1
Fork 0
rust/src/librustc/middle/trans/base.rs

3004 lines
109 KiB
Rust
Raw Normal View History

// Copyright 2012-2014 The Rust Project Developers. See the COPYRIGHT
// file at the top-level directory of this distribution and at
// http://rust-lang.org/COPYRIGHT.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.
2011-06-01 11:34:52 -07:00
// trans.rs: Translate the completed AST to the LLVM IR.
//
// Some functions here, such as trans_block and trans_expr, return a value --
// the result of the translation to LLVM -- while others, such as trans_fn,
// trans_impl, and trans_item, are called only for the side effect of adding a
2011-06-01 11:34:52 -07:00
// particular definition to the LLVM IR output we're producing.
2011-06-01 16:33:03 -07:00
//
// Hopefully useful general knowledge about trans:
//
2011-06-01 16:33:03 -07:00
// * There's no way to find out the ty::t type of a ValueRef. Doing so
// would be "trying to get the eggs out of an omelette" (credit:
// pcwalton). You can, instead, find out its TypeRef by calling val_ty,
// but one TypeRef corresponds to many `ty::t`s; for instance, tup(int, int,
2011-06-01 16:33:03 -07:00
// int) and rec(x=int, y=int, z=int) will have the same TypeRef.
#![allow(non_camel_case_types)]
use back::link::{mangle_exported_name};
use back::{link, abi};
use driver::config;
use driver::config::{NoDebugInfo, FullDebugInfo};
use driver::driver::{CrateAnalysis, CrateTranslation};
use driver::session::Session;
use lint;
use llvm::{BasicBlockRef, ModuleRef, ValueRef, Vector, get_param};
use llvm;
use metadata::{csearch, encoder, loader};
use middle::astencode;
use middle::lang_items::{LangItem, ExchangeMallocFnLangItem, StartFnLangItem};
use middle::subst;
use middle::weak_lang_items;
use middle::subst::Subst;
use middle::trans::_match;
use middle::trans::adt;
use middle::trans::build::*;
use middle::trans::builder::{Builder, noname};
use middle::trans::callee;
use middle::trans::cleanup::{CleanupMethods, ScopeId};
use middle::trans::cleanup;
use middle::trans::common::{Block, C_bool, C_bytes, C_i32, C_integral, C_nil};
use middle::trans::common::{C_null, C_struct, C_u64, C_u8, C_uint, C_undef};
use middle::trans::common::{CrateContext, ExternMap, FunctionContext};
use middle::trans::common::{NodeInfo, Result, SubstP, monomorphize_type};
use middle::trans::common::{node_id_type, param_substs, return_type_is_void};
use middle::trans::common::{tydesc_info, type_is_immediate};
use middle::trans::common::{type_is_zero_size, val_ty};
use middle::trans::common;
use middle::trans::consts;
use middle::trans::controlflow;
use middle::trans::datum;
use middle::trans::debuginfo;
use middle::trans::expr;
use middle::trans::foreign;
use middle::trans::glue;
use middle::trans::inline;
use middle::trans::intrinsic;
use middle::trans::machine;
2014-08-06 11:59:40 +02:00
use middle::trans::machine::{llsize_of, llsize_of_real};
use middle::trans::meth;
use middle::trans::monomorphize;
use middle::trans::tvec;
use middle::trans::type_::Type;
use middle::trans::type_of;
use middle::trans::type_of::*;
use middle::trans::value::Value;
use middle::ty;
use middle::typeck;
use util::common::indenter;
use util::ppaux::{Repr, ty_to_string};
use util::sha2::Sha256;
use util::nodemap::NodeMap;
2013-06-16 22:52:44 +12:00
use arena::TypedArena;
use libc::{c_uint, uint64_t};
use std::c_str::ToCStr;
use std::cell::{Cell, RefCell};
2014-04-22 02:21:52 +03:00
use std::rc::Rc;
use std::{i8, i16, i32, i64};
use syntax::abi::{X86, X86_64, Arm, Mips, Mipsel, Rust, RustCall};
use syntax::abi::{RustIntrinsic, Abi, OsWindows};
use syntax::ast_util::{local_def, is_local};
use syntax::attr::AttrMetaMethods;
2014-01-21 10:08:10 -08:00
use syntax::attr;
use syntax::codemap::Span;
2014-01-21 10:08:10 -08:00
use syntax::parse::token::InternedString;
use syntax::visit::Visitor;
use syntax::visit;
2014-01-21 10:08:10 -08:00
use syntax::{ast, ast_util, ast_map};
2014-02-19 19:08:12 -08:00
use time;
local_data_key!(task_local_insn_key: RefCell<Vec<&'static str>>)
pub fn with_insn_ctxt(blk: |&[&'static str]|) {
match task_local_insn_key.get() {
Some(ctx) => blk(ctx.borrow().as_slice()),
None => ()
}
}
pub fn init_insn_ctxt() {
task_local_insn_key.replace(Some(RefCell::new(Vec::new())));
2012-09-05 15:58:43 -07:00
}
pub struct _InsnCtxt {
_cannot_construct_outside_of_this_module: ()
}
#[unsafe_destructor]
impl Drop for _InsnCtxt {
2013-09-16 21:18:07 -04:00
fn drop(&mut self) {
match task_local_insn_key.get() {
Some(ctx) => { ctx.borrow_mut().pop(); }
None => {}
}
}
}
pub fn push_ctxt(s: &'static str) -> _InsnCtxt {
debug!("new InsnCtxt: {}", s);
match task_local_insn_key.get() {
Some(ctx) => ctx.borrow_mut().push(s),
None => {}
}
_InsnCtxt { _cannot_construct_outside_of_this_module: () }
}
2014-03-06 05:07:47 +02:00
pub struct StatRecorder<'a> {
ccx: &'a CrateContext,
name: Option<String>,
start: u64,
istart: uint,
}
2014-03-06 05:07:47 +02:00
impl<'a> StatRecorder<'a> {
pub fn new(ccx: &'a CrateContext, name: String) -> StatRecorder<'a> {
2014-03-05 16:36:01 +02:00
let start = if ccx.sess().trans_stats() {
time::precise_time_ns()
} else {
0
};
let istart = ccx.stats.n_llvm_insns.get();
StatRecorder {
ccx: ccx,
name: Some(name),
start: start,
istart: istart,
}
}
}
#[unsafe_destructor]
2014-03-06 05:07:47 +02:00
impl<'a> Drop for StatRecorder<'a> {
2013-09-16 21:18:07 -04:00
fn drop(&mut self) {
2014-03-05 16:36:01 +02:00
if self.ccx.sess().trans_stats() {
let end = time::precise_time_ns();
let elapsed = ((end - self.start) / 1_000_000) as uint;
let iend = self.ccx.stats.n_llvm_insns.get();
2014-08-18 17:52:38 -07:00
self.ccx.stats.fn_stats.borrow_mut().push((self.name.take().unwrap(),
2014-03-20 19:49:20 -07:00
elapsed,
iend - self.istart));
self.ccx.stats.n_fns.set(self.ccx.stats.n_fns.get() + 1);
// Reset LLVM insn count to avoid compound costs.
self.ccx.stats.n_llvm_insns.set(self.istart);
}
}
}
// only use this for foreign function ABIs and glue, use `decl_rust_fn` for Rust functions
pub fn decl_fn(ccx: &CrateContext, name: &str, cc: llvm::CallConv,
ty: Type, output: ty::t) -> ValueRef {
let llfn: ValueRef = name.with_c_str(|buf| {
unsafe {
llvm::LLVMGetOrInsertFunction(ccx.llmod, buf, ty.to_ref())
}
});
match ty::get(output).sty {
// functions returning bottom may unwind, but can never return normally
ty::ty_bot => {
unsafe {
llvm::LLVMAddFunctionAttribute(llfn,
llvm::FunctionIndex as c_uint,
llvm::NoReturnAttribute as uint64_t)
}
}
_ => {}
}
if ccx.tcx.sess.opts.cg.no_redzone {
unsafe {
llvm::LLVMAddFunctionAttribute(llfn,
llvm::FunctionIndex as c_uint,
llvm::NoRedZoneAttribute as uint64_t)
}
}
llvm::SetFunctionCallConv(llfn, cc);
// Function addresses in Rust are never significant, allowing functions to be merged.
llvm::SetUnnamedAddr(llfn, true);
if ccx.is_split_stack_supported() {
set_split_stack(llfn);
}
llfn
}
// only use this for foreign function ABIs and glue, use `decl_rust_fn` for Rust functions
pub fn decl_cdecl_fn(ccx: &CrateContext,
name: &str,
ty: Type,
output: ty::t) -> ValueRef {
decl_fn(ccx, name, llvm::CCallConv, ty, output)
}
// only use this for foreign function ABIs and glue, use `get_extern_rust_fn` for Rust functions
pub fn get_extern_fn(ccx: &CrateContext,
externs: &mut ExternMap,
name: &str,
cc: llvm::CallConv,
ty: Type,
output: ty::t)
-> ValueRef {
match externs.find_equiv(&name) {
Some(n) => return *n,
None => {}
}
let f = decl_fn(ccx, name, cc, ty, output);
externs.insert(name.to_string(), f);
f
}
fn get_extern_rust_fn(ccx: &CrateContext, fn_ty: ty::t, name: &str, did: ast::DefId) -> ValueRef {
2014-03-20 19:49:20 -07:00
match ccx.externs.borrow().find_equiv(&name) {
Some(n) => return *n,
None => ()
}
let f = decl_rust_fn(ccx, fn_ty, name);
csearch::get_item_attrs(&ccx.sess().cstore, did, |attrs| {
set_llvm_fn_attrs(attrs.as_slice(), f)
});
ccx.externs.borrow_mut().insert(name.to_string(), f);
f
}
pub fn self_type_for_unboxed_closure(ccx: &CrateContext,
closure_id: ast::DefId)
-> ty::t {
let unboxed_closure_type = ty::mk_unboxed_closure(ccx.tcx(),
closure_id,
ty::ReStatic);
let unboxed_closures = ccx.tcx.unboxed_closures.borrow();
let unboxed_closure = unboxed_closures.get(&closure_id);
match unboxed_closure.kind {
ty::FnUnboxedClosureKind => {
ty::mk_imm_rptr(&ccx.tcx, ty::ReStatic, unboxed_closure_type)
}
ty::FnMutUnboxedClosureKind => {
ty::mk_mut_rptr(&ccx.tcx, ty::ReStatic, unboxed_closure_type)
}
ty::FnOnceUnboxedClosureKind => unboxed_closure_type,
}
}
pub fn kind_for_unboxed_closure(ccx: &CrateContext, closure_id: ast::DefId)
-> ty::UnboxedClosureKind {
let unboxed_closures = ccx.tcx.unboxed_closures.borrow();
unboxed_closures.get(&closure_id).kind
}
pub fn decl_rust_fn(ccx: &CrateContext, fn_ty: ty::t, name: &str) -> ValueRef {
let (inputs, output, abi, env) = match ty::get(fn_ty).sty {
ty::ty_bare_fn(ref f) => {
(f.sig.inputs.clone(), f.sig.output, f.abi, None)
}
ty::ty_closure(ref f) => {
(f.sig.inputs.clone(), f.sig.output, f.abi, Some(Type::i8p(ccx)))
}
ty::ty_unboxed_closure(closure_did, _) => {
let unboxed_closures = ccx.tcx.unboxed_closures.borrow();
let unboxed_closure = unboxed_closures.get(&closure_did);
let function_type = unboxed_closure.closure_type.clone();
let self_type = self_type_for_unboxed_closure(ccx, closure_did);
let llenvironment_type = type_of_explicit_arg(ccx, self_type);
(function_type.sig.inputs.clone(),
function_type.sig.output,
RustCall,
Some(llenvironment_type))
}
_ => fail!("expected closure or fn")
};
let llfty = type_of_rust_fn(ccx, env, inputs.as_slice(), output, abi);
debug!("decl_rust_fn(input count={},type={})",
inputs.len(),
ccx.tn.type_to_string(llfty));
let llfn = decl_fn(ccx, name, llvm::CCallConv, llfty, output);
let attrs = get_fn_llvm_attributes(ccx, fn_ty);
attrs.apply_llfn(llfn);
llfn
}
pub fn decl_internal_rust_fn(ccx: &CrateContext, fn_ty: ty::t, name: &str) -> ValueRef {
let llfn = decl_rust_fn(ccx, fn_ty, name);
llvm::SetLinkage(llfn, llvm::InternalLinkage);
llfn
}
pub fn get_extern_const(externs: &mut ExternMap, llmod: ModuleRef,
name: &str, ty: Type) -> ValueRef {
match externs.find_equiv(&name) {
Some(n) => return *n,
None => ()
}
unsafe {
let c = name.with_c_str(|buf| {
2013-06-16 02:29:52 +12:00
llvm::LLVMAddGlobal(llmod, ty.to_ref(), buf)
});
externs.insert(name.to_string(), c);
return c;
}
}
// Returns a pointer to the body for the box. The box may be an opaque
// box. The result will be casted to the type of body_t, if it is statically
// known.
pub fn at_box_body(bcx: &Block, body_t: ty::t, boxptr: ValueRef) -> ValueRef {
let _icx = push_ctxt("at_box_body");
let ccx = bcx.ccx();
let ty = Type::at_box(ccx, type_of(ccx, body_t));
2013-06-16 02:29:52 +12:00
let boxptr = PointerCast(bcx, boxptr, ty.ptr_to());
GEPi(bcx, boxptr, [0u, abi::box_field_body])
}
2014-04-06 13:54:41 +03:00
fn require_alloc_fn(bcx: &Block, info_ty: ty::t, it: LangItem) -> ast::DefId {
match bcx.tcx().lang_items.require(it) {
Ok(id) => id,
Err(s) => {
bcx.sess().fatal(format!("allocation of `{}` {}",
bcx.ty_to_string(info_ty),
s).as_slice());
}
}
2014-04-06 13:54:41 +03:00
}
2014-04-06 13:54:41 +03:00
// The following malloc_raw_dyn* functions allocate a box to contain
// a given type, but with a potentially dynamic size.
2014-04-06 13:54:41 +03:00
pub fn malloc_raw_dyn<'a>(bcx: &'a Block<'a>,
2014-04-25 15:14:52 +12:00
llty_ptr: Type,
info_ty: ty::t,
size: ValueRef,
align: ValueRef)
2014-04-06 13:54:41 +03:00
-> Result<'a> {
let _icx = push_ctxt("malloc_raw_exchange");
2014-04-06 13:54:41 +03:00
// Allocate space:
let r = callee::trans_lang_call(bcx,
2014-04-25 15:14:52 +12:00
require_alloc_fn(bcx, info_ty, ExchangeMallocFnLangItem),
[size, align],
2014-04-06 13:54:41 +03:00
None);
2014-05-03 23:14:56 +12:00
Result::new(r.bcx, PointerCast(r.bcx, r.val, llty_ptr))
}
2014-04-06 13:54:41 +03:00
pub fn malloc_raw_dyn_managed<'a>(
bcx: &'a Block<'a>,
t: ty::t,
alloc_fn: LangItem,
size: ValueRef)
-> Result<'a> {
let _icx = push_ctxt("malloc_raw_managed");
let ccx = bcx.ccx();
2014-04-06 13:54:41 +03:00
let langcall = require_alloc_fn(bcx, t, alloc_fn);
// Grab the TypeRef type of box_ptr_ty.
let box_ptr_ty = ty::mk_box(bcx.tcx(), t);
let llty = type_of(ccx, box_ptr_ty);
2014-08-06 11:59:40 +02:00
let llalign = C_uint(ccx, type_of::align_of(ccx, box_ptr_ty) as uint);
2014-04-06 13:54:41 +03:00
// Allocate space:
let drop_glue = glue::get_drop_glue(ccx, t);
let r = callee::trans_lang_call(
bcx,
langcall,
[
PointerCast(bcx, drop_glue, Type::glue_fn(ccx, Type::i8p(ccx)).ptr_to()),
size,
llalign
],
None);
2014-05-03 23:14:56 +12:00
Result::new(r.bcx, PointerCast(r.bcx, r.val, llty))
}
// Type descriptor and type glue stuff
2014-04-22 03:45:16 +03:00
pub fn get_tydesc(ccx: &CrateContext, t: ty::t) -> Rc<tydesc_info> {
2014-03-20 19:49:20 -07:00
match ccx.tydescs.borrow().find(&t) {
2014-04-22 03:45:16 +03:00
Some(inf) => return inf.clone(),
2014-03-20 19:49:20 -07:00
_ => { }
}
2013-03-15 15:24:24 -04:00
ccx.stats.n_static_tydescs.set(ccx.stats.n_static_tydescs.get() + 1u);
2014-04-22 03:45:16 +03:00
let inf = Rc::new(glue::declare_tydesc(ccx, t));
2014-04-22 03:45:16 +03:00
ccx.tydescs.borrow_mut().insert(t, inf.clone());
inf
}
2014-04-01 10:27:40 -04:00
#[allow(dead_code)] // useful
2013-02-14 14:48:15 -05:00
pub fn set_optimize_for_size(f: ValueRef) {
llvm::SetFunctionAttribute(f, llvm::OptimizeForSizeAttribute)
2013-02-14 14:48:15 -05:00
}
pub fn set_no_inline(f: ValueRef) {
llvm::SetFunctionAttribute(f, llvm::NoInlineAttribute)
}
2014-04-01 10:27:40 -04:00
#[allow(dead_code)] // useful
pub fn set_no_unwind(f: ValueRef) {
llvm::SetFunctionAttribute(f, llvm::NoUnwindAttribute)
}
2011-07-06 17:02:56 -07:00
// Tell LLVM to emit the information necessary to unwind the stack for the
// function f.
