#c_literal{val=infinity},#c_literal{}} -> + {_Type,Arity} = TypeSig, + Es = lists:duplicate(Arity, A0), + core_lib:make_values(Es); + _ -> + delay_build_expr(A0, TypeSig) end, Rec#c_receive{clauses=Cs,action=A}; delay_build_expr_1(#c_seq{body=B0}=Seq, TypeSig) -> diff --git a/lib/compiler/src/v3_codegen.erl b/lib/compiler/src/v3_codegen.erl index 9bcd6987bf..74d64e1456 100644 --- a/lib/compiler/src/v3_codegen.erl +++ b/lib/compiler/src/v3_codegen.erl @@ -51,7 +51,7 @@ set_kanno(Kthing, Anno) -> setelement(2, Kthing, Anno). %% Stack/register state record. -record(sr, {reg=[], %Register table stk=[], %Stack table - res=[]}). %Reserved regs: [{reserved,I,V}] + res=[]}). %Registers to reserve %% Internal records. -record(cg_need_heap, {anno=[] :: term(), @@ -77,10 +77,15 @@ functions(Forms, AtomMod) -> function(#k_fdef{anno=#k{a=Anno},func=Name,arity=Arity, vars=As,body=Kb}, AtomMod, St0) -> try - %% Annotate kernel records with variable usage. #k_match{} = Kb, %Assertion. + + %% Try to suppress the stack frame unless it is + %% really needed. + Body0 = avoid_stack_frame(Kb), + + %% Annotate kernel records with variable usage. Vdb0 = init_vars(As), - {Body,_,Vdb} = body(Kb, 1, Vdb0), + {Body,_,Vdb} = body(Body0, 1, Vdb0), %% Generate the BEAM assembly code. {Asm,EntryLabel,St} = cg_fun(Body, As, Vdb, AtomMod, @@ -94,6 +99,112 @@ function(#k_fdef{anno=#k{a=Anno},func=Name,arity=Arity, erlang:raise(Class, Error, Stack) end. + +%% avoid_stack_frame(Kernel) -> Kernel' +%% If possible, avoid setting up a stack frame. Functions +%% that only do matching, calls to guard BIFs, and tail-recursive +%% calls don't need a stack frame. + +avoid_stack_frame(#k_match{body=Body}=M) -> + try + M#k_match{body=avoid_stack_frame_1(Body)} + catch + impossible -> + M + end. + +avoid_stack_frame_1(#k_alt{first=First0,then=Then0}=Alt) -> + First = avoid_stack_frame_1(First0), + Then = avoid_stack_frame_1(Then0), + Alt#k_alt{first=First,then=Then}; +avoid_stack_frame_1(#k_bif{op=Op}=Bif) -> + case Op of + #k_internal{} -> + %% Most internal BIFs clobber the X registers. + throw(impossible); + _ -> + Bif + end; +avoid_stack_frame_1(#k_break{anno=Anno,args=Args}) -> + #k_guard_break{anno=Anno,args=Args}; +avoid_stack_frame_1(#k_guard_break{}=Break) -> + Break; +avoid_stack_frame_1(#k_enter{}=Enter) -> + %% Tail-recursive calls don't need a stack frame. + Enter; +avoid_stack_frame_1(#k_guard{clauses=Cs0}=Guard) -> + Cs = avoid_stack_frame_list(Cs0), + Guard#k_guard{clauses=Cs}; +avoid_stack_frame_1(#k_guard_clause{guard=G0,body=B0}=C) -> + G = avoid_stack_frame_1(G0), + B = avoid_stack_frame_1(B0), + C#k_guard_clause{guard=G,body=B}; +avoid_stack_frame_1(#k_match{anno=A,vars=Vs,body=B0,ret=Ret}) -> + %% Use #k_guard_match{} instead to avoid saving the X registers + %% to the stack before matching. + B = avoid_stack_frame_1(B0), + #k_guard_match{anno=A,vars=Vs,body=B,ret=Ret}; +avoid_stack_frame_1(#k_guard_match{body=B0}=M) -> + B = avoid_stack_frame_1(B0), + M#k_guard_match{body=B}; +avoid_stack_frame_1(#k_protected{arg=Arg0}=Prot) -> + Arg = avoid_stack_frame_1(Arg0), + Prot#k_protected{arg=Arg}; +avoid_stack_frame_1(#k_put{}=Put) -> + Put; +avoid_stack_frame_1(#k_return{}=Ret) -> + Ret; +avoid_stack_frame_1(#k_select{var=#k_var{anno=Vanno},types=Types0}=Select) -> + case