rust/compiler/rustc_codegen_llvm/src/builder.rs

1431 lines
49 KiB
Rust
Raw Normal View History

2019-02-17 18:58:58 +00:00
use crate::common::Funclet;
use crate::context::CodegenCx;
2019-12-22 22:42:04 +00:00
use crate::llvm::{self, BasicBlock, False};
use crate::llvm::{AtomicOrdering, AtomicRmwBinOp, SynchronizationScope};
2019-02-17 18:58:58 +00:00
use crate::type_::Type;
use crate::type_of::LayoutLlvmExt;
use crate::value::Value;
2021-02-13 11:17:15 +00:00
use cstr::cstr;
2019-12-22 22:42:04 +00:00
use libc::{c_char, c_uint};
use rustc_codegen_ssa::common::{IntPredicate, RealPredicate, TypeKind};
use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
use rustc_codegen_ssa::mir::place::PlaceRef;
2019-12-22 22:42:04 +00:00
use rustc_codegen_ssa::traits::*;
use rustc_codegen_ssa::MemFlags;
use rustc_data_structures::small_c_str::SmallCStr;
use rustc_hir::def_id::DefId;
use rustc_middle::ty::layout::TyAndLayout;
2020-03-29 14:41:09 +00:00
use rustc_middle::ty::{self, Ty, TyCtxt};
use rustc_span::{sym, Span};
use rustc_target::abi::{self, Align, Size};
2019-05-14 15:53:01 +00:00
use rustc_target::spec::{HasTargetSpec, Target};
use std::borrow::Cow;
use std::ffi::CStr;
use std::ops::{Deref, Range};
use std::ptr;
2020-08-05 11:35:53 +00:00
use tracing::debug;
2016-12-31 23:00:24 +00:00
// All Builders must have an llfn associated with them
#[must_use]
pub struct Builder<'a, 'll, 'tcx> {
pub llbuilder: &'ll mut llvm::Builder<'ll>,
pub cx: &'a CodegenCx<'ll, 'tcx>,
}
impl Drop for Builder<'a, 'll, 'tcx> {
2016-12-17 00:39:35 +00:00
fn drop(&mut self) {
unsafe {
llvm::LLVMDisposeBuilder(&mut *(self.llbuilder as *mut _));
2016-12-17 00:39:35 +00:00
}
}
}
// FIXME(eddyb) use a checked constructor when they become `const fn`.
2019-12-22 22:42:04 +00:00
const EMPTY_C_STR: &CStr = unsafe { CStr::from_bytes_with_nul_unchecked(b"\0") };
/// Empty string, to be used where LLVM expects an instruction name, indicating
/// that the instruction is to be left unnamed (i.e. numbered, in textual IR).
// FIXME(eddyb) pass `&CStr` directly to FFI once it's a thin pointer.
const UNNAMED: *const c_char = EMPTY_C_STR.as_ptr();
impl BackendTypes for Builder<'_, 'll, 'tcx> {
type Value = <CodegenCx<'ll, 'tcx> as BackendTypes>::Value;
2019-10-13 09:28:19 +00:00
type Function = <CodegenCx<'ll, 'tcx> as BackendTypes>::Function;
type BasicBlock = <CodegenCx<'ll, 'tcx> as BackendTypes>::BasicBlock;
type Type = <CodegenCx<'ll, 'tcx> as BackendTypes>::Type;
type Funclet = <CodegenCx<'ll, 'tcx> as BackendTypes>::Funclet;
type DIScope = <CodegenCx<'ll, 'tcx> as BackendTypes>::DIScope;
type DILocation = <CodegenCx<'ll, 'tcx> as BackendTypes>::DILocation;
type DIVariable = <CodegenCx<'ll, 'tcx> as BackendTypes>::DIVariable;
2018-09-13 12:58:19 +00:00
}
impl abi::HasDataLayout for Builder<'_, '_, '_> {
fn data_layout(&self) -> &abi::TargetDataLayout {
2018-09-13 12:58:19 +00:00
self.cx.data_layout()
}
}
impl ty::layout::HasTyCtxt<'tcx> for Builder<'_, '_, 'tcx> {
2019-06-13 21:48:52 +00:00
fn tcx(&self) -> TyCtxt<'tcx> {
2018-09-13 12:58:19 +00:00
self.cx.tcx
}
}
2019-05-04 09:32:22 +00:00
impl ty::layout::HasParamEnv<'tcx> for Builder<'_, '_, 'tcx> {
fn param_env(&self) -> ty::ParamEnv<'tcx> {
self.cx.param_env()
}
}
2019-05-14 15:53:01 +00:00
impl HasTargetSpec for Builder<'_, '_, 'tcx> {
