rust/src/librustc_codegen_llvm/abi.rs

863 lines
34 KiB
Rust
Raw Normal View History

2019-02-17 18:58:58 +00:00
use crate::llvm::{self, AttributePlace};
use crate::builder::Builder;
use crate::context::CodegenCx;
use crate::type_::Type;
use crate::type_of::{LayoutLlvmExt, PointerKind};
use crate::value::Value;
use rustc_codegen_ssa::MemFlags;
use rustc_codegen_ssa::mir::place::PlaceRef;
use rustc_codegen_ssa::mir::operand::OperandValue;
use rustc_target::abi::call::ArgType;
use rustc_codegen_ssa::traits::*;
use rustc_target::abi::{HasDataLayout, LayoutOf, Size, TyLayout, Abi as LayoutAbi};
use rustc::ty::{self, Ty, Instance};
use rustc::ty::layout;
use libc::c_uint;
pub use rustc_target::spec::abi::Abi;
pub use rustc::ty::layout::{FAT_PTR_ADDR, FAT_PTR_EXTRA};
pub use rustc_target::abi::call::*;
macro_rules! for_each_kind {
($flags: ident, $f: ident, $($kind: ident),+) => ({
$(if $flags.contains(ArgAttribute::$kind) { $f(llvm::Attribute::$kind) })+
})
}
trait ArgAttributeExt {
fn for_each_kind<F>(&self, f: F) where F: FnMut(llvm::Attribute);
}
impl ArgAttributeExt for ArgAttribute {
fn for_each_kind<F>(&self, mut f: F) where F: FnMut(llvm::Attribute) {
for_each_kind!(self, f,
ByVal, NoAlias, NoCapture, NonNull, ReadOnly, SExt, StructRet, ZExt, InReg)
}
}
pub trait ArgAttributesExt {
fn apply_llfn(&self, idx: AttributePlace, llfn: &Value);
fn apply_callsite(&self, idx: AttributePlace, callsite: &Value);
}
impl ArgAttributesExt for ArgAttributes {
fn apply_llfn(&self, idx: AttributePlace, llfn: &Value) {
let mut regular = self.regular;
unsafe {
let deref = self.pointee_size.bytes();
if deref != 0 {
if regular.contains(ArgAttribute::NonNull) {
llvm::LLVMRustAddDereferenceableAttr(llfn,
idx.as_uint(),
deref);
} else {
llvm::LLVMRustAddDereferenceableOrNullAttr(llfn,
idx.as_uint(),
deref);
}
regular -= ArgAttribute::NonNull;
}
if let Some(align) = self.pointee_align {
llvm::LLVMRustAddAlignmentAttr(llfn,
idx.as_uint(),
align.bytes() as u32);
}
regular.for_each_kind(|attr| attr.apply_llfn(idx, llfn));
}
}
fn apply_callsite(&self, idx: AttributePlace, callsite: &Value) {
let mut regular = self.regular;
unsafe {
let deref = self.pointee_size.bytes();
if deref != 0 {
if regular.contains(ArgAttribute::NonNull) {
llvm::LLVMRustAddDereferenceableCallSiteAttr(callsite,
idx.as_uint(),
deref);
} else {
llvm::LLVMRustAddDereferenceableOrNullCallSiteAttr(callsite,
idx.as_uint(),
deref);
}
regular -= ArgAttribute::NonNull;
}
if let Some(align) = self.pointee_align {
llvm::LLVMRustAddAlignmentCallSiteAttr(callsite,
idx.as_uint(),
align.bytes() as u32);
}
regular.for_each_kind(|attr| attr.apply_callsite(idx, callsite));
}
}
}
pub trait LlvmType {
fn llvm_type(&self, cx: &CodegenCx<'ll, '_>) -> &'ll Type;
}
impl LlvmType for Reg {
fn llvm_type(&self, cx: &CodegenCx<'ll, '_>) -> &'ll Type {
match self.kind {
RegKind::Integer => cx.type_ix(self.size.bits()),
RegKind::Float => {
match self.size.bits() {
32 => cx.type_f32(),
64 => cx.type_f64(),
_ => bug!("unsupported float: {:?}", self)
}
}
RegKind::Vector => {
cx.type_vector(cx.type_i8(), self.size.bytes())
}
}
}
}
impl LlvmType for CastTarget {
fn llvm_type(&self, cx: &CodegenCx<'ll, '_>) -> &'ll Type {
let rest_ll_unit = self.rest.unit.llvm_type(cx);
let (rest_count, rem_bytes) = if self.rest.unit.size.bytes() == 0 {
(0, 0)
} else {
(self.rest.total.bytes() / self.rest.unit.size.bytes(),
self.rest.total.bytes() % self.rest.unit.size.bytes())
};
if self.prefix.iter().all(|x| x.is_none()) {
// Simplify to a single unit when there is no prefix and size <= unit size
if self.rest.total <= self.rest.unit.size {
return rest_ll_unit;
}
// Simplify to array when all chunks are the same size and type
if rem_bytes == 0 {
return cx.type_array(rest_ll_unit, rest_count);
}
}
// Create list of fields in the main structure
let mut args: Vec<_> =
self.prefix.iter().flat_map(|option_kind| option_kind.map(
|kind| Reg { kind: kind, size: self.prefix_chunk }.llvm_type(cx)))
.chain((0..rest_count).map(|_| rest_ll_unit))
.collect();
// Append final integer
if rem_bytes != 0 {
// Only integers can be really split further.
