path: root/src/comp/middle/shape.rs

// A "shape" is a compact encoding of a type that is used by interpreted glue.
// This substitutes for the runtime tags used by e.g. MLs.

import lib::llvm::True;
import lib::llvm::llvm::{ModuleRef, TypeRef, ValueRef};
import driver::session;
import middle::{trans, trans_common};
import middle::trans_common::{crate_ctxt, val_ty, C_bytes,
                              C_named_struct, C_struct, T_tag_variant};
import middle::ty;
import middle::ty::field;
import syntax::ast;
import syntax::ast_util::dummy_sp;
import syntax::util::interner;
import util::common;

import core::{vec, str};
import std::map::hashmap;
import option::{none, some};

import ty_ctxt = middle::ty::ctxt;

type res_info = {did: ast::def_id, t: ty::t};

type ctxt =
    {mutable next_tag_id: u16,
     pad: u16,
     tag_id_to_index: hashmap<ast::def_id, u16>,
     mutable tag_order: [ast::def_id],
     resources: interner::interner<res_info>,
     llshapetablesty: TypeRef,
     llshapetables: ValueRef};

const shape_u8: u8 = 0u8;
const shape_u16: u8 = 1u8;
const shape_u32: u8 = 2u8;
const shape_u64: u8 = 3u8;
const shape_i8: u8 = 4u8;
const shape_i16: u8 = 5u8;
const shape_i32: u8 = 6u8;
const shape_i64: u8 = 7u8;
const shape_f32: u8 = 8u8;
const shape_f64: u8 = 9u8;
// (10 is currently unused, was evec)
const shape_vec: u8 = 11u8;
const shape_tag: u8 = 12u8;
const shape_box: u8 = 13u8;
const shape_struct: u8 = 17u8;
const shape_fn: u8 = 18u8;
const shape_obj: u8 = 19u8;
const shape_res: u8 = 20u8;
const shape_var: u8 = 21u8;
const shape_uniq: u8 = 22u8;
const shape_opaque_closure: u8 = 23u8; // the closure itself.

// FIXME: This is a bad API in trans_common.
fn C_u8(n: u8) -> ValueRef { ret trans_common::C_u8(n as uint); }

fn hash_res_info(ri: res_info) -> uint {
    let h = 5381u;
    h *= 33u;
    h += ri.did.crate as uint;
    h *= 33u;
    h += ri.did.node as uint;
    h *= 33u;
    h += ri.t as uint;
    ret h;
}

fn eq_res_info(a: res_info, b: res_info) -> bool {
    ret a.did.crate == b.did.crate && a.did.node == b.did.node && a.t == b.t;
}

fn mk_global(ccx: @crate_ctxt, name: str, llval: ValueRef, internal: bool) ->
   ValueRef {
    let llglobal =
        str::as_buf(name,
                    {|buf|
                        lib::llvm::llvm::LLVMAddGlobal(ccx.llmod,
                                                       val_ty(llval), buf)
                    });
    lib::llvm::llvm::LLVMSetInitializer(llglobal, llval);
    lib::llvm::llvm::LLVMSetGlobalConstant(llglobal, True);

    if internal {
        lib::llvm::llvm::LLVMSetLinkage(llglobal,
                                        lib::llvm::LLVMInternalLinkage as
                                            lib::llvm::llvm::Linkage);
    }

    ret llglobal;
}


// Computes a set of variants of a tag that are guaranteed to have size and
// alignment at least as large as any other variant of the tag. This is an
// important performance optimization.
//
// TODO: Use this in dynamic_size_of() as well.

fn largest_variants(ccx: @crate_ctxt, tag_id: ast::def_id) -> [uint] {
    // Compute the minimum and maximum size and alignment for each variant.
    //
    // TODO: We could do better here; e.g. we know that any variant that
    // contains (T,T) must be as least as large as any variant that contains
    // just T.
    let ranges = [];
    let variants = ty::tag_variants(ccx.tcx, tag_id);
    for variant: ty::variant_info in *variants {
        let bounded = true;
        let {a: min_size, b: min_align} = {a: 0u, b: 0u};
        for elem_t: ty::t in variant.args {
            if ty::type_contains_params(ccx.tcx, elem_t) {
                // TODO: We could do better here; this causes us to
                // conservatively assume that (int, T) has minimum size 0,
                // when in fact it has minimum size sizeof(int).
                bounded = false;
            } else {
                // Could avoid this check: the constraint should
                // follow from how elem_t doesn't contain params.
                // (Could add a postcondition to type_contains_params,
                // once we implement Issue #586.)
                check (trans_common::type_has_static_size(ccx, elem_t));
                let llty = trans::type_of(ccx, dummy_sp(), elem_t);
                min_size += trans::llsize_of_real(ccx, llty);
                min_align += trans::llalign_of_real(ccx, llty);
            }
        }