member(reuse_for_context, Vanno) of + false -> + Types = avoid_stack_frame_list(Types0), + Select#k_select{types=Types}; + true -> + %% Including binary patterns that overwrite the register containing + %% the binary with the match context may not be safe. For example, + %% bs_match_SUITE:bin_tail_e/1 with inlining will be rejected by + %% beam_validator. + %% + %% Essentially the following code is produced: + %% + %% bs_match {x,0} => {x,0} + %% ... + %% bs_match {x,0} => {x,1} %% ILLEGAL + %% + %% A bs_match instruction will only accept a match context as the + %% source operand if the source and destination registers are the + %% the same (as in the first bs_match instruction above). + %% The second bs_match instruction is therefore illegal. + %% + %% This situation is avoided if there is a stack frame: + %% + %% move {x,0} => {y,0} + %% bs_match {x,0} => {x,0} + %% ... + %% bs_match {y,0} => {x,1} %% LEGAL + %% + throw(impossible) + end; +avoid_stack_frame_1(#k_seq{arg=A0,body=B0}=Seq) -> + A = avoid_stack_frame_1(A0), + B = avoid_stack_frame_1(B0), + Seq#k_seq{arg=A,body=B}; +avoid_stack_frame_1(#k_test{}=Test) -> + Test; +avoid_stack_frame_1(#k_type_clause{values=Values0}=TC) -> + Values = avoid_stack_frame_list(Values0), + TC#k_type_clause{values=Values}; +avoid_stack_frame_1(#k_val_clause{body=B0}=VC) -> + B = avoid_stack_frame_1(B0), + VC#k_val_clause{body=B}; +avoid_stack_frame_1(_Body) -> + throw(impossible). + +avoid_stack_frame_list([H|T]) -> + [avoid_stack_frame_1(H)|avoid_stack_frame_list(T)]; +avoid_stack_frame_list([]) -> []. + + %% This pass creates beam format annotated with variable lifetime %% information. Each thing is given an index and for each variable we %% store the first and last index for its occurrence. The variable @@ -219,10 +330,8 @@ expr(#k_put{anno=A}=Put, I, _Vdb) -> Put#k_put{anno=#l{i=I,a=A#k.a}}; expr(#k_break{anno=A}=Break, I, _Vdb) -> Break#k_break{anno=#l{i=I,a=A#k.a}}; -expr(#k_guard_break{anno=A}=Break, I, Vdb) -> - Locked = [V || {V,_,_} <- Vdb], - L = #l{i=I,a=A#k.a}, - Break#k_guard_break{anno=L,locked=Locked}; +expr(#k_guard_break{anno=A}=Break, I, _Vdb) -> + Break#k_guard_break{anno=#l{i=I,a=A#k.a}}; expr(#k_return{anno=A}=Ret, I, _Vdb) -> Ret#k_return{anno=#l{i=I,a=A#k.a}}. @@ -246,14 +355,9 @@ match(#k_alt{anno=A,first=Kf,then=Kt}, Ls, I, Vdb0) -> F = match(Kf, Ls, I+1, Vdb1), T = match(Kt, Ls, I+1, Vdb1), #k_alt{anno=[],first=F,then=T}; -match(#k_select{anno=A,var=V,types=Kts}=Select, Ls0, I, Vdb0) -> - Vanno = get_kanno(V), - Ls1 = case member(no_usage, Vanno) of - false -> add_element(V#k_var.name, Ls0); - true -> Ls0 - end, - Vdb1 = use_vars(union(A#k.us, Ls1), I, Vdb0), - Ts = [type_clause(Tc, Ls1, I+1, Vdb1) || Tc <- Kts], +match(#k_select{anno=A,types=Kts}=Select, Ls, I, Vdb0) -> + Vdb1 = use_vars(union(A#k.us, Ls), I, Vdb0), + Ts = [type_clause(Tc, Ls, I+1, Vdb1) || Tc <- Kts], Select#k_select{anno=[],types=Ts}; match(#k_guard{anno=A,clauses=Kcs}, Ls, I, Vdb0) -> Vdb1 = use_vars(union(A#k.us, Ls), I, Vdb0), @@ -388,8 +492,8 @@ cg(#k_return{anno=Le,args=Rs}, Vdb, Bef, St) -> return_cg(Rs, Le, Vdb, Bef, St); cg(#k_break{anno=Le,args=Bs}, Vdb, Bef, St) -> break_cg(Bs, Le, Vdb, Bef, St); -cg(#k_guard_break{anno=Le,args=Bs,locked=N}, Vdb, Bef, St) -> - guard_break_cg(Bs, N, Le, Vdb, Bef, St); +cg(#k_guard_break{anno=Le,args=Bs}, Vdb, Bef, St) -> + guard_break_cg(Bs, Le, Vdb, Bef, St); cg(#cg_need_heap{h=H}, _Vdb, Bef, St) -> {[{test_heap,H,max_reg(Bef#sr.reg)}],Bef,St}. @@ -487,7 +591,10 @@ match_cg(M, Rs, Le, Vdb, Bef, St0) -> guard_match_cg(M, Rs, Le, Vdb, Bef, St0) -> I = Le#l.i, {B,St1} = new_label(St0), - #cg{bfail=Fail} = St1, + Fail = case St0 of + #cg{bfail=0,ultimate_failure=Fail0} -> Fail0; + #cg{bfail=Fail0} -> Fail0 + end, {Mis,Aft,St2} = match_cg(M, Fail, Bef, St1#cg{break=B}), %% Update the register descriptors for the return registers. Reg = guard_match_regs(Aft#sr.reg, Rs), @@ -593,9 +700,6 @@ bsm_rename_ctx(#k_protected{arg=Ts0}=Prot, Old, New, _InProt) -> InProt = true, Ts = bsm_rename_ctx_list(Ts0, Old, New, InProt), bsm_forget_var(Prot#k_protected{arg=Ts}, Old); -bsm_rename_ctx(#k_match{body=Ms0}=Match, Old, New, InProt) -> - Ms = bsm_rename_ctx(Ms0, Old, New, InProt), - Match#k_match{body=Ms}; bsm_rename_ctx(#k_guard_match{body=Ms0}=Match, Old, New, InProt) -> Ms = bsm_rename_ctx(Ms0, Old, New, InProt), Match#k_guard_match{body=Ms}; @@ -612,9 +716,8 @@ bsm_rename_ctx(#cg_block{es=Es0}=Block, Old, New, true) -> %% inside the block. Es = bsm_rename_ctx_list(Es0, Old, New, true), bsm_forget_var(Block#cg_block{es=Es}, Old); -bsm_rename_ctx(#k_guard_break{locked=Locked0}=Break, Old, _New, _InProt) -> - Locked = Locked0 -- [Old], - bsm_forget_var(Break#k_guard_break{locked=Locked}, Old). +bsm_rename_ctx(#k_guard_break{}=Break, Old, _New, _InProt) -> + bsm_forget_var(Break, Old). bsm_rename_ctx_list([C|Cs], Old, New, InProt) -> [bsm_rename_ctx(C, Old, New, InProt)| @@ -679,129 +782,158 @@ basic_block([Ke|Kes], Acc) -> no_block -> {reverse(Acc, [Ke]),Kes} end. -%% #k_put{} instructions that may garbage collect are not allowed in basic blocks. - -collect_block(#k_put{arg=#k_binary{}}) -> - no_block; -collect_block(#k_put{arg=#k_map{}}) -> - no_block; -collect_block(#k_put{}) -> - include; -collect_block(#k_call{op=#k_var{}=Var,args=As}) -> - {block_end,As++[Var]}; +collect_block(#k_put{arg=Arg}) -> + %% #k_put{} instructions that may garbage collect are not allowed + %% in basic blocks. + case Arg of + #k_binary{} -> no_block; + #k_map{} -> no_block; + _ -> include + end; collect_block(#k_call{op=Func,args=As}) -> {block_end,As++func_vars(Func)}; -collect_block(#k_enter{op=#k_var{}=Var,args=As}) -> - {block_end,As++[Var]}; collect_block(#k_enter{op=Func,args=As}) -> {block_end,As++func_vars(Func)}; collect_block(#k_return{args=Rs}) -> {block_end,Rs}; collect_block(#k_break{args=Bs}) -> {block_end,Bs}; -collect_block(_) -> no_block. +collect_block(_) -> no_block. +func_vars(#k_var{}=Var) -> + [Var]; func_vars(#k_remote{mod=M,name=F}) when is_record(M, k_var); is_record(F, k_var) -> [M,F]; func_vars(_) -> []. -%% cg_basic_block([Kexpr], FirstI, LastI, As, Vdb, StackReg, State) -> +%% cg_basic_block([Kexpr], FirstI, LastI, Arguments, Vdb, StackReg, State) -> %% {[Ainstr],StackReg,State}. +%% +%% Do a specialized code