assert_eq!(self.rest.unit.kind, RegKind::Integer);
args.push(cx.type_ix(rem_bytes * 8));
}
cx.type_struct(&args, false)
}
}
pub trait ArgTypeExt<'ll, 'tcx> {
fn memory_ty(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
fn store(
&self,
bx: &mut Builder<'_, 'll, 'tcx>,
val: &'ll Value,
dst: PlaceRef<'tcx, &'ll Value>,
);
fn store_fn_arg(
&self,
bx: &mut Builder<'_, 'll, 'tcx>,
idx: &mut usize,
dst: PlaceRef<'tcx, &'ll Value>,
);
}
impl ArgTypeExt<'ll, 'tcx> for ArgType<'tcx, Ty<'tcx>> {
2019-02-08 13:53:55 +00:00
/// Gets the LLVM type for a place of the original Rust type of
/// this argument/return, i.e., the result of `type_of::type_of`.
fn memory_ty(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type {
2018-01-05 05:04:08 +00:00
self.layout.llvm_type(cx)
}
2019-02-08 13:53:55 +00:00
/// Stores a direct/indirect value described by this ArgType into a
/// place for the original Rust type of this argument/return.
/// Can be used for both storing formal arguments into Rust variables
/// or results of call/invoke instructions into their destinations.
fn store(
&self,
bx: &mut Builder<'_, 'll, 'tcx>,
val: &'ll Value,
dst: PlaceRef<'tcx, &'ll Value>,
) {
if self.is_ignore() {
return;
}
if self.is_sized_indirect() {
OperandValue::Ref(val, None, self.layout.align.abi).store(bx, dst)
} else if self.is_unsized_indirect() {
bug!("unsized ArgType must be handled through store_fn_arg");
} else if let PassMode::Cast(cast) = self.mode {
// FIXME(eddyb): Figure out when the simpler Store is safe, clang
// uses it for i16 -> {i8, i8}, but not for i24 -> {i8, i8, i8}.
let can_store_through_cast_ptr = false;
if can_store_through_cast_ptr {
let cast_ptr_llty = bx.type_ptr_to(cast.llvm_type(bx));
let cast_dst = bx.pointercast(dst.llval, cast_ptr_llty);
bx.store(val, cast_dst, self.layout.align.abi);
} else {
// The actual return type is a struct, but the ABI
// adaptation code has cast it into some scalar type. The
// code that follows is the only reliable way I have
// found to do a transform like i64 -> {i32,i32}.
// Basically we dump the data onto the stack then memcpy it.
//
// Other approaches I tried:
// - Casting rust ret pointer to the foreign type and using Store
// is (a) unsafe if size of foreign type > size of rust type and
// (b) runs afoul of strict aliasing rules, yielding invalid
// assembly under -O (specifically, the store gets removed).
// - Truncating foreign type to correct integral type and then
// bitcasting to the struct type yields invalid cast errors.
// We instead thus allocate some scratch space...
let scratch_size = cast.size(bx);
let scratch_align = cast.align(bx);
let llscratch = bx.alloca(cast.llvm_type(bx), "abi_cast", scratch_align);
2018-01-05 05:12:32 +00:00
bx.lifetime_start(llscratch, scratch_size);
// ...where we first store the value...
2018-01-05 05:12:32 +00:00
bx.store(val, llscratch, scratch_align);
// ...and then memcpy it to the intended destination.
