This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch] Get rid of stack trampolines for nested functions
- From: "Richard Earnshaw (lists)" <Richard dot Earnshaw at arm dot com>
- To: Eric Botcazou <ebotcazou at adacore dot com>, gcc-patches at gcc dot gnu dot org
- Cc: Ramana Radhakrishnan <Ramana dot Radhakrishnan at arm dot com>
- Date: Mon, 25 Jul 2016 14:05:53 +0100
- Subject: Re: [patch] Get rid of stack trampolines for nested functions
- Authentication-results: sourceware.org; auth=none
- References: <6524055.MLc1u6juUE@polaris>
On 29/06/16 23:08, Eric Botcazou wrote:
> Index: config/aarch64/aarch64.h
> ===================================================================
> --- config/aarch64/aarch64.h (revision 237789)
> +++ config/aarch64/aarch64.h (working copy)
> @@ -779,6 +779,9 @@ typedef struct
> correctly. */
> #define TRAMPOLINE_SECTION text_section
>
> +/* Use custom descriptors instead of trampolines when possible. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 1
> +
Eric,
If I understand how this is supposed to work then this is not
future-proof against changes to the architecture. The bottom two bits
in both AArch32 (arm) and AArch64 are reserved for future use by the
architecture; they must not be used by software for tricks like this.
As has already been seen in AArch32 state, bit-0 is used to indicate the
ARM/Thumb ISA selection.
The patch to arm.h is similarly problematic in this regard.
R.
> Hi,
>
> this patch implements generic support for the elimination of stack trampolines
> and, consequently, of the need to make the stack executable when pointers to
> nested functions are used. That's done on a per-language and per-target basis
> (i.e. there is 1 language hook and 1 target hook to parameterize it) and there
> are no changes whatsoever in code generation if both are not turned on (and
> the patch implements a -ftrampolines option to let the user override them).
>
> The idea is based on the fact that, for targets using function descriptors as
> per their ABI like IA-64, AIX or VMS platforms, stack trampolines "degenerate"
> into descriptors built at run time on the stack and thus made up of data only,
> which in turn means that the stack doesn't need to be made executable.
>
> This descriptor-based scheme is implemented generically for nested functions,
> i.e. the nested function lowering pass builds generic descriptors instead of
> trampolines on the stack when encountering pointers to nested functions, which
> means that there are 2 kinds of pointers to functions and therefore a run-time
> identification mechanism is needed for indirect calls to distinguish them.
>
> Because of that, enabling the support breaks binary compatibility (for code
> manipulating pointers to nested functions). That's OK for Ada and nested
> functions are first-class citizens in the language anyway so we really need
> this, but not for C so for example Ada doesn't use it at the interface with C
> (when objects have "convention C" in Ada parlance).
>
> This was bootstrapped/regtested on x86_64-suse-linux but AdaCore has been
> using it on native platforms (Linux, Windows, Solaris, etc) for years.
>
> OK for the mainline?
>
>
> 2016-06-29 Eric Botcazou <ebotcazou@adacore.com>
>
> PR ada/37139
> PR ada/67205
> * common.opt (-ftrampolines): New option.
> * doc/invoke.texi (Code Gen Options): Document it.
> * doc/tm.texi.in (Trampolines): Add TARGET_CUSTOM_FUNCTION_DESCRIPTORS
> * doc/tm.texi: Regenerate.
> * builtins.def: Add init_descriptor and adjust_descriptor.
> * builtins.c (expand_builtin_init_trampoline): Do not issue a warning
> on platforms with descriptors.
> (expand_builtin_init_descriptor): New function.
> (expand_builtin_adjust_descriptor): Likewise.
> (expand_builtin) <BUILT_IN_INIT_DESCRIPTOR>: New case.
> <BUILT_IN_ADJUST_DESCRIPTOR>: Likewise.
> * calls.c (prepare_call_address): Remove SIBCALLP parameter and add
> FLAGS parameter. Deal with indirect calls by descriptor and adjust.
> Set STATIC_CHAIN_REG_P on the static chain register, if any.
> (call_expr_flags): Set ECF_BY_DESCRIPTOR for calls by descriptor.
> (expand_call): Likewise. Move around call to prepare_call_address
> and pass all flags to it.
> * cfgexpand.c (expand_call_stmt): Reinstate CALL_EXPR_BY_DESCRIPTOR.
> * gimple.h (enum gf_mask): New GF_CALL_BY_DESCRIPTOR value.
> (gimple_call_set_by_descriptor): New setter.
> (gimple_call_by_descriptor_p): New getter.
> * gimple.c (gimple_build_call_from_tree): Set CALL_EXPR_BY_DESCRIPTOR.
> (gimple_call_flags): Deal with GF_CALL_BY_DESCRIPTOR.
> * langhooks.h (struct lang_hooks): Add custom_function_descriptors.
> * langhooks-def.h (LANG_HOOKS_CUSTOM_FUNCTION_DESCRIPTORS): Define.
> (LANG_HOOKS_INITIALIZER): Add LANG_HOOKS_CUSTOM_FUNCTION_DESCRIPTORS.
> * rtl.h (STATIC_CHAIN_REG_P): New macro.
> * rtlanal.c (find_first_parameter_load): Skip static chain registers.
> * target.def (custom_function_descriptors): New POD hook.
> * tree.h (FUNC_ADDR_BY_DESCRIPTOR): New flag on ADDR_EXPR.
> (CALL_EXPR_BY_DESCRIPTOR): New flag on CALL_EXPR.
> * tree-core.h (ECF_BY_DESCRIPTOR): New mask.
> Document FUNC_ADDR_BY_DESCRIPTOR and CALL_EXPR_BY_DESCRIPTOR.
> * tree.c (make_node_stat) <tcc_declaration>: Set function alignment to
> DEFAULT_FUNCTION_ALIGNMENT instead of FUNCTION_BOUNDARY.
> (build_common_builtin_nodes): Initialize init_descriptor and
> adjust_descriptor.
> * tree-nested.c: Include target.h.
> (struct nesting_info): Add 'any_descr_created' field.
> (get_descriptor_type): New function.
> (lookup_element_for_decl): New function extracted from...
> (create_field_for_decl): Likewise.
> (lookup_tramp_for_decl): ...here. Adjust.
> (lookup_descr_for_decl): New function.
> (convert_tramp_reference_op): Deal with descriptors.
> (build_init_call_stmt): New function extracted from...
> (finalize_nesting_tree_1): ...here. Adjust and deal with descriptors.
> * defaults.h (DEFAULT_FUNCTION_ALIGNMENT): Define.
> (TRAMPOLINE_ALIGNMENT): Set to above instead of FUNCTION_BOUNDARY.
> * config/aarch64/aarch64.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS):Define
> * config/alpha/alpha.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS): Likewise.
> * config/arm/arm.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS): Likewise.
> * config/arm/arm.c (arm_function_ok_for_sibcall): Return false for an
> indirect call by descriptor if all the argument registers are used.
> * config/i386/i386.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS): Define.
> * config/ia64/ia64.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS): Likewise.
> * config/mips/mips.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS): Likewise.
> * config/pa/pa.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS): Likewise.
> * config/rs6000/rs6000.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS):Likewise
> * config/sparc/sparc.h (TARGET_CUSTOM_FUNCTION_DESCRIPTORS): Likewise.
> ada/
> * gcc-interface/misc.c (LANG_HOOKS_CUSTOM_FUNCTION_DESCRIPTORS):Define
> * gcc-interface/trans.c (Attribute_to_gnu) <Attr_Access>: Deal with
> a zero TARGET_CUSTOM_FUNCTION_DESCRIPTORSspecially for 'Code_Address.