pub fn set_uwtable(f: ValueRef) {
llvm::SetFunctionAttribute(f, llvm::UWTableAttribute)
}
pub fn set_inline_hint(f: ValueRef) {
llvm::SetFunctionAttribute(f, llvm::InlineHintAttribute)
}
pub fn set_llvm_fn_attrs(attrs: &[ast::Attribute], llfn: ValueRef) {
use syntax::attr::*;
// Set the inline hint if there is one
match find_inline_attr(attrs) {
InlineHint => set_inline_hint(llfn),
InlineAlways => set_always_inline(llfn),
InlineNever => set_no_inline(llfn),
InlineNone => { /* fallthrough */ }
}
// Add the no-split-stack attribute if requested
if contains_name(attrs, "no_split_stack") {
unset_split_stack(llfn);
}
if contains_name(attrs, "cold") {
unsafe {
llvm::LLVMAddFunctionAttribute(llfn,
llvm::FunctionIndex as c_uint,
llvm::ColdAttribute as uint64_t)
}
}
}
pub fn set_always_inline(f: ValueRef) {
llvm::SetFunctionAttribute(f, llvm::AlwaysInlineAttribute)
}
pub fn set_split_stack(f: ValueRef) {
"split-stack".with_c_str(|buf| {
unsafe { llvm::LLVMAddFunctionAttrString(f, llvm::FunctionIndex as c_uint, buf); }
})
}
pub fn unset_split_stack(f: ValueRef) {
"split-stack".with_c_str(|buf| {
unsafe { llvm::LLVMRemoveFunctionAttrString(f, llvm::FunctionIndex as c_uint, buf); }
})
}
2012-03-20 14:21:02 -07:00
// Double-check that we never ask LLVM to declare the same symbol twice. It
// silently mangles such symbols, breaking our linkage model.
pub fn note_unique_llvm_symbol(ccx: &CrateContext, sym: String) {
2014-03-20 19:49:20 -07:00
if ccx.all_llvm_symbols.borrow().contains(&sym) {
ccx.sess().bug(format!("duplicate LLVM symbol: {}", sym).as_slice());
2012-03-20 14:21:02 -07:00
}
2014-03-20 19:49:20 -07:00
ccx.all_llvm_symbols.borrow_mut().insert(sym);
2012-03-20 14:21:02 -07:00
}
2014-03-06 18:47:24 +02:00
pub fn get_res_dtor(ccx: &CrateContext,
did: ast::DefId,
t: ty::t,
parent_id: ast::DefId,
substs: &subst::Substs)
-> ValueRef {
let _icx = push_ctxt("trans_res_dtor");
let did = if did.krate != ast::LOCAL_CRATE {
inline::maybe_instantiate_inline(ccx, did)
} else {
did
};
if !substs.types.is_empty() {
assert_eq!(did.krate, ast::LOCAL_CRATE);
// Since we're in trans we don't care for any region parameters
let ref substs = subst::Substs::erased(substs.types.clone());
let vtables = typeck::check::vtable::trans_resolve_method(ccx.tcx(), did.node, substs);
let (val, _) = monomorphize::monomorphic_fn(ccx, did, substs, vtables, None);
2013-02-19 00:14:56 -08:00
val
} else if did.krate == ast::LOCAL_CRATE {
get_item_val(ccx, did.node)
} else {
let tcx = ccx.tcx();
2014-03-09 15:20:44 +02:00
let name = csearch::get_symbol(&ccx.sess().cstore, did);
let class_ty = ty::lookup_item_type(tcx, parent_id).ty.subst(tcx, substs);
let llty = type_of_dtor(ccx, class_ty);
let dtor_ty = ty::mk_ctor_fn(ccx.tcx(), ast::DUMMY_NODE_ID,
[glue::get_drop_glue_type(ccx, t)], ty::mk_nil());
get_extern_fn(ccx,
&mut *ccx.externs.borrow_mut(),
name.as_slice(),
llvm::CCallConv,
llty,
dtor_ty)
}
}
// Structural comparison: a rather involved form of glue.
pub fn maybe_name_value(cx: &CrateContext, v: ValueRef, s: &str) {
2014-03-05 16:36:01 +02:00
if cx.sess().opts.cg.save_temps {
s.with_c_str(|buf| {
unsafe {
llvm::LLVMSetValueName(v, buf)
}
})
}
}
2011-08-02 12:26:52 -07:00
// Used only for creating scalar comparison glue.
pub enum scalar_type { nil_type, signed_int, unsigned_int, floating_point, }
pub fn compare_scalar_types<'a>(
cx: &'a Block<'a>,
lhs: ValueRef,
rhs: ValueRef,
t: ty::t,
op: ast::BinOp)
-> Result<'a> {
2014-05-03 23:14:56 +12:00
let f = |a| Result::new(cx, compare_scalar_values(cx, lhs, rhs, a, op));
2011-07-27 14:19:39 +02:00
match ty::get(t).sty {
ty::ty_nil => f(nil_type),
ty::ty_bool | ty::ty_uint(_) | ty::ty_char => f(unsigned_int),
ty::ty_ptr(mt) if ty::type_is_sized(cx.tcx(), mt.ty) => f(unsigned_int),
ty::ty_int(_) => f(signed_int),
ty::ty_float(_) => f(floating_point),
// Should never get here, because t is scalar.
_ => cx.sess().bug("non-scalar type passed to compare_scalar_types")
}
}
// A helper function to do the actual comparison of scalar values.
pub fn compare_scalar_values<'a>(
cx: &'a Block<'a>,
lhs: ValueRef,
rhs: ValueRef,
nt: scalar_type,
op: ast::BinOp)
-> ValueRef {
let _icx = push_ctxt("compare_scalar_values");
fn die(cx: &Block) -> ! {
2014-03-05 16:36:01 +02:00
cx.sess().bug("compare_scalar_values: must be a comparison operator");
}
2012-08-06 12:34:08 -07:00
match nt {
2012-08-03 19:59:04 -07:00
nil_type => {
2011-07-27 14:19:39 +02:00
// We don't need to do actual comparisons for nil.
// () == () holds but () < () does not.
2012-08-06 12:34:08 -07:00
match op {
ast::BiEq | ast::BiLe | ast::BiGe => return C_bool(cx.ccx(), true),
ast::BiNe | ast::BiLt | ast::BiGt => return C_bool(cx.ccx(), false),
// refinements would be nice
_ => die(cx)
}
2011-07-27 14:19:39 +02:00
}
2012-08-03 19:59:04 -07:00
floating_point => {
2012-08-06 12:34:08 -07:00
let cmp = match op {
ast::BiEq => llvm::RealOEQ,
ast::BiNe => llvm::RealUNE,
ast::BiLt => llvm::RealOLT,
ast::BiLe => llvm::RealOLE,
ast::BiGt => llvm::RealOGT,
ast::BiGe => llvm::RealOGE,
_ => die(cx)
};
2012-08-01 17:30:05 -07:00
return FCmp(cx, cmp, lhs, rhs);
2011-07-27 14:19:39 +02:00
}
2012-08-03 19:59:04 -07:00
signed_int => {
2012-08-06 12:34:08 -07:00
let cmp = match op {
ast::BiEq => llvm::IntEQ,
ast::BiNe => llvm::IntNE,
ast::BiLt => llvm::IntSLT,
ast::BiLe => llvm::IntSLE,
ast::BiGt => llvm::IntSGT,
ast::BiGe => llvm::IntSGE,
_ => die(cx)
};
2012-08-01 17:30:05 -07:00
return ICmp(cx, cmp, lhs, rhs);
2011-07-27 14:19:39 +02:00
}
2012-08-03 19:59:04 -07:00
unsigned_int => {
2012-08-06 12:34:08 -07:00
let cmp = match op {
ast::BiEq => llvm::IntEQ,
ast::BiNe => llvm::IntNE,
ast::BiLt => llvm::IntULT,
ast::BiLe => llvm::IntULE,
ast::BiGt => llvm::IntUGT,
ast::BiGe => llvm::IntUGE,
_ => die(cx)
};
2012-08-01 17:30:05 -07:00
return ICmp(cx, cmp, lhs, rhs);
2011-07-27 14:19:39 +02:00
}
}
}
pub fn compare_simd_types(
cx: &Block,
lhs: ValueRef,
rhs: ValueRef,
t: ty::t,
size: uint,
op: ast::BinOp)
-> ValueRef {
match ty::get(t).sty {
ty::ty_float(_) => {
// The comparison operators for floating point vectors are challenging.
// LLVM outputs a `< size x i1 >`, but if we perform a sign extension
// then bitcast to a floating point vector, the result will be `-NaN`
// for each truth value. Because of this they are unsupported.
cx.sess().bug("compare_simd_types: comparison operators \
not supported for floating point SIMD types")
},
ty::ty_uint(_) | ty::ty_int(_) => {
let cmp = match op {
ast::BiEq => llvm::IntEQ,
ast::BiNe => llvm::IntNE,
ast::BiLt => llvm::IntSLT,
ast::BiLe => llvm::IntSLE,
ast::BiGt => llvm::IntSGT,
ast::BiGe => llvm::IntSGE,
_ => cx.sess().bug("compare_simd_types: must be a comparison operator"),
};
let return_ty = Type::vector(&type_of(cx.ccx(), t), size as u64);
// LLVM outputs an `< size x i1 >`, so we need to perform a sign extension
// to get the correctly sized type. This will compile to a single instruction
// once the IR is converted to assembly if the SIMD instruction is supported
// by the target architecture.
SExt(cx, ICmp(cx, cmp, lhs, rhs), return_ty)
},
_ => cx.sess().bug("compare_simd_types: invalid SIMD type"),
}
}
pub type val_and_ty_fn<'r,'b> =
2014-04-07 13:30:48 -07:00
|&'b Block<'b>, ValueRef, ty::t|: 'r -> &'b Block<'b>;
// Iterates through the elements of a structural type.
pub fn iter_structural_ty<'r,
'b>(
cx: &'b Block<'b>,
av: ValueRef,
t: ty::t,
f: val_and_ty_fn<'r,'b>)
-> &'b Block<'b> {
let _icx = push_ctxt("iter_structural_ty");
fn iter_variant<'r,
'b>(
cx: &'b Block<'b>,
repr: &adt::Repr,
av: ValueRef,
2014-04-22 02:21:52 +03:00
variant: &ty::VariantInfo,
substs: &subst::Substs,
f: val_and_ty_fn<'r,'b>)
-> &'b Block<'b> {
let _icx = push_ctxt("iter_variant");
let tcx = cx.tcx();
let mut cx = cx;
for (i, &arg) in variant.args.iter().enumerate() {
cx = f(cx,
adt::trans_field_ptr(cx, repr, av, variant.disr_val, i),
arg.subst(tcx, substs));
}
2012-08-01 17:30:05 -07:00
return cx;
}
2014-08-06 11:59:40 +02:00
let (data_ptr, info) = if ty::type_is_sized(cx.tcx(), t) {
(av, None)
} else {
let data = GEPi(cx, av, [0, abi::slice_elt_base]);
let info = GEPi(cx, av, [0, abi::slice_elt_len]);
(Load(cx, data), Some(Load(cx, info)))
};
let mut cx = cx;
match ty::get(t).sty {
2013-11-28 12:22:53 -08:00
ty::ty_struct(..) => {
let repr = adt::represent_type(cx.ccx(), t);
expr::with_field_tys(cx.tcx(), t, None, |discr, field_tys| {
for (i, field_ty) in field_tys.iter().enumerate() {
2014-08-06 11:59:40 +02:00
let field_ty = field_ty.mt.ty;
let llfld_a = adt::trans_field_ptr(cx, &*repr, data_ptr, discr, i);
let val = if ty::type_is_sized(cx.tcx(), field_ty) {
llfld_a
} else {
let boxed_ty = ty::mk_open(cx.tcx(), field_ty);
let scratch = datum::rvalue_scratch_datum(cx, boxed_ty, "__fat_ptr_iter");
Store(cx, llfld_a, GEPi(cx, scratch.val, [0, abi::slice_elt_base]));
Store(cx, info.unwrap(), GEPi(cx, scratch.val, [0, abi::slice_elt_len]));
scratch.val
};
cx = f(cx, val, field_ty);
}
})
2011-07-27 14:19:39 +02:00
}
ty::ty_unboxed_closure(def_id, _) => {
let repr = adt::represent_type(cx.ccx(), t);
let upvars = ty::unboxed_closure_upvars(cx.tcx(), def_id);
for (i, upvar) in upvars.iter().enumerate() {
2014-08-06 11:59:40 +02:00
let llupvar = adt::trans_field_ptr(cx, &*repr, data_ptr, 0, i);
cx = f(cx, llupvar, upvar.ty);
}
}
ty::ty_vec(_, Some(n)) => {
2014-08-06 11:59:40 +02:00
let (base, len) = tvec::get_fixed_base_and_len(cx, data_ptr, n);
2014-04-06 13:54:41 +03:00
let unit_ty = ty::sequence_element_type(cx.tcx(), t);
cx = tvec::iter_vec_raw(cx, base, unit_ty, len, f);
2012-04-09 17:32:49 -07:00
}
ty::ty_tup(ref args) => {
2013-02-24 17:32:09 -08:00
let repr = adt::represent_type(cx.ccx(), t);
for (i, arg) in args.iter().enumerate() {
2014-08-06 11:59:40 +02:00
let llfld_a = adt::trans_field_ptr(cx, &*repr, data_ptr, 0, i);
2013-02-24 17:32:09 -08:00
cx = f(cx, llfld_a, *arg);
}
}
ty::ty_enum(tid, ref substs) => {
let fcx = cx.fcx;
let ccx = fcx.ccx;
2011-07-27 14:19:39 +02:00
let repr = adt::represent_type(ccx, t);
let variants = ty::enum_variants(ccx.tcx(), tid);
let n_variants = (*variants).len();
// NB: we must hit the discriminant first so that structural
// comparison know not to proceed when the discriminants differ.
2014-04-22 03:03:02 +03:00
match adt::trans_switch(cx, &*repr, av) {
(_match::single, None) => {
2014-04-22 03:03:02 +03:00
cx = iter_variant(cx, &*repr, av, &**variants.get(0),
substs, f);
}
(_match::switch, Some(lldiscrim_a)) => {
2013-04-22 20:19:05 -07:00
cx = f(cx, lldiscrim_a, ty::mk_int());
let unr_cx = fcx.new_temp_block("enum-iter-unr");
Unreachable(unr_cx);
let llswitch = Switch(cx, lldiscrim_a, unr_cx.llbb,
n_variants);
let next_cx = fcx.new_temp_block("enum-iter-next");
for variant in (*variants).iter() {
let variant_cx =
fcx.new_temp_block(
format!("enum-iter-variant-{}",
variant.disr_val.to_string().as_slice())
.as_slice());
2014-04-22 03:03:02 +03:00
match adt::trans_case(cx, &*repr, variant.disr_val) {
_match::single_result(r) => {
AddCase(llswitch, r.val, variant_cx.llbb)
}
2014-03-05 16:36:01 +02:00
_ => ccx.sess().unimpl("value from adt::trans_case \
in iter_structural_ty")
}
let variant_cx =
iter_variant(variant_cx,
2014-04-22 03:03:02 +03:00
&*repr,
2014-08-06 11:59:40 +02:00
data_ptr,
2014-04-22 02:21:52 +03:00
&**variant,
substs,
|x,y,z| f(x,y,z));
Br(variant_cx, next_cx.llbb);
}
cx = next_cx;
}
2014-03-05 16:36:01 +02:00
_ => ccx.sess().unimpl("value from adt::trans_switch \
in iter_structural_ty")
}
2011-07-27 14:19:39 +02:00
}
_ => cx.sess().unimpl("type in iter_structural_ty")
}
2012-08-01 17:30:05 -07:00
return cx;
}
pub fn cast_shift_expr_rhs<'a>(
cx: &'a Block<'a>,
op: ast::BinOp,
lhs: ValueRef,
rhs: ValueRef)
-> ValueRef {
cast_shift_rhs(op, lhs, rhs,
2012-06-30 16:19:07 -07:00
|a,b| Trunc(cx, a, b),
|a,b| ZExt(cx, a, b))
}
pub fn cast_shift_const_rhs(op: ast::BinOp,
lhs: ValueRef, rhs: ValueRef) -> ValueRef {
cast_shift_rhs(op, lhs, rhs,
2013-06-16 22:52:44 +12:00
|a, b| unsafe { llvm::LLVMConstTrunc(a, b.to_ref()) },
|a, b| unsafe { llvm::LLVMConstZExt(a, b.to_ref()) })
}
pub fn cast_shift_rhs(op: ast::BinOp,
lhs: ValueRef,
rhs: ValueRef,
trunc: |ValueRef, Type| -> ValueRef,
zext: |ValueRef, Type| -> ValueRef)
-> ValueRef {
// Shifts may have any size int on the rhs
unsafe {
if ast_util::is_shift_binop(op) {
2014-01-30 00:28:29 +09:00
let mut rhs_llty = val_ty(rhs);
let mut lhs_llty = val_ty(lhs);
if rhs_llty.kind() == Vector { rhs_llty = rhs_llty.element_type() }
if lhs_llty.kind() == Vector { lhs_llty = lhs_llty.element_type() }
2013-06-16 22:52:44 +12:00
let rhs_sz = llvm::LLVMGetIntTypeWidth(rhs_llty.to_ref());
let lhs_sz = llvm::LLVMGetIntTypeWidth(lhs_llty.to_ref());
if lhs_sz < rhs_sz {
trunc(rhs, lhs_llty)
} else if lhs_sz > rhs_sz {
// FIXME (#1877: If shifting by negative
// values becomes not undefined then this is wrong.
zext(rhs, lhs_llty)
} else {
rhs
}
} else {
rhs
}
}
}
pub fn fail_if_zero_or_overflows<'a>(
cx: &'a Block<'a>,
span: Span,
divrem: ast::BinOp,
lhs: ValueRef,
rhs: ValueRef,
rhs_t: ty::t)
-> &'a Block<'a> {
let (zero_text, overflow_text) = if divrem == ast::BiDiv {
("attempted to divide by zero",
"attempted to divide with overflow")
} else {
("attempted remainder with a divisor of zero",
"attempted remainder with overflow")
};
let (is_zero, is_signed) = match ty::get(rhs_t).sty {
ty::ty_int(t) => {
let zero = C_integral(Type::int_from_ty(cx.ccx(), t), 0u64, false);
(ICmp(cx, llvm::IntEQ, rhs, zero), true)
}
ty::ty_uint(t) => {
let zero = C_integral(Type::uint_from_ty(cx.ccx(), t), 0u64, false);
(ICmp(cx, llvm::IntEQ, rhs, zero), false)
}
_ => {
cx.sess().bug(format!("fail-if-zero on unexpected type: {}",
ty_to_string(cx.tcx(), rhs_t)).as_slice());
}
};
let bcx = with_cond(cx, is_zero, |bcx| {
controlflow::trans_fail(bcx, span, InternedString::new(zero_text))
});
// To quote LLVM's documentation for the sdiv instruction:
//
// Division by zero leads to undefined behavior. Overflow also leads
// to undefined behavior; this is a rare case, but can occur, for
// example, by doing a 32-bit division of -2147483648 by -1.