bx.memcpy(
dst.llval,
self.layout.align.abi,
llscratch,
scratch_align,
bx.const_usize(self.layout.size.bytes()),
MemFlags::empty()
);
2018-01-05 05:12:32 +00:00
bx.lifetime_end(llscratch, scratch_size);
}
} else {
2018-01-05 05:12:32 +00:00
OperandValue::Immediate(val).store(bx, dst);
}
}
fn store_fn_arg(
&self,
bx: &mut Builder<'a, 'll, 'tcx>,
idx: &mut usize,
dst: PlaceRef<'tcx, &'ll Value>,
) {
let mut next = || {
2018-01-05 05:12:32 +00:00
let val = llvm::get_param(bx.llfn(), *idx as c_uint);
*idx += 1;
val
};
match self.mode {
PassMode::Ignore(_) => {}
PassMode::Pair(..) => {
2018-01-05 05:12:32 +00:00
OperandValue::Pair(next(), next()).store(bx, dst);
}
PassMode::Indirect(_, Some(_)) => {
OperandValue::Ref(next(), Some(next()), self.layout.align.abi).store(bx, dst);
}
PassMode::Direct(_) | PassMode::Indirect(_, None) | PassMode::Cast(_) => {
2018-01-05 05:12:32 +00:00
self.store(bx, next(), dst);
}
}
}
}
impl ArgTypeMethods<'tcx> for Builder<'a, 'll, 'tcx> {
fn store_fn_arg(
&mut self,
ty: &ArgType<'tcx, Ty<'tcx>>,
idx: &mut usize, dst: PlaceRef<'tcx, Self::Value>
) {
ty.store_fn_arg(self, idx, dst)
}
fn store_arg_ty(
&mut self,
ty: &ArgType<'tcx, Ty<'tcx>>,
val: &'ll Value,
dst: PlaceRef<'tcx, &'ll Value>
) {
ty.store(self, val, dst)
}
fn memory_ty(&self, ty: &ArgType<'tcx, Ty<'tcx>>) -> &'ll Type {
ty.memory_ty(self)
}
}
pub trait FnTypeExt<'tcx> {
fn of_instance(cx: &CodegenCx<'ll, 'tcx>, instance: &ty::Instance<'tcx>) -> Self;
fn new(cx: &CodegenCx<'ll, 'tcx>,
sig: ty::FnSig<'tcx>,
extra_args: &[Ty<'tcx>]) -> Self;
fn new_vtable(cx: &CodegenCx<'ll, 'tcx>,
sig: ty::FnSig<'tcx>,
extra_args: &[Ty<'tcx>]) -> Self;
fn new_internal(
cx: &CodegenCx<'ll, 'tcx>,
sig: ty::FnSig<'tcx>,
extra_args: &[Ty<'tcx>],
mk_arg_type: impl Fn(Ty<'tcx>, Option<usize>) -> ArgType<'tcx, Ty<'tcx>>,
) -> Self;
fn adjust_for_abi(&mut self,
cx: &CodegenCx<'ll, 'tcx>,
abi: Abi);
fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
fn ptr_to_llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
fn llvm_cconv(&self) -> llvm::CallConv;
fn apply_attrs_llfn(&self, llfn: &'ll Value);
fn apply_attrs_callsite(&self, bx: &mut Builder<'a, 'll, 'tcx>, callsite: &'ll Value);
}
impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
fn of_instance(cx: &CodegenCx<'ll, 'tcx>, instance: &ty::Instance<'tcx>) -> Self {
let sig = instance.fn_sig(cx.tcx);
let sig = cx.tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &sig);
2018-01-05 05:04:08 +00:00
FnType::new(cx, sig, &[])
}
fn new(cx: &CodegenCx<'ll, 'tcx>,
sig: ty::FnSig<'tcx>,
extra_args: &[Ty<'tcx>]) -> Self {
FnType::new_internal(cx, sig, extra_args, |ty, _| {
ArgType::new(cx.layout_of(ty))
})
}
fn new_vtable(cx: &CodegenCx<'ll, 'tcx>,
sig: ty::FnSig<'tcx>,
extra_args: &[Ty<'tcx>]) -> Self {
FnType::new_internal(cx, sig, extra_args, |ty, arg_idx| {
let mut layout = cx.layout_of(ty);
// Don't pass the vtable, it's not an argument of the virtual fn.
// Instead, pass just the data pointer, but give it the type `*const/mut dyn Trait`
// or `&/&mut dyn Trait` because this is special-cased elsewhere in codegen
if arg_idx == Some(0) {
let fat_pointer_ty = if layout.is_unsized() {
// unsized `self` is passed as a pointer to `self`
// FIXME (mikeyhew) change this to use &own if it is ever added to the language
cx.tcx.mk_mut_ptr(layout.ty)
} else {
match layout.abi {
LayoutAbi::ScalarPair(..) => (),
_ => bug!("receiver type has unsupported layout: {:?}", layout)
}
// In the case of Rc<Self>, we need to explicitly pass a *mut RcBox<Self>
// with a Scalar (not ScalarPair) ABI. This is a hack that is understood
// elsewhere in the compiler as a method on a `dyn Trait`.