> Otherwise, if TARGET_CUSTOM_FUNCTION_DESCRIPTORS is positive, set
> FUNC_ADDR_BY_DESCRIPTOR for 'Access/'Unrestricted_Access of nested
> subprograms if the type can use an internal representation.
> (call_to_gnu): Likewise, but set CALL_EXPR_BY_DESCRIPTOR on indirect
> calls if the type can use an internal representation.
>
>
> 2016-06-29 Eric Botcazou <ebotcazou@adacore.com>
>
> * gnat.dg/trampoline3.adb: New test.
> * gnat.dg/trampoline4.adb: Likewise.
>
>
> p.diff
>
>
> Index: ada/gcc-interface/misc.c
> ===================================================================
> --- ada/gcc-interface/misc.c (revision 237848)
> +++ ada/gcc-interface/misc.c (working copy)
> @@ -1416,6 +1416,8 @@ get_lang_specific (tree node)
> #define LANG_HOOKS_EH_PERSONALITY gnat_eh_personality
> #undef LANG_HOOKS_DEEP_UNSHARING
> #define LANG_HOOKS_DEEP_UNSHARING true
> +#undef LANG_HOOKS_CUSTOM_FUNCTION_DESCRIPTORS
> +#define LANG_HOOKS_CUSTOM_FUNCTION_DESCRIPTORS true
>
> struct lang_hooks lang_hooks = LANG_HOOKS_INITIALIZER;
>
> Index: ada/gcc-interface/trans.c
> ===================================================================
> --- ada/gcc-interface/trans.c (revision 237850)
> +++ ada/gcc-interface/trans.c (working copy)
> @@ -1702,6 +1702,17 @@ Attribute_to_gnu (Node_Id gnat_node, tre
>
> if (TREE_CODE (gnu_expr) == ADDR_EXPR)
> TREE_NO_TRAMPOLINE (gnu_expr) = TREE_CONSTANT (gnu_expr) = 1;
> +
> + /* On targets for which function symbols denote a descriptor, the
> + code address is stored within the first slot of the descriptor
> + so we do an additional dereference:
> + result = *((result_type *) result)
> + where we expect result to be of some pointer type already. */
> + if (targetm.calls.custom_function_descriptors == 0)
> + gnu_result
> + = build_unary_op (INDIRECT_REF, NULL_TREE,
> + convert (build_pointer_type (gnu_result_type),
> + gnu_result));
> }
>
> /* For 'Access, issue an error message if the prefix is a C++ method
> @@ -1728,10 +1739,19 @@ Attribute_to_gnu (Node_Id gnat_node, tre
> /* Also check the inlining status. */
> check_inlining_for_nested_subprog (TREE_OPERAND (gnu_expr, 0));
>
> - /* Check that we're not violating the No_Implicit_Dynamic_Code
> - restriction. Be conservative if we don't know anything
> - about the trampoline strategy for the target. */
> - Check_Implicit_Dynamic_Code_Allowed (gnat_node);
> + /* Moreover, for 'Access or 'Unrestricted_Access with non-
> + foreign-compatible representation, mark the ADDR_EXPR so
> + that we can build a descriptor instead of a trampoline. */
> + if ((attribute == Attr_Access
> + || attribute == Attr_Unrestricted_Access)
> + && targetm.calls.custom_function_descriptors > 0
> + && Can_Use_Internal_Rep (Etype (gnat_node)))
> + FUNC_ADDR_BY_DESCRIPTOR (gnu_expr) = 1;
> +
> + /* Otherwise, we need to check that we are not violating the
> + No_Implicit_Dynamic_Code restriction. */
> + else if (targetm.calls.custom_function_descriptors != 0)
> + Check_Implicit_Dynamic_Code_Allowed (gnat_node);
> }
> }
> break;
> @@ -4228,6 +4248,7 @@ Call_to_gnu (Node_Id gnat_node, tree *gn
> tree gnu_after_list = NULL_TREE;
> tree gnu_retval = NULL_TREE;
> tree gnu_call, gnu_result;
> + bool by_descriptor = false;
> bool went_into_elab_proc = false;
> bool pushed_binding_level = false;
> Entity_Id gnat_formal;
> @@ -4267,7 +4288,15 @@ Call_to_gnu (Node_Id gnat_node, tree *gn
> type the access type is pointing to. Otherwise, get the formals from the
> entity being called. */
> if (Nkind (Name (gnat_node)) == N_Explicit_Dereference)
> - gnat_formal = First_Formal_With_Extras (Etype (Name (gnat_node)));
> + {
> + gnat_formal = First_Formal_With_Extras (Etype (Name (gnat_node)));
> +
> + /* If the access type doesn't require foreign-compatible representation,
> + be prepared for descriptors. */
> + if (targetm.calls.custom_function_descriptors > 0
> + && Can_Use_Internal_Rep (Etype (Prefix (Name (gnat_node)))))
> + by_descriptor = true;
> + }
> else if (Nkind (Name (gnat_node)) == N_Attribute_Reference)
> /* Assume here that this must be 'Elab_Body or 'Elab_Spec. */
> gnat_formal = Empty;
> @@ -4668,6 +4697,7 @@ Call_to_gnu (Node_Id gnat_node, tree *gn
>
> gnu_call
> = build_call_vec (gnu_result_type, gnu_subprog_addr, gnu_actual_vec);
> + CALL_EXPR_BY_DESCRIPTOR (gnu_call) = by_descriptor;
> set_expr_location_from_node (gnu_call, gnat_node);
>
> /* If we have created a temporary for the return value, initialize it. */
> Index: builtins.c
> ===================================================================
> --- builtins.c (revision 237789)
> +++ builtins.c (working copy)
> @@ -4621,8 +4621,9 @@ expand_builtin_init_trampoline (tree exp
> {
> trampolines_created = 1;
>
> - warning_at (DECL_SOURCE_LOCATION (t_func), OPT_Wtrampolines,
> - "trampoline generated for nested function %qD", t_func);
> + if (targetm.calls.custom_function_descriptors != 0)
> + warning_at (DECL_SOURCE_LOCATION (t_func), OPT_Wtrampolines,
> + "trampoline generated for nested function %qD", t_func);
> }
>
> return const0_rtx;
> @@ -4644,6 +4645,57 @@ expand_builtin_adjust_trampoline (tree e
> return tramp;
> }
>
> +/* Expand a call to the builtin descriptor initialization routine.