//
// In order to avoid undefined behavior, we perform runtime checks for
// signed division/remainder which would trigger overflow. For unsigned
// integers, no action beyond checking for zero need be taken.
if is_signed {
let (llty, min) = match ty::get(rhs_t).sty {
ty::ty_int(t) => {
let llty = Type::int_from_ty(cx.ccx(), t);
let min = match t {
ast::TyI if llty == Type::i32(cx.ccx()) => i32::MIN as u64,
ast::TyI => i64::MIN as u64,
ast::TyI8 => i8::MIN as u64,
ast::TyI16 => i16::MIN as u64,
ast::TyI32 => i32::MIN as u64,
ast::TyI64 => i64::MIN as u64,
};
(llty, min)
}
_ => unreachable!(),
};
let minus_one = ICmp(bcx, llvm::IntEQ, rhs,
C_integral(llty, -1, false));
with_cond(bcx, minus_one, |bcx| {
let is_min = ICmp(bcx, llvm::IntEQ, lhs,
C_integral(llty, min, true));
with_cond(bcx, is_min, |bcx| {
controlflow::trans_fail(bcx, span,
InternedString::new(overflow_text))
})
})
} else {
bcx
}
}
2013-12-19 16:47:15 -08:00
pub fn trans_external_path(ccx: &CrateContext, did: ast::DefId, t: ty::t) -> ValueRef {
2014-03-09 15:20:44 +02:00
let name = csearch::get_symbol(&ccx.sess().cstore, did);
match ty::get(t).sty {
ty::ty_bare_fn(ref fn_ty) => {
match fn_ty.abi.for_target(ccx.sess().targ_cfg.os,
ccx.sess().targ_cfg.arch) {
Some(Rust) | Some(RustCall) => {
get_extern_rust_fn(ccx, t, name.as_slice(), did)
}
Some(RustIntrinsic) => {
ccx.sess().bug("unexpected intrinsic in trans_external_path")
}
2013-11-28 12:22:53 -08:00
Some(..) | None => {
foreign::register_foreign_item_fn(ccx, fn_ty.abi, t,
name.as_slice(), None)
}
}
}
ty::ty_closure(_) => {
get_extern_rust_fn(ccx, t, name.as_slice(), did)
}
_ => {
let llty = type_of(ccx, t);
get_extern_const(&mut *ccx.externs.borrow_mut(),
ccx.llmod,
name.as_slice(),
2014-03-20 19:49:20 -07:00
llty)
}
}
}
pub fn invoke<'a>(
bcx: &'a Block<'a>,
llfn: ValueRef,
llargs: Vec<ValueRef> ,
fn_ty: ty::t,
DST coercions and DST structs [breaking-change] 1. The internal layout for traits has changed from (vtable, data) to (data, vtable). If you were relying on this in unsafe transmutes, you might get some very weird and apparently unrelated errors. You should not be doing this! Prefer not to do this at all, but if you must, you should use raw::TraitObject rather than hardcoding rustc's internal representation into your code. 2. The minimal type of reference-to-vec-literals (e.g., `&[1, 2, 3]`) is now a fixed size vec (e.g., `&[int, ..3]`) where it used to be an unsized vec (e.g., `&[int]`). If you want the unszied type, you must explicitly give the type (e.g., `let x: &[_] = &[1, 2, 3]`). Note in particular where multiple blocks must have the same type (e.g., if and else clauses, vec elements), the compiler will not coerce to the unsized type without a hint. E.g., `[&[1], &[1, 2]]` used to be a valid expression of type '[&[int]]'. It no longer type checks since the first element now has type `&[int, ..1]` and the second has type &[int, ..2]` which are incompatible. 3. The type of blocks (including functions) must be coercible to the expected type (used to be a subtype). Mostly this makes things more flexible and not less (in particular, in the case of coercing function bodies to the return type). However, in some rare cases, this is less flexible. TBH, I'm not exactly sure of the exact effects. I think the change causes us to resolve inferred type variables slightly earlier which might make us slightly more restrictive. Possibly it only affects blocks with unreachable code. E.g., `if ... { fail!(); "Hello" }` used to type check, it no longer does. The fix is to add a semicolon after the string.
2014-08-04 14:20:11 +02:00
call_info: Option<NodeInfo>,
// FIXME(15064) is_lang_item is a horrible hack, please remove it
// at the soonest opportunity.
is_lang_item: bool)
-> (ValueRef, &'a Block<'a>) {
let _icx = push_ctxt("invoke_");
if bcx.unreachable.get() {
return (C_null(Type::i8(bcx.ccx())), bcx);
}
DST coercions and DST structs [breaking-change] 1. The internal layout for traits has changed from (vtable, data) to (data, vtable). If you were relying on this in unsafe transmutes, you might get some very weird and apparently unrelated errors. You should not be doing this! Prefer not to do this at all, but if you must, you should use raw::TraitObject rather than hardcoding rustc's internal representation into your code. 2. The minimal type of reference-to-vec-literals (e.g., `&[1, 2, 3]`) is now a fixed size vec (e.g., `&[int, ..3]`) where it used to be an unsized vec (e.g., `&[int]`). If you want the unszied type, you must explicitly give the type (e.g., `let x: &[_] = &[1, 2, 3]`). Note in particular where multiple blocks must have the same type (e.g., if and else clauses, vec elements), the compiler will not coerce to the unsized type without a hint. E.g., `[&[1], &[1, 2]]` used to be a valid expression of type '[&[int]]'. It no longer type checks since the first element now has type `&[int, ..1]` and the second has type &[int, ..2]` which are incompatible. 3. The type of blocks (including functions) must be coercible to the expected type (used to be a subtype). Mostly this makes things more flexible and not less (in particular, in the case of coercing function bodies to the return type). However, in some rare cases, this is less flexible. TBH, I'm not exactly sure of the exact effects. I think the change causes us to resolve inferred type variables slightly earlier which might make us slightly more restrictive. Possibly it only affects blocks with unreachable code. E.g., `if ... { fail!(); "Hello" }` used to type check, it no longer does. The fix is to add a semicolon after the string.
2014-08-04 14:20:11 +02:00
// FIXME(15064) Lang item methods may (in the reflect case) not have proper
// types, so doing an attribute lookup will fail.
let attributes = if is_lang_item {
2014-08-06 11:59:40 +02:00
llvm::AttrBuilder::new()
DST coercions and DST structs [breaking-change] 1. The internal layout for traits has changed from (vtable, data) to (data, vtable). If you were relying on this in unsafe transmutes, you might get some very weird and apparently unrelated errors. You should not be doing this! Prefer not to do this at all, but if you must, you should use raw::TraitObject rather than hardcoding rustc's internal representation into your code. 2. The minimal type of reference-to-vec-literals (e.g., `&[1, 2, 3]`) is now a fixed size vec (e.g., `&[int, ..3]`) where it used to be an unsized vec (e.g., `&[int]`). If you want the unszied type, you must explicitly give the type (e.g., `let x: &[_] = &[1, 2, 3]`). Note in particular where multiple blocks must have the same type (e.g., if and else clauses, vec elements), the compiler will not coerce to the unsized type without a hint. E.g., `[&[1], &[1, 2]]` used to be a valid expression of type '[&[int]]'. It no longer type checks since the first element now has type `&[int, ..1]` and the second has type &[int, ..2]` which are incompatible. 3. The type of blocks (including functions) must be coercible to the expected type (used to be a subtype). Mostly this makes things more flexible and not less (in particular, in the case of coercing function bodies to the return type). However, in some rare cases, this is less flexible. TBH, I'm not exactly sure of the exact effects. I think the change causes us to resolve inferred type variables slightly earlier which might make us slightly more restrictive. Possibly it only affects blocks with unreachable code. E.g., `if ... { fail!(); "Hello" }` used to type check, it no longer does. The fix is to add a semicolon after the string.
2014-08-04 14:20:11 +02:00
} else {
get_fn_llvm_attributes(bcx.ccx(), fn_ty)
};
match bcx.opt_node_id {
None => {
debug!("invoke at ???");
}
Some(id) => {
debug!("invoke at {}", bcx.tcx().map.node_to_string(id));
}
}
if need_invoke(bcx) {
debug!("invoking {} at {}", llfn, bcx.llbb);
for &llarg in llargs.iter() {
debug!("arg: {}", llarg);
}
let normal_bcx = bcx.fcx.new_temp_block("normal-return");
let landing_pad = bcx.fcx.get_landing_pad();
match call_info {
Some(info) => debuginfo::set_source_location(bcx.fcx, info.id, info.span),
None => debuginfo::clear_source_location(bcx.fcx)
};
let llresult = Invoke(bcx,
llfn,
llargs.as_slice(),
normal_bcx.llbb,
landing_pad,
Some(attributes));
return (llresult, normal_bcx);
} else {
debug!("calling {} at {}", llfn, bcx.llbb);
for &llarg in llargs.iter() {
debug!("arg: {}", llarg);
}
match call_info {
Some(info) => debuginfo::set_source_location(bcx.fcx, info.id, info.span),
None => debuginfo::clear_source_location(bcx.fcx)
};
let llresult = Call(bcx, llfn, llargs.as_slice(), Some(attributes));
return (llresult, bcx);
}
}
pub fn need_invoke(bcx: &Block) -> bool {
2014-03-05 16:36:01 +02:00
if bcx.sess().no_landing_pads() {
2012-08-01 17:30:05 -07:00
return false;
}
// Avoid using invoke if we are already inside a landing pad.
if bcx.is_lpad {
2012-08-01 17:30:05 -07:00
return false;
}
bcx.fcx.needs_invoke()
}
pub fn load_if_immediate(cx: &Block, v: ValueRef, t: ty::t) -> ValueRef {
let _icx = push_ctxt("load_if_immediate");
if type_is_immediate(cx.ccx(), t) { return load_ty(cx, v, t); }
2012-08-01 17:30:05 -07:00
return v;
2010-12-03 13:03:07 -08:00
}
pub fn load_ty(cx: &Block, ptr: ValueRef, t: ty::t) -> ValueRef {
/*!
* Helper for loading values from memory. Does the necessary conversion if
* the in-memory type differs from the type used for SSA values. Also
* handles various special cases where the type gives us better information
* about what we are loading.
*/
if type_is_zero_size(cx.ccx(), t) {
C_undef(type_of::type_of(cx.ccx(), t))
} else if ty::type_is_bool(t) {
Trunc(cx, LoadRangeAssert(cx, ptr, 0, 2, llvm::False), Type::i1(cx.ccx()))
} else if ty::type_is_char(t) {
// a char is a unicode codepoint, and so takes values from 0
// to 0x10FFFF inclusive only.
LoadRangeAssert(cx, ptr, 0, 0x10FFFF + 1, llvm::False)
} else {
Load(cx, ptr)
}
}
pub fn store_ty(cx: &Block, v: ValueRef, dst: ValueRef, t: ty::t) {
/*!
* Helper for storing values in memory. Does the necessary conversion if
* the in-memory type differs from the type used for SSA values.
*/
if ty::type_is_bool(t) {
Store(cx, ZExt(cx, v, Type::i8(cx.ccx())), dst);
} else {
Store(cx, v, dst);
};
}
pub fn ignore_lhs(_bcx: &Block, local: &ast::Local) -> bool {
match local.pat.node {
ast::PatWild(ast::PatWildSingle) => true, _ => false
}
}
pub fn init_local<'a>(bcx: &'a Block<'a>, local: &ast::Local)
-> &'a Block<'a> {
debug!("init_local(bcx={}, local.id={:?})", bcx.to_str(), local.id);
let _indenter = indenter();
let _icx = push_ctxt("init_local");
_match::store_local(bcx, local)
}
pub fn raw_block<'a>(
fcx: &'a FunctionContext<'a>,
is_lpad: bool,
llbb: BasicBlockRef)
-> &'a Block<'a> {
common::Block::new(llbb, is_lpad, None, fcx)
}
pub fn with_cond<'a>(
bcx: &'a Block<'a>,
val: ValueRef,
f: |&'a Block<'a>| -> &'a Block<'a>)
-> &'a Block<'a> {
let _icx = push_ctxt("with_cond");
let fcx = bcx.fcx;
let next_cx = fcx.new_temp_block("next");
let cond_cx = fcx.new_temp_block("cond");
CondBr(bcx, val, cond_cx.llbb, next_cx.llbb);
let after_cx = f(cond_cx);
if !after_cx.terminated.get() {
Br(after_cx, next_cx.llbb);
}
next_cx
}
Emit LLVM lifetime intrinsics to improve stack usage and codegen in general Lifetime intrinsics help to reduce stack usage, because LLVM can apply stack coloring to reuse the stack slots of dead allocas for new ones. For example these functions now both use the same amount of stack, while previous `bar()` used five times as much as `foo()`: ````rust fn foo() { println("{}", 5); } fn bar() { println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); } ```` On top of that, LLVM can also optimize out certain operations when it knows that memory is dead after a certain point. For example, it can sometimes remove the zeroing used to cancel the drop glue. This is possible when the glue drop itself was already removed because the zeroing dominated the drop glue call. For example in: ````rust pub fn bar(x: (Box<int>, int)) -> (Box<int>, int) { x } ```` With optimizations, this currently results in: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 16, i32 8, i1 false) ret void } ```` But with lifetime intrinsics we get: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.lifetime.end(i64 16, i8* %2) ret void } ```` Fixes #15665
2014-05-01 19:32:07 +02:00
pub fn call_lifetime_start(cx: &Block, ptr: ValueRef) {
if cx.sess().opts.optimize == config::No {
return;
}
let _icx = push_ctxt("lifetime_start");
let ccx = cx.ccx();
let llsize = C_u64(ccx, machine::llsize_of_alloc(ccx, val_ty(ptr).element_type()));
let ptr = PointerCast(cx, ptr, Type::i8p(ccx));
let lifetime_start = ccx.get_intrinsic(&"llvm.lifetime.start");
Call(cx, lifetime_start, [llsize, ptr], None);
Emit LLVM lifetime intrinsics to improve stack usage and codegen in general Lifetime intrinsics help to reduce stack usage, because LLVM can apply stack coloring to reuse the stack slots of dead allocas for new ones. For example these functions now both use the same amount of stack, while previous `bar()` used five times as much as `foo()`: ````rust fn foo() { println("{}", 5); } fn bar() { println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); } ```` On top of that, LLVM can also optimize out certain operations when it knows that memory is dead after a certain point. For example, it can sometimes remove the zeroing used to cancel the drop glue. This is possible when the glue drop itself was already removed because the zeroing dominated the drop glue call. For example in: ````rust pub fn bar(x: (Box<int>, int)) -> (Box<int>, int) { x } ```` With optimizations, this currently results in: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 16, i32 8, i1 false) ret void } ```` But with lifetime intrinsics we get: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.lifetime.end(i64 16, i8* %2) ret void } ```` Fixes #15665
2014-05-01 19:32:07 +02:00
}
pub fn call_lifetime_end(cx: &Block, ptr: ValueRef) {
if cx.sess().opts.optimize == config::No {
return;
}
let _icx = push_ctxt("lifetime_end");
let ccx = cx.ccx();
let llsize = C_u64(ccx, machine::llsize_of_alloc(ccx, val_ty(ptr).element_type()));
let ptr = PointerCast(cx, ptr, Type::i8p(ccx));
let lifetime_end = ccx.get_intrinsic(&"llvm.lifetime.end");
Call(cx, lifetime_end, [llsize, ptr], None);
Emit LLVM lifetime intrinsics to improve stack usage and codegen in general Lifetime intrinsics help to reduce stack usage, because LLVM can apply stack coloring to reuse the stack slots of dead allocas for new ones. For example these functions now both use the same amount of stack, while previous `bar()` used five times as much as `foo()`: ````rust fn foo() { println("{}", 5); } fn bar() { println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); } ```` On top of that, LLVM can also optimize out certain operations when it knows that memory is dead after a certain point. For example, it can sometimes remove the zeroing used to cancel the drop glue. This is possible when the glue drop itself was already removed because the zeroing dominated the drop glue call. For example in: ````rust pub fn bar(x: (Box<int>, int)) -> (Box<int>, int) { x } ```` With optimizations, this currently results in: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 16, i32 8, i1 false) ret void } ```` But with lifetime intrinsics we get: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.lifetime.end(i64 16, i8* %2) ret void } ```` Fixes #15665
2014-05-01 19:32:07 +02:00
}
pub fn call_memcpy(cx: &Block, dst: ValueRef, src: ValueRef, n_bytes: ValueRef, align: u32) {
let _icx = push_ctxt("call_memcpy");
let ccx = cx.ccx();
2014-03-05 16:36:01 +02:00
let key = match ccx.sess().targ_cfg.arch {
2014-06-17 09:16:03 +02:00
X86 | Arm | Mips | Mipsel => "llvm.memcpy.p0i8.p0i8.i32",
X86_64 => "llvm.memcpy.p0i8.p0i8.i64"
};
2014-04-09 19:56:31 -04:00
let memcpy = ccx.get_intrinsic(&key);
let src_ptr = PointerCast(cx, src, Type::i8p(ccx));
let dst_ptr = PointerCast(cx, dst, Type::i8p(ccx));
let size = IntCast(cx, n_bytes, ccx.int_type);
let align = C_i32(ccx, align as i32);
let volatile = C_bool(ccx, false);
Call(cx, memcpy, [dst_ptr, src_ptr, size, align, volatile], None);
}
pub fn memcpy_ty(bcx: &Block, dst: ValueRef, src: ValueRef, t: ty::t) {
let _icx = push_ctxt("memcpy_ty");
let ccx = bcx.ccx();
if ty::type_is_structural(t) {
2013-05-25 09:52:25 -04:00
let llty = type_of::type_of(ccx, t);
let llsz = llsize_of(ccx, llty);
2014-08-06 11:59:40 +02:00
let llalign = type_of::align_of(ccx, t);
2013-05-25 09:52:25 -04:00
call_memcpy(bcx, dst, src, llsz, llalign as u32);
} else {
store_ty(bcx, Load(bcx, src), dst, t);
2011-07-27 14:19:39 +02:00
}
}
pub fn zero_mem(cx: &Block, llptr: ValueRef, t: ty::t) {
if cx.unreachable.get() { return; }
let _icx = push_ctxt("zero_mem");
let bcx = cx;
2014-08-06 11:59:40 +02:00
memzero(&B(bcx), llptr, t);
}
// Always use this function instead of storing a zero constant to the memory
// in question. If you store a zero constant, LLVM will drown in vreg
// allocation for large data structures, and the generated code will be
// awful. (A telltale sign of this is large quantities of
// `mov [byte ptr foo],0` in the generated code.)