// To get the type `*mut RcBox<Self>`, we just keep unwrapping newtypes until we
// get a built-in pointer type
let mut fat_pointer_layout = layout;
'descend_newtypes: while !fat_pointer_layout.ty.is_unsafe_ptr()
&& !fat_pointer_layout.ty.is_region_ptr()
{
'iter_fields: for i in 0..fat_pointer_layout.fields.count() {
let field_layout = fat_pointer_layout.field(cx, i);
if !field_layout.is_zst() {
fat_pointer_layout = field_layout;
continue 'descend_newtypes
}
}
bug!("receiver has no non-zero-sized fields {:?}", fat_pointer_layout);
}
fat_pointer_layout.ty
};
// we now have a type like `*mut RcBox<dyn Trait>`
// change its layout to that of `*mut ()`, a thin pointer, but keep the same type
// this is understood as a special case elsewhere in the compiler
let unit_pointer_ty = cx.tcx.mk_mut_ptr(cx.tcx.mk_unit());
layout = cx.layout_of(unit_pointer_ty);
layout.ty = fat_pointer_ty;
}
ArgType::new(layout)
})
}
fn new_internal(
cx: &CodegenCx<'ll, 'tcx>,
sig: ty::FnSig<'tcx>,
extra_args: &[Ty<'tcx>],
mk_arg_type: impl Fn(Ty<'tcx>, Option<usize>) -> ArgType<'tcx, Ty<'tcx>>,
) -> Self {
debug!("FnType::new_internal({:?}, {:?})", sig, extra_args);
use self::Abi::*;
let conv = match cx.sess().target.target.adjust_abi(sig.abi) {
RustIntrinsic | PlatformIntrinsic |
Rust | RustCall => Conv::C,
// It's the ABI's job to select this, not ours.
System => bug!("system abi should be selected elsewhere"),
Stdcall => Conv::X86Stdcall,
Fastcall => Conv::X86Fastcall,
Vectorcall => Conv::X86VectorCall,
Thiscall => Conv::X86ThisCall,
C => Conv::C,
Unadjusted => Conv::C,
Win64 => Conv::X86_64Win64,
SysV64 => Conv::X86_64SysV,
Aapcs => Conv::ArmAapcs,
PtxKernel => Conv::PtxKernel,
Msp430Interrupt => Conv::Msp430Intr,
X86Interrupt => Conv::X86Intr,
AmdGpuKernel => Conv::AmdGpuKernel,
// These API constants ought to be more specific...
Cdecl => Conv::C,
};
let mut inputs = sig.inputs();
let extra_args = if sig.abi == RustCall {
assert!(!sig.variadic && extra_args.is_empty());
match sig.inputs().last().unwrap().sty {
ty::Tuple(ref tupled_arguments) => {
inputs = &sig.inputs()[0..sig.inputs().len() - 1];
tupled_arguments
}
_ => {
bug!("argument to function with \"rust-call\" ABI \
is not a tuple");
}
}
} else {
assert!(sig.variadic || extra_args.is_empty());
extra_args
};
2018-01-05 05:04:08 +00:00
let target = &cx.sess().target.target;
let win_x64_gnu = target.target_os == "windows"
&& target.arch == "x86_64"
&& target.target_env == "gnu";
Add s390x support This adds support for building the Rust compiler and standard library for s390x-linux, allowing a full cross-bootstrap sequence to complete. This includes: - Makefile/configure changes to allow native s390x builds - Full Rust compiler support for the s390x C ABI (only the non-vector ABI is supported at this point) - Port of the standard library to s390x - Update the liblibc submodule to a version including s390x support - Testsuite fixes to allow clean "make check" on s390x Caveats: - Resets base cpu to "z10" to bring support in sync with the default behaviour of other compilers on the platforms. (Usually, upstream supports all older processors; a distribution build may then chose to require a more recent base version.) (Also, using zEC12 causes failures in the valgrind tests since valgrind doesn't fully support this CPU yet.) - z13 vector ABI is not yet supported. To ensure compatible code generation, the -vector feature is passed to LLVM. Note that this means that even when compiling for z13, no vector instructions will be used. In the future, support for the vector ABI should be added (this will require common code support for different ABIs that need different data_layout strings on the same platform). - Two test cases are (temporarily) ignored on s390x to allow passing the test suite. The underlying issues still need to be fixed: * debuginfo/simd.rs fails because of incorrect debug information. This seems to be a LLVM bug (also seen with C code). * run-pass/union/union-basic.rs simply seems to be incorrect for all big-endian platforms. Signed-off-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
2016-09-09 21:00:23 +00:00
let linux_s390x = target.target_os == "linux"
&& target.arch == "s390x"
&& target.target_env == "gnu";
let linux_sparc64 = target.target_os == "linux"
&& target.arch == "sparc64"
&& target.target_env == "gnu";
let rust_abi = match sig.abi {
RustIntrinsic | PlatformIntrinsic | Rust | RustCall => true,
_ => false
};
// Handle safe Rust thin and fat pointers.
let adjust_for_rust_scalar = |attrs: &mut ArgAttributes,
scalar: &layout::Scalar,
layout: TyLayout<'tcx, Ty<'tcx>>,
offset: Size,
is_return: bool| {
// Booleans are always an i1 that needs to be zero-extended.