        ranges +=
            [{size: {min: min_size, bounded: bounded},
              align: {min: min_align, bounded: bounded}}];
    }

    // Initialize the candidate set to contain all variants.
    let candidates = [mutable];
    for variant in *variants { candidates += [mutable true]; }

    // Do a pairwise comparison among all variants still in the candidate set.
    // Throw out any variant that we know has size and alignment at least as
    // small as some other variant.
    let i = 0u;
    while i < vec::len(ranges) - 1u {
        if candidates[i] {
            let j = i + 1u;
            while j < vec::len(ranges) {
                if candidates[j] {
                    if ranges[i].size.bounded && ranges[i].align.bounded &&
                           ranges[j].size.bounded && ranges[j].align.bounded {
                        if ranges[i].size >= ranges[j].size &&
                               ranges[i].align >= ranges[j].align {
                            // Throw out j.
                            candidates[j] = false;
                        } else if ranges[j].size >= ranges[i].size &&
                                      ranges[j].align >= ranges[j].align {
                            // Throw out i.
                            candidates[i] = false;
                        }
                    }
                }
                j += 1u;
            }
        }
        i += 1u;
    }

    // Return the resulting set.
    let result = [];
    i = 0u;
    while i < vec::len(candidates) {
        if candidates[i] { result += [i]; }
        i += 1u;
    }
    ret result;
}

// Computes the static size of a tag, without using mk_tup(), which is
// bad for performance.
//
// TODO: Migrate trans over to use this.

fn round_up(size: u16, align: u8) -> u16 {
    assert (align >= 1u8);
    let alignment = align as u16;
    ret size - 1u16 + alignment & !(alignment - 1u16);
}

type size_align = {size: u16, align: u8};

fn compute_static_tag_size(ccx: @crate_ctxt, largest_variants: [uint],
                           did: ast::def_id) -> size_align {
    let max_size = 0u16;
    let max_align = 1u8;
    let variants = ty::tag_variants(ccx.tcx, did);
    for vid: uint in largest_variants {
        // We increment a "virtual data pointer" to compute the size.
        let lltys = [];
        for typ: ty::t in variants[vid].args {
            // FIXME: there should really be a postcondition
            // on tag_variants that would obviate the need for
            // this check. (Issue #586)
            check (trans_common::type_has_static_size(ccx, typ));
            lltys += [trans::type_of(ccx, dummy_sp(), typ)];
        }

        let llty = trans_common::T_struct(lltys);
        let dp = trans::llsize_of_real(ccx, llty) as u16;
        let variant_align = trans::llalign_of_real(ccx, llty) as u8;

        if max_size < dp { max_size = dp; }
        if max_align < variant_align { max_align = variant_align; }
    }

    // Add space for the tag if applicable.
    // FIXME (issue #792): This is wrong. If the tag starts with an 8 byte
    // aligned quantity, we don't align it.
    if vec::len(*variants) > 1u {
        let variant_t = T_tag_variant(ccx);
        max_size += trans::llsize_of_real(ccx, variant_t) as u16;
        let align = trans::llalign_of_real(ccx, variant_t) as u8;
        if max_align < align { max_align = align; }
    }

    ret {size: max_size, align: max_align};
}

tag tag_kind { tk_unit; tk_enum; tk_complex; }

fn tag_kind(ccx: @crate_ctxt, did: ast::def_id) -> tag_kind {
    let variants = ty::tag_variants(ccx.tcx, did);
    if vec::len(*variants) == 0u { ret tk_complex; }
    for v: ty::variant_info in *variants {
        if vec::len(v.args) > 0u { ret tk_complex; }
    }
    if vec::len(*variants) == 1u { ret tk_unit; }
    ret tk_enum;
}