> + A descriptor is made up of a couple of pointers to the static
> + chain and the code entry in this order. */
> +
> +static rtx
> +expand_builtin_init_descriptor (tree exp)
> +{
> + tree t_descr, t_func, t_chain;
> + rtx m_descr, r_descr, r_func, r_chain;
> +
> + if (!validate_arglist (exp, POINTER_TYPE, POINTER_TYPE, POINTER_TYPE,
> + VOID_TYPE))
> + return NULL_RTX;
> +
> + t_descr = CALL_EXPR_ARG (exp, 0);
> + t_func = CALL_EXPR_ARG (exp, 1);
> + t_chain = CALL_EXPR_ARG (exp, 2);
> +
> + r_descr = expand_normal (t_descr);
> + m_descr = gen_rtx_MEM (BLKmode, r_descr);
> + MEM_NOTRAP_P (m_descr) = 1;
> +
> + r_func = expand_normal (t_func);
> + r_chain = expand_normal (t_chain);
> +
> + /* Generate insns to initialize the descriptor. */
> + emit_move_insn (adjust_address_nv (m_descr, Pmode, 0), r_chain);
> + emit_move_insn (adjust_address_nv (m_descr, Pmode, UNITS_PER_WORD), r_func);
> +
> + return const0_rtx;
> +}
> +
> +/* Expand a call to the builtin descriptor adjustment routine. */
> +
> +static rtx
> +expand_builtin_adjust_descriptor (tree exp)
> +{
> + rtx tramp;
> +
> + if (!validate_arglist (exp, POINTER_TYPE, VOID_TYPE))
> + return NULL_RTX;
> +
> + tramp = expand_normal (CALL_EXPR_ARG (exp, 0));
> +
> + /* Unalign the descriptor to allow runtime identification. */
> + tramp
> + = plus_constant (Pmode, tramp, targetm.calls.custom_function_descriptors);
> +
> + return force_operand (tramp, NULL_RTX);
> +}
> +
> /* Expand the call EXP to the built-in signbit, signbitf or signbitl
> function. The function first checks whether the back end provides
> an insn to implement signbit for the respective mode. If not, it
> @@ -6221,6 +6273,11 @@ expand_builtin (tree exp, rtx target, rt
> case BUILT_IN_ADJUST_TRAMPOLINE:
> return expand_builtin_adjust_trampoline (exp);
>
> + case BUILT_IN_INIT_DESCRIPTOR:
> + return expand_builtin_init_descriptor (exp);
> + case BUILT_IN_ADJUST_DESCRIPTOR:
> + return expand_builtin_adjust_descriptor (exp);
> +
> case BUILT_IN_FORK:
> case BUILT_IN_EXECL:
> case BUILT_IN_EXECV:
> Index: builtins.def
> ===================================================================
> --- builtins.def (revision 237789)
> +++ builtins.def (working copy)
> @@ -856,6 +856,8 @@ DEF_C99_BUILTIN (BUILT_IN__EXIT2,
> DEF_BUILTIN_STUB (BUILT_IN_INIT_TRAMPOLINE, "__builtin_init_trampoline")
> DEF_BUILTIN_STUB (BUILT_IN_INIT_HEAP_TRAMPOLINE, "__builtin_init_heap_trampoline")
> DEF_BUILTIN_STUB (BUILT_IN_ADJUST_TRAMPOLINE, "__builtin_adjust_trampoline")
> +DEF_BUILTIN_STUB (BUILT_IN_INIT_DESCRIPTOR, "__builtin_init_descriptor")
> +DEF_BUILTIN_STUB (BUILT_IN_ADJUST_DESCRIPTOR, "__builtin_adjust_descriptor")
> DEF_BUILTIN_STUB (BUILT_IN_NONLOCAL_GOTO, "__builtin_nonlocal_goto")
>
> /* Implementing __builtin_setjmp. */
> Index: calls.c
> ===================================================================
> --- calls.c (revision 237789)
> +++ calls.c (working copy)
> @@ -183,18 +183,73 @@ static void restore_fixed_argument_area
>
> rtx
> prepare_call_address (tree fndecl_or_type, rtx funexp, rtx static_chain_value,
> - rtx *call_fusage, int reg_parm_seen, int sibcallp)
> + rtx *call_fusage, int reg_parm_seen, int flags)
> {
> /* Make a valid memory address and copy constants through pseudo-regs,
> but not for a constant address if -fno-function-cse. */
> if (GET_CODE (funexp) != SYMBOL_REF)
> - /* If we are using registers for parameters, force the
> - function address into a register now. */
> - funexp = ((reg_parm_seen
> - && targetm.small_register_classes_for_mode_p (FUNCTION_MODE))
> - ? force_not_mem (memory_address (FUNCTION_MODE, funexp))
> - : memory_address (FUNCTION_MODE, funexp));
> - else if (! sibcallp)
> + {
> + /* If it's an indirect call by descriptor, generate code to perform
> + runtime identification of the pointer and load the descriptor. */
> + if ((flags & ECF_BY_DESCRIPTOR) && !flag_trampolines)
> + {
> + const int bit_val = targetm.calls.custom_function_descriptors;
> + rtx call_lab = gen_label_rtx ();
> +
> + gcc_assert (fndecl_or_type && TYPE_P (fndecl_or_type));
> + fndecl_or_type
> + = build_decl (UNKNOWN_LOCATION, FUNCTION_DECL, NULL_TREE,
> + fndecl_or_type);
> + DECL_STATIC_CHAIN (fndecl_or_type) = 1;
> + rtx chain = targetm.calls.static_chain (fndecl_or_type, false);
> +
> + /* Avoid long live ranges around function calls. */
> + funexp = copy_to_mode_reg (Pmode, funexp);
> +
> + if (REG_P (chain))
> + emit_insn (gen_rtx_CLOBBER (VOIDmode, chain));
> +
> + /* Emit the runtime identification pattern. */
> + rtx mask = gen_rtx_AND (Pmode, funexp, GEN_INT (bit_val));
> + emit_cmp_and_jump_insns (mask, const0_rtx, EQ, NULL_RTX, Pmode, 1,
> + call_lab);
> +
> + /* Statically predict the branch to very likely taken. */
> + rtx_insn *insn = get_last_insn ();
> + if (JUMP_P (insn))
> + predict_insn_def (insn, PRED_BUILTIN_EXPECT, TAKEN);
> +
> + /* Load the descriptor. */
> + rtx mem = gen_rtx_MEM (Pmode,
> + plus_constant (Pmode, funexp, - bit_val));
> + MEM_NOTRAP_P (mem) = 1;
> + emit_move_insn (chain, mem);
> + mem = gen_rtx_MEM (Pmode,
> + plus_constant (Pmode, funexp,
> + UNITS_PER_WORD - bit_val));
> + MEM_NOTRAP_P (mem) = 1;
> + emit_move_insn (funexp, mem);
> +
> + emit_label (call_lab);
> +
> + if (REG_P (chain))
> + {
> + use_reg (call_fusage, chain);
> + STATIC_CHAIN_REG_P (chain) = 1;
> + }
> +
> + /* Make sure we're not going to be overwritten below. */
> + gcc_assert (!static_chain_value);
> + }
> +
> + /* If we are using registers for parameters, force the
> + function address into a register now. */
> + funexp = ((reg_parm_seen
> + && targetm.small_register_classes_for_mode_p (FUNCTION_MODE))
> + ? force_not_mem (memory_address (FUNCTION_MODE, funexp))
> + : memory_address (FUNCTION_MODE, funexp));
> + }
> + else if (!(flags & ECF_SIBCALL))
> {
> if (!NO_FUNCTION_CSE && optimize && ! flag_no_function_cse)
> funexp = force_reg (Pmode, funexp);
> @@ -211,7 +266,10 @@ prepare_call_address (tree fndecl_or_typ
>
> emit_move_insn (chain, static_chain_value);
> if (REG_P (chain))
> - use_reg (call_fusage, chain);
> + {
> + use_reg (call_fusage, chain);
> + STATIC_CHAIN_REG_P (chain) = 1;
> + }
> }
>
> return funexp;
> @@ -792,11 +850,13 @@ call_expr_flags (const_tree t)
> flags = internal_fn_flags (CALL_EXPR_IFN (t));
> else
> {
> - t = TREE_TYPE (CALL_EXPR_FN (t));
> - if (t && TREE_CODE (t) == POINTER_TYPE)
> - flags = flags_from_decl_or_type (TREE_TYPE (t));
> + tree type = TREE_TYPE (CALL_EXPR_FN (t));
> + if (type && TREE_CODE (type) == POINTER_TYPE)
> + flags = flags_from_decl_or_type (TREE_TYPE (type));
> else
> flags = 0;
> + if (CALL_EXPR_BY_DESCRIPTOR (t))
> + flags |= ECF_BY_DESCRIPTOR;
> }
>
> return flags;
> @@ -2633,6 +2693,8 @@ expand_call (tree exp, rtx target, int i
> {
> fntype = TREE_TYPE (TREE_TYPE (addr));
> flags |= flags_from_decl_or_type (fntype);
> + if (CALL_EXPR_BY_DESCRIPTOR (exp))
> + flags |= ECF_BY_DESCRIPTOR;
> }
> rettype = TREE_TYPE (exp);
>
> @@ -3344,6 +3406,13 @@ expand_call (tree exp, rtx target, int i
> if (STRICT_ALIGNMENT)
> store_unaligned_arguments_into_pseudos (args, num_actuals);
>
> + /* Prepare the address of the call. This must be done before any
> + register parameters is loaded for find_first_parameter_load to
> + work properly in the presence of descriptors. */
> + funexp = prepare_call_address (fndecl ? fndecl : fntype, funexp,
> + static_chain_value, &call_fusage,
> + reg_parm_seen, flags);
> +
> /* Now store any partially-in-registers parm.