if scalar.is_bool() {
attrs.set(ArgAttribute::ZExt);
return;
}
// Only pointer types handled below.
if scalar.value != layout::Pointer {
return;
}
if scalar.valid_range.start() < scalar.valid_range.end() {
if *scalar.valid_range.start() > 0 {
attrs.set(ArgAttribute::NonNull);
}
}
2018-01-05 05:04:08 +00:00
if let Some(pointee) = layout.pointee_info_at(cx, offset) {
if let Some(kind) = pointee.safe {
attrs.pointee_size = pointee.size;
attrs.pointee_align = Some(pointee.align);
// `Box` pointer parameters never alias because ownership is transferred
// `&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
let no_alias = match kind {
PointerKind::Shared => false,
PointerKind::UniqueOwned => true,
PointerKind::Frozen |
PointerKind::UniqueBorrowed => !is_return
};
if no_alias {
attrs.set(ArgAttribute::NoAlias);
}
if kind == PointerKind::Frozen && !is_return {
attrs.set(ArgAttribute::ReadOnly);
}
}
}
};
// Store the index of the last argument. This is useful for working with
// C-compatible variadic arguments.
let last_arg_idx = if sig.inputs().is_empty() {
None
} else {
Some(sig.inputs().len() - 1)
};
let arg_of = |ty: Ty<'tcx>, arg_idx: Option<usize>| {
let is_return = arg_idx.is_none();
let mut arg = mk_arg_type(ty, arg_idx);
if arg.layout.is_zst() {
// For some forsaken reason, x86_64-pc-windows-gnu
// doesn't ignore zero-sized struct arguments.
// The same is true for s390x-unknown-linux-gnu
// and sparc64-unknown-linux-gnu.
if is_return || rust_abi || (!win_x64_gnu && !linux_s390x && !linux_sparc64) {
arg.mode = PassMode::Ignore(IgnoreMode::Zst);
}
}
// If this is a C-variadic function, this is not the return value,
// and there is one or more fixed arguments; ensure that the `VaList`
// is ignored as an argument.
if sig.variadic {
match (last_arg_idx, arg_idx) {
(Some(last_idx), Some(cur_idx)) if last_idx == cur_idx => {
let va_list_did = match cx.tcx.lang_items().va_list() {
Some(did) => did,
None => bug!("`va_list` lang item required for C-variadic functions"),
};
match ty.sty {
ty::Adt(def, _) if def.did == va_list_did => {
// This is the "spoofed" `VaList`. Set the arguments mode
// so that it will be ignored.
arg.mode = PassMode::Ignore(IgnoreMode::CVarArgs);
},
_ => (),
}
}
_ => {}
}
}
// FIXME(eddyb) other ABIs don't have logic for scalar pairs.
if !is_return && rust_abi {
if let layout::Abi::ScalarPair(ref a, ref b) = arg.layout.abi {
let mut a_attrs = ArgAttributes::new();
let mut b_attrs = ArgAttributes::new();
adjust_for_rust_scalar(&mut a_attrs,
a,
arg.layout,
2018-05-20 12:14:39 +00:00
Size::ZERO,
false);
adjust_for_rust_scalar(&mut b_attrs,
b,
arg.layout,
a.value.size(cx).align_to(b.value.align(cx).abi),
false);
arg.mode = PassMode::Pair(a_attrs, b_attrs);
return arg;
}
}
if let layout::Abi::Scalar(ref scalar) = arg.layout.abi {
if let PassMode::Direct(ref mut attrs) = arg.mode {
adjust_for_rust_scalar(attrs,
scalar,
arg.layout,
2018-05-20 12:14:39 +00:00
Size::ZERO,
is_return);
}
}
arg
};
let mut fn_ty = FnType {
ret: arg_of(sig.output(), None),
args: inputs.iter().chain(extra_args).enumerate().map(|(i, ty)| {
arg_of(ty, Some(i))
}).collect(),
variadic: sig.variadic,
conv,
};
fn_ty.adjust_for_abi(cx, sig.abi);
fn_ty
}
fn adjust_for_abi(&mut self,
cx: &CodegenCx<'ll, 'tcx>,
abi: Abi) {
if abi == Abi::Unadjusted { return }
if abi == Abi::Rust || abi == Abi::RustCall ||
abi == Abi::RustIntrinsic || abi == Abi::PlatformIntrinsic {
let fixup = |arg: &mut ArgType<'tcx, Ty<'tcx>>| {
if arg.is_ignore() { return; }
match arg.layout.abi {
layout::Abi::Aggregate { .. } => {}
rustc: SIMD types use pointers in Rust's ABI This commit changes the ABI of SIMD types in the "Rust" ABI to unconditionally be passed via pointers instead of being passed as immediates. This should fix a longstanding issue, #44367, where SIMD-using programs ended up showing very odd behavior at runtime because the ABI between functions was mismatched. As a bit of a recap, this is sort of an LLVM bug and sort of an LLVM feature (today's behavior). LLVM will generate code for a function solely looking at the function it's generating, including calls to other functions. Let's then say you've got something that looks like: ```llvm define void @foo() { ; no target features enabled call void @bar(<i64 