// Returns the code corresponding to the pointer size on this architecture.
fn s_int(tcx: ty_ctxt) -> u8 {
    ret alt tcx.sess.get_targ_cfg().arch {
        session::arch_x86. { shape_i32 }
        session::arch_x86_64. { shape_i64 }
        session::arch_arm. { shape_i32 }
    };
}

fn s_uint(tcx: ty_ctxt) -> u8 {
    ret alt tcx.sess.get_targ_cfg().arch {
        session::arch_x86. { shape_u32 }
        session::arch_x86_64. { shape_u64 }
        session::arch_arm. { shape_u32 }
    };
}

fn s_float(tcx: ty_ctxt) -> u8 {
    ret alt tcx.sess.get_targ_cfg().arch {
        session::arch_x86. { shape_f64 }
        session::arch_x86_64. { shape_f64 }
        session::arch_arm. { shape_f64 }
    };
}

fn s_variant_tag_t(tcx: ty_ctxt) -> u8 {
    ret s_int(tcx);
}

fn mk_ctxt(llmod: ModuleRef) -> ctxt {
    let llshapetablesty = trans_common::T_named_struct("shapes");
    let llshapetables =
        str::as_buf("shapes",
                    {|buf|
                        lib::llvm::llvm::LLVMAddGlobal(llmod, llshapetablesty,
                                                       buf)
                    });

    ret {mutable next_tag_id: 0u16,
         pad: 0u16,
         tag_id_to_index: common::new_def_hash(),
         mutable tag_order: [],
         resources: interner::mk(hash_res_info, eq_res_info),
         llshapetablesty: llshapetablesty,
         llshapetables: llshapetables};
}

fn add_bool(&dest: [u8], val: bool) { dest += [if val { 1u8 } else { 0u8 }]; }

fn add_u16(&dest: [u8], val: u16) {
    dest += [val & 0xffu16 as u8, val >> 8u16 as u8];
}

fn add_substr(&dest: [u8], src: [u8]) {
    add_u16(dest, vec::len(src) as u16);
    dest += src;
}

fn shape_of(ccx: @crate_ctxt, t: ty::t, ty_param_map: [uint],
            is_obj_body: bool) -> [u8] {
    let s = [];

    alt ty::struct(ccx.tcx, t) {
      ty::ty_nil. | ty::ty_bool. | ty::ty_uint(ast::ty_u8.) |
      ty::ty_bot. { s += [shape_u8]; }
      ty::ty_int(ast::ty_i.) { s += [s_int(ccx.tcx)]; }
      ty::ty_float(ast::ty_f.) { s += [s_float(ccx.tcx)]; }
      ty::ty_uint(ast::ty_u.) | ty::ty_ptr(_) | ty::ty_type. |
      ty::ty_send_type. | ty::ty_native(_) { s += [s_uint(ccx.tcx)]; }
      ty::ty_int(ast::ty_i8.) { s += [shape_i8]; }
      ty::ty_uint(ast::ty_u16.) { s += [shape_u16]; }
      ty::ty_int(ast::ty_i16.) { s += [shape_i16]; }
      ty::ty_uint(ast::ty_u32.) { s += [shape_u32]; }
      ty::ty_int(ast::ty_i32.) | ty::ty_int(ast::ty_char.) {s += [shape_i32];}
      ty::ty_uint(ast::ty_u64.) { s += [shape_u64]; }
      ty::ty_int(ast::ty_i64.) { s += [shape_i64]; }
      ty::ty_float(ast::ty_f32.) { s += [shape_f32]; }
      ty::ty_float(ast::ty_f64.) { s += [shape_f64]; }
      ty::ty_str. {
        s += [shape_vec];
        add_bool(s, true); // type is POD
        let unit_ty = ty::mk_mach_uint(ccx.tcx, ast::ty_u8);
        add_substr(s, shape_of(ccx, unit_ty, ty_param_map, is_obj_body));
      }
      ty::ty_tag(did, tps) {
        alt tag_kind(ccx, did) {
          tk_unit. {
            // FIXME: For now we do this.
            s += [s_variant_tag_t(ccx.tcx)];
          }
          tk_enum. { s += [s_variant_tag_t(ccx.tcx)]; }
          tk_complex. {
            s += [shape_tag];