> This is the last place a block-move can happen. */
> if (reg_parm_seen)
> @@ -3454,10 +3523,6 @@ expand_call (tree exp, rtx target, int i
> }
>
> after_args = get_last_insn ();
> - funexp = prepare_call_address (fndecl ? fndecl : fntype, funexp,
> - static_chain_value, &call_fusage,
> - reg_parm_seen, pass == 0);
> -
> load_register_parameters (args, num_actuals, &call_fusage, flags,
> pass == 0, &sibcall_failure);
>
> Index: cfgexpand.c
> ===================================================================
> --- cfgexpand.c (revision 237789)
> +++ cfgexpand.c (working copy)
> @@ -2636,6 +2636,7 @@ expand_call_stmt (gcall *stmt)
> else
> CALL_FROM_THUNK_P (exp) = gimple_call_from_thunk_p (stmt);
> CALL_EXPR_VA_ARG_PACK (exp) = gimple_call_va_arg_pack_p (stmt);
> + CALL_EXPR_BY_DESCRIPTOR (exp) = gimple_call_by_descriptor_p (stmt);
> SET_EXPR_LOCATION (exp, gimple_location (stmt));
> CALL_WITH_BOUNDS_P (exp) = gimple_call_with_bounds_p (stmt);
>
> Index: common.opt
> ===================================================================
> --- common.opt (revision 237789)
> +++ common.opt (working copy)
> @@ -2303,6 +2303,10 @@ ftracer
> Common Report Var(flag_tracer) Optimization
> Perform superblock formation via tail duplication.
>
> +ftrampolines
> +Common Report Var(flag_trampolines) Init(0)
> +Always generate trampolines for pointers to nested functions
> +
> ; Zero means that floating-point math operations cannot generate a
> ; (user-visible) trap. This is the case, for example, in nonstop
> ; IEEE 754 arithmetic.
> Index: config/aarch64/aarch64.h
> ===================================================================
> --- config/aarch64/aarch64.h (revision 237789)
> +++ config/aarch64/aarch64.h (working copy)
> @@ -779,6 +779,9 @@ typedef struct
> correctly. */
> #define TRAMPOLINE_SECTION text_section
>
> +/* Use custom descriptors instead of trampolines when possible. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 1
> +
> /* To start with. */
> #define BRANCH_COST(SPEED_P, PREDICTABLE_P) \
> (aarch64_branch_cost (SPEED_P, PREDICTABLE_P))
> Index: config/alpha/alpha.h
> ===================================================================
> --- config/alpha/alpha.h (revision 237789)
> +++ config/alpha/alpha.h (working copy)
> @@ -996,3 +996,6 @@ extern long alpha_auto_offset;
> #define NO_IMPLICIT_EXTERN_C
>
> #define TARGET_SUPPORTS_WIDE_INT 1
> +
> +/* Use custom descriptors instead of trampolines when possible if not VMS. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS (TARGET_ABI_OPEN_VMS ? 0 : 1)
> Index: config/arm/arm.c
> ===================================================================
> --- config/arm/arm.c (revision 237789)
> +++ config/arm/arm.c (working copy)
> @@ -6781,6 +6781,29 @@ arm_function_ok_for_sibcall (tree decl,
> && DECL_WEAK (decl))
> return false;
>
> + /* We cannot do a tailcall for an indirect call by descriptor if all the
> + argument registers are used because the only register left to load the
> + address is IP and it will already contain the static chain. */
> + if (!decl && CALL_EXPR_BY_DESCRIPTOR (exp) && !flag_trampolines)
> + {
> + tree fntype = TREE_TYPE (TREE_TYPE (CALL_EXPR_FN (exp)));
> + CUMULATIVE_ARGS cum;
> + cumulative_args_t cum_v;
> +
> + arm_init_cumulative_args (&cum, fntype, NULL_RTX, NULL_TREE);
> + cum_v = pack_cumulative_args (&cum);
> +
> + for (tree t = TYPE_ARG_TYPES (fntype); t; t = TREE_CHAIN (t))
> + {
> + tree type = TREE_VALUE (t);
> + if (!VOID_TYPE_P (type))
> + arm_function_arg_advance (cum_v, TYPE_MODE (type), type, true);
> + }
> +
> + if (!arm_function_arg (cum_v, SImode, integer_type_node, true))
> + return false;
> + }
> +
> /* Everything else is ok. */
> return true;
> }
> Index: config/arm/arm.h
> ===================================================================
> --- config/arm/arm.h (revision 237789)
> +++ config/arm/arm.h (working copy)
> @@ -1632,6 +1632,10 @@ typedef struct
>
> /* Alignment required for a trampoline in bits. */
> #define TRAMPOLINE_ALIGNMENT 32
> +
> +/* Use custom descriptors instead of trampolines when possible, but
> + we cannot use bit #0 because it is the ARM/Thumb selection bit. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 2
>
> /* Addressing modes, and classification of registers for them. */
> #define HAVE_POST_INCREMENT 1
> Index: config/i386/i386.h
> ===================================================================
> --- config/i386/i386.h (revision 237789)
> +++ config/i386/i386.h (working copy)
> @@ -2660,6 +2660,9 @@ extern void debug_dispatch_window (int);
>
> #define TARGET_SUPPORTS_WIDE_INT 1
>
> +/* Use custom descriptors instead of trampolines when possible. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 1
> +
> /*
> Local variables:
> version-control: t
> Index: config/ia64/ia64.h
> ===================================================================
> --- config/ia64/ia64.h (revision 237789)
> +++ config/ia64/ia64.h (working copy)
> @@ -1714,4 +1714,7 @@ struct GTY(()) machine_function
> /* Switch on code for querying unit reservations. */
> #define CPU_UNITS_QUERY 1
>
> +/* IA-64 already uses descriptors for its standard calling sequence. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 0
> +
> /* End of ia64.h */
> Index: config/mips/mips.h
> ===================================================================
> --- config/mips/mips.h (revision 237789)
> +++ config/mips/mips.h (working copy)
> @@ -3413,3 +3413,6 @@ struct GTY(()) machine_function {
> #define ENABLE_LD_ST_PAIRS \
> (TARGET_LOAD_STORE_PAIRS && (TUNE_P5600 || TUNE_I6400) \
> && !TARGET_MICROMIPS && !TARGET_FIX_24K)
> +
> +/* Use custom descriptors instead of trampolines when possible. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 1
> Index: config/pa/pa.h
> ===================================================================
> --- config/pa/pa.h (revision 237789)
> +++ config/pa/pa.h (working copy)
> @@ -1313,3 +1313,6 @@ do { \
> seven and four instructions, respectively. */
> #define MAX_PCREL17F_OFFSET \
> (flag_pic ? (TARGET_HPUX ? 198164 : 221312) : 240000)
> +
> +/* HP-PA already uses descriptors for its standard calling sequence. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 0