generation for a basic block of #put{} +%% instructions (that don't do any garbage collection) followed by a +%% call, break, or return. +%% +%% 'Arguments' is a list of the variables that must be loaded into +%% consecutive X registers before the last instruction in the block. +%% The point of this specialized code generation is to try put the +%% all of the variables in 'Arguments' into the correct X register +%% to begin with, instead of putting them into the first available +%% X register and having to move them to the correct X register +%% later. +%% +%% To achieve that, we attempt to reserve the X registers that the +%% variables in 'Arguments' will need to be in when the block ends. +%% +%% To make it more likely that reservations will be successful, we +%% will try to save variables that need to be saved to the stack as +%% early as possible (if an X register needed by a variable in +%% Arguments is occupied by another variable, the value in the +%% X register can be evicted if it is saved on the stack). +%% +%% We will take care not to increase the size of stack frame compared +%% to what the standard code generator would have done (that is, to +%% save all X registers at the last possible moment). We will do that +%% by extending the stack frame to the minimal size needed to save +%% all that needs to be saved using extend_stack/4, and use +%% save_carefully/4 during code generation to only save the variables +%% that can be saved without growing the stack frame. cg_basic_block(Kes, Fb, Lf, As, Vdb, Bef, St0) -> - Res = make_reservation(As, 0), - Regs0 = reserve(Res, Bef#sr.reg, Bef#sr.stk), - Stk = extend_stack(Bef, Lf, Lf+1, Vdb), - Int0 = Bef#sr{reg=Regs0,stk=Stk,res=Res}, - X0_v0 = x0_vars(As, Fb, Lf, Vdb), - {Keis,{Aft,_,St1}} = + Int0 = reserve_arg_regs(As, Bef), + Int = extend_stack(Int0, Lf, Lf+1, Vdb), + {Keis,{Aft,St1}} = flatmapfoldl(fun(Ke, St) -> cg_basic_block(Ke, St, Lf, Vdb) end, - {Int0,X0_v0,St0}, need_heap(Kes, Fb)), + {Int,St0}, need_heap(Kes, Fb)), {Keis,Aft,St1}. -cg_basic_block(#cg_need_heap{}=Ke, {Inta,X0v,Sta}, _Lf, Vdb) -> - {Keis,Intb,Stb} = cg(Ke, Vdb, Inta, Sta), - {Keis, {Intb,X0v,Stb}}; -cg_basic_block(Ke, {Inta,X0_v1,Sta}, Lf, Vdb) -> +cg_basic_block(#cg_need_heap{}=Ke, {Bef,St0}, _Lf, Vdb) -> + {Keis,Aft,St1} = cg(Ke, Vdb, Bef, St0), + {Keis,{Aft,St1}}; +cg_basic_block(Ke, {Bef,St0}, Lf, Vdb) -> #l{i=I} = get_kanno(Ke), - {Sis,Intb} = save_carefully(Inta, I, Lf+1, Vdb), - {X0_v2,Intc} = allocate_x0(X0_v1, I, Intb), - Intd = reserve(Intc), - {Keis,Inte,Stb} = cg(Ke, Vdb, Intd, Sta), - {Sis ++ Keis, {Inte,X0_v2,Stb}}. -make_reservation([], _) -> []; -make_reservation([#k_var{name=V}|As], I) -> [{I,V}|make_reservation(As, I+1)]; -make_reservation([A|As], I) -> [{I,A}|make_reservation(As, I+1)]. - -reserve(Sr) -> Sr#sr{reg=reserve(Sr#sr.res, Sr#sr.reg, Sr#sr.stk)}. - -reserve([{I,V}|Rs], [free|Regs], Stk) -> [{reserved,I,V}|reserve(Rs, Regs, Stk)]; -reserve([{I,V}|Rs], [{I,V}|Regs], Stk) -> [{I,V}|reserve(Rs, Regs, Stk)]; -reserve([{I,V}|Rs], [{I,Var}|Regs], Stk) -> + %% Save all we can to increase the possibility that reserving + %% registers will succeed. + {Sis,Int0} = save_carefully(Bef, I, Lf+1, Vdb), + Int1 = reserve(Int0), + {Keis,Aft,St1} = cg(Ke, Vdb, Int1, St0), + {Sis ++ Keis,{Aft,St1}}. + +%% reserve_arg_regs([Argument], Bef) -> Aft. +%% Try to reserve the X registers for all arguments. All registers +%% that we wish to reserve will be saved in Bef#sr.res. + +reserve_arg_regs(As, Bef) -> + Res = reserve_arg_regs_1(As, 0), + reserve(Bef#sr{res=Res}). + +reserve_arg_regs_1([#k_var{name=V}|As], I) -> + [{I,V}|reserve_arg_regs_1(As, I+1)]; +reserve_arg_regs_1([A|As], I) -> + [{I,A}|reserve_arg_regs_1(As, I+1)]; +reserve_arg_regs_1([], _) -> []. + +%% reserve(Bef) -> Aft. +%% Try to reserve more registers. The registers we wish to reserve +%% are found in Bef#sr.res. + +reserve(#sr{reg=Regs,stk=Stk,res=Res}=Sr) -> + Sr#sr{reg=reserve_1(Res, Regs, Stk)}. + +reserve_1([{I,V}|Rs], [free|Regs], Stk) -> + [{reserved,I,V}|reserve_1(Rs, Regs, Stk)]; +reserve_1([{I,V}|Rs], [{I,V}|Regs], Stk) -> + [{I,V}|reserve_1(Rs, Regs, Stk)]; +reserve_1([{I,V}|Rs], [{I,Var}|Regs], Stk) -> case on_stack(Var, Stk) of - true -> [{reserved,I,V}|reserve(Rs, Regs, Stk)]; - false -> [{I,Var}|reserve(Rs, Regs, Stk)] + true -> [{reserved,I,V}|reserve_1(Rs, Regs, Stk)]; + false -> [{I,Var}|reserve_1(Rs, Regs, Stk)] end; -reserve([{I,V}|Rs], [{reserved,I,_}|Regs], Stk) -> - [{reserved,I,V}|reserve(Rs, Regs, Stk)]; -%reserve([{I,V}|Rs], [Other|Regs], Stk) -> [Other|reserve(Rs, Regs, Stk)]; -reserve([{I,V}|Rs], [], Stk) -> [{reserved,I,V}|reserve(Rs, [], Stk)]; -reserve([], Regs, _) -> Regs. - -extend_stack(Bef, Fb, Lf, Vdb) -> - Stk0 = clear_dead_stk(Bef#sr.stk, Fb, Vdb), - Saves = [V || {V,F,L} <- Vdb, - F < Fb, - L >= Lf, - not on_stack(V, Stk0)], - Stk1 = foldl(fun (V, Stk) -> put_stack(V, Stk) end, Stk0, Saves), - Bef#sr.stk ++ lists:duplicate(length(Stk1) - length(Bef#sr.stk), free). - -save_carefully(Bef, Fb, Lf, Vdb) -> - Stk = Bef#sr.stk, - %% New variables that are in use but not on stack. - New = [VFL || {V,F,L} = VFL <- Vdb, - F < Fb, - L >= Lf, - not on_stack(V, Stk)], - Saves = [V || {V,_,_} <- keysort(2, New)], - save_carefully(Saves, Bef, []). - -save_carefully([], Bef, Acc) -> {reverse(Acc),Bef}; -save_carefully([V|Vs], Bef, Acc) -> - case put_stack_carefully(V, Bef#sr.stk) of - error -> {reverse(Acc),Bef}; +reserve_1([{I,V}|Rs], [{reserved,I,_}|Regs], Stk) -> + [{reserved,I,V}|reserve_1(Rs, Regs, Stk)]; +reserve_1([{I,V}|Rs], [], Stk) -> + [{reserved,I,V}|reserve_1(Rs, [], Stk)]; +reserve_1([], Regs, _) -> Regs. + +%% extend_stack(Bef, FirstBefore, LastFrom, Vdb) -> Aft. +%% Extend the stack enough to fit all variables alive past LastFrom +%% and not already on the stack. + +extend_stack(#sr{stk=Stk0}=Bef, Fb, Lf, Vdb) -> + Stk1 = clear_dead_stk(Stk0, Fb, Vdb), + New = new_not_on_stack(Stk1, Fb, Lf, Vdb), + Stk2 = foldl(fun ({V,_,_}, Stk) -> put_stack(V, Stk) end, Stk1, New), + Stk = Stk0 ++ lists:duplicate(length(Stk2) - length(Stk0), free), + Bef#sr{stk=Stk}. + +%% save_carefully(Bef, FirstBefore, LastFrom, Vdb) -> {[SaveVar],Aft}. +%% Save variables which are used past current point and which are not +%% already on the stack, but only if the variables can be saved without +%% growing the stack frame. + +save_carefully(#sr{stk=Stk}=Bef, Fb, Lf, Vdb) -> + New0 = new_not_on_stack(Stk, Fb, Lf, Vdb), + New = keysort(2, New0), + save_carefully_1(New, Bef, []). + +save_carefully_1([{V,_,_}|Vs], #sr{reg=Regs,stk=Stk0}=Bef, Acc) -> + case put_stack_carefully(V, Stk0) of + error -> + {reverse(Acc),Bef}; Stk1 -> - SrcReg = fetch_reg(V, Bef#sr.reg), + SrcReg = fetch_reg(V, Regs), Move = {move,SrcReg,fetch_stack(V, Stk1)}, {x,_} = SrcReg, %Assertion - must be X register. - save_carefully(Vs, Bef#sr{stk=Stk1}, [Move|Acc]) - end. + save_carefully_1(Vs, Bef#sr{stk=Stk1}, [Move|Acc]) + end; +save_carefully_1([], Bef, Acc) -> + {reverse(Acc),Bef}. -x0_vars([], _Fb, _Lf, _Vdb) -> []; -x0_vars([#k_var{name=V}|_], Fb, _Lf, Vdb) -> - {V,F,_L} = VFL = vdb_find(V, Vdb), - x0_vars1([VFL], Fb, F, Vdb); -x0_vars([X0|_], Fb, Lf, Vdb) -> - x0_vars1([{X0,Lf,Lf}], Fb, Lf, Vdb). - -x0_vars1(X0, Fb, Xf, Vdb) -> - Vs0 = [VFL || {_V,F,L}=VFL <- Vdb, - F >= Fb, - L < Xf], - Vs1 = keysort(3, Vs0), - keysort(2, X0++Vs1). - -allocate_x0([], _, Bef) -> {[],Bef#sr{res=[]}}; -allocate_x0([{_,_,L}|Vs], I, Bef) when L =< I -> - allocate_x0(Vs, I, Bef); -allocate_x0([{V,_F,_L}=VFL|Vs], _, Bef) -> - {[VFL|Vs],Bef#sr{res=reserve_x0(V, Bef#sr.res)}}. - -reserve_x0(V, [_|Res]) -> [{0,V}|Res]; -reserve_x0(V, []) -> [{0,V}]. +%% top_level_block([Instruction], Bef, MaxRegs, St) -> [Instruction]. +%% For the top-level block, allocate a stack frame a necessary, +%% adjust Y register numbering and instructions that return +%% from the function. top_level_block(Keis, #sr{stk=[]}, _MaxRegs, #cg{need_frame=false}) -> Keis; @@ -1349,8 +1481,6 @@ guard_clause_cg(#k_guard_clause{anno=#l{vdb=Vdb},guard=G,body=B}, Fail, Bef, St0 guard_cg(#k_protected{arg=Ts,ret=Rs,anno=#l{i=I,vdb=Pdb}}, Fail, _Vdb, Bef, St) -> protected_cg(Ts, Rs, Fail, I, Pdb, Bef, St); -guard_cg(#cg_block{es=Ts,anno=#l{i=I,vdb=Bdb}}, Fail, _Vdb, Bef, St) -> - guard_cg_list(Ts, Fail, I, Bdb, Bef, St); guard_cg(#k_test{anno=#l{i=I},op=Test0,args=As,inverted=Inverted}, Fail, Vdb, Bef, St0) -> #k_remote{mod=#k_atom{val=erlang},name=#k_atom{val=Test}} = Test0, @@ -1368,6 +1498,18 @@ guard_cg(G, _Fail, Vdb, Bef, St) -> %%ok = io:fwrite("cg ~w: ~p~n", [?LINE,{Aft}]), {Gis,Aft,St1}. +%% guard_cg_list([Kexpr], Fail, I, Vdb, StackReg, St) -> +%% {[Ainstr],StackReg,St}. + +guard_cg_list(Kes, Fail, I, Vdb, Bef, St0) -> + {Keis,{Aft,St1}} = + flatmapfoldl(fun (Ke, {Inta,Sta}) -> + {Keis,Intb,Stb} = + guard_cg(Ke, Fail, Vdb, Inta, Sta), + {Keis,{Intb,Stb}} + end, {Bef,St0}, need_heap(Kes, I)), + {Keis,Aft,St1}. + %% protected_cg([Kexpr], [Ret], Fail, I, Vdb, Bef, St) -> {[Ainstr],Aft,St}. %% Do a protected. Protecteds without return values are just done %% for effect, the return value is not checked, success passes on to @@ -1424,18 +1566,6 @@ test_cg(Test, As, Fail, I, Vdb, Bef, St) -> Aft = clear_dead(Bef, I, Vdb), {[beam_utils:bif_to_test(Test, Args, {f,Fail})],Aft,St}. -%% guard_cg_list([Kexpr], Fail, I, Vdb, StackReg, St) -> -%% {[Ainstr],StackReg,St}. - -guard_cg_list(Kes, Fail, I, Vdb, Bef, St0) -> - {Keis,{Aft,St1}} = - flatmapfoldl(fun (Ke, {Inta,Sta}) -> - {Keis,Intb,Stb} = - guard_cg(Ke, Fail, Vdb, Inta, Sta), - {Keis,{Intb,Stb}} - end, {Bef,St0}, need_heap(Kes, I)), - {Keis,Aft,St1}. - %% match_fmf(Fun, LastFail, State, [Clause]) -> {Is,Aft,State}. %% This