x 4> zeroinitializer) ret void } define void @bar(<i64 x 4>) #0 { ; enables the AVX feature ... } ``` LLVM will codegen the call to `bar` *without* using AVX registers becauase `foo` doesn't have access to these registers. Instead it's generated with emulation that uses two 128-bit registers. The `bar` function, on the other hand, will expect its argument in an AVX register (as it has AVX enabled). This means we've got a codegen problem! Comments on #44367 have some more contexutal information but the crux of the issue is that if we want SIMD to work in general we'll need to ensure that whenever a function calls another they ABI of the arguments being passed is in agreement. One possible solution to this would be to insert "shim functions" where whenever a `target_feature` mismatch is detected the compiler inserts a shim function where you pass arguments via memory to the shim and then the shim loads the values and calls the target function (where the shim and the target have the same target features enabled). This unfortunately is quite nontrivial to implement in rustc today (especially when accounting for function pointers and such). This commit takes a different solution, *always* passing SIMD arguments through memory instead of passing as immediates. This strategy solves the problem at the LLVM layer because the ABI between two functions never uses SIMD registers. This also shouldn't be a hit to performance because SIMD performance is thought to often rely on inlining anyway, where a `call` instruction, even if using SIMD registers, would be disastrous to performance regardless. LLVM should then be more than capable of fixing all our memory usage to use registers instead after enough inlining has been performed. Note that there's a few caveats to this commit though: * The "platform intrinsic" ABI is omitted from "always pass via memory". This ABI is used to define intrinsics like `simd_shuffle4` where LLVM and rustc need to have the arguments as an immediate. * Additionally this commit does *not* fix the `extern` ("C") ABI. This means that the bug in #44367 can still happen when using non-Rust-ABI functions. My hope is that before stabilization we can ban and/or warn about SIMD types in these functions (as AFAIK there's not much motivation to belong there anyway), but I'll leave that for a later commit and if this is merged I'll file a follow-up issue. All in all this... Closes #44367
2018-01-25 16:00:22 +00:00
// This is a fun case! The gist of what this is doing is
// that we want callers and callees to always agree on the
// ABI of how they pass SIMD arguments. If we were to *not*
// make these arguments indirect then they'd be immediates
// in LLVM, which means that they'd used whatever the
// appropriate ABI is for the callee and the caller. That
// means, for example, if the caller doesn't have AVX
// enabled but the callee does, then passing an AVX argument
// across this boundary would cause corrupt data to show up.
//
// This problem is fixed by unconditionally passing SIMD
// arguments through memory between callers and callees
// which should get them all to agree on ABI regardless of
// target feature sets. Some more information about this
// issue can be found in #44367.
//
// Note that the platform intrinsic ABI is exempt here as
// that's how we connect up to LLVM and it's unstable
// anyway, we control all calls to it in libstd.
layout::Abi::Vector { .. }
if abi != Abi::PlatformIntrinsic &&
cx.sess().target.target.options.simd_types_indirect =>
{
rustc: SIMD types use pointers in Rust's ABI This commit changes the ABI of SIMD types in the "Rust" ABI to unconditionally be passed via pointers instead of being passed as immediates. This should fix a longstanding issue, #44367, where SIMD-using programs ended up showing very odd behavior at runtime because the ABI between functions was mismatched. As a bit of a recap, this is sort of an LLVM bug and sort of an LLVM feature (today's behavior). LLVM will generate code for a function solely looking at the function it's generating, including calls to other functions. Let's then say you've got something that looks like: ```llvm define void @foo() { ; no target features enabled call void @bar(<i64 x 4> zeroinitializer) ret