2014-08-06 11:59:40 +02:00
fn memzero(b: &Builder, llptr: ValueRef, ty: ty::t) {
let _icx = push_ctxt("memzero");
let ccx = b.ccx;
2014-08-06 11:59:40 +02:00
let llty = type_of::type_of(ccx, ty);
2014-03-05 16:36:01 +02:00
let intrinsic_key = match ccx.sess().targ_cfg.arch {
2014-06-17 09:16:03 +02:00
X86 | Arm | Mips | Mipsel => "llvm.memset.p0i8.i32",
X86_64 => "llvm.memset.p0i8.i64"
};
2014-04-09 19:56:31 -04:00
let llintrinsicfn = ccx.get_intrinsic(&intrinsic_key);
let llptr = b.pointercast(llptr, Type::i8(ccx).ptr_to());
let llzeroval = C_u8(ccx, 0);
2014-08-06 11:59:40 +02:00
let size = machine::llsize_of(ccx, llty);
let align = C_i32(ccx, type_of::align_of(ccx, ty) as i32);
let volatile = C_bool(ccx, false);
b.call(llintrinsicfn, [llptr, llzeroval, size, align, volatile], None);
}
pub fn alloc_ty(bcx: &Block, t: ty::t, name: &str) -> ValueRef {
let _icx = push_ctxt("alloc_ty");
let ccx = bcx.ccx();
2013-06-16 02:29:52 +12:00
let ty = type_of::type_of(ccx, t);
2013-09-27 22:38:08 -07:00
assert!(!ty::type_has_params(t));
let val = alloca(bcx, ty, name);
return val;
}
pub fn alloca(cx: &Block, ty: Type, name: &str) -> ValueRef {
let p = alloca_no_lifetime(cx, ty, name);
call_lifetime_start(cx, p);
p
}
pub fn alloca_no_lifetime(cx: &Block, ty: Type, name: &str) -> ValueRef {
let _icx = push_ctxt("alloca");
if cx.unreachable.get() {
unsafe {
return llvm::LLVMGetUndef(ty.ptr_to().to_ref());
}
}
debuginfo::clear_source_location(cx.fcx);
Alloca(cx, ty, name)
}
2014-08-06 11:59:40 +02:00
pub fn alloca_zeroed(cx: &Block, ty: ty::t, name: &str) -> ValueRef {
let llty = type_of::type_of(cx.ccx(), ty);
if cx.unreachable.get() {
unsafe {
2014-08-06 11:59:40 +02:00
return llvm::LLVMGetUndef(llty.ptr_to().to_ref());
}
}
2014-08-06 11:59:40 +02:00
let p = alloca_no_lifetime(cx, llty, name);
let b = cx.fcx.ccx.builder();
b.position_before(cx.fcx.alloca_insert_pt.get().unwrap());
memzero(&b, p, ty);
2013-06-16 02:29:52 +12:00
p
}
pub fn arrayalloca(cx: &Block, ty: Type, v: ValueRef) -> ValueRef {
let _icx = push_ctxt("arrayalloca");
if cx.unreachable.get() {
unsafe {
2013-06-16 22:52:44 +12:00
return llvm::LLVMGetUndef(ty.to_ref());
}
}
debuginfo::clear_source_location(cx.fcx);
Emit LLVM lifetime intrinsics to improve stack usage and codegen in general Lifetime intrinsics help to reduce stack usage, because LLVM can apply stack coloring to reuse the stack slots of dead allocas for new ones. For example these functions now both use the same amount of stack, while previous `bar()` used five times as much as `foo()`: ````rust fn foo() { println("{}", 5); } fn bar() { println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); println("{}", 5); } ```` On top of that, LLVM can also optimize out certain operations when it knows that memory is dead after a certain point. For example, it can sometimes remove the zeroing used to cancel the drop glue. This is possible when the glue drop itself was already removed because the zeroing dominated the drop glue call. For example in: ````rust pub fn bar(x: (Box<int>, int)) -> (Box<int>, int) { x } ```` With optimizations, this currently results in: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 16, i32 8, i1 false) ret void } ```` But with lifetime intrinsics we get: ````llvm define void @_ZN3bar20h330fa42547df8179niaE({ i64*, i64 }* noalias nocapture nonnull sret, { i64*, i64 }* noalias nocapture nonnull) unnamed_addr #0 { "_ZN29_$LP$Box$LT$int$GT$$C$int$RP$39glue_drop.$x22glue_drop$x22$LP$1347$RP$17h88cf42702e5a322aE.exit": %2 = bitcast { i64*, i64 }* %1 to i8* %3 = bitcast { i64*, i64 }* %0 to i8* tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %2, i64 16, i32 8, i1 false) tail call void @llvm.lifetime.end(i64 16, i8* %2) ret void } ```` Fixes #15665
2014-05-01 19:32:07 +02:00
let p = ArrayAlloca(cx, ty, v);
call_lifetime_start(cx, p);
p
}
// Creates the alloca slot which holds the pointer to the slot for the final return value
pub fn make_return_slot_pointer(fcx: &FunctionContext, output_type: ty::t) -> ValueRef {
let lloutputtype = type_of::type_of(fcx.ccx, output_type);
// We create an alloca to hold a pointer of type `output_type`
// which will hold the pointer to the right alloca which has the
// final ret value
if fcx.needs_ret_allocas {
// Let's create the stack slot
let slot = AllocaFcx(fcx, lloutputtype.ptr_to(), "llretslotptr");
// and if we're using an out pointer, then store that in our newly made slot
if type_of::return_uses_outptr(fcx.ccx, output_type) {
let outptr = get_param(fcx.llfn, 0);
let b = fcx.ccx.builder();
b.position_before(fcx.alloca_insert_pt.get().unwrap());
b.store(outptr, slot);
}
slot
// But if there are no nested returns, we skip the indirection and have a single
// retslot
} else {
if type_of::return_uses_outptr(fcx.ccx, output_type) {
get_param(fcx.llfn, 0)
} else {
AllocaFcx(fcx, lloutputtype, "sret_slot")
}
}
}
struct CheckForNestedReturnsVisitor {
found: bool
}
impl Visitor<bool> for CheckForNestedReturnsVisitor {
fn visit_expr(&mut self, e: &ast::Expr, in_return: bool) {
match e.node {
ast::ExprRet(..) if in_return => {
self.found = true;
return;
}
ast::ExprRet(..) => visit::walk_expr(self, e, true),
_ => visit::walk_expr(self, e, in_return)
}
}
}
fn has_nested_returns(tcx: &ty::ctxt, id: ast::NodeId) -> bool {
match tcx.map.find(id) {
Some(ast_map::NodeItem(i)) => {
match i.node {
ast::ItemFn(_, _, _, _, blk) => {
let mut explicit = CheckForNestedReturnsVisitor { found: false };
let mut implicit = CheckForNestedReturnsVisitor { found: false };
visit::walk_item(&mut explicit, &*i, false);
visit::walk_expr_opt(&mut implicit, blk.expr, true);
explicit.found || implicit.found
}
_ => tcx.sess.bug("unexpected item variant in has_nested_returns")
}
}
Some(ast_map::NodeTraitItem(trait_method)) => {
match *trait_method {
ast::ProvidedMethod(m) => {
match m.node {
ast::MethDecl(_, _, _, _, _, _, blk, _) => {
let mut explicit = CheckForNestedReturnsVisitor { found: false };
let mut implicit = CheckForNestedReturnsVisitor { found: false };
visit::walk_method_helper(&mut explicit, &*m, false);
visit::walk_expr_opt(&mut implicit, blk.expr, true);
explicit.found || implicit.found
}
ast::MethMac(_) => tcx.sess.bug("unexpanded macro")
}
}
ast::RequiredMethod(_) => {
tcx.sess.bug("unexpected variant: required trait method \
in has_nested_returns")
}
}
}
Some(ast_map::NodeImplItem(ref ii)) => {
match **ii {
ast::MethodImplItem(ref m) => {
match m.node {
ast::MethDecl(_, _, _, _, _, _, blk, _) => {
let mut explicit = CheckForNestedReturnsVisitor {
found: false,
};
let mut implicit = CheckForNestedReturnsVisitor {
found: false,
};
visit::walk_method_helper(&mut explicit,
&**m,
false);
visit::walk_expr_opt(&mut implicit,
blk.expr,
true);
explicit.found || implicit.found
}
ast::MethMac(_) => tcx.sess.bug("unexpanded macro")
}
}
}
}
Some(ast_map::NodeExpr(e)) => {
match e.node {
ast::ExprFnBlock(_, _, blk) |
ast::ExprProc(_, blk) |
ast::ExprUnboxedFn(_, _, _, blk) => {
let mut explicit = CheckForNestedReturnsVisitor { found: false };
let mut implicit = CheckForNestedReturnsVisitor { found: false };
visit::walk_expr(&mut explicit, &*e, false);
visit::walk_expr_opt(&mut implicit, blk.expr, true);
explicit.found || implicit.found
}
_ => tcx.sess.bug("unexpected expr variant in has_nested_returns")
}
}
Some(ast_map::NodeVariant(..)) | Some(ast_map::NodeStructCtor(..)) => false,
// glue, shims, etc
None if id == ast::DUMMY_NODE_ID => false,
_ => tcx.sess.bug(format!("unexpected variant in has_nested_returns: {}",
tcx.map.path_to_string(id)).as_slice())
}
}
// NB: must keep 4 fns in sync:
//
// - type_of_fn
// - create_datums_for_fn_args.
// - new_fn_ctxt
// - trans_args
//
// Be warned! You must call `init_function` before doing anything with the
// returned function context.
pub fn new_fn_ctxt<'a>(ccx: &'a CrateContext,
llfndecl: ValueRef,
id: ast::NodeId,
has_env: bool,
output_type: ty::t,
param_substs: &'a param_substs,
sp: Option<Span>,
block_arena: &'a TypedArena<Block<'a>>)
-> FunctionContext<'a> {
param_substs.validate();
debug!("new_fn_ctxt(path={}, id={}, param_substs={})",
if id == -1 {
2014-05-25 03:10:11 -07:00
"".to_string()
} else {
ccx.tcx.map.path_to_string(id).to_string()
},
id, param_substs.repr(ccx.tcx()));
let substd_output_type = output_type.substp(ccx.tcx(), param_substs);
let uses_outptr = type_of::return_uses_outptr(ccx, substd_output_type);
let debug_context = debuginfo::create_function_debug_context(ccx, id, param_substs, llfndecl);
let nested_returns = has_nested_returns(ccx.tcx(), id);
let mut fcx = FunctionContext {
llfn: llfndecl,
llenv: None,
llretslotptr: Cell::new(None),
alloca_insert_pt: Cell::new(None),
llreturn: Cell::new(None),
needs_ret_allocas: nested_returns,
personality: Cell::new(None),
caller_expects_out_pointer: uses_outptr,
llargs: RefCell::new(NodeMap::new()),
lllocals: RefCell::new(NodeMap::new()),
llupvars: RefCell::new(NodeMap::new()),
id: id,
param_substs: param_substs,
span: sp,
block_arena: block_arena,
ccx: ccx,
debug_context: debug_context,
scopes: RefCell::new(Vec::new())
};
if has_env {
fcx.llenv = Some(get_param(fcx.llfn, fcx.env_arg_pos() as c_uint))
}
fcx
}
/// Performs setup on a newly created function, creating the entry scope block
/// and allocating space for the return pointer.
2014-04-20 20:04:57 +03:00
pub fn init_function<'a>(fcx: &'a FunctionContext<'a>,
skip_retptr: bool,
output_type: ty::t) -> &'a Block<'a> {
let entry_bcx = fcx.new_temp_block("entry-block");
// Use a dummy instruction as the insertion point for all allocas.
// This is later removed in FunctionContext::cleanup.
fcx.alloca_insert_pt.set(Some(unsafe {
Load(entry_bcx, C_null(Type::i8p(fcx.ccx)));
llvm::LLVMGetFirstInstruction(entry_bcx.llbb)
}));
2014-04-20 20:04:57 +03:00
// This shouldn't need to recompute the return type,
// as new_fn_ctxt did it already.
let substd_output_type = output_type.substp(fcx.ccx.tcx(), fcx.param_substs);
if !return_type_is_void(fcx.ccx, substd_output_type) {
// If the function returns nil/bot, there is no real return
// value, so do not set `llretslotptr`.
if !skip_retptr || fcx.caller_expects_out_pointer {
// Otherwise, we normally allocate the llretslotptr, unless we
// have been instructed to skip it for immediate return
// values.
fcx.llretslotptr.set(Some(make_return_slot_pointer(fcx, substd_output_type)));
}
}
entry_bcx
}
// NB: must keep 4 fns in sync:
//
// - type_of_fn
// - create_datums_for_fn_args.
// - new_fn_ctxt
// - trans_args
pub fn arg_kind(cx: &FunctionContext, t: ty::t) -> datum::Rvalue {
use middle::trans::datum::{ByRef, ByValue};
datum::Rvalue {
mode: if arg_is_indirect(cx.ccx, t) { ByRef } else { ByValue }
}
}
// work around bizarre resolve errors
2014-03-02 15:26:39 -08:00
pub type RvalueDatum = datum::Datum<datum::Rvalue>;
pub type LvalueDatum = datum::Datum<datum::Lvalue>;
// create_datums_for_fn_args: creates rvalue datums for each of the
// incoming function arguments. These will later be stored into
// appropriate lvalue datums.
pub fn create_datums_for_fn_args(fcx: &FunctionContext,
arg_tys: &[ty::t])
-> Vec<RvalueDatum> {
let _icx = push_ctxt("create_datums_for_fn_args");
// Return an array wrapping the ValueRefs that we get from `get_param` for
// each argument into datums.
arg_tys.iter().enumerate().map(|(i, &arg_ty)| {
let llarg = get_param(fcx.llfn, fcx.arg_pos(i) as c_uint);
datum::Datum::new(llarg, arg_ty, arg_kind(fcx, arg_ty))
}).collect()
}
/// Creates rvalue datums for each of the incoming function arguments and
/// tuples the arguments. These will later be stored into appropriate lvalue
/// datums.
///
/// FIXME(pcwalton): Reduce the amount of code bloat this is responsible for.
fn create_datums_for_fn_args_under_call_abi<
'a>(
mut bcx: &'a Block<'a>,
arg_scope: cleanup::CustomScopeIndex,
arg_tys: &[ty::t])
-> Vec<RvalueDatum> {
let mut result = Vec::new();
for (i, &arg_ty) in arg_tys.iter().enumerate() {
if i < arg_tys.len() - 1 {
// Regular argument.
let llarg = get_param(bcx.fcx.llfn, bcx.fcx.arg_pos(i) as c_uint);
result.push(datum::Datum::new(llarg, arg_ty, arg_kind(bcx.fcx,
arg_ty)));
continue
}
// This is the last argument. Tuple it.
match ty::get(arg_ty).sty {
ty::ty_tup(ref tupled_arg_tys) => {
let tuple_args_scope_id = cleanup::CustomScope(arg_scope);
let tuple =
unpack_datum!(bcx,
datum::lvalue_scratch_datum(bcx,
arg_ty,
"tupled_args",
false,
tuple_args_scope_id,
(),
|(),
mut bcx,
llval| {
for (j, &tupled_arg_ty) in
tupled_arg_tys.iter().enumerate() {
let llarg =
get_param(bcx.fcx.llfn,
bcx.fcx.arg_pos(i + j) as c_uint);
let lldest = GEPi(bcx, llval, [0, j]);
let datum = datum::Datum::new(
llarg,
tupled_arg_ty,
arg_kind(bcx.fcx, tupled_arg_ty));
bcx = datum.store_to(bcx, lldest);
}
bcx
}));
let tuple = unpack_datum!(bcx,
tuple.to_expr_datum()
.to_rvalue_datum(bcx,
"argtuple"));
result.push(tuple);
}
ty::ty_nil => {
let mode = datum::Rvalue::new(datum::ByValue);
result.push(datum::Datum::new(C_nil(bcx.ccx()),
ty::mk_nil(),
mode))
}
_ => {
bcx.tcx().sess.bug("last argument of a function with \
`rust-call` ABI isn't a tuple?!")
}
};
}
result
}
fn copy_args_to_allocas<'a>(fcx: &FunctionContext<'a>,
arg_scope: cleanup::CustomScopeIndex,
bcx: &'a Block<'a>,
args: &[ast::Arg],
arg_datums: Vec<RvalueDatum> )
-> &'a Block<'a> {
debug!("copy_args_to_allocas");
let _icx = push_ctxt("copy_args_to_allocas");
let mut bcx = bcx;
let arg_scope_id = cleanup::CustomScope(arg_scope);
for (i, arg_datum) in arg_datums.move_iter().enumerate() {
// For certain mode/type combinations, the raw llarg values are passed
// by value. However, within the fn body itself, we want to always
// have all locals and arguments be by-ref so that we can cancel the
// cleanup and for better interaction with LLVM's debug info. So, if
// the argument would be passed by value, we store it into an alloca.