            let sub = [];

            let id;
            alt ccx.shape_cx.tag_id_to_index.find(did) {
              none. {
                id = ccx.shape_cx.next_tag_id;
                ccx.shape_cx.tag_id_to_index.insert(did, id);
                ccx.shape_cx.tag_order += [did];
                ccx.shape_cx.next_tag_id += 1u16;
              }
              some(existing_id) { id = existing_id; }
            }
            add_u16(sub, id as u16);

            add_u16(sub, vec::len(tps) as u16);
            for tp: ty::t in tps {
                let subshape = shape_of(ccx, tp, ty_param_map, is_obj_body);
                add_u16(sub, vec::len(subshape) as u16);
                sub += subshape;
            }

            s += sub;
          }
        }
      }
      ty::ty_box(mt) {
        s += [shape_box];
        add_substr(s, shape_of(ccx, mt.ty, ty_param_map, is_obj_body));
      }
      ty::ty_uniq(mt) {
        s += [shape_uniq];
        add_substr(s, shape_of(ccx, mt.ty, ty_param_map, is_obj_body));
      }
      ty::ty_vec(mt) {
        s += [shape_vec];
        add_bool(s, ty::type_is_pod(ccx.tcx, mt.ty));
        add_substr(s, shape_of(ccx, mt.ty, ty_param_map, is_obj_body));
      }
      ty::ty_rec(fields) {
        s += [shape_struct];
        let sub = [];
        for f: field in fields {
            sub += shape_of(ccx, f.mt.ty, ty_param_map, is_obj_body);
        }
        add_substr(s, sub);
      }
      ty::ty_tup(elts) {
        s += [shape_struct];
        let sub = [];
        for elt in elts {
            sub += shape_of(ccx, elt, ty_param_map, is_obj_body);
        }
        add_substr(s, sub);
      }
      ty::ty_native_fn(_, _) { s += [shape_u32]; }
      ty::ty_obj(_) { s += [shape_obj]; }
      ty::ty_res(did, raw_subt, tps) {
        let subt = ty::substitute_type_params(ccx.tcx, tps, raw_subt);
        let ri = {did: did, t: subt};
        let id = interner::intern(ccx.shape_cx.resources, ri);

        s += [shape_res];
        add_u16(s, id as u16);
        add_u16(s, vec::len(tps) as u16);
        for tp: ty::t in tps {
            add_substr(s, shape_of(ccx, tp, ty_param_map, is_obj_body));
        }
        add_substr(s, shape_of(ccx, subt, ty_param_map, is_obj_body));

      }
      ty::ty_var(n) {
        fail "shape_of ty_var";
      }
      ty::ty_param(n, _) {
        if is_obj_body {
            // Just write in the parameter number.
            s += [shape_var, n as u8];
        } else {
            // Find the type parameter in the parameter list.
            alt vec::position(n, ty_param_map) {
              some(i) { s += [shape_var, i as u8]; }
              none. { fail "ty param not found in ty_param_map"; }
            }
        }
      }
      ty::ty_fn(_, _, _, _, _) {
        s += [shape_fn];
      }
      ty::ty_opaque_closure. {
        s += [shape_opaque_closure];
      }
    }

    ret s;
}

// FIXME: We might discover other variants as we traverse these. Handle this.
fn shape_of_variant(ccx: @crate_ctxt, v: ty::variant_info,
                    ty_param_count: uint) -> [u8] {
    let ty_param_map = [];
    let i = 0u;
    while i < ty_param_count { ty_param_map += [i]; i += 1u; }

    let s = [];
    for t: ty::t in v.args { s += shape_of(ccx, t, ty_param_map, false); }
    ret s;
}

fn gen_tag_shapes(ccx: @crate_ctxt) -> ValueRef {
    // Loop over all the tag variants and write their shapes into a data
    // buffer. As we do this, it's possible for us to discover new tags, so we
    // must do this first.
    let i = 0u;
    let data = [];
    let offsets = [];
    while i < vec::len(ccx.shape_cx.tag_order) {
        let did = ccx.shape_cx.tag_order[i];
        let variants = ty::tag_variants(ccx.tcx, did);
        let item_tyt = ty::lookup_item_type(ccx.tcx, did);
        let ty_param_count = vec::len(item_tyt.kinds);

        for v: ty::variant_info in *variants {
            offsets += [vec::len(data) as u16];

            let variant_shape = shape_of_variant(ccx, v, ty_param_count);
            add_substr(data, variant_shape);
        }

        i += 1u;
    }

    // Now calculate the sizes of the header space (which contains offsets to
    // info records for each tag) and the info space (which contains offsets
    // to each variant shape). As we do so, build up the header.