> Index: config/rs6000/rs6000.h
> ===================================================================
> --- config/rs6000/rs6000.h (revision 237789)
> +++ config/rs6000/rs6000.h (working copy)
> @@ -2894,3 +2894,6 @@ extern GTY(()) tree rs6000_builtin_types
> extern GTY(()) tree rs6000_builtin_decls[RS6000_BUILTIN_COUNT];
>
> #define TARGET_SUPPORTS_WIDE_INT 1
> +
> +/* Use custom descriptors instead of trampolines when possible if not AIX. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS (DEFAULT_ABI == ABI_AIX ? 0 : 1)
> Index: config/sparc/sparc.h
> ===================================================================
> --- config/sparc/sparc.h (revision 237789)
> +++ config/sparc/sparc.h (working copy)
> @@ -1817,3 +1817,6 @@ extern int sparc_indent_opcode;
> #define SPARC_LOW_FE_EXCEPT_VALUES 0
>
> #define TARGET_SUPPORTS_WIDE_INT 1
> +
> +/* Use custom descriptors instead of trampolines when possible. */
> +#define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 1
> Index: defaults.h
> ===================================================================
> --- defaults.h (revision 237789)
> +++ defaults.h (working copy)
> @@ -1080,9 +1080,18 @@ see the files COPYING3 and COPYING.RUNTI
> #define CASE_VECTOR_PC_RELATIVE 0
> #endif
>
> +/* Force minimum alignment to be able to use the least significant bits
> + for distinguishing descriptor addresses from code addresses. */
> +#define DEFAULT_FUNCTION_ALIGNMENT \
> + (lang_hooks.custom_function_descriptors \
> + && targetm.calls.custom_function_descriptors > 0 \
> + ? MAX (FUNCTION_BOUNDARY, \
> + 2 * targetm.calls.custom_function_descriptors * BITS_PER_UNIT)\
> + : FUNCTION_BOUNDARY)
> +
> /* Assume that trampolines need function alignment. */
> #ifndef TRAMPOLINE_ALIGNMENT
> -#define TRAMPOLINE_ALIGNMENT FUNCTION_BOUNDARY
> +#define TRAMPOLINE_ALIGNMENT DEFAULT_FUNCTION_ALIGNMENT
> #endif
>
> /* Register mappings for target machines without register windows. */
> Index: doc/invoke.texi
> ===================================================================
> --- doc/invoke.texi (revision 237789)
> +++ doc/invoke.texi (working copy)
> @@ -498,7 +498,7 @@ Objective-C and Objective-C++ Dialects}.
> -fverbose-asm -fpack-struct[=@var{n}] @gol
> -fleading-underscore -ftls-model=@var{model} @gol
> -fstack-reuse=@var{reuse_level} @gol
> --ftrapv -fwrapv @gol
> +-ftrampolines -ftrapv -fwrapv @gol
> -fvisibility=@r{[}default@r{|}internal@r{|}hidden@r{|}protected@r{]} @gol
> -fstrict-volatile-bitfields -fsync-libcalls}
>
> @@ -11546,6 +11546,31 @@ unit, or if @option{-fpic} is not given
> The default without @option{-fpic} is @samp{initial-exec}; with
> @option{-fpic} the default is @samp{global-dynamic}.
>
> +@item -ftrampolines
> +@opindex ftrampolines
> +Always generate trampolines for pointers to nested functions.
> +
> +A trampoline is a small piece of data or code that is created at run
> +time on the stack when the address of a nested function is taken, and
> +is used to call the nested function indirectly. For some targets, it
> +is made up of data only and thus requires no special treatment. But,
> +for most targets, it is made up of code and thus requires the stack
> +to be made executable in order for the program to work properly.
> +
> +@option{-fno-trampolines} is enabled by default to let the compiler avoid
> +generating them if it computes that this is safe, on a case by case basis,
> +and replace them with descriptors. Descriptors are always made up of data
> +only, but the generated code must be prepared to deal with them.
> +
> +This option has no effects for any other languages than Ada as of this
> +writing. Moreover, code compiled with @option{-ftrampolines} and code
> +compiled with @option{-fno-trampolines} are not binary compatible if
> +nested functions are present. This option must therefore be used on
> +a program-wide basis and be manipulated with extreme care.
> +
> +This option has no effects for targets whose trampolines are made up of
> +data only, for example IA-64 targets, AIX or VMS platforms.
> +
> @item -fvisibility=@r{[}default@r{|}internal@r{|}hidden@r{|}protected@r{]}
> @opindex fvisibility
> Set the default ELF image symbol visibility to the specified option---all
> Index: doc/tm.texi
> ===================================================================
> --- doc/tm.texi (revision 237789)
> +++ doc/tm.texi (working copy)
> @@ -5181,6 +5181,25 @@ be returned; otherwise @var{addr} should
> If this hook is not defined, @var{addr} will be used for function calls.
> @end deftypefn
>
> +@deftypevr {Target Hook} int TARGET_CUSTOM_FUNCTION_DESCRIPTORS
> +This hook should be defined to a power of 2 if the target will benefit
> +from the use of custom descriptors for nested functions instead of the
> +standard trampolines. Such descriptors are created at run time on the
> +stack and made up of data only, but they are non-standard so the generated
> +code must be prepared to deal with them. This hook should be defined to 0
> +if the target uses function descriptors for its standard calling sequence,
> +like for example HP-PA or IA-64. Using descriptors for nested functions
> +eliminates the need for trampolines that reside on the stack and require
> +it to be made executable.
> +
> +The value of the macro is used to parameterize the run-time identification
> +scheme implemented to distinguish descriptors from function addresses: it
> +gives the number of bytes by which their address is shifted in comparison
> +with function addresses. The value of 1 will generally work, unless it is
> +already used by the target for a similar purpose, like for example on ARM
> +where it is used to distinguish Thumb functions from ARM ones.
> +@end deftypevr
> +
> Implementing trampolines is difficult on many machines because they have
> separate instruction and data caches. Writing into a stack location
> fails to clear the memory in the instruction cache, so when the program
> Index: doc/tm.texi.in
> ===================================================================
> --- doc/tm.texi.in (revision 237789)
> +++ doc/tm.texi.in (working copy)
> @@ -3947,6 +3947,8 @@ is used for aligning trampolines.