// This alloca should be optimized away by LLVM's mem-to-reg pass in
// the event it's not truly needed.
bcx = _match::store_arg(bcx, args[i].pat, arg_datum, arg_scope_id);
2014-03-05 16:36:01 +02:00
if fcx.ccx.sess().opts.debuginfo == FullDebugInfo {
debuginfo::create_argument_metadata(bcx, &args[i]);
}
}
bcx
}
fn copy_unboxed_closure_args_to_allocas<'a>(
mut bcx: &'a Block<'a>,
arg_scope: cleanup::CustomScopeIndex,
args: &[ast::Arg],
arg_datums: Vec<RvalueDatum>,
monomorphized_arg_types: &[ty::t])
-> &'a Block<'a> {
let _icx = push_ctxt("copy_unboxed_closure_args_to_allocas");
let arg_scope_id = cleanup::CustomScope(arg_scope);
assert_eq!(arg_datums.len(), 1);
let arg_datum = arg_datums.move_iter().next().unwrap();
// Untuple the rest of the arguments.
let tuple_datum =
unpack_datum!(bcx,
arg_datum.to_lvalue_datum_in_scope(bcx,
"argtuple",
arg_scope_id));
let empty = Vec::new();
let untupled_arg_types = match ty::get(monomorphized_arg_types[0]).sty {
ty::ty_tup(ref types) => types.as_slice(),
ty::ty_nil => empty.as_slice(),
_ => {
bcx.tcx().sess.span_bug(args[0].pat.span,
"first arg to `rust-call` ABI function \
wasn't a tuple?!")
}
};
for j in range(0, args.len()) {
let tuple_element_type = untupled_arg_types[j];
let tuple_element_datum =
2014-08-06 11:59:40 +02:00
tuple_datum.get_element(bcx,
tuple_element_type,
|llval| GEPi(bcx, llval, [0, j]));
let tuple_element_datum = tuple_element_datum.to_expr_datum();
let tuple_element_datum =
unpack_datum!(bcx,
tuple_element_datum.to_rvalue_datum(bcx,
"arg"));
bcx = _match::store_arg(bcx,
args[j].pat,
tuple_element_datum,
arg_scope_id);
if bcx.fcx.ccx.sess().opts.debuginfo == FullDebugInfo {
debuginfo::create_argument_metadata(bcx, &args[j]);
}
}
bcx
}
// Ties up the llstaticallocas -> llloadenv -> lltop edges,
// and builds the return block.
pub fn finish_fn<'a>(fcx: &'a FunctionContext<'a>,
last_bcx: &'a Block<'a>,
retty: ty::t) {
let _icx = push_ctxt("finish_fn");
// This shouldn't need to recompute the return type,
// as new_fn_ctxt did it already.
let substd_retty = retty.substp(fcx.ccx.tcx(), fcx.param_substs);
let ret_cx = match fcx.llreturn.get() {
Some(llreturn) => {
if !last_bcx.terminated.get() {
Br(last_bcx, llreturn);
}
raw_block(fcx, false, llreturn)
}
None => last_bcx
};
build_return_block(fcx, ret_cx, substd_retty);
debuginfo::clear_source_location(fcx);
fcx.cleanup();
}
// Builds the return block for a function.
pub fn build_return_block(fcx: &FunctionContext, ret_cx: &Block, retty: ty::t) {
if fcx.llretslotptr.get().is_none() ||
(!fcx.needs_ret_allocas && fcx.caller_expects_out_pointer) {
return RetVoid(ret_cx);
}
let retslot = if fcx.needs_ret_allocas {
Load(ret_cx, fcx.llretslotptr.get().unwrap())
} else {
fcx.llretslotptr.get().unwrap()
};
let retptr = Value(retslot);
match retptr.get_dominating_store(ret_cx) {
// If there's only a single store to the ret slot, we can directly return
// the value that was stored and omit the store and the alloca
Some(s) => {
let retval = s.get_operand(0).unwrap().get();
s.erase_from_parent();
if retptr.has_no_uses() {
retptr.erase_from_parent();
}
let retval = if ty::type_is_bool(retty) {
Trunc(ret_cx, retval, Type::i1(fcx.ccx))
} else {
retval
};
if fcx.caller_expects_out_pointer {
store_ty(ret_cx, retval, get_param(fcx.llfn, 0), retty);
return RetVoid(ret_cx);
} else {
return Ret(ret_cx, retval);
}
}
// Otherwise, copy the return value to the ret slot
None => {
if fcx.caller_expects_out_pointer {
memcpy_ty(ret_cx, get_param(fcx.llfn, 0), retslot, retty);
return RetVoid(ret_cx);
} else {
return Ret(ret_cx, load_ty(ret_cx, retslot, retty));
}
}
}
2012-02-13 16:06:56 -08:00
}
#[deriving(Clone, Eq, PartialEq)]
pub enum IsUnboxedClosureFlag {
NotUnboxedClosure,
IsUnboxedClosure,
}
2011-06-29 19:50:50 -07:00
// trans_closure: Builds an LLVM function out of a source function.
// If the function closes over its environment a closure will be
// returned.
2014-03-06 18:47:24 +02:00
pub fn trans_closure(ccx: &CrateContext,
decl: &ast::FnDecl,
body: &ast::Block,
llfndecl: ValueRef,
param_substs: &param_substs,
2014-03-06 18:47:24 +02:00
id: ast::NodeId,
_attributes: &[ast::Attribute],
arg_types: Vec<ty::t>,
2014-03-06 18:47:24 +02:00
output_type: ty::t,
abi: Abi,
has_env: bool,
is_unboxed_closure: IsUnboxedClosureFlag,
maybe_load_env: <'a>|&'a Block<'a>, ScopeId|
-> &'a Block<'a>) {
ccx.stats.n_closures.set(ccx.stats.n_closures.get() + 1);
let _icx = push_ctxt("trans_closure");
set_uwtable(llfndecl);
debug!("trans_closure(..., param_substs={})",
param_substs.repr(ccx.tcx()));
let arena = TypedArena::new();
let fcx = new_fn_ctxt(ccx,
llfndecl,
id,
has_env,
output_type,
param_substs,
Some(body.span),
&arena);
let mut bcx = init_function(&fcx, false, output_type);
// cleanup scope for the incoming arguments
let arg_scope = fcx.push_custom_cleanup_scope();
let block_ty = node_id_type(bcx, body.id);
// Set up arguments to the function.
let monomorphized_arg_types =
arg_types.iter()
.map(|at| monomorphize_type(bcx, *at))
.collect::<Vec<_>>();
for monomorphized_arg_type in monomorphized_arg_types.iter() {
debug!("trans_closure: monomorphized_arg_type: {}",
ty_to_string(ccx.tcx(), *monomorphized_arg_type));
}
debug!("trans_closure: function lltype: {}",
bcx.fcx.ccx.tn.val_to_string(bcx.fcx.llfn));
let arg_datums = if abi != RustCall {
create_datums_for_fn_args(&fcx,
monomorphized_arg_types.as_slice())
} else {
create_datums_for_fn_args_under_call_abi(
bcx,
arg_scope,
monomorphized_arg_types.as_slice())
};
bcx = match is_unboxed_closure {
NotUnboxedClosure => {
copy_args_to_allocas(&fcx,
arg_scope,
bcx,
decl.inputs.as_slice(),
arg_datums)
}
IsUnboxedClosure => {
copy_unboxed_closure_args_to_allocas(
bcx,
arg_scope,
decl.inputs.as_slice(),
arg_datums,
monomorphized_arg_types.as_slice())
}
};
2011-06-28 18:54:05 -07:00
bcx = maybe_load_env(bcx, cleanup::CustomScope(arg_scope));
2011-06-28 18:54:05 -07:00
// Up until here, IR instructions for this function have explicitly not been annotated with
// source code location, so we don't step into call setup code. From here on, source location
// emitting should be enabled.
debuginfo::start_emitting_source_locations(&fcx);
let dest = match fcx.llretslotptr.get() {
Some(_) => expr::SaveIn(fcx.get_ret_slot(bcx, block_ty, "iret_slot")),
None => {
assert!(type_is_zero_size(bcx.ccx(), block_ty));
expr::Ignore
}
};
// This call to trans_block is the place where we bridge between
// translation calls that don't have a return value (trans_crate,
// trans_mod, trans_item, et cetera) and those that do
// (trans_block, trans_expr, et cetera).
bcx = controlflow::trans_block(bcx, body, dest);
match dest {
expr::SaveIn(slot) if fcx.needs_ret_allocas => {
Store(bcx, slot, fcx.llretslotptr.get().unwrap());
}
_ => {}
}
match fcx.llreturn.get() {
Some(_) => {
Br(bcx, fcx.return_exit_block());
fcx.pop_custom_cleanup_scope(arg_scope);
}
None => {
// Microoptimization writ large: avoid creating a separate
// llreturn basic block
bcx = fcx.pop_and_trans_custom_cleanup_scope(bcx, arg_scope);
}
};
2013-06-14 11:59:49 -07:00
// Put return block after all other blocks.
// This somewhat improves single-stepping experience in debugger.
unsafe {
let llreturn = fcx.llreturn.get();
for &llreturn in llreturn.iter() {
llvm::LLVMMoveBasicBlockAfter(llreturn, bcx.llbb);
}
2013-06-14 11:59:49 -07:00
}
2011-06-29 17:29:24 -07:00
// Insert the mandatory first few basic blocks before lltop.
finish_fn(&fcx, bcx, output_type);
2011-06-29 19:50:50 -07:00
}
// trans_fn: creates an LLVM function corresponding to a source language
// function.
2014-03-06 18:47:24 +02:00
pub fn trans_fn(ccx: &CrateContext,
decl: &ast::FnDecl,
body: &ast::Block,
llfndecl: ValueRef,
param_substs: &param_substs,
id: ast::NodeId,
attrs: &[ast::Attribute]) {
let _s = StatRecorder::new(ccx, ccx.tcx.map.path_to_string(id).to_string());
debug!("trans_fn(param_substs={})", param_substs.repr(ccx.tcx()));
let _icx = push_ctxt("trans_fn");
let fn_ty = ty::node_id_to_type(ccx.tcx(), id);
let arg_types = ty::ty_fn_args(fn_ty);
let output_type = ty::ty_fn_ret(fn_ty);
let abi = ty::ty_fn_abi(fn_ty);
trans_closure(ccx,
decl,
body,
llfndecl,
param_substs,
id,
attrs,
arg_types,
output_type,
abi,
false,
NotUnboxedClosure,
|bcx, _| bcx);
}
2014-03-06 18:47:24 +02:00
pub fn trans_enum_variant(ccx: &CrateContext,
_enum_id: ast::NodeId,
variant: &ast::Variant,
_args: &[ast::VariantArg],
disr: ty::Disr,
param_substs: &param_substs,
llfndecl: ValueRef) {
let _icx = push_ctxt("trans_enum_variant");
trans_enum_variant_or_tuple_like_struct(
ccx,
variant.node.id,
disr,
param_substs,
llfndecl);
}
pub fn trans_named_tuple_constructor<'a>(mut bcx: &'a Block<'a>,
ctor_ty: ty::t,
disr: ty::Disr,
args: callee::CallArgs,
dest: expr::Dest) -> Result<'a> {
let ccx = bcx.fcx.ccx;
let tcx = &ccx.tcx;
let result_ty = match ty::get(ctor_ty).sty {
ty::ty_bare_fn(ref bft) => bft.sig.output,
_ => ccx.sess().bug(
format!("trans_enum_variant_constructor: \
unexpected ctor return type {}",
ctor_ty.repr(tcx)).as_slice())
};
// Get location to store the result. If the user does not care about
// the result, just make a stack slot
let llresult = match dest {
expr::SaveIn(d) => d,
expr::Ignore => {
if !type_is_zero_size(ccx, result_ty) {
alloc_ty(bcx, result_ty, "constructor_result")
} else {
C_undef(type_of::type_of(ccx, result_ty))
}
}
};
if !type_is_zero_size(ccx, result_ty) {
match args {
callee::ArgExprs(exprs) => {
let fields = exprs.iter().map(|x| *x).enumerate().collect::<Vec<_>>();
bcx = expr::trans_adt(bcx, result_ty, disr, fields.as_slice(),
None, expr::SaveIn(llresult));
}
_ => ccx.sess().bug("expected expr as arguments for variant/struct tuple constructor")
}
}
// If the caller doesn't care about the result
// drop the temporary we made
let bcx = match dest {
expr::SaveIn(_) => bcx,
expr::Ignore => glue::drop_ty(bcx, llresult, result_ty)
};
Result::new(bcx, llresult)
}
2014-03-06 18:47:24 +02:00
pub fn trans_tuple_struct(ccx: &CrateContext,
_fields: &[ast::StructField],
ctor_id: ast::NodeId,
param_substs: &param_substs,
llfndecl: ValueRef) {
let _icx = push_ctxt("trans_tuple_struct");
trans_enum_variant_or_tuple_like_struct(
ccx,
ctor_id,
0,
param_substs,
llfndecl);
}
2014-03-06 18:47:24 +02:00
fn trans_enum_variant_or_tuple_like_struct(ccx: &CrateContext,
ctor_id: ast::NodeId,
disr: ty::Disr,
param_substs: &param_substs,
llfndecl: ValueRef) {
let ctor_ty = ty::node_id_to_type(ccx.tcx(), ctor_id);
let ctor_ty = ctor_ty.substp(ccx.tcx(), param_substs);
let result_ty = match ty::get(ctor_ty).sty {
ty::ty_bare_fn(ref bft) => bft.sig.output,
2014-03-05 16:36:01 +02:00
_ => ccx.sess().bug(
2013-09-27 22:38:08 -07:00
format!("trans_enum_variant_or_tuple_like_struct: \
unexpected ctor return type {}",
ty_to_string(ccx.tcx(), ctor_ty)).as_slice())
};
let arena = TypedArena::new();
let fcx = new_fn_ctxt(ccx, llfndecl, ctor_id, false, result_ty,
param_substs, None, &arena);
let bcx = init_function(&fcx, false, result_ty);
assert!(!fcx.needs_ret_allocas);
let arg_tys = ty::ty_fn_args(ctor_ty);
let arg_datums = create_datums_for_fn_args(&fcx, arg_tys.as_slice());
if !type_is_zero_size(fcx.ccx, result_ty) {
let dest = fcx.get_ret_slot(bcx, result_ty, "eret_slot");
let repr = adt::represent_type(ccx, result_ty);
for (i, arg_datum) in arg_datums.move_iter().enumerate() {
let lldestptr = adt::trans_field_ptr(bcx,
2014-04-22 03:03:02 +03:00
&*repr,
dest,
disr,
i);
arg_datum.store_to(bcx, lldestptr);
}
adt::trans_set_discr(bcx, &*repr, dest, disr);
}
finish_fn(&fcx, bcx, result_ty);
}
fn enum_variant_size_lint(ccx: &CrateContext, enum_def: &ast::EnumDef, sp: Span, id: ast::NodeId) {
let mut sizes = Vec::new(); // does no allocation if no pushes, thankfully
let levels = ccx.tcx.node_lint_levels.borrow();
let lint_id = lint::LintId::of(lint::builtin::VARIANT_SIZE_DIFFERENCE);
let lvlsrc = match levels.find(&(id, lint_id)) {
None | Some(&(lint::Allow, _)) => return,
Some(&lvlsrc) => lvlsrc,
};
let avar = adt::represent_type(ccx, ty::node_id_to_type(ccx.tcx(), id));
match *avar {
adt::General(_, ref variants, _) => {
for var in variants.iter() {
let mut size = 0;
for field in var.fields.iter().skip(1) {
// skip the discriminant
size += llsize_of_real(ccx, sizing_type_of(ccx, *field));
}
sizes.push(size);
}
},
_ => { /* its size is either constant or unimportant */ }
}
let (largest, slargest, largest_index) = sizes.iter().enumerate().fold((0, 0, 0),
|(l, s, li), (idx, &size)|
if size > l {
(size, l, idx)
} else if size > s {
(l, size, li)
} else {
(l, s, li)
}
);
// we only warn if the largest variant is at least thrice as large as
// the second-largest.
if largest > slargest * 3 && slargest > 0 {
// Use lint::raw_emit_lint rather than sess.add_lint because the lint-printing
// pass for the latter already ran.
lint::raw_emit_lint(&ccx.tcx().sess, lint::builtin::VARIANT_SIZE_DIFFERENCE,
lvlsrc, Some(sp),
format!("enum variant is more than three times larger \
({} bytes) than the next largest (ignoring padding)",
largest).as_slice());
ccx.sess().span_note(enum_def.variants.get(largest_index).span,
"this variant is the largest");
}
}
2014-03-06 05:07:47 +02:00
pub struct TransItemVisitor<'a> {
pub ccx: &'a CrateContext,
2013-09-25 10:58:40 +02:00
}
2014-03-06 05:07:47 +02:00
impl<'a> Visitor<()> for TransItemVisitor<'a> {
fn visit_item(&mut self, i: &ast::Item, _:()) {
2013-09-25 10:58:40 +02:00
trans_item(self.ccx, i);
}
}
2014-03-06 18:47:24 +02:00
pub fn trans_item(ccx: &CrateContext, item: &ast::Item) {
let _icx = push_ctxt("trans_item");
match item.node {
2014-05-16 10:15:33 -07:00
ast::ItemFn(ref decl, _fn_style, abi, ref generics, ref body) => {
if !generics.is_type_parameterized() {
let llfn = get_item_val(ccx, item.id);
if abi != Rust {
foreign::trans_rust_fn_with_foreign_abi(ccx,
&**decl,
&**body,
item.attrs.as_slice(),
llfn,
&param_substs::empty(),
item.id,
None);
} else {
trans_fn(ccx,
&**decl,
&**body,
llfn,
&param_substs::empty(),
item.id,
item.attrs.as_slice());
}
}
// Be sure to travel more than just one layer deep to catch nested
// items in blocks and such.
let mut v = TransItemVisitor{ ccx: ccx };
v.visit_block(&**body, ());
2011-07-27 14:19:39 +02:00
}
ast::ItemImpl(ref generics, _, _, ref impl_items) => {
meth::trans_impl(ccx,
item.ident,
impl_items.as_slice(),
generics,
item.id);
}
ast::ItemMod(ref m) => {
trans_mod(ccx, m);
2011-07-27 14:19:39 +02:00
}
ast::ItemEnum(ref enum_definition, _) => {
enum_variant_size_lint(ccx, enum_definition, item.span, item.id);
2011-07-27 14:19:39 +02:00
}
2014-05-16 10:15:33 -07:00
ast::ItemStatic(_, m, ref expr) => {
// Recurse on the expression to catch items in blocks
let mut v = TransItemVisitor{ ccx: ccx };
2014-05-16 10:15:33 -07:00
v.visit_expr(&**expr, ());
2013-06-21 18:46:34 -07:00
consts::trans_const(ccx, m, item.id);
// Do static_assert checking. It can't really be done much earlier
// because we need to get the value of the bool out of LLVM
if attr::contains_name(item.attrs.as_slice(), "static_assert") {
if m == ast::MutMutable {
2014-03-05 16:36:01 +02:00
ccx.sess().span_fatal(expr.span,
"cannot have static_assert on a mutable \
static");
}
2013-12-18 17:08:56 -08:00
2014-03-20 19:49:20 -07:00
let v = ccx.const_values.borrow().get_copy(&item.id);
unsafe {
if !(llvm::LLVMConstIntGetZExtValue(v) != 0) {
2014-03-05 16:36:01 +02:00
ccx.sess().span_fatal(expr.span, "static assertion failed");
}
}
}
},
ast::ItemForeignMod(ref foreign_mod) => {
foreign::trans_foreign_mod(ccx, foreign_mod);
}
ast::ItemTrait(..) => {
// Inside of this trait definition, we won't be actually translating any
// functions, but the trait still needs to be walked. Otherwise default
// methods with items will not get translated and will cause ICE's when
// metadata time comes around.