fn target_spec(&self) -> &Target {
&self.cx.target_spec()
}
}
impl abi::LayoutOf for Builder<'_, '_, 'tcx> {
2018-09-13 12:58:19 +00:00
type Ty = Ty<'tcx>;
2020-03-04 14:50:21 +00:00
type TyAndLayout = TyAndLayout<'tcx>;
2018-09-13 12:58:19 +00:00
2020-03-04 14:50:21 +00:00
fn layout_of(&self, ty: Ty<'tcx>) -> Self::TyAndLayout {
2018-09-13 12:58:19 +00:00
self.cx.layout_of(ty)
}
}
impl Deref for Builder<'_, 'll, 'tcx> {
type Target = CodegenCx<'ll, 'tcx>;
fn deref(&self) -> &Self::Target {
self.cx
}
}
2018-09-13 12:58:19 +00:00
impl HasCodegen<'tcx> for Builder<'_, 'll, 'tcx> {
2018-09-11 09:46:03 +00:00
type CodegenCx = CodegenCx<'ll, 'tcx>;
}
macro_rules! builder_methods_for_value_instructions {
($($name:ident($($arg:ident),*) => $llvm_capi:ident),+ $(,)?) => {
$(fn $name(&mut self, $($arg: &'ll Value),*) -> &'ll Value {
unsafe {
llvm::$llvm_capi(self.llbuilder, $($arg,)* UNNAMED)
}
})+
}
}
impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
2019-12-22 22:42:04 +00:00
fn new_block<'b>(cx: &'a CodegenCx<'ll, 'tcx>, llfn: &'ll Value, name: &'b str) -> Self {
let mut bx = Builder::with_cx(cx);
2016-12-31 23:00:24 +00:00
let llbb = unsafe {
let name = SmallCStr::new(name);
2019-12-22 22:42:04 +00:00
llvm::LLVMAppendBasicBlockInContext(cx.llcx, llfn, name.as_ptr())
2016-12-31 23:00:24 +00:00
};
2018-01-05 05:12:32 +00:00
bx.position_at_end(llbb);
bx
2016-12-31 23:00:24 +00:00
}
fn with_cx(cx: &'a CodegenCx<'ll, 'tcx>) -> Self {
2016-12-17 00:39:35 +00:00
// Create a fresh builder from the crate context.
2019-12-22 22:42:04 +00:00
let llbuilder = unsafe { llvm::LLVMCreateBuilderInContext(cx.llcx) };
Builder { llbuilder, cx }
}
2019-06-21 18:27:44 +00:00
fn build_sibling_block(&self, name: &str) -> Self {
2018-01-05 05:04:08 +00:00
Builder::new_block(self.cx, self.llfn(), name)
2016-12-31 23:00:24 +00:00
}
2018-08-23 13:23:48 +00:00
fn llbb(&self) -> &'ll BasicBlock {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMGetInsertBlock(self.llbuilder) }
2016-12-31 23:00:24 +00:00
}
2020-10-17 11:28:58 +00:00
fn set_span(&mut self, _span: Span) {}
fn position_at_end(&mut self, llbb: &'ll BasicBlock) {
unsafe {
llvm::LLVMPositionBuilderAtEnd(self.llbuilder, llbb);
}
}
fn ret_void(&mut self) {
unsafe {
llvm::LLVMBuildRetVoid(self.llbuilder);
}
}
fn ret(&mut self, v: &'ll Value) {
unsafe {
llvm::LLVMBuildRet(self.llbuilder, v);
}
}
fn br(&mut self, dest: &'ll BasicBlock) {
unsafe {
llvm::LLVMBuildBr(self.llbuilder, dest);
}
}
fn cond_br(
&mut self,
2018-08-23 13:23:48 +00:00
cond: &'ll Value,
then_llbb: &'ll BasicBlock,
else_llbb: &'ll BasicBlock,
2018-07-17 15:26:58 +00:00
) {
unsafe {
llvm::LLVMBuildCondBr(self.llbuilder, cond, then_llbb, else_llbb);
}
}
fn switch(
&mut self,
2018-08-23 13:23:48 +00:00
v: &'ll Value,
else_llbb: &'ll BasicBlock,
cases: impl ExactSizeIterator<Item = (u128, &'ll BasicBlock)>,
2018-12-08 17:42:31 +00:00
) {
2019-12-22 22:42:04 +00:00
let switch =
unsafe { llvm::LLVMBuildSwitch(self.llbuilder, v, else_llbb, cases.len() as c_uint) };
2018-12-08 17:42:31 +00:00
for (on_val, dest) in cases {
let on_val = self.const_uint_big(self.val_ty(v), on_val);
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMAddCase(switch, on_val, dest) }
}
}
fn invoke(
&mut self,
llfn: &'ll Value,
args: &[&'ll Value],