void } define void @bar(<i64 x 4>) #0 { ; enables the AVX feature ... } ``` LLVM will codegen the call to `bar` *without* using AVX registers becauase `foo` doesn't have access to these registers. Instead it's generated with emulation that uses two 128-bit registers. The `bar` function, on the other hand, will expect its argument in an AVX register (as it has AVX enabled). This means we've got a codegen problem! Comments on #44367 have some more contexutal information but the crux of the issue is that if we want SIMD to work in general we'll need to ensure that whenever a function calls another they ABI of the arguments being passed is in agreement. One possible solution to this would be to insert "shim functions" where whenever a `target_feature` mismatch is detected the compiler inserts a shim function where you pass arguments via memory to the shim and then the shim loads the values and calls the target function (where the shim and the target have the same target features enabled). This unfortunately is quite nontrivial to implement in rustc today (especially when accounting for function pointers and such). This commit takes a different solution, *always* passing SIMD arguments through memory instead of passing as immediates. This strategy solves the problem at the LLVM layer because the ABI between two functions never uses SIMD registers. This also shouldn't be a hit to performance because SIMD performance is thought to often rely on inlining anyway, where a `call` instruction, even if using SIMD registers, would be disastrous to performance regardless. LLVM should then be more than capable of fixing all our memory usage to use registers instead after enough inlining has been performed. Note that there's a few caveats to this commit though: * The "platform intrinsic" ABI is omitted from "always pass via memory". This ABI is used to define intrinsics like `simd_shuffle4` where LLVM and rustc need to have the arguments as an immediate. * Additionally this commit does *not* fix the `extern` ("C") ABI. This means that the bug in #44367 can still happen when using non-Rust-ABI functions. My hope is that before stabilization we can ban and/or warn about SIMD types in these functions (as AFAIK there's not much motivation to belong there anyway), but I'll leave that for a later commit and if this is merged I'll file a follow-up issue. All in all this... Closes #44367
2018-01-25 16:00:22 +00:00
arg.make_indirect();
return
}
_ => return
}
let size = arg.layout.size;
if arg.layout.is_unsized() || size > layout::Pointer.size(cx) {
arg.make_indirect();
} else {
// We want to pass small aggregates as immediates, but using
// a LLVM aggregate type for this leads to bad optimizations,
// so we pick an appropriately sized integer type instead.
arg.cast_to(Reg {
kind: RegKind::Integer,
size
});
}
};
fixup(&mut self.ret);
for arg in &mut self.args {
fixup(arg);
}
if let PassMode::Indirect(ref mut attrs, _) = self.ret.mode {
attrs.set(ArgAttribute::StructRet);
}
return;
}
if let Err(msg) = self.adjust_for_cabi(cx, abi) {
cx.sess().fatal(&msg);
}
}
fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type {
let args_capacity: usize = self.args.iter().map(|arg|
if arg.pad.is_some() { 1 } else { 0 } +
if let PassMode::Pair(_, _) = arg.mode { 2 } else { 1 }
).sum();
let mut llargument_tys = Vec::with_capacity(
if let PassMode::Indirect(..) = self.ret.mode { 1 } else { 0 } + args_capacity
);
let llreturn_ty = match self.ret.mode {
PassMode::Ignore(IgnoreMode::Zst) => cx.type_void(),
PassMode::Ignore(IgnoreMode::CVarArgs) =>
bug!("`va_list` should never be a return type"),
PassMode::Direct(_) | PassMode::Pair(..) => {
2018-01-05 05:04:08 +00:00
self.ret.layout.immediate_llvm_type(cx)
}
2018-01-05 05:04:08 +00:00
PassMode::Cast(cast) => cast.llvm_type(cx),
PassMode::Indirect(..) => {
llargument_tys.push(cx.type_ptr_to(self.ret.memory_ty(cx)));
cx.type_void()
}
};
for arg in &self.args {
// add padding
if let Some(ty) = arg.pad {
2018-01-05 05:04:08 +00:00
llargument_tys.push(ty.llvm_type(cx));
}
let llarg_ty = match arg.mode {
PassMode::Ignore(_) => continue,
2018-01-05 05:04:08 +00:00
PassMode::Direct(_) => arg.layout.immediate_llvm_type(cx),
PassMode::Pair(..) => {