2013-09-25 10:58:40 +02:00
let mut v = TransItemVisitor{ ccx: ccx };
visit::walk_item(&mut v, item, ());
}
2012-08-03 19:59:04 -07:00
_ => {/* fall through */ }
}
}
// Translate a module. Doing this amounts to translating the items in the
// module; there ends up being no artifact (aside from linkage names) of
// separate modules in the compiled program. That's because modules exist
// only as a convenience for humans working with the code, to organize names
// and control visibility.
2014-03-06 18:47:24 +02:00
pub fn trans_mod(ccx: &CrateContext, m: &ast::Mod) {
let _icx = push_ctxt("trans_mod");
for item in m.items.iter() {
2014-05-16 10:15:33 -07:00
trans_item(ccx, &**item);
}
}
fn finish_register_fn(ccx: &CrateContext, sp: Span, sym: String, node_id: ast::NodeId,
llfn: ValueRef) {
2014-03-20 19:49:20 -07:00
ccx.item_symbols.borrow_mut().insert(node_id, sym);
2014-03-09 13:42:22 +02:00
if !ccx.reachable.contains(&node_id) {
llvm::SetLinkage(llfn, llvm::InternalLinkage);
}
rustc: Add official support for weak failure This commit is part of the ongoing libstd facade efforts (cc #13851). The compiler now recognizes some language items as "extern { fn foo(...); }" and will automatically perform the following actions: 1. The foreign function has a pre-defined name. 2. The crate and downstream crates can only be built as rlibs until a crate defines the lang item itself. 3. The actual lang item has a pre-defined name. This is essentially nicer compiler support for the hokey core-depends-on-std-failure scheme today, but it is implemented the same way. The details are a little more hidden under the covers. In addition to failure, this commit promotes the eh_personality and rust_stack_exhausted functions to official lang items. The compiler can generate calls to these functions, causing linkage errors if they are left undefined. The checking for these items is not as precise as it could be. Crates compiling with `-Z no-landing-pads` will not need the eh_personality lang item, and crates compiling with no split stacks won't need the stack exhausted lang item. For ease, however, these items are checked for presence in all final outputs of the compiler. It is quite easy to define dummy versions of the functions necessary: #[lang = "stack_exhausted"] extern fn stack_exhausted() { /* ... */ } #[lang = "eh_personality"] extern fn eh_personality() { /* ... */ } cc #11922, rust_stack_exhausted is now a lang item cc #13851, libcollections is blocked on eh_personality becoming weak
2014-05-19 09:30:09 -07:00
// The stack exhaustion lang item shouldn't have a split stack because
// otherwise it would continue to be exhausted (bad), and both it and the
// eh_personality functions need to be externally linkable.
let def = ast_util::local_def(node_id);
if ccx.tcx.lang_items.stack_exhausted() == Some(def) {
unset_split_stack(llfn);
llvm::SetLinkage(llfn, llvm::ExternalLinkage);
rustc: Add official support for weak failure This commit is part of the ongoing libstd facade efforts (cc #13851). The compiler now recognizes some language items as "extern { fn foo(...); }" and will automatically perform the following actions: 1. The foreign function has a pre-defined name. 2. The crate and downstream crates can only be built as rlibs until a crate defines the lang item itself. 3. The actual lang item has a pre-defined name. This is essentially nicer compiler support for the hokey core-depends-on-std-failure scheme today, but it is implemented the same way. The details are a little more hidden under the covers. In addition to failure, this commit promotes the eh_personality and rust_stack_exhausted functions to official lang items. The compiler can generate calls to these functions, causing linkage errors if they are left undefined. The checking for these items is not as precise as it could be. Crates compiling with `-Z no-landing-pads` will not need the eh_personality lang item, and crates compiling with no split stacks won't need the stack exhausted lang item. For ease, however, these items are checked for presence in all final outputs of the compiler. It is quite easy to define dummy versions of the functions necessary: #[lang = "stack_exhausted"] extern fn stack_exhausted() { /* ... */ } #[lang = "eh_personality"] extern fn eh_personality() { /* ... */ } cc #11922, rust_stack_exhausted is now a lang item cc #13851, libcollections is blocked on eh_personality becoming weak
2014-05-19 09:30:09 -07:00
}
if ccx.tcx.lang_items.eh_personality() == Some(def) {
llvm::SetLinkage(llfn, llvm::ExternalLinkage);
rustc: Add official support for weak failure This commit is part of the ongoing libstd facade efforts (cc #13851). The compiler now recognizes some language items as "extern { fn foo(...); }" and will automatically perform the following actions: 1. The foreign function has a pre-defined name. 2. The crate and downstream crates can only be built as rlibs until a crate defines the lang item itself. 3. The actual lang item has a pre-defined name. This is essentially nicer compiler support for the hokey core-depends-on-std-failure scheme today, but it is implemented the same way. The details are a little more hidden under the covers. In addition to failure, this commit promotes the eh_personality and rust_stack_exhausted functions to official lang items. The compiler can generate calls to these functions, causing linkage errors if they are left undefined. The checking for these items is not as precise as it could be. Crates compiling with `-Z no-landing-pads` will not need the eh_personality lang item, and crates compiling with no split stacks won't need the stack exhausted lang item. For ease, however, these items are checked for presence in all final outputs of the compiler. It is quite easy to define dummy versions of the functions necessary: #[lang = "stack_exhausted"] extern fn stack_exhausted() { /* ... */ } #[lang = "eh_personality"] extern fn eh_personality() { /* ... */ } cc #11922, rust_stack_exhausted is now a lang item cc #13851, libcollections is blocked on eh_personality becoming weak
2014-05-19 09:30:09 -07:00
}
if is_entry_fn(ccx.sess(), node_id) {
create_entry_wrapper(ccx, sp, llfn);
}
}
2014-03-06 18:47:24 +02:00
fn register_fn(ccx: &CrateContext,
sp: Span,
sym: String,
node_id: ast::NodeId,
node_type: ty::t)
-> ValueRef {
match ty::get(node_type).sty {
ty::ty_bare_fn(ref f) => {
assert!(f.abi == Rust || f.abi == RustCall);
}
_ => fail!("expected bare rust fn")
};
let llfn = decl_rust_fn(ccx, node_type, sym.as_slice());
finish_register_fn(ccx, sp, sym, node_id, llfn);
llfn
2012-02-13 16:06:56 -08:00
}
pub fn get_fn_llvm_attributes(ccx: &CrateContext, fn_ty: ty::t)
-> llvm::AttrBuilder {
use middle::ty::{BrAnon, ReLateBound};
let (fn_sig, abi, has_env) = match ty::get(fn_ty).sty {
ty::ty_closure(ref f) => (f.sig.clone(), f.abi, true),
ty::ty_bare_fn(ref f) => (f.sig.clone(), f.abi, false),
ty::ty_unboxed_closure(closure_did, _) => {
let unboxed_closures = ccx.tcx.unboxed_closures.borrow();
let ref function_type = unboxed_closures.get(&closure_did)
.closure_type;
(function_type.sig.clone(), RustCall, true)
}
_ => ccx.sess().bug("expected closure or function.")
};
// Since index 0 is the return value of the llvm func, we start
// at either 1 or 2 depending on whether there's an env slot or not
let mut first_arg_offset = if has_env { 2 } else { 1 };
let mut attrs = llvm::AttrBuilder::new();
let ret_ty = fn_sig.output;
// These have an odd calling convention, so we need to manually
// unpack the input ty's
let input_tys = match ty::get(fn_ty).sty {
ty::ty_unboxed_closure(_, _) => {
assert!(abi == RustCall);
match ty::get(fn_sig.inputs[0]).sty {
ty::ty_nil => Vec::new(),
ty::ty_tup(ref inputs) => inputs.clone(),
_ => ccx.sess().bug("expected tuple'd inputs")
}
},
ty::ty_bare_fn(_) if abi == RustCall => {
let inputs = vec![fn_sig.inputs[0]];
match ty::get(fn_sig.inputs[1]).sty {
ty::ty_nil => inputs,
ty::ty_tup(ref t_in) => inputs.append(t_in.as_slice()),
_ => ccx.sess().bug("expected tuple'd inputs")
}
}
_ => fn_sig.inputs.clone()
};
// A function pointer is called without the declaration
// available, so we have to apply any attributes with ABI
// implications directly to the call instruction. Right now,
// the only attribute we need to worry about is `sret`.
if type_of::return_uses_outptr(ccx, ret_ty) {
let llret_sz = llsize_of_real(ccx, type_of::type_of(ccx, ret_ty));
// The outptr can be noalias and nocapture because it's entirely
// invisible to the program. We also know it's nonnull as well
// as how many bytes we can dereference
attrs.arg(1, llvm::StructRetAttribute)
.arg(1, llvm::NoAliasAttribute)
.arg(1, llvm::NoCaptureAttribute)
.arg(1, llvm::DereferenceableAttribute(llret_sz));
// Add one more since there's an outptr
first_arg_offset += 1;
} else {
// The `noalias` attribute on the return value is useful to a
// function ptr caller.
match ty::get(ret_ty).sty {
// `~` pointer return values never alias because ownership
// is transferred
DST coercions and DST structs [breaking-change] 1. The internal layout for traits has changed from (vtable, data) to (data, vtable). If you were relying on this in unsafe transmutes, you might get some very weird and apparently unrelated errors. You should not be doing this! Prefer not to do this at all, but if you must, you should use raw::TraitObject rather than hardcoding rustc's internal representation into your code. 2. The minimal type of reference-to-vec-literals (e.g., `&[1, 2, 3]`) is now a fixed size vec (e.g., `&[int, ..3]`) where it used to be an unsized vec (e.g., `&[int]`). If you want the unszied type, you must explicitly give the type (e.g., `let x: &[_] = &[1, 2, 3]`). Note in particular where multiple blocks must have the same type (e.g., if and else clauses, vec elements), the compiler will not coerce to the unsized type without a hint. E.g., `[&[1], &[1, 2]]` used to be a valid expression of type '[&[int]]'. It no longer type checks since the first element now has type `&[int, ..1]` and the second has type &[int, ..2]` which are incompatible. 3. The type of blocks (including functions) must be coercible to the expected type (used to be a subtype). Mostly this makes things more flexible and not less (in particular, in the case of coercing function bodies to the return type). However, in some rare cases, this is less flexible. TBH, I'm not exactly sure of the exact effects. I think the change causes us to resolve inferred type variables slightly earlier which might make us slightly more restrictive. Possibly it only affects blocks with unreachable code. E.g., `if ... { fail!(); "Hello" }` used to type check, it no longer does. The fix is to add a semicolon after the string.
2014-08-04 14:20:11 +02:00
ty::ty_uniq(it) if !ty::type_is_sized(ccx.tcx(), it) => {}
ty::ty_uniq(_) => {
attrs.ret(llvm::NoAliasAttribute);
}
_ => {}
}
// We can also mark the return value as `dereferenceable` in certain cases
match ty::get(ret_ty).sty {
// These are not really pointers but pairs, (pointer, len)
ty::ty_uniq(it) |
DST coercions and DST structs [breaking-change] 1. The internal layout for traits has changed from (vtable, data) to (data, vtable). If you were relying on this in unsafe transmutes, you might get some very weird and apparently unrelated errors. You should not be doing this! Prefer not to do this at all, but if you must, you should use raw::TraitObject rather than hardcoding rustc's internal representation into your code. 2. The minimal type of reference-to-vec-literals (e.g., `&[1, 2, 3]`) is now a fixed size vec (e.g., `&[int, ..3]`) where it used to be an unsized vec (e.g., `&[int]`). If you want the unszied type, you must explicitly give the type (e.g., `let x: &[_] = &[1, 2, 3]`). Note in particular where multiple blocks must have the same type (e.g., if and else clauses, vec elements), the compiler will not coerce to the unsized type without a hint. E.g., `[&[1], &[1, 2]]` used to be a valid expression of type '[&[int]]'. It no longer type checks since the first element now has type `&[int, ..1]` and the second has type &[int, ..2]` which are incompatible. 3. The type of blocks (including functions) must be coercible to the expected type (used to be a subtype). Mostly this makes things more flexible and not less (in particular, in the case of coercing function bodies to the return type). However, in some rare cases, this is less flexible. TBH, I'm not exactly sure of the exact effects. I think the change causes us to resolve inferred type variables slightly earlier which might make us slightly more restrictive. Possibly it only affects blocks with unreachable code. E.g., `if ... { fail!(); "Hello" }` used to type check, it no longer does. The fix is to add a semicolon after the string.
2014-08-04 14:20:11 +02:00
ty::ty_rptr(_, ty::mt { ty: it, .. }) if !ty::type_is_sized(ccx.tcx(), it) => {}
ty::ty_uniq(inner) | ty::ty_rptr(_, ty::mt { ty: inner, .. }) => {
let llret_sz = llsize_of_real(ccx, type_of::type_of(ccx, inner));
attrs.ret(llvm::DereferenceableAttribute(llret_sz));
}
_ => {}
}
match ty::get(ret_ty).sty {
ty::ty_bool => {
attrs.ret(llvm::ZExtAttribute);
}
_ => {}
}
}
for (idx, &t) in input_tys.iter().enumerate().map(|(i, v)| (i + first_arg_offset, v)) {
match ty::get(t).sty {
// this needs to be first to prevent fat pointers from falling through
_ if !type_is_immediate(ccx, t) => {
let llarg_sz = llsize_of_real(ccx, type_of::type_of(ccx, t));
// For non-immediate arguments the callee gets its own copy of
// the value on the stack, so there are no aliases. It's also
// program-invisible so can't possibly capture
attrs.arg(idx, llvm::NoAliasAttribute)
.arg(idx, llvm::NoCaptureAttribute)
.arg(idx, llvm::DereferenceableAttribute(llarg_sz));
}
ty::ty_bool => {
attrs.arg(idx, llvm::ZExtAttribute);
}
// `~` pointer parameters never alias because ownership is transferred
ty::ty_uniq(inner) => {
let llsz = llsize_of_real(ccx, type_of::type_of(ccx, inner));
attrs.arg(idx, llvm::NoAliasAttribute)
.arg(idx, llvm::DereferenceableAttribute(llsz));
}
// The visit glue deals only with opaque pointers so we don't
// actually know the concrete type of Self thus we don't know how
// many bytes to mark as dereferenceable so instead we just mark
// it as nonnull which still holds true
ty::ty_rptr(b, ty::mt { ty: it, mutbl }) if match ty::get(it).sty {
ty::ty_param(_) => true, _ => false
} && mutbl == ast::MutMutable => {
attrs.arg(idx, llvm::NoAliasAttribute)
.arg(idx, llvm::NonNullAttribute);
match b {
ReLateBound(_, BrAnon(_)) => {
attrs.arg(idx, llvm::NoCaptureAttribute);
}
_ => {}
}
}
// `&mut` pointer parameters never alias other parameters, or mutable global data
//
// `&T` where `T` contains no `UnsafeCell<U>` is immutable, and can be marked as both
// `readonly` and `noalias`, as LLVM's definition of `noalias` is based solely on
// memory dependencies rather than pointer equality
ty::ty_rptr(b, mt) if mt.mutbl == ast::MutMutable ||
!ty::type_contents(ccx.tcx(), mt.ty).interior_unsafe() => {
let llsz = llsize_of_real(ccx, type_of::type_of(ccx, mt.ty));
attrs.arg(idx, llvm::NoAliasAttribute)
.arg(idx, llvm::DereferenceableAttribute(llsz));
if mt.mutbl == ast::MutImmutable {
attrs.arg(idx, llvm::ReadOnlyAttribute);
}
match b {
ReLateBound(_, BrAnon(_)) => {
attrs.arg(idx, llvm::NoCaptureAttribute);
}
_ => {}
}
}
// When a reference in an argument has no named lifetime, it's impossible for that
// reference to escape this function (returned or stored beyond the call by a closure).
ty::ty_rptr(ReLateBound(_, BrAnon(_)), mt) => {
let llsz = llsize_of_real(ccx, type_of::type_of(ccx, mt.ty));
attrs.arg(idx, llvm::NoCaptureAttribute)
.arg(idx, llvm::DereferenceableAttribute(llsz));
}
// & pointer parameters are also never null and we know exactly how
// many bytes we can dereference
ty::ty_rptr(_, mt) => {
let llsz = llsize_of_real(ccx, type_of::type_of(ccx, mt.ty));
attrs.arg(idx, llvm::DereferenceableAttribute(llsz));
}
_ => ()
}
}
attrs
}
// only use this for foreign function ABIs and glue, use `register_fn` for Rust functions
2014-03-06 18:47:24 +02:00
pub fn register_fn_llvmty(ccx: &CrateContext,
sp: Span,
sym: String,
node_id: ast::NodeId,
cc: llvm::CallConv,
llfty: Type) -> ValueRef {
debug!("register_fn_llvmty id={} sym={}", node_id, sym);
let llfn = decl_fn(ccx, sym.as_slice(), cc, llfty, ty::mk_nil());
finish_register_fn(ccx, sp, sym, node_id, llfn);
llfn
}
pub fn is_entry_fn(sess: &Session, node_id: ast::NodeId) -> bool {
match *sess.entry_fn.borrow() {
Some((entry_id, _)) => node_id == entry_id,
None => false
}
}
// Create a _rust_main(args: ~[str]) function which will be called from the
// runtime rust_start function
2014-03-06 18:47:24 +02:00
pub fn create_entry_wrapper(ccx: &CrateContext,
_sp: Span,
main_llfn: ValueRef) {
2014-03-05 16:36:01 +02:00
let et = ccx.sess().entry_type.get().unwrap();
match et {
config::EntryMain => {
create_entry_fn(ccx, main_llfn, true);
}
config::EntryStart => create_entry_fn(ccx, main_llfn, false),
config::EntryNone => {} // Do nothing.