then: &'ll BasicBlock,
catch: &'ll BasicBlock,
funclet: Option<&Funclet<'ll>>,
) -> &'ll Value {
2019-12-22 22:42:04 +00:00
debug!("invoke {:?} with args ({:?})", llfn, args);
let args = self.check_call("invoke", llfn, args);
let bundle = funclet.map(|funclet| funclet.bundle());
let bundle = bundle.as_ref().map(|b| &*b.raw);
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMRustBuildInvoke(
self.llbuilder,
llfn,
args.as_ptr(),
args.len() as c_uint,
then,
catch,
bundle,
UNNAMED,
)
}
}
fn unreachable(&mut self) {
unsafe {
llvm::LLVMBuildUnreachable(self.llbuilder);
}
}
builder_methods_for_value_instructions! {
add(a, b) => LLVMBuildAdd,
fadd(a, b) => LLVMBuildFAdd,
sub(a, b) => LLVMBuildSub,
fsub(a, b) => LLVMBuildFSub,
mul(a, b) => LLVMBuildMul,
fmul(a, b) => LLVMBuildFMul,
udiv(a, b) => LLVMBuildUDiv,
exactudiv(a, b) => LLVMBuildExactUDiv,
sdiv(a, b) => LLVMBuildSDiv,
exactsdiv(a, b) => LLVMBuildExactSDiv,
fdiv(a, b) => LLVMBuildFDiv,
urem(a, b) => LLVMBuildURem,
srem(a, b) => LLVMBuildSRem,
frem(a, b) => LLVMBuildFRem,
shl(a, b) => LLVMBuildShl,
lshr(a, b) => LLVMBuildLShr,
ashr(a, b) => LLVMBuildAShr,
and(a, b) => LLVMBuildAnd,
or(a, b) => LLVMBuildOr,
xor(a, b) => LLVMBuildXor,
neg(x) => LLVMBuildNeg,
fneg(x) => LLVMBuildFNeg,
not(x) => LLVMBuildNot,
2019-06-03 10:59:17 +00:00
unchecked_sadd(x, y) => LLVMBuildNSWAdd,
unchecked_uadd(x, y) => LLVMBuildNUWAdd,
unchecked_ssub(x, y) => LLVMBuildNSWSub,
unchecked_usub(x, y) => LLVMBuildNUWSub,
unchecked_smul(x, y) => LLVMBuildNSWMul,
unchecked_umul(x, y) => LLVMBuildNUWMul,
}
fn fadd_fast(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMBuildFAdd(self.llbuilder, lhs, rhs, UNNAMED);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
fn fsub_fast(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMBuildFSub(self.llbuilder, lhs, rhs, UNNAMED);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
fn fmul_fast(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMBuildFMul(self.llbuilder, lhs, rhs, UNNAMED);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
fn fdiv_fast(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMBuildFDiv(self.llbuilder, lhs, rhs, UNNAMED);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
fn frem_fast(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMBuildFRem(self.llbuilder, lhs, rhs, UNNAMED);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
fn checked_binop(
&mut self,
oop: OverflowOp,
ty: Ty<'_>,
lhs: Self::Value,
rhs: Self::Value,
) -> (Self::Value, Self::Value) {
2020-03-29 14:41:09 +00:00
use rustc_middle::ty::{Int, Uint};
use rustc_middle::ty::{IntTy::*, UintTy::*};
2020-08-02 22:49:11 +00:00
let new_kind = match ty.kind() {
Int(t @ Isize) => Int(t.normalize(self.tcx.sess.target.pointer_width)),
Uint(t @ Usize) => Uint(t.normalize(self.tcx.sess.target.pointer_width)),
2020-08-02 22:49:11 +00:00
t @ (Uint(_) | Int(_)) => t.clone(),
2019-12-22 22:42:04 +00:00
_ => panic!("tried to get overflow intrinsic for op applied to non-int type"),
};
let name = match oop {
OverflowOp::Add => match new_kind {
Int(I8) => "llvm.sadd.with.overflow.i8",
Int(I16) => "llvm.sadd.with.overflow.i16",
Int(I32) => "llvm.sadd.with.overflow.i32",
Int(I64) => "llvm.sadd.with.overflow.i64",
Int(I128) => "llvm.sadd.with.overflow.i128",
Uint(U8) => "llvm.uadd.with.overflow.i8",