is a special flatmapfoldl for match code gen where we %% generate a "failure" label for each clause. The last clause uses @@ -2202,13 +2332,12 @@ break_cg(Bs, Le, Vdb, Bef, St) -> {Ms ++ [{jump,{f,St#cg.break}}], Int#sr{reg=clear_regs(Int#sr.reg)},St}. -guard_break_cg(Bs, Locked, #l{i=I}, Vdb, #sr{reg=Reg0}=Bef, St) -> - RegLocked = get_locked_regs(Reg0, Locked), - #sr{reg=Reg1} = Int = clear_dead(Bef#sr{reg=RegLocked}, I, Vdb), +guard_break_cg(Bs, #l{i=I}, Vdb, #sr{reg=Reg0}=Bef, St) -> + #sr{reg=Reg1} = Int = clear_dead(Bef, I, Vdb), Reg2 = trim_free(Reg1), NumLocked = length(Reg2), Moves0 = gen_moves(Bs, Bef, NumLocked, []), - Moves = order_moves(Moves0, find_scratch_reg(RegLocked)), + Moves = order_moves(Moves0, find_scratch_reg(Reg0)), {BreakVars,_} = mapfoldl(fun(_, RegNum) -> {{RegNum,gbreakvar},RegNum+1} end, length(Reg2), Bs), @@ -2216,20 +2345,6 @@ guard_break_cg(Bs, Locked, #l{i=I}, Vdb, #sr{reg=Reg0}=Bef, St) -> Aft = Int#sr{reg=Reg}, {Moves ++ [{jump,{f,St#cg.break}}],Aft,St}. -get_locked_regs([R|Rs0], Preserve) -> - case {get_locked_regs(Rs0, Preserve),R} of - {[],{_,V}} -> - case lists:member(V, Preserve) of - true -> [R]; - false -> [] - end; - {[],_} -> - []; - {Rs,_} -> - [R|Rs] - end; -get_locked_regs([], _) -> []. - %% cg_reg_arg(Arg0, Info) -> Arg %% cg_reg_args([Arg0], Info) -> [Arg] %% Convert argument[s] into registers. Literal values are returned unchanged. @@ -2370,21 +2485,21 @@ break_up_cycle1(Dst, [M|Path], LastMove) -> %% clear_dead(Sr, Until, Vdb) -> Aft. %% Remove all variables in Sr which have died AT ALL so far. -clear_dead(Sr, Until, Vdb) -> - Sr#sr{reg=clear_dead_reg(Sr, Until, Vdb), - stk=clear_dead_stk(Sr#sr.stk, Until, Vdb)}. +clear_dead(#sr{stk=Stk}=Sr0, Until, Vdb) -> + Sr = Sr0#sr{reg=clear_dead_reg(Sr0, Until, Vdb), + stk=clear_dead_stk(Stk, Until, Vdb)}, + reserve(Sr). clear_dead_reg(Sr, Until, Vdb) -> - Reg = [case R of - {_I,V} = IV -> - case vdb_find(V, Vdb) of - {V,_,L} when L > Until -> IV; - _ -> free %Remove anything else - end; - {reserved,_I,_V} = Reserved -> Reserved; - free -> free - end || R <- Sr#sr.reg], - reserve(Sr#sr.res, Reg, Sr#sr.stk). + [case R of + {_I,V} = IV -> + case vdb_find(V, Vdb) of + {V,_,L} when L > Until -> IV; + _ -> free %Remove anything else + end; + {reserved,_I,_V}=Reserved -> Reserved; + free -> free + end || R <- Sr#sr.reg]. clear_dead_stk(Stk, Until, Vdb) -> [case S of @@ -2456,16 +2571,25 @@ adjust_stack(Bef, Fb, Lf, Vdb) -> save_stack(Stk0, Fb, Lf, Vdb) -> %% New variables that are in use but not on stack. - New = [VFL || {V,F,L} = VFL <- Vdb, - F < Fb, - L >= Lf, - not on_stack(V, Stk0)], + New = new_not_on_stack(Stk0, Fb, Lf, Vdb), + %% Add new variables that are not just dropped immediately. %% N.B. foldr works backwards from the end!! Saves = [V || {V,_,_} <- keysort(3, New)], Stk1 = foldr(fun (V, Stk) -> put_stack(V, Stk) end, Stk0, Saves), {Stk1,Saves}. +%% new_not_on_stack(Stack, FirstBefore, LastFrom, Vdb) -> +%% [{Variable,First,Last}] +%% Return information about all variables that are used past current +%% point and that are not already on the stack. + +new_not_on_stack(Stk, Fb, Lf, Vdb) -> + [VFL || {V,F,L} = VFL <- Vdb, + F < Fb, + L >= Lf, + not on_stack(V, Stk)]. + %% saves([SaveVar], Reg, Stk) -> [{move,Reg,Stk}]. %% Generate move instructions to save variables onto stack. The %% stack/reg info used is that after the new stack has been made. diff --git a/lib/compiler/src/v3_core.erl b/lib/compiler/src/v3_core.erl index cc8ea475d2..6029b91cdc 100644 --- a/lib/compiler/src/v3_core.erl +++ b/lib/compiler/src/v3_core.erl @@ -2482,9 +2482,11 @@ cexpr(#icase{anno=A,args=Largs,clauses=Lcs,fc=Lfc}, As, St0) -> cexpr(#ireceive1{anno=A,clauses=Lcs}, As, St0) -> Exp = intersection(A#a.ns, As), %Exports {Ccs,St1} = cclauses(Lcs, Exp, St0), + True = #c_literal{val=true}, + Action = core_lib:make_values(lists:duplicate(1+length(Exp), True)), {#c_receive{anno=A#a.anno, clauses=Ccs, - timeout=#c_literal{val=infinity},action=#c_literal{val=true}}, + timeout=#c_literal{val=infinity},action=Action}, Exp,A#a.us,St1}; cexpr(#ireceive2{anno=A,clauses=Lcs,timeout=Lto,action=Les}, As, St0) -> Exp = intersection(A#a.ns, As), %Exports diff --git a/lib/compiler/src/v3_kernel.erl b/lib/compiler/src/v3_kernel.erl index 3eea058153..23625b1f2e 100644 --- a/lib/compiler/src/v3_kernel.erl +++ b/lib/compiler/src/v3_kernel.erl @@ -108,6 +108,7 @@ copy_anno(Kdst, Ksrc) -> -record(iclause, {anno=[],isub,osub,pats,guard,body}). -record(ireceive_accept, {anno=[],arg}). -record(ireceive_next, {anno=[],arg}). +-record(ignored, {anno=[]}). -type warning() :: term(). % XXX: REFINE @@ -489,7 +490,7 @@ make_alt(First0, Then0) -> Then1 = pre_seq(droplast(Then0), last(Then0)), First2 = make_protected(First1), Then2 = make_protected(Then1), - Body = #k_atom{val=ignored}, + Body = #ignored{}, First3 = #k_guard_clause{guard=First2,body=Body}, Then3 = #k_guard_clause{guard=Then2,body=Body}, First = #k_guard{clauses=[First3]}, @@ -2225,7 +2226,9 @@ ubody(E, return, St0) -> {Ea,Pa,St1} = force_atomic(E, St0), ubody(pre_seq(Pa, #ivalues{args=[Ea]}), return, St1) end; -ubody(E, {break,_Rs} = Break, St0) -> +ubody(#ignored{}, {break,_} = Break, St) -> + ubody(#ivalues{args=[]}, Break, St); +ubody(E, {break,[_]} = Break, St0) -> %%ok = io:fwrite("ubody ~w:~p~n", [?LINE,{E,Br}]), %% Exiting expressions need no trailing break. case is_exit_expr(E) of @@ -2233,6 +2236,16 @@ ubody(E, {break,_Rs} = Break, St0) -> false -> {Ea,Pa,St1} = force_atomic(E, St0), ubody(pre_seq(Pa, #ivalues{args=[Ea]}), Break, St1) + end; +ubody(E, {break,Rs}=Break, St0) -> + case is_exit_expr(E) of + true -> + uexpr(E, return, St0); + false -> + {Vs,St1} = new_vars(length(Rs), St0), + Iset = #iset{vars=Vs,arg=E}, + PreSeq = pre_seq([Iset], #ivalues{args=Vs}), + ubody(PreSeq, Break, St1) end. iletrec_funs(#iletrec{defs=Fs}, St0) -> diff --git a/lib/compiler/src/v3_kernel.hrl b/lib/compiler/src/v3_kernel.hrl index 87011b7680..7cd30b25a8 100644 --- a/lib/compiler/src/v3_kernel.hrl +++ b/lib/compiler/src/v3_kernel.hrl @@ -79,7 +79,7 @@ -record(k_guard_clause, {anno=[],guard,body}). -record(k_break, {anno=[],args=[]}). --record(k_guard_break, {anno=[],args=[],locked=[]}). +-record(k_guard_break, {anno=[],args=[]}). -record(k_return, {anno=[],args=[]}). %%k_get_anno(Thing) -> element(2, Thing). |