>
> @hook TARGET_TRAMPOLINE_ADJUST_ADDRESS
>
> +@hook TARGET_CUSTOM_FUNCTION_DESCRIPTORS
> +
> Implementing trampolines is difficult on many machines because they have
> separate instruction and data caches. Writing into a stack location
> fails to clear the memory in the instruction cache, so when the program
> Index: gimple.c
> ===================================================================
> --- gimple.c (revision 237789)
> +++ gimple.c (working copy)
> @@ -373,6 +373,7 @@ gimple_build_call_from_tree (tree t)
> gimple_call_set_from_thunk (call, CALL_FROM_THUNK_P (t));
> gimple_call_set_va_arg_pack (call, CALL_EXPR_VA_ARG_PACK (t));
> gimple_call_set_nothrow (call, TREE_NOTHROW (t));
> + gimple_call_set_by_descriptor (call, CALL_EXPR_BY_DESCRIPTOR (t));
> gimple_set_no_warning (call, TREE_NO_WARNING (t));
> gimple_call_set_with_bounds (call, CALL_WITH_BOUNDS_P (t));
>
> @@ -1386,6 +1387,9 @@ gimple_call_flags (const gimple *stmt)
> if (stmt->subcode & GF_CALL_NOTHROW)
> flags |= ECF_NOTHROW;
>
> + if (stmt->subcode & GF_CALL_BY_DESCRIPTOR)
> + flags |= ECF_BY_DESCRIPTOR;
> +
> return flags;
> }
>
> Index: gimple.h
> ===================================================================
> --- gimple.h (revision 237789)
> +++ gimple.h (working copy)
> @@ -146,6 +146,7 @@ enum gf_mask {
> GF_CALL_CTRL_ALTERING = 1 << 7,
> GF_CALL_WITH_BOUNDS = 1 << 8,
> GF_CALL_MUST_TAIL_CALL = 1 << 9,
> + GF_CALL_BY_DESCRIPTOR = 1 << 10,
> GF_OMP_PARALLEL_COMBINED = 1 << 0,
> GF_OMP_PARALLEL_GRID_PHONY = 1 << 1,
> GF_OMP_TASK_TASKLOOP = 1 << 0,
> @@ -3357,6 +3358,26 @@ gimple_call_alloca_for_var_p (gcall *s)
> return (s->subcode & GF_CALL_ALLOCA_FOR_VAR) != 0;
> }
>
> +/* If BY_DESCRIPTOR_P is true, GIMPLE_CALL S is an indirect call for which
> + pointers to nested function are descriptors instead of trampolines. */
> +
> +static inline void
> +gimple_call_set_by_descriptor (gcall *s, bool by_descriptor_p)
> +{
> + if (by_descriptor_p)
> + s->subcode |= GF_CALL_BY_DESCRIPTOR;
> + else
> + s->subcode &= ~GF_CALL_BY_DESCRIPTOR;
> +}
> +
> +/* Return true if S is a by-descriptor call. */
> +
> +static inline bool
> +gimple_call_by_descriptor_p (gcall *s)
> +{
> + return (s->subcode & GF_CALL_BY_DESCRIPTOR) != 0;
> +}
> +
> /* Copy all the GF_CALL_* flags from ORIG_CALL to DEST_CALL. */
>
> static inline void
> Index: langhooks-def.h
> ===================================================================
> --- langhooks-def.h (revision 237789)
> +++ langhooks-def.h (working copy)
> @@ -120,6 +120,7 @@ extern bool lhd_omp_mappable_type (tree)
> #define LANG_HOOKS_BLOCK_MAY_FALLTHRU hook_bool_const_tree_true
> #define LANG_HOOKS_EH_USE_CXA_END_CLEANUP false
> #define LANG_HOOKS_DEEP_UNSHARING false
> +#define LANG_HOOKS_CUSTOM_FUNCTION_DESCRIPTORS false
>
> /* Attribute hooks. */
> #define LANG_HOOKS_ATTRIBUTE_TABLE NULL
> @@ -319,7 +320,8 @@ extern void lhd_end_section (void);
> LANG_HOOKS_EH_PROTECT_CLEANUP_ACTIONS, \
> LANG_HOOKS_BLOCK_MAY_FALLTHRU, \
> LANG_HOOKS_EH_USE_CXA_END_CLEANUP, \
> - LANG_HOOKS_DEEP_UNSHARING \
> + LANG_HOOKS_DEEP_UNSHARING, \
> + LANG_HOOKS_CUSTOM_FUNCTION_DESCRIPTORS \
> }
>
> #endif /* GCC_LANG_HOOKS_DEF_H */
> Index: langhooks.h
> ===================================================================
> --- langhooks.h (revision 237789)
> +++ langhooks.h (working copy)
> @@ -505,6 +505,10 @@ struct lang_hooks
> gimplification. */
> bool deep_unsharing;
>
> + /* True if this language may use custom descriptors for nested functions
> + instead of trampolines. */
> + bool custom_function_descriptors;
> +
> /* Whenever you add entries here, make sure you adjust langhooks-def.h
> and langhooks.c accordingly. */
> };
> Index: rtl.h
> ===================================================================
> --- rtl.h (revision 237789)
> +++ rtl.h (working copy)
> @@ -317,6 +317,7 @@ struct GTY((desc("0"), tag("0"),
> 1 in a CONCAT is VAL_EXPR_IS_COPIED in var-tracking.c.
> 1 in a VALUE is SP_BASED_VALUE_P in cselib.c.
> 1 in a SUBREG generated by LRA for reload insns.
> + 1 in a REG if this is a static chain register.
> 1 in a CALL for calls instrumented by Pointer Bounds Checker. */
> unsigned int jump : 1;
> /* In a CODE_LABEL, part of the two-bit alternate entry field.
> @@ -2264,6 +2265,10 @@ do { \
> : (SIGN) == SRP_SIGNED ? SUBREG_PROMOTED_SIGNED_P (RTX) \
> : SUBREG_PROMOTED_UNSIGNED_P (RTX))
>
> +/* True if the REG is the static chain register for some CALL_INSN. */
> +#define STATIC_CHAIN_REG_P(RTX) \
> + (RTL_FLAG_CHECK1 ("STATIC_CHAIN_REG_P", (RTX), REG)->jump)
> +
> /* True if the subreg was generated by LRA for reload insns. Such
> subregs are valid only during LRA. */
> #define LRA_SUBREG_P(RTX) \
> Index: rtlanal.c
> ===================================================================
> --- rtlanal.c (revision 237789)
> +++ rtlanal.c (working copy)
> @@ -3914,7 +3914,8 @@ find_first_parameter_load (rtx_insn *cal
> parm.nregs = 0;
> for (p = CALL_INSN_FUNCTION_USAGE (call_insn); p; p = XEXP (p, 1))
> if (GET_CODE (XEXP (p, 0)) == USE
> - && REG_P (XEXP (XEXP (p, 0), 0)))
> + && REG_P (XEXP (XEXP (p, 0), 0))
> + && !STATIC_CHAIN_REG_P (XEXP (XEXP (p, 0), 0)))
> {
> gcc_assert (REGNO (XEXP (XEXP (p, 0), 0)) < FIRST_PSEUDO_REGISTER);
>
> Index: target.def
> ===================================================================
> --- target.def (revision 237789)
> +++ target.def (working copy)
> @@ -4723,6 +4723,26 @@ be returned; otherwise @var{addr} should
> If this hook is not defined, @var{addr} will be used for function calls.",
> rtx, (rtx addr), NULL)
>
> +DEFHOOKPOD
> +(custom_function_descriptors,
> + "This hook should be defined to a power of 2 if the target will benefit\n\
> +from the use of custom descriptors for nested functions instead of the\n\
> +standard trampolines. Such descriptors are created at run time on the\n\
> +stack and made up of data only, but they are non-standard so the generated\n\
> +code must be prepared to deal with them. This hook should be defined to 0\n\
> +if the target uses function descriptors for its standard calling sequence,\n\
> +like for example HP-PA or IA-64. Using descriptors for nested functions\n\
> +eliminates the need for trampolines that reside on the stack and require\n\
> +it to be made executable.\n\
> +\n\
> +The value of the macro is used to parameterize the run-time identification\n\