Uint(U16) => "llvm.uadd.with.overflow.i16",
Uint(U32) => "llvm.uadd.with.overflow.i32",
Uint(U64) => "llvm.uadd.with.overflow.i64",
Uint(U128) => "llvm.uadd.with.overflow.i128",
_ => unreachable!(),
},
OverflowOp::Sub => match new_kind {
Int(I8) => "llvm.ssub.with.overflow.i8",
Int(I16) => "llvm.ssub.with.overflow.i16",
Int(I32) => "llvm.ssub.with.overflow.i32",
Int(I64) => "llvm.ssub.with.overflow.i64",
Int(I128) => "llvm.ssub.with.overflow.i128",
Uint(U8) => "llvm.usub.with.overflow.i8",
Uint(U16) => "llvm.usub.with.overflow.i16",
Uint(U32) => "llvm.usub.with.overflow.i32",
Uint(U64) => "llvm.usub.with.overflow.i64",
Uint(U128) => "llvm.usub.with.overflow.i128",
_ => unreachable!(),
},
OverflowOp::Mul => match new_kind {
Int(I8) => "llvm.smul.with.overflow.i8",
Int(I16) => "llvm.smul.with.overflow.i16",
Int(I32) => "llvm.smul.with.overflow.i32",
Int(I64) => "llvm.smul.with.overflow.i64",
Int(I128) => "llvm.smul.with.overflow.i128",
Uint(U8) => "llvm.umul.with.overflow.i8",
Uint(U16) => "llvm.umul.with.overflow.i16",
Uint(U32) => "llvm.umul.with.overflow.i32",
Uint(U64) => "llvm.umul.with.overflow.i64",
Uint(U128) => "llvm.umul.with.overflow.i128",
_ => unreachable!(),
},
};
let intrinsic = self.get_intrinsic(&name);
let res = self.call(intrinsic, &[lhs, rhs], None);
2019-12-22 22:42:04 +00:00
(self.extract_value(res, 0), self.extract_value(res, 1))
}
fn from_immediate(&mut self, val: Self::Value) -> Self::Value {
if self.cx().val_ty(val) == self.cx().type_i1() {
self.zext(val, self.cx().type_i8())
} else {
val
}
}
fn to_immediate_scalar(&mut self, val: Self::Value, scalar: &abi::Scalar) -> Self::Value {
if scalar.is_bool() {
return self.trunc(val, self.cx().type_i1());
}
val
}
fn alloca(&mut self, ty: &'ll Type, align: Align) -> &'ll Value {
let mut bx = Builder::with_cx(self.cx);
2019-12-22 22:42:04 +00:00
bx.position_at_start(unsafe { llvm::LLVMGetFirstBasicBlock(self.llfn()) });
bx.dynamic_alloca(ty, align)
2016-12-31 23:00:24 +00:00
}
fn dynamic_alloca(&mut self, ty: &'ll Type, align: Align) -> &'ll Value {
unsafe {
let alloca = llvm::LLVMBuildAlloca(self.llbuilder, ty, UNNAMED);
llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint);
alloca
}
}
2019-12-22 22:42:04 +00:00
fn array_alloca(&mut self, ty: &'ll Type, len: &'ll Value, align: Align) -> &'ll Value {
2018-05-28 15:07:23 +00:00
unsafe {
let alloca = llvm::LLVMBuildArrayAlloca(self.llbuilder, ty, len, UNNAMED);
llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint);
2018-05-28 15:07:23 +00:00
alloca
}
}
fn load(&mut self, ptr: &'ll Value, align: Align) -> &'ll Value {
unsafe {
let load = llvm::LLVMBuildLoad(self.llbuilder, ptr, UNNAMED);
llvm::LLVMSetAlignment(load, align.bytes() as c_uint);
load
}
}
fn volatile_load(&mut self, ptr: &'ll Value) -> &'ll Value {
unsafe {
let load = llvm::LLVMBuildLoad(self.llbuilder, ptr, UNNAMED);
llvm::LLVMSetVolatile(load, llvm::True);
load
}
}
fn atomic_load(
&mut self,
2018-08-23 13:23:48 +00:00
ptr: &'ll Value,
order: rustc_codegen_ssa::common::AtomicOrdering,
size: Size,
2018-08-23 13:23:48 +00:00
) -> &'ll Value {
unsafe {
let load = llvm::LLVMRustBuildAtomicLoad(
self.llbuilder,
ptr,
UNNAMED,
AtomicOrdering::from_generic(order),
);
// LLVM requires the alignment of atomic loads to be at least the size of the type.