}
2014-03-06 18:47:24 +02:00
fn create_entry_fn(ccx: &CrateContext,
rust_main: ValueRef,
use_start_lang_item: bool) {
let llfty = Type::func([ccx.int_type, Type::i8p(ccx).ptr_to()],
&ccx.int_type);
let llfn = decl_cdecl_fn(ccx, "main", llfty, ty::mk_nil());
// FIXME: #16581: Marking a symbol in the executable with `dllexport`
// linkage forces MinGW's linker to output a `.reloc` section for ASLR
if ccx.sess().targ_cfg.os == OsWindows {
unsafe { llvm::LLVMRustSetDLLExportStorageClass(llfn) }
}
let llbb = "top".with_c_str(|buf| {
unsafe {
llvm::LLVMAppendBasicBlockInContext(ccx.llcx, llfn, buf)
}
});
let bld = ccx.builder.b;
unsafe {
llvm::LLVMPositionBuilderAtEnd(bld, llbb);
let (start_fn, args) = if use_start_lang_item {
let start_def_id = match ccx.tcx.lang_items.require(StartFnLangItem) {
Ok(id) => id,
Err(s) => { ccx.sess().fatal(s.as_slice()); }
};
let start_fn = if start_def_id.krate == ast::LOCAL_CRATE {
get_item_val(ccx, start_def_id.node)
} else {
let start_fn_type = csearch::get_type(ccx.tcx(),
start_def_id).ty;
trans_external_path(ccx, start_def_id, start_fn_type)
};
let args = {
let opaque_rust_main = "rust_main".with_c_str(|buf| {
llvm::LLVMBuildPointerCast(bld, rust_main, Type::i8p(ccx).to_ref(), buf)
});
vec!(
opaque_rust_main,
get_param(llfn, 0),
get_param(llfn, 1)
)
};
(start_fn, args)
} else {
debug!("using user-defined start fn");
let args = vec!(
get_param(llfn, 0 as c_uint),
get_param(llfn, 1 as c_uint)
);
(rust_main, args)
};
let result = llvm::LLVMBuildCall(bld,
start_fn,
args.as_ptr(),
args.len() as c_uint,
noname());
llvm::LLVMBuildRet(bld, result);
}
}
}
fn exported_name(ccx: &CrateContext, id: ast::NodeId,
ty: ty::t, attrs: &[ast::Attribute]) -> String {
match attr::first_attr_value_str_by_name(attrs, "export_name") {
// Use provided name
Some(name) => name.get().to_string(),
_ => ccx.tcx.map.with_path(id, |mut path| {
if attr::contains_name(attrs, "no_mangle") {
// Don't mangle
path.last().unwrap().to_string()
} else {
rustc: Add official support for weak failure This commit is part of the ongoing libstd facade efforts (cc #13851). The compiler now recognizes some language items as "extern { fn foo(...); }" and will automatically perform the following actions: 1. The foreign function has a pre-defined name. 2. The crate and downstream crates can only be built as rlibs until a crate defines the lang item itself. 3. The actual lang item has a pre-defined name. This is essentially nicer compiler support for the hokey core-depends-on-std-failure scheme today, but it is implemented the same way. The details are a little more hidden under the covers. In addition to failure, this commit promotes the eh_personality and rust_stack_exhausted functions to official lang items. The compiler can generate calls to these functions, causing linkage errors if they are left undefined. The checking for these items is not as precise as it could be. Crates compiling with `-Z no-landing-pads` will not need the eh_personality lang item, and crates compiling with no split stacks won't need the stack exhausted lang item. For ease, however, these items are checked for presence in all final outputs of the compiler. It is quite easy to define dummy versions of the functions necessary: #[lang = "stack_exhausted"] extern fn stack_exhausted() { /* ... */ } #[lang = "eh_personality"] extern fn eh_personality() { /* ... */ } cc #11922, rust_stack_exhausted is now a lang item cc #13851, libcollections is blocked on eh_personality becoming weak
2014-05-19 09:30:09 -07:00
match weak_lang_items::link_name(attrs) {
Some(name) => name.get().to_string(),
rustc: Add official support for weak failure This commit is part of the ongoing libstd facade efforts (cc #13851). The compiler now recognizes some language items as "extern { fn foo(...); }" and will automatically perform the following actions: 1. The foreign function has a pre-defined name. 2. The crate and downstream crates can only be built as rlibs until a crate defines the lang item itself. 3. The actual lang item has a pre-defined name. This is essentially nicer compiler support for the hokey core-depends-on-std-failure scheme today, but it is implemented the same way. The details are a little more hidden under the covers. In addition to failure, this commit promotes the eh_personality and rust_stack_exhausted functions to official lang items. The compiler can generate calls to these functions, causing linkage errors if they are left undefined. The checking for these items is not as precise as it could be. Crates compiling with `-Z no-landing-pads` will not need the eh_personality lang item, and crates compiling with no split stacks won't need the stack exhausted lang item. For ease, however, these items are checked for presence in all final outputs of the compiler. It is quite easy to define dummy versions of the functions necessary: #[lang = "stack_exhausted"] extern fn stack_exhausted() { /* ... */ } #[lang = "eh_personality"] extern fn eh_personality() { /* ... */ } cc #11922, rust_stack_exhausted is now a lang item cc #13851, libcollections is blocked on eh_personality becoming weak
2014-05-19 09:30:09 -07:00
None => {
// Usual name mangling
mangle_exported_name(ccx, path, ty, id)
}
}
}
})
}
}
2014-03-06 18:47:24 +02:00
pub fn get_item_val(ccx: &CrateContext, id: ast::NodeId) -> ValueRef {
debug!("get_item_val(id=`{:?}`)", id);
2014-03-20 19:49:20 -07:00
match ccx.item_vals.borrow().find_copy(&id) {
Some(v) => return v,
None => {}
}
2014-03-20 19:49:20 -07:00
let mut foreign = false;
let item = ccx.tcx.map.get(id);
let val = match item {
ast_map::NodeItem(i) => {
let ty = ty::node_id_to_type(ccx.tcx(), i.id);
let sym = exported_name(ccx, id, ty, i.attrs.as_slice());
let v = match i.node {
ast::ItemStatic(_, mutbl, ref expr) => {
2014-03-20 19:49:20 -07:00
// If this static came from an external crate, then
// we need to get the symbol from csearch instead of
// using the current crate's name/version
// information in the hash of the symbol
debug!("making {}", sym);
let (sym, is_local) = {
match ccx.external_srcs.borrow().find(&i.id) {
Some(&did) => {
debug!("but found in other crate...");
(csearch::get_symbol(&ccx.sess().cstore,
did), false)
}
2014-03-20 19:49:20 -07:00
None => (sym, true)
}
2014-03-20 19:49:20 -07:00
};
2014-03-20 19:49:20 -07:00
// We need the translated value here, because for enums the
// LLVM type is not fully determined by the Rust type.
DST coercions and DST structs [breaking-change] 1. The internal layout for traits has changed from (vtable, data) to (data, vtable). If you were relying on this in unsafe transmutes, you might get some very weird and apparently unrelated errors. You should not be doing this! Prefer not to do this at all, but if you must, you should use raw::TraitObject rather than hardcoding rustc's internal representation into your code. 2. The minimal type of reference-to-vec-literals (e.g., `&[1, 2, 3]`) is now a fixed size vec (e.g., `&[int, ..3]`) where it used to be an unsized vec (e.g., `&[int]`). If you want the unszied type, you must explicitly give the type (e.g., `let x: &[_] = &[1, 2, 3]`). Note in particular where multiple blocks must have the same type (e.g., if and else clauses, vec elements), the compiler will not coerce to the unsized type without a hint. E.g., `[&[1], &[1, 2]]` used to be a valid expression of type '[&[int]]'. It no longer type checks since the first element now has type `&[int, ..1]` and the second has type &[int, ..2]` which are incompatible. 3. The type of blocks (including functions) must be coercible to the expected type (used to be a subtype). Mostly this makes things more flexible and not less (in particular, in the case of coercing function bodies to the return type). However, in some rare cases, this is less flexible. TBH, I'm not exactly sure of the exact effects. I think the change causes us to resolve inferred type variables slightly earlier which might make us slightly more restrictive. Possibly it only affects blocks with unreachable code. E.g., `if ... { fail!(); "Hello" }` used to type check, it no longer does. The fix is to add a semicolon after the string.
2014-08-04 14:20:11 +02:00
let (v, inlineable, _) = consts::const_expr(ccx, &**expr, is_local);
2014-03-20 19:49:20 -07:00
ccx.const_values.borrow_mut().insert(id, v);
let mut inlineable = inlineable;
2014-03-20 19:49:20 -07:00
unsafe {
let llty = llvm::LLVMTypeOf(v);
let g = sym.as_slice().with_c_str(|buf| {
2014-03-20 19:49:20 -07:00
llvm::LLVMAddGlobal(ccx.llmod, llty, buf)
});
2014-03-20 19:49:20 -07:00
if !ccx.reachable.contains(&id) {
llvm::SetLinkage(g, llvm::InternalLinkage);
2014-03-20 19:49:20 -07:00
}
2014-03-20 19:49:20 -07:00
// Apply the `unnamed_addr` attribute if
// requested
if !ast_util::static_has_significant_address(
mutbl,
i.attrs.as_slice()) {
llvm::SetUnnamedAddr(g, true);
2014-03-20 19:49:20 -07:00
// This is a curious case where we must make
// all of these statics inlineable. If a
// global is not tagged as `#[inline(never)]`,
// then LLVM won't coalesce globals unless they
// have an internal linkage type. This means that
2014-03-20 19:49:20 -07:00
// external crates cannot use this global.
// This is a problem for things like inner
// statics in generic functions, because the
// function will be inlined into another
// crate and then attempt to link to the
// static in the original crate, only to
// find that it's not there. On the other
// side of inlining, the crates knows to
2014-03-20 19:49:20 -07:00
// not declare this static as
// available_externally (because it isn't)
inlineable = true;
}
2014-03-20 19:49:20 -07:00
if attr::contains_name(i.attrs.as_slice(),
"thread_local") {
llvm::set_thread_local(g, true);
}
2014-03-20 19:49:20 -07:00
if !inlineable {
debug!("{} not inlined", sym);
ccx.non_inlineable_statics.borrow_mut()
.insert(id);
}
2014-03-20 19:49:20 -07:00
ccx.item_symbols.borrow_mut().insert(i.id, sym);
g
}
}
ast::ItemFn(_, _, abi, _, _) => {
let llfn = if abi == Rust {
2014-03-20 19:49:20 -07:00
register_fn(ccx, i.span, sym, i.id, ty)
} else {
foreign::register_rust_fn_with_foreign_abi(ccx,
i.span,
sym,
i.id)
};
set_llvm_fn_attrs(i.attrs.as_slice(), llfn);
llfn
}
2014-03-20 19:49:20 -07:00
_ => fail!("get_item_val: weird result in table")
};
2014-03-20 19:49:20 -07:00
match attr::first_attr_value_str_by_name(i.attrs.as_slice(),
"link_section") {
Some(sect) => unsafe {
sect.get().with_c_str(|buf| {
llvm::LLVMSetSection(v, buf);
})
},
None => ()
}
2014-03-20 19:49:20 -07:00
v
}
ast_map::NodeTraitItem(trait_method) => {
debug!("get_item_val(): processing a NodeTraitItem");
2014-03-20 19:49:20 -07:00
match *trait_method {
ast::RequiredMethod(_) => {
2014-03-20 19:49:20 -07:00
ccx.sess().bug("unexpected variant: required trait method in \
get_item_val()");
}
ast::ProvidedMethod(m) => {
2014-05-16 10:15:33 -07:00
register_method(ccx, id, &*m)
}
2014-03-20 19:49:20 -07:00
}
}
ast_map::NodeImplItem(ii) => {
match *ii {
ast::MethodImplItem(m) => register_method(ccx, id, &*m),
}
2014-03-20 19:49:20 -07:00
}
ast_map::NodeForeignItem(ni) => {
foreign = true;
match ni.node {
ast::ForeignItemFn(..) => {
let abi = ccx.tcx.map.get_foreign_abi(id);
let ty = ty::node_id_to_type(ccx.tcx(), ni.id);
2014-05-16 10:15:33 -07:00
let name = foreign::link_name(&*ni);
foreign::register_foreign_item_fn(ccx, abi, ty,
name.get().as_slice(),
Some(ni.span))
2014-03-20 19:49:20 -07:00
}
ast::ForeignItemStatic(..) => {
2014-05-16 10:15:33 -07:00
foreign::register_static(ccx, &*ni)
}
2014-03-20 19:49:20 -07:00
}
}
2014-03-20 19:49:20 -07:00
ast_map::NodeVariant(ref v) => {
let llfn;
let args = match v.node.kind {
ast::TupleVariantKind(ref args) => args,
ast::StructVariantKind(_) => {
fail!("struct variant kind unexpected in get_item_val")
}
};
assert!(args.len() != 0u);
let ty = ty::node_id_to_type(ccx.tcx(), id);
let parent = ccx.tcx.map.get_parent(id);
let enm = ccx.tcx.map.expect_item(parent);
let sym = exported_name(ccx,
id,
ty,
enm.attrs.as_slice());
llfn = match enm.node {
ast::ItemEnum(_, _) => {
register_fn(ccx, (*v).span, sym, id, ty)
}
2014-03-20 19:49:20 -07:00
_ => fail!("NodeVariant, shouldn't happen")
};
2014-03-20 19:49:20 -07:00
set_inline_hint(llfn);
llfn
}
2014-03-20 19:49:20 -07:00
ast_map::NodeStructCtor(struct_def) => {
// Only register the constructor if this is a tuple-like struct.
let ctor_id = match struct_def.ctor_id {
None => {
ccx.sess().bug("attempt to register a constructor of \
a non-tuple-like struct")
}
Some(ctor_id) => ctor_id,
};
let parent = ccx.tcx.map.get_parent(id);
let struct_item = ccx.tcx.map.expect_item(parent);
let ty = ty::node_id_to_type(ccx.tcx(), ctor_id);
let sym = exported_name(ccx,
id,
ty,
struct_item.attrs
.as_slice());
let llfn = register_fn(ccx, struct_item.span,
sym, ctor_id, ty);
set_inline_hint(llfn);
llfn
}
2014-03-20 19:49:20 -07:00
ref variant => {
ccx.sess().bug(format!("get_item_val(): unexpected variant: {:?}",
variant).as_slice())
}
2014-03-20 19:49:20 -07:00
};
// foreign items (extern fns and extern statics) don't have internal
// linkage b/c that doesn't quite make sense. Otherwise items can
// have internal linkage if they're not reachable.
if !foreign && !ccx.reachable.contains(&id) {
llvm::SetLinkage(val, llvm::InternalLinkage);
}
2014-03-20 19:49:20 -07:00
ccx.item_vals.borrow_mut().insert(id, val);
val
}
2014-03-06 18:47:24 +02:00
fn register_method(ccx: &CrateContext, id: ast::NodeId,
m: &ast::Method) -> ValueRef {
let mty = ty::node_id_to_type(ccx.tcx(), id);
2013-06-21 20:28:33 +12:00
let sym = exported_name(ccx, id, mty, m.attrs.as_slice());
let llfn = register_fn(ccx, m.span, sym, id, mty);
set_llvm_fn_attrs(m.attrs.as_slice(), llfn);
llfn
}
pub fn p2i(ccx: &CrateContext, v: ValueRef) -> ValueRef {
unsafe {
2013-06-16 22:52:44 +12:00
return llvm::LLVMConstPtrToInt(v, ccx.int_type.to_ref());
}
}
pub fn crate_ctxt_to_encode_parms<'r>(cx: &'r CrateContext, ie: encoder::EncodeInlinedItem<'r>)
-> encoder::EncodeParams<'r> {
encoder::EncodeParams {
2014-04-22 19:06:43 +03:00
diag: cx.sess().diagnostic(),
tcx: cx.tcx(),
2014-04-22 19:06:43 +03:00
reexports2: &cx.exp_map2,
item_symbols: &cx.item_symbols,
non_inlineable_statics: &cx.non_inlineable_statics,
2014-04-22 19:06:43 +03:00
link_meta: &cx.link_meta,
2014-03-09 15:20:44 +02:00
cstore: &cx.sess().cstore,
encode_inlined_item: ie,
reachable: &cx.reachable,
}
}
pub fn write_metadata(cx: &CrateContext, krate: &ast::Crate) -> Vec<u8> {
use flate;
Store metadata separately in rlib files Right now whenever an rlib file is linked against, all of the metadata from the rlib is pulled in to the final staticlib or binary. The reason for this is that the metadata is currently stored in a section of the object file. Note that this is intentional for dynamic libraries in order to distribute metadata bundled with static libraries. This commit alters the situation for rlib libraries to instead store the metadata in a separate file in the archive. In doing so, when the archive is passed to the linker, none of the metadata will get pulled into the result executable. Furthermore, the metadata file is skipped when assembling rlibs into an archive. The snag in this implementation comes with multiple output formats. When generating a dylib, the metadata needs to be in the object file, but when generating an rlib this needs to be separate. In order to accomplish this, the metadata variable is inserted into an entirely separate LLVM Module which is then codegen'd into a different location (foo.metadata.o). This is then linked into dynamic libraries and silently ignored for rlib files. While changing how metadata is inserted into archives, I have also stopped compressing metadata when inserted into rlib files. We have wanted to stop compressing metadata, but the sections it creates in object file sections are apparently too large. Thankfully if it's just an arbitrary file it doesn't matter how large it is. I have seen massive reductions in executable sizes, as well as staticlib output sizes (to confirm that this is all working).