llargument_tys.push(arg.layout.scalar_pair_element_llvm_type(cx, 0, true));
llargument_tys.push(arg.layout.scalar_pair_element_llvm_type(cx, 1, true));
continue;
}
PassMode::Indirect(_, Some(_)) => {
let ptr_ty = cx.tcx.mk_mut_ptr(arg.layout.ty);
let ptr_layout = cx.layout_of(ptr_ty);
llargument_tys.push(ptr_layout.scalar_pair_element_llvm_type(cx, 0, true));
llargument_tys.push(ptr_layout.scalar_pair_element_llvm_type(cx, 1, true));
continue;
}
2018-01-05 05:04:08 +00:00
PassMode::Cast(cast) => cast.llvm_type(cx),
PassMode::Indirect(_, None) => cx.type_ptr_to(arg.memory_ty(cx)),
};
llargument_tys.push(llarg_ty);
}
if self.variadic {
cx.type_variadic_func(&llargument_tys, llreturn_ty)
} else {
cx.type_func(&llargument_tys, llreturn_ty)
}
}
fn ptr_to_llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type {
unsafe {
llvm::LLVMPointerType(self.llvm_type(cx),
cx.data_layout().instruction_address_space as c_uint)
}
}
fn llvm_cconv(&self) -> llvm::CallConv {
match self.conv {
Conv::C => llvm::CCallConv,
Conv::AmdGpuKernel => llvm::AmdGpuKernel,
Conv::ArmAapcs => llvm::ArmAapcsCallConv,
Conv::Msp430Intr => llvm::Msp430Intr,
Conv::PtxKernel => llvm::PtxKernel,
Conv::X86Fastcall => llvm::X86FastcallCallConv,
Conv::X86Intr => llvm::X86_Intr,
Conv::X86Stdcall => llvm::X86StdcallCallConv,
Conv::X86ThisCall => llvm::X86_ThisCall,
Conv::X86VectorCall => llvm::X86_VectorCall,
Conv::X86_64SysV => llvm::X86_64_SysV,
Conv::X86_64Win64 => llvm::X86_64_Win64,
}
}
fn apply_attrs_llfn(&self, llfn: &'ll Value) {
let mut i = 0;
let mut apply = |attrs: &ArgAttributes| {
attrs.apply_llfn(llvm::AttributePlace::Argument(i), llfn);
i += 1;
};
match self.ret.mode {
PassMode::Direct(ref attrs) => {
attrs.apply_llfn(llvm::AttributePlace::ReturnValue, llfn);
}
PassMode::Indirect(ref attrs, _) => apply(attrs),
_ => {}
}
for arg in &self.args {
if arg.pad.is_some() {
apply(&ArgAttributes::new());
}
match arg.mode {
PassMode::Ignore(_) => {}
PassMode::Direct(ref attrs) |
PassMode::Indirect(ref attrs, None) => apply(attrs),
PassMode::Indirect(ref attrs, Some(ref extra_attrs)) => {
apply(attrs);
apply(extra_attrs);
}
PassMode::Pair(ref a, ref b) => {
apply(a);
apply(b);
}
PassMode::Cast(_) => apply(&ArgAttributes::new()),
}
}
}
fn apply_attrs_callsite(&self, bx: &mut Builder<'a, 'll, 'tcx>, callsite: &'ll Value) {
let mut i = 0;
let mut apply = |attrs: &ArgAttributes| {
attrs.apply_callsite(llvm::AttributePlace::Argument(i), callsite);
i += 1;
};
match self.ret.mode {
PassMode::Direct(ref attrs) => {
attrs.apply_callsite(llvm::AttributePlace::ReturnValue, callsite);
}
PassMode::Indirect(ref attrs, _) => apply(attrs),
_ => {}
}
if let layout::Abi::Scalar(ref scalar) = self.ret.layout.abi {
// If the value is a boolean, the range is 0..2 and that ultimately
// become 0..0 when the type becomes i1, which would be rejected
// by the LLVM verifier.
2018-10-08 14:58:26 +00:00
if let layout::Int(..) = scalar.value {
if !scalar.is_bool() {
let range = scalar.valid_range_exclusive(bx);
if range.start != range.end {
2018-06-29 08:28:51 +00:00
bx.range_metadata(callsite, range);
}
}
}
}
for arg in &self.args {
if arg.pad.is_some() {
apply(&ArgAttributes::new());
}
match arg.mode {
PassMode::Ignore(_) => {}
PassMode::Direct(ref attrs) |
PassMode::Indirect(ref attrs, None) => apply(attrs),
PassMode::Indirect(ref attrs, Some(ref extra_attrs)) => {
apply(attrs);
apply(extra_attrs);
}
PassMode::Pair(ref a, ref b) => {
apply(a);
apply(b);
}
PassMode::Cast(_) => apply(&ArgAttributes::new()),
}
}
let cconv = self.llvm_cconv();
if cconv != llvm::CCallConv {
llvm::SetInstructionCallConv(callsite, cconv);
}
}
}
impl AbiMethods<'tcx> for CodegenCx<'ll, 'tcx> {
fn new_fn_type(&self, sig: ty::FnSig<'tcx>, extra_args: &[Ty<'tcx>]) -> FnType<'tcx, Ty<'tcx>> {
FnType::new(&self, sig, extra_args)
}
fn new_vtable(
&self,
sig: ty::FnSig<'tcx>,
extra_args: &[Ty<'tcx>]
) -> FnType<'tcx, Ty<'tcx>> {
FnType::new_vtable(&self, sig, extra_args)
}
fn fn_type_of_instance(&self, instance: &Instance<'tcx>) -> FnType<'tcx, Ty<'tcx>> {
FnType::of_instance(&self, instance)
}
}
impl AbiBuilderMethods<'tcx> for Builder<'a, 'll, 'tcx> {
fn apply_attrs_callsite(
&mut self,
ty: &FnType<'tcx, Ty<'tcx>>,
callsite: Self::Value
) {
ty.apply_attrs_callsite(self, callsite)
}
}