> +scheme implemented to distinguish descriptors from function addresses: it\n\
> +gives the number of bytes by which their address is shifted in comparison\n\
> +with function addresses. The value of 1 will generally work, unless it is\n\
> +already used by the target for a similar purpose, like for example on ARM\n\
> +where it is used to distinguish Thumb functions from ARM ones.",
> + int, -1)
> +
> /* Return the number of bytes of its own arguments that a function
> pops on returning, or 0 if the function pops no arguments and the
> caller must therefore pop them all after the function returns. */
> Index: testsuite/gnat.dg/trampoline3.adb
> ===================================================================
> --- testsuite/gnat.dg/trampoline3.adb (revision 0)
> +++ testsuite/gnat.dg/trampoline3.adb (working copy)
> @@ -0,0 +1,22 @@
> +-- { dg-do compile { target *-*-linux* } }
> +-- { dg-options "-gnatws" }
> +
> +procedure Trampoline3 is
> +
> + A : Integer;
> +
> + type FuncPtr is access function (I : Integer) return Integer;
> +
> + function F (I : Integer) return Integer is
> + begin
> + return A + I;
> + end F;
> +
> + P : FuncPtr := F'Access;
> + I : Integer;
> +
> +begin
> + I := P(0);
> +end;
> +
> +-- { dg-final { scan-assembler-not "GNU-stack.*x" } }
> Index: testsuite/gnat.dg/trampoline4.adb
> ===================================================================
> --- testsuite/gnat.dg/trampoline4.adb (revision 0)
> +++ testsuite/gnat.dg/trampoline4.adb (working copy)
> @@ -0,0 +1,22 @@
> +-- { dg-do compile { target *-*-linux* } }
> +-- { dg-options "-ftrampolines -gnatws" }
> +
> +procedure Trampoline4 is
> +
> + A : Integer;
> +
> + type FuncPtr is access function (I : Integer) return Integer;
> +
> + function F (I : Integer) return Integer is
> + begin
> + return A + I;
> + end F;
> +
> + P : FuncPtr := F'Access;
> + I : Integer;
> +
> +begin
> + I := P(0);
> +end;
> +
> +-- { dg-final { scan-assembler "GNU-stack.*x" } }
> Index: tree-core.h
> ===================================================================
> --- tree-core.h (revision 237789)
> +++ tree-core.h (working copy)
> @@ -90,6 +90,9 @@ struct die_struct;
> /* Nonzero if this call is into the transaction runtime library. */
> #define ECF_TM_BUILTIN (1 << 13)
>
> +/* Nonzero if this is an indirect call by descriptor. */
> +#define ECF_BY_DESCRIPTOR (1 << 14)
> +
> /* Call argument flags. */
> /* Nonzero if the argument is not dereferenced recursively, thus only
> directly reachable memory is read or written. */
> @@ -1216,6 +1219,12 @@ struct GTY(()) tree_base {
>
> REF_REVERSE_STORAGE_ORDER in
> BIT_FIELD_REF, MEM_REF
> +
> + FUNC_ADDR_BY_DESCRIPTOR in
> + ADDR_EXPR
> +
> + CALL_EXPR_BY_DESCRIPTOR in
> + CALL_EXPR
> */
>
> struct GTY(()) tree_typed {
> Index: tree-nested.c
> ===================================================================
> --- tree-nested.c (revision 237789)
> +++ tree-nested.c (working copy)
> @@ -21,6 +21,7 @@
> #include "system.h"
> #include "coretypes.h"
> #include "backend.h"
> +#include "target.h"
> #include "rtl.h"
> #include "tree.h"
> #include "gimple.h"
> @@ -103,6 +104,7 @@ struct nesting_info
>
> bool any_parm_remapped;
> bool any_tramp_created;
> + bool any_descr_created;
> char static_chain_added;
> };
>
> @@ -486,12 +488,40 @@ get_trampoline_type (struct nesting_info
> return trampoline_type;
> }
>
> -/* Given DECL, a nested function, find or create a field in the non-local
> - frame structure for a trampoline for this function. */
> +/* Build or return the type used to represent a nested function descriptor. */
> +
> +static GTY(()) tree descriptor_type;
>
> static tree
> -lookup_tramp_for_decl (struct nesting_info *info, tree decl,
> - enum insert_option insert)
> +get_descriptor_type (struct nesting_info *info)
> +{
> + tree t;
> +
> + if (descriptor_type)
> + return descriptor_type;
> +
> + t = build_index_type (build_int_cst (NULL_TREE, 2 * UNITS_PER_WORD - 1));
> + t = build_array_type (char_type_node, t);
> + t = build_decl (DECL_SOURCE_LOCATION (info->context),
> + FIELD_DECL, get_identifier ("__data"), t);
> + SET_DECL_ALIGN (t, BITS_PER_WORD);
> + DECL_USER_ALIGN (t) = 1;
> +
> + descriptor_type = make_node (RECORD_TYPE);
> + TYPE_NAME (descriptor_type) = get_identifier ("__builtin_descriptor");
> + TYPE_FIELDS (descriptor_type) = t;
> + layout_type (descriptor_type);
> + DECL_CONTEXT (t) = descriptor_type;
> +
> + return descriptor_type;
> +}
> +
> +/* Given DECL, a nested function, find or create an element in the
> + var map for this function. */
> +
> +static tree
> +lookup_element_for_decl (struct nesting_info *info, tree decl,
> + enum insert_option insert)
> {
> if (insert == NO_INSERT)
> {
> @@ -501,19 +531,73 @@ lookup_tramp_for_decl (struct nesting_in
>
> tree *slot = &info->var_map->get_or_insert (decl);
> if (!*slot)
> - {
> - tree field = make_node (FIELD_DECL);
> - DECL_NAME (field) = DECL_NAME (decl);
> - TREE_TYPE (field) = get_trampoline_type (info);
> - TREE_ADDRESSABLE (field) = 1;
> + *slot = build_tree_list (NULL_TREE, NULL_TREE);
>
> - insert_field_into_struct (get_frame_type (info), field);
> - *slot = field;
> + return (tree) *slot;
> +}
> +
> +/* Given DECL, a nested function, create a field in the non-local
> + frame structure for this function. */
> +
> +static tree
> +create_field_for_decl (struct nesting_info *info, tree decl, tree type)
> +{
> + tree field = make_node (FIELD_DECL);
> + DECL_NAME (field) = DECL_NAME (decl);
> + TREE_TYPE (field) = type;
> + TREE_ADDRESSABLE (field) = 1;
> + insert_field_into_struct (get_frame_type (info), field);
> + return field;
> +}
> +
> +/* Given DECL, a nested function, find or create a field in the non-local
> + frame structure for a trampoline for this function. */
> +
> +static tree
> +lookup_tramp_for_decl (struct nesting_info *info, tree decl,
> + enum insert_option insert)
> +{
> + tree elt, field;
> +
> + elt = lookup_element_for_decl (info, decl, insert);
> + if (!elt)
> + return NULL_TREE;
> +
> + field = TREE_PURPOSE (elt);
>
> + if (!field && insert == INSERT)
> + {
> + field = create_field_for_decl (info, decl, get_trampoline_type (info));
> + TREE_PURPOSE (elt) = field;
> info->any_tramp_created = true;
> }
>
> - return *slot;
> + return field;
> +}
> +
> +/* Given DECL, a nested function, find or create a field in the non-local
> + frame structure for a descriptor for this function. */
> +
> +static tree