2013-12-03 17:41:01 -08:00
let any_library = cx.sess().crate_types.borrow().iter().any(|ty| {
*ty != config::CrateTypeExecutable
});
if !any_library {
return Vec::new()
}
let encode_inlined_item: encoder::EncodeInlinedItem =
|ecx, rbml_w, ii| astencode::encode_inlined_item(ecx, rbml_w, ii);
let encode_parms = crate_ctxt_to_encode_parms(cx, encode_inlined_item);
let metadata = encoder::encode_metadata(encode_parms, krate);
let compressed = Vec::from_slice(encoder::metadata_encoding_version)
.append(match flate::deflate_bytes(metadata.as_slice()) {
Some(compressed) => compressed,
None => {
cx.sess().fatal("failed to compress metadata")
}
}.as_slice());
2014-05-03 23:43:38 -07:00
let llmeta = C_bytes(cx, compressed.as_slice());
let llconst = C_struct(cx, [llmeta], false);
let name = format!("rust_metadata_{}_{}",
cx.link_meta.crate_name,
cx.link_meta.crate_hash);
Implement LTO This commit implements LTO for rust leveraging LLVM's passes. What this means is: * When compiling an rlib, in addition to insdering foo.o into the archive, also insert foo.bc (the LLVM bytecode) of the optimized module. * When the compiler detects the -Z lto option, it will attempt to perform LTO on a staticlib or binary output. The compiler will emit an error if a dylib or rlib output is being generated. * The actual act of performing LTO is as follows: 1. Force all upstream libraries to have an rlib version available. 2. Load the bytecode of each upstream library from the rlib. 3. Link all this bytecode into the current LLVM module (just using llvm apis) 4. Run an internalization pass which internalizes all symbols except those found reachable for the local crate of compilation. 5. Run the LLVM LTO pass manager over this entire module 6a. If assembling an archive, then add all upstream rlibs into the output archive. This ignores all of the object/bitcode/metadata files rust generated and placed inside the rlibs. 6b. If linking a binary, create copies of all upstream rlibs, remove the rust-generated object-file, and then link everything as usual. As I have explained in #10741, this process is excruciatingly slow, so this is *not* turned on by default, and it is also why I have decided to hide it behind a -Z flag for now. The good news is that the binary sizes are about as small as they can be as a result of LTO, so it's definitely working. Closes #10741 Closes #10740
2013-12-02 23:19:29 -08:00
let llglobal = name.with_c_str(|buf| {
unsafe {
Store metadata separately in rlib files Right now whenever an rlib file is linked against, all of the metadata from the rlib is pulled in to the final staticlib or binary. The reason for this is that the metadata is currently stored in a section of the object file. Note that this is intentional for dynamic libraries in order to distribute metadata bundled with static libraries. This commit alters the situation for rlib libraries to instead store the metadata in a separate file in the archive. In doing so, when the archive is passed to the linker, none of the metadata will get pulled into the result executable. Furthermore, the metadata file is skipped when assembling rlibs into an archive. The snag in this implementation comes with multiple output formats. When generating a dylib, the metadata needs to be in the object file, but when generating an rlib this needs to be separate. In order to accomplish this, the metadata variable is inserted into an entirely separate LLVM Module which is then codegen'd into a different location (foo.metadata.o). This is then linked into dynamic libraries and silently ignored for rlib files. While changing how metadata is inserted into archives, I have also stopped compressing metadata when inserted into rlib files. We have wanted to stop compressing metadata, but the sections it creates in object file sections are apparently too large. Thankfully if it's just an arbitrary file it doesn't matter how large it is. I have seen massive reductions in executable sizes, as well as staticlib output sizes (to confirm that this is all working).
2013-12-03 17:41:01 -08:00
llvm::LLVMAddGlobal(cx.metadata_llmod, val_ty(llconst).to_ref(), buf)
}
});
unsafe {
llvm::LLVMSetInitializer(llglobal, llconst);
let name = loader::meta_section_name(cx.sess().targ_cfg.os);
name.unwrap_or("rust_metadata").with_c_str(|buf| {
llvm::LLVMSetSection(llglobal, buf)
});
}
Store metadata separately in rlib files Right now whenever an rlib file is linked against, all of the metadata from the rlib is pulled in to the final staticlib or binary. The reason for this is that the metadata is currently stored in a section of the object file. Note that this is intentional for dynamic libraries in order to distribute metadata bundled with static libraries. This commit alters the situation for rlib libraries to instead store the metadata in a separate file in the archive. In doing so, when the archive is passed to the linker, none of the metadata will get pulled into the result executable. Furthermore, the metadata file is skipped when assembling rlibs into an archive. The snag in this implementation comes with multiple output formats. When generating a dylib, the metadata needs to be in the object file, but when generating an rlib this needs to be separate. In order to accomplish this, the metadata variable is inserted into an entirely separate LLVM Module which is then codegen'd into a different location (foo.metadata.o). This is then linked into dynamic libraries and silently ignored for rlib files. While changing how metadata is inserted into archives, I have also stopped compressing metadata when inserted into rlib files. We have wanted to stop compressing metadata, but the sections it creates in object file sections are apparently too large. Thankfully if it's just an arbitrary file it doesn't matter how large it is. I have seen massive reductions in executable sizes, as well as staticlib output sizes (to confirm that this is all working).
2013-12-03 17:41:01 -08:00
return metadata;
}
2014-03-05 16:36:01 +02:00
pub fn trans_crate(krate: ast::Crate,
analysis: CrateAnalysis) -> (ty::ctxt, CrateTranslation) {
let CrateAnalysis { ty_cx: tcx, exp_map2, reachable, name, .. } = analysis;
2013-06-25 23:46:26 -07:00
// Before we touch LLVM, make sure that multithreading is enabled.
unsafe {
use std::sync::{Once, ONCE_INIT};
static mut INIT: Once = ONCE_INIT;
static mut POISONED: bool = false;
INIT.doit(|| {
if llvm::LLVMStartMultithreaded() != 1 {
// use an extra bool to make sure that all future usage of LLVM
// cannot proceed despite the Once not running more than once.
POISONED = true;
}
});
if POISONED {
tcx.sess.bug("couldn't enable multi-threaded LLVM");
}
2013-06-25 23:46:26 -07:00
}
let link_meta = link::build_link_meta(&tcx.sess, &krate, name);
// Append ".rs" to crate name as LLVM module identifier.
//
// LLVM code generator emits a ".file filename" directive
// for ELF backends. Value of the "filename" is set as the
// LLVM module identifier. Due to a LLVM MC bug[1], LLVM
// crashes if the module identifier is same as other symbols
2011-12-05 11:21:44 -08:00
// such as a function name in the module.
// 1. http://llvm.org/bugs/show_bug.cgi?id=11479
let mut llmod_id = link_meta.crate_name.clone();
llmod_id.push_str(".rs");
let ccx = CrateContext::new(llmod_id.as_slice(), tcx, exp_map2,
Sha256::new(), link_meta, reachable);
// First, verify intrinsics.
intrinsic::check_intrinsics(&ccx);
// Next, translate the module.
{
let _icx = push_ctxt("text");
trans_mod(&ccx, &krate.module);
}
glue::emit_tydescs(&ccx);
2014-03-05 16:36:01 +02:00
if ccx.sess().opts.debuginfo != NoDebugInfo {
debuginfo::finalize(&ccx);
2013-06-17 08:42:05 -07:00
}
2013-06-14 11:59:49 -07:00
// Translate the metadata.
let metadata = write_metadata(&ccx, &krate);
2014-03-05 16:36:01 +02:00
if ccx.sess().trans_stats() {
println!("--- trans stats ---");
println!("n_static_tydescs: {}", ccx.stats.n_static_tydescs.get());
println!("n_glues_created: {}", ccx.stats.n_glues_created.get());
println!("n_null_glues: {}", ccx.stats.n_null_glues.get());
2013-12-22 13:41:16 -08:00
println!("n_real_glues: {}", ccx.stats.n_real_glues.get());
println!("n_fns: {}", ccx.stats.n_fns.get());
println!("n_monos: {}", ccx.stats.n_monos.get());
println!("n_inlines: {}", ccx.stats.n_inlines.get());
println!("n_closures: {}", ccx.stats.n_closures.get());
println!("fn stats:");
2014-03-20 19:49:20 -07:00
ccx.stats.fn_stats.borrow_mut().sort_by(|&(_, _, insns_a), &(_, _, insns_b)| {
insns_b.cmp(&insns_a)
});
2014-03-20 19:49:20 -07:00
for tuple in ccx.stats.fn_stats.borrow().iter() {
match *tuple {
(ref name, ms, insns) => {
println!("{} insns, {} ms, {}", insns, ms, *name);
2013-07-08 11:05:52 -04:00
}
}
}
}
2014-03-05 16:36:01 +02:00
if ccx.sess().count_llvm_insns() {
2014-03-20 19:49:20 -07:00
for (k, v) in ccx.stats.llvm_insns.borrow().iter() {
println!("{:7u} {}", *v, *k);
}
}
2013-06-14 17:38:17 +12:00
let llcx = ccx.llcx;
let link_meta = ccx.link_meta.clone();
2013-06-14 17:38:17 +12:00
let llmod = ccx.llmod;
let mut reachable: Vec<String> = ccx.reachable.iter().filter_map(|id| {
ccx.item_symbols.borrow().find(id).map(|s| s.to_string())
2014-03-09 13:42:22 +02:00
}).collect();
Implement LTO This commit implements LTO for rust leveraging LLVM's passes. What this means is: * When compiling an rlib, in addition to insdering foo.o into the archive, also insert foo.bc (the LLVM bytecode) of the optimized module. * When the compiler detects the -Z lto option, it will attempt to perform LTO on a staticlib or binary output. The compiler will emit an error if a dylib or rlib output is being generated. * The actual act of performing LTO is as follows: 1. Force all upstream libraries to have an rlib version available. 2. Load the bytecode of each upstream library from the rlib. 3. Link all this bytecode into the current LLVM module (just using llvm apis) 4. Run an internalization pass which internalizes all symbols except those found reachable for the local crate of compilation. 5. Run the LLVM LTO pass manager over this entire module 6a. If assembling an archive, then add all upstream rlibs into the output archive. This ignores all of the object/bitcode/metadata files rust generated and placed inside the rlibs. 6b. If linking a binary, create copies of all upstream rlibs, remove the rust-generated object-file, and then link everything as usual. As I have explained in #10741, this process is excruciatingly slow, so this is *not* turned on by default, and it is also why I have decided to hide it behind a -Z flag for now. The good news is that the binary sizes are about as small as they can be as a result of LTO, so it's definitely working. Closes #10741 Closes #10740
2013-12-02 23:19:29 -08:00
// For the purposes of LTO, we add to the reachable set all of the upstream
// reachable extern fns. These functions are all part of the public ABI of
// the final product, so LTO needs to preserve them.
ccx.sess().cstore.iter_crate_data(|cnum, _| {
let syms = csearch::get_reachable_extern_fns(&ccx.sess().cstore, cnum);
reachable.extend(syms.move_iter().map(|did| {
csearch::get_symbol(&ccx.sess().cstore, did)
}));
});
Implement LTO This commit implements LTO for rust leveraging LLVM's passes. What this means is: * When compiling an rlib, in addition to insdering foo.o into the archive, also insert foo.bc (the LLVM bytecode) of the optimized module. * When the compiler detects the -Z lto option, it will attempt to perform LTO on a staticlib or binary output. The compiler will emit an error if a dylib or rlib output is being generated. * The actual act of performing LTO is as follows: 1. Force all upstream libraries to have an rlib version available. 2. Load the bytecode of each upstream library from the rlib. 3. Link all this bytecode into the current LLVM module (just using llvm apis) 4. Run an internalization pass which internalizes all symbols except those found reachable for the local crate of compilation. 5. Run the LLVM LTO pass manager over this entire module 6a. If assembling an archive, then add all upstream rlibs into the output archive. This ignores all of the object/bitcode/metadata files rust generated and placed inside the rlibs. 6b. If linking a binary, create copies of all upstream rlibs, remove the rust-generated object-file, and then link everything as usual. As I have explained in #10741, this process is excruciatingly slow, so this is *not* turned on by default, and it is also why I have decided to hide it behind a -Z flag for now. The good news is that the binary sizes are about as small as they can be as a result of LTO, so it's definitely working. Closes #10741 Closes #10740
2013-12-02 23:19:29 -08:00
// Make sure that some other crucial symbols are not eliminated from the
// module. This includes the main function, the crate map (used for debug
// log settings and I/O), and finally the curious rust_stack_exhausted
// symbol. This symbol is required for use by the libmorestack library that
// we link in, so we must ensure that this symbol is not internalized (if
// defined in the crate).
reachable.push("main".to_string());
reachable.push("rust_stack_exhausted".to_string());
2014-04-15 18:17:48 -07:00
// referenced from .eh_frame section on some platforms
reachable.push("rust_eh_personality".to_string());
// referenced from rt/rust_try.ll
reachable.push("rust_eh_personality_catch".to_string());
let metadata_module = ccx.metadata_llmod;
let formats = ccx.tcx.dependency_formats.borrow().clone();
let no_builtins = attr::contains_name(krate.attrs.as_slice(), "no_builtins");
(ccx.tcx, CrateTranslation {
context: llcx,
module: llmod,
Add generation of static libraries to rustc This commit implements the support necessary for generating both intermediate and result static rust libraries. This is an implementation of my thoughts in https://mail.mozilla.org/pipermail/rust-dev/2013-November/006686.html. When compiling a library, we still retain the "lib" option, although now there are "rlib", "staticlib", and "dylib" as options for crate_type (and these are stackable). The idea of "lib" is to generate the "compiler default" instead of having too choose (although all are interchangeable). For now I have left the "complier default" to be a dynamic library for size reasons. Of the rust libraries, lib{std,extra,rustuv} will bootstrap with an rlib/dylib pair, but lib{rustc,syntax,rustdoc,rustpkg} will only be built as a dynamic object. I chose this for size reasons, but also because you're probably not going to be embedding the rustc compiler anywhere any time soon. Other than the options outlined above, there are a few defaults/preferences that are now opinionated in the compiler: * If both a .dylib and .rlib are found for a rust library, the compiler will prefer the .rlib variant. This is overridable via the -Z prefer-dynamic option * If generating a "lib", the compiler will generate a dynamic library. This is overridable by explicitly saying what flavor you'd like (rlib, staticlib, dylib). * If no options are passed to the command line, and no crate_type is found in the destination crate, then an executable is generated With this change, you can successfully build a rust program with 0 dynamic dependencies on rust libraries. There is still a dynamic dependency on librustrt, but I plan on removing that in a subsequent commit. This change includes no tests just yet. Our current testing infrastructure/harnesses aren't very amenable to doing flavorful things with linking, so I'm planning on adding a new mode of testing which I believe belongs as a separate commit. Closes #552
2013-11-15 14:03:29 -08:00
link: link_meta,
metadata_module: metadata_module,
Store metadata separately in rlib files Right now whenever an rlib file is linked against, all of the metadata from the rlib is pulled in to the final staticlib or binary. The reason for this is that the metadata is currently stored in a section of the object file. Note that this is intentional for dynamic libraries in order to distribute metadata bundled with static libraries. This commit alters the situation for rlib libraries to instead store the metadata in a separate file in the archive. In doing so, when the archive is passed to the linker, none of the metadata will get pulled into the result executable. Furthermore, the metadata file is skipped when assembling rlibs into an archive. The snag in this implementation comes with multiple output formats. When generating a dylib, the metadata needs to be in the object file, but when generating an rlib this needs to be separate. In order to accomplish this, the metadata variable is inserted into an entirely separate LLVM Module which is then codegen'd into a different location (foo.metadata.o). This is then linked into dynamic libraries and silently ignored for rlib files. While changing how metadata is inserted into archives, I have also stopped compressing metadata when inserted into rlib files. We have wanted to stop compressing metadata, but the sections it creates in object file sections are apparently too large. Thankfully if it's just an arbitrary file it doesn't matter how large it is. I have seen massive reductions in executable sizes, as well as staticlib output sizes (to confirm that this is all working).
2013-12-03 17:41:01 -08:00
metadata: metadata,
Implement LTO This commit implements LTO for rust leveraging LLVM's passes. What this means is: * When compiling an rlib, in addition to insdering foo.o into the archive, also insert foo.bc (the LLVM bytecode) of the optimized module. * When the compiler detects the -Z lto option, it will attempt to perform LTO on a staticlib or binary output. The compiler will emit an error if a dylib or rlib output is being generated. * The actual act of performing LTO is as follows: 1. Force all upstream libraries to have an rlib version available. 2. Load the bytecode of each upstream library from the rlib. 3. Link all this bytecode into the current LLVM module (just using llvm apis) 4. Run an internalization pass which internalizes all symbols except those found reachable for the local crate of compilation. 5. Run the LLVM LTO pass manager over this entire module 6a. If assembling an archive, then add all upstream rlibs into the output archive. This ignores all of the object/bitcode/metadata files rust generated and placed inside the rlibs. 6b. If linking a binary, create copies of all upstream rlibs, remove the rust-generated object-file, and then link everything as usual. As I have explained in #10741, this process is excruciatingly slow, so this is *not* turned on by default, and it is also why I have decided to hide it behind a -Z flag for now. The good news is that the binary sizes are about as small as they can be as a result of LTO, so it's definitely working. Closes #10741 Closes #10740
2013-12-02 23:19:29 -08:00
reachable: reachable,
crate_formats: formats,
no_builtins: no_builtins,
})
}