    let header = [];
    let info = [];
    let header_sz = 2u16 * ccx.shape_cx.next_tag_id;
    let data_sz = vec::len(data) as u16;

    let info_sz = 0u16;
    for did_: ast::def_id in ccx.shape_cx.tag_order {
        let did = did_; // Satisfy alias checker.
        let variants = ty::tag_variants(ccx.tcx, did);
        add_u16(header, header_sz + info_sz);
        info_sz += 2u16 * ((vec::len(*variants) as u16) + 2u16) + 3u16;
    }

    // Construct the info tables, which contain offsets to the shape of each
    // variant. Also construct the largest-variant table for each tag, which
    // contains the variants that the size-of operation needs to look at.

    let lv_table = [];
    i = 0u;
    for did_: ast::def_id in ccx.shape_cx.tag_order {
        let did = did_; // Satisfy alias checker.
        let variants = ty::tag_variants(ccx.tcx, did);
        add_u16(info, vec::len(*variants) as u16);

        // Construct the largest-variants table.
        add_u16(info,
                header_sz + info_sz + data_sz + (vec::len(lv_table) as u16));

        let lv = largest_variants(ccx, did);
        add_u16(lv_table, vec::len(lv) as u16);
        for v: uint in lv { add_u16(lv_table, v as u16); }

        // Determine whether the tag has dynamic size.
        let dynamic = false;
        for variant: ty::variant_info in *variants {
            for typ: ty::t in variant.args {
                if ty::type_has_dynamic_size(ccx.tcx, typ) { dynamic = true; }
            }
        }

        // If we can, write in the static size and alignment of the tag.
        // Otherwise, write a placeholder.
        let size_align;
        if dynamic {
            size_align = {size: 0u16, align: 0u8};
        } else { size_align = compute_static_tag_size(ccx, lv, did); }
        add_u16(info, size_align.size);
        info += [size_align.align];

        // Now write in the offset of each variant.
        for v: ty::variant_info in *variants {
            add_u16(info, header_sz + info_sz + offsets[i]);
            i += 1u;
        }
    }

    assert (i == vec::len(offsets));
    assert (header_sz == vec::len(header) as u16);
    assert (info_sz == vec::len(info) as u16);
    assert (data_sz == vec::len(data) as u16);

    header += info;
    header += data;
    header += lv_table;

    ret mk_global(ccx, "tag_shapes", C_bytes(header), true);
}

fn gen_resource_shapes(ccx: @crate_ctxt) -> ValueRef {
    let dtors = [];
    let i = 0u;
    let len = interner::len(ccx.shape_cx.resources);
    while i < len {
        let ri = interner::get(ccx.shape_cx.resources, i);
        dtors += [trans_common::get_res_dtor(ccx, dummy_sp(), ri.did, ri.t)];
        i += 1u;
    }

    ret mk_global(ccx, "resource_shapes", C_struct(dtors), true);
}

fn gen_shape_tables(ccx: @crate_ctxt) {
    let lltagstable = gen_tag_shapes(ccx);
    let llresourcestable = gen_resource_shapes(ccx);
    trans_common::set_struct_body(ccx.shape_cx.llshapetablesty,
                                  [val_ty(lltagstable),
                                   val_ty(llresourcestable)]);

    let lltables =
        C_named_struct(ccx.shape_cx.llshapetablesty,
                       [lltagstable, llresourcestable]);
    lib::llvm::llvm::LLVMSetInitializer(ccx.shape_cx.llshapetables, lltables);
    lib::llvm::llvm::LLVMSetGlobalConstant(ccx.shape_cx.llshapetables, True);
    lib::llvm::llvm::LLVMSetLinkage(ccx.shape_cx.llshapetables,
                                    lib::llvm::LLVMInternalLinkage as
                                        lib::llvm::llvm::Linkage);
}