> +lookup_descr_for_decl (struct nesting_info *info, tree decl,
> + enum insert_option insert)
> +{
> + tree elt, field;
> +
> + elt = lookup_element_for_decl (info, decl, insert);
> + if (!elt)
> + return NULL_TREE;
> +
> + field = TREE_VALUE (elt);
> +
> + if (!field && insert == INSERT)
> + {
> + field = create_field_for_decl (info, decl, get_descriptor_type (info));
> + TREE_VALUE (elt) = field;
> + info->any_descr_created = true;
> + }
> +
> + return field;
> }
>
> /* Build or return the field within the non-local frame state that holds
> @@ -2303,6 +2387,7 @@ convert_tramp_reference_op (tree *tp, in
> struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
> struct nesting_info *const info = (struct nesting_info *) wi->info, *i;
> tree t = *tp, decl, target_context, x, builtin;
> + bool descr;
> gcall *call;
>
> *walk_subtrees = 0;
> @@ -2337,7 +2422,14 @@ convert_tramp_reference_op (tree *tp, in
> we need to insert the trampoline. */
> for (i = info; i->context != target_context; i = i->outer)
> continue;
> - x = lookup_tramp_for_decl (i, decl, INSERT);
> +
> + /* Decide whether to generate a descriptor or a trampoline. */
> + descr = FUNC_ADDR_BY_DESCRIPTOR (t) && !flag_trampolines;
> +
> + if (descr)
> + x = lookup_descr_for_decl (i, decl, INSERT);
> + else
> + x = lookup_tramp_for_decl (i, decl, INSERT);
>
> /* Compute the address of the field holding the trampoline. */
> x = get_frame_field (info, target_context, x, &wi->gsi);
> @@ -2346,7 +2438,10 @@ convert_tramp_reference_op (tree *tp, in
>
> /* Do machine-specific ugliness. Normally this will involve
> computing extra alignment, but it can really be anything. */
> - builtin = builtin_decl_implicit (BUILT_IN_ADJUST_TRAMPOLINE);
> + if (descr)
> + builtin = builtin_decl_implicit (BUILT_IN_ADJUST_DESCRIPTOR);
> + else
> + builtin = builtin_decl_implicit (BUILT_IN_ADJUST_TRAMPOLINE);
> call = gimple_build_call (builtin, 1, x);
> x = init_tmp_var_with_call (info, &wi->gsi, call);
>
> @@ -2820,6 +2915,27 @@ fold_mem_refs (tree *const &e, void *dat
> return true;
> }
>
> +/* Given DECL, a nested function, build an initialization call for FIELD,
> + the trampoline or descriptor for DECL, using FUNC as the function. */
> +
> +static gcall *
> +build_init_call_stmt (struct nesting_info *info, tree decl, tree field,
> + tree func)
> +{
> + tree arg1, arg2, arg3, x;
> +
> + gcc_assert (DECL_STATIC_CHAIN (decl));
> + arg3 = build_addr (info->frame_decl);
> +
> + arg2 = build_addr (decl);
> +
> + x = build3 (COMPONENT_REF, TREE_TYPE (field),
> + info->frame_decl, field, NULL_TREE);
> + arg1 = build_addr (x);
> +
> + return gimple_build_call (func, 3, arg1, arg2, arg3);
> +}
> +
> /* Do "everything else" to clean up or complete state collected by the various
> walking passes -- create a field to hold the frame base address, lay out the
> types and decls, generate code to initialize the frame decl, store critical
> @@ -2965,23 +3081,32 @@ finalize_nesting_tree_1 (struct nesting_
> struct nesting_info *i;
> for (i = root->inner; i ; i = i->next)
> {
> - tree arg1, arg2, arg3, x, field;
> + tree field, x;
>
> field = lookup_tramp_for_decl (root, i->context, NO_INSERT);
> if (!field)
> continue;
>
> - gcc_assert (DECL_STATIC_CHAIN (i->context));
> - arg3 = build_addr (root->frame_decl);
> + x = builtin_decl_implicit (BUILT_IN_INIT_TRAMPOLINE);
> + stmt = build_init_call_stmt (root, i->context, field, x);
> + gimple_seq_add_stmt (&stmt_list, stmt);
> + }
> + }
>
> - arg2 = build_addr (i->context);
> + /* If descriptors were created, then we need to initialize them. */
> + if (root->any_descr_created)
> + {
> + struct nesting_info *i;
> + for (i = root->inner; i ; i = i->next)
> + {
> + tree field, x;
>
> - x = build3 (COMPONENT_REF, TREE_TYPE (field),
> - root->frame_decl, field, NULL_TREE);
> - arg1 = build_addr (x);
> + field = lookup_descr_for_decl (root, i->context, NO_INSERT);
> + if (!field)
> + continue;
>
> - x = builtin_decl_implicit (BUILT_IN_INIT_TRAMPOLINE);
> - stmt = gimple_build_call (x, 3, arg1, arg2, arg3);
> + x = builtin_decl_implicit (BUILT_IN_INIT_DESCRIPTOR);
> + stmt = build_init_call_stmt (root, i->context, field, x);
> gimple_seq_add_stmt (&stmt_list, stmt);
> }
> }
> Index: tree.c
> ===================================================================
> --- tree.c (revision 237789)
> +++ tree.c (working copy)
> @@ -1019,7 +1019,7 @@ make_node_stat (enum tree_code code MEM_
> {
> if (code == FUNCTION_DECL)
> {
> - SET_DECL_ALIGN (t, FUNCTION_BOUNDARY);
> + SET_DECL_ALIGN (t, DEFAULT_FUNCTION_ALIGNMENT);
> DECL_MODE (t) = FUNCTION_MODE;
> }
> else
> @@ -10567,12 +10567,19 @@ build_common_builtin_nodes (void)
> BUILT_IN_INIT_HEAP_TRAMPOLINE,
> "__builtin_init_heap_trampoline",
> ECF_NOTHROW | ECF_LEAF);
> + local_define_builtin ("__builtin_init_descriptor", ftype,
> + BUILT_IN_INIT_DESCRIPTOR,
> + "__builtin_init_descriptor", ECF_NOTHROW | ECF_LEAF);
>
> ftype = build_function_type_list (ptr_type_node, ptr_type_node, NULL_TREE);
> local_define_builtin ("__builtin_adjust_trampoline", ftype,
> BUILT_IN_ADJUST_TRAMPOLINE,
> "__builtin_adjust_trampoline",
> ECF_CONST | ECF_NOTHROW);
> + local_define_builtin ("__builtin_adjust_descriptor", ftype,
> + BUILT_IN_ADJUST_DESCRIPTOR,
> + "__builtin_adjust_descriptor",
> + ECF_CONST | ECF_NOTHROW);
>
> ftype = build_function_type_list (void_type_node,
> ptr_type_node, ptr_type_node, NULL_TREE);
> Index: tree.h
> ===================================================================
> --- tree.h (revision 237789)
> +++ tree.h (working copy)
> @@ -967,6 +967,16 @@ extern void omp_clause_range_check_faile
> #define REF_REVERSE_STORAGE_ORDER(NODE) \
> (TREE_CHECK2 (NODE, BIT_FIELD_REF, MEM_REF)->base.default_def_flag)
>
> + /* In an ADDR_EXPR, indicates that this is a pointer to nested function
> + represented by a descriptor instead of a trampoline. */
> +#define FUNC_ADDR_BY_DESCRIPTOR(NODE) \
> + (TREE_CHECK (NODE, ADDR_EXPR)->base.default_def_flag)
> +
> +/* In a CALL_EXPR, indicates that this is an indirect call for which
> + pointers to nested function are descriptors instead of trampolines. */
> +#define CALL_EXPR_BY_DESCRIPTOR(NODE) \
> + (TREE_CHECK (NODE, CALL_EXPR)->base.default_def_flag)
> +
> /* These flags are available for each language front end to use internally. */
> #define TREE_LANG_FLAG_0(NODE) \
> (TREE_NOT_CHECK2 (NODE, TREE_VEC, SSA_NAME)->base.u.bits.lang_flag_0)
>