llvm::LLVMSetAlignment(load, size.bytes() as c_uint);
load
}
}
2019-12-22 22:42:04 +00:00
fn load_operand(&mut self, place: PlaceRef<'tcx, &'ll Value>) -> OperandRef<'tcx, &'ll Value> {
2018-09-14 15:48:57 +00:00
debug!("PlaceRef::load: {:?}", place);
assert_eq!(place.llextra.is_some(), place.layout.is_unsized());
if place.layout.is_zst() {
return OperandRef::new_zst(self, place.layout);
2018-09-14 15:48:57 +00:00
}
fn scalar_load_metadata<'a, 'll, 'tcx>(
bx: &mut Builder<'a, 'll, 'tcx>,
load: &'ll Value,
scalar: &abi::Scalar,
) {
2018-09-14 15:48:57 +00:00
let vr = scalar.valid_range.clone();
match scalar.value {
abi::Int(..) => {
let range = scalar.valid_range_exclusive(bx);
2018-09-14 15:48:57 +00:00
if range.start != range.end {
bx.range_metadata(load, range);
2018-09-14 15:48:57 +00:00
}
}
abi::Pointer if vr.start() < vr.end() && !vr.contains(&0) => {
bx.nonnull_metadata(load);
2018-09-14 15:48:57 +00:00
}
_ => {}
}
}
2018-09-14 15:48:57 +00:00
let val = if let Some(llextra) = place.llextra {
OperandValue::Ref(place.llval, Some(llextra), place.align)
} else if place.layout.is_llvm_immediate() {
let mut const_llval = None;
unsafe {
if let Some(global) = llvm::LLVMIsAGlobalVariable(place.llval) {
if llvm::LLVMIsGlobalConstant(global) == llvm::True {
const_llval = llvm::LLVMGetInitializer(global);
}
}
}
let llval = const_llval.unwrap_or_else(|| {
let load = self.load(place.llval, place.align);
if let abi::Abi::Scalar(ref scalar) = place.layout.abi {
scalar_load_metadata(self, load, scalar);
2018-09-14 15:48:57 +00:00
}
load
});
OperandValue::Immediate(self.to_immediate(llval, place.layout))
} else if let abi::Abi::ScalarPair(ref a, ref b) = place.layout.abi {
let b_offset = a.value.size(self).align_to(b.value.align(self).abi);
let mut load = |i, scalar: &abi::Scalar, align| {
2018-09-14 15:48:57 +00:00
let llptr = self.struct_gep(place.llval, i as u64);
let load = self.load(llptr, align);
scalar_load_metadata(self, load, scalar);
self.to_immediate_scalar(load, scalar)
2018-09-14 15:48:57 +00:00
};
OperandValue::Pair(
load(0, a, place.align),
load(1, b, place.align.restrict_for_offset(b_offset)),
)
2018-09-14 15:48:57 +00:00
} else {
OperandValue::Ref(place.llval, None, place.align)
};
OperandRef { val, layout: place.layout }
}
fn write_operand_repeatedly(
mut self,
cg_elem: OperandRef<'tcx, &'ll Value>,
count: u64,
dest: PlaceRef<'tcx, &'ll Value>,
) -> Self {
let zero = self.const_usize(0);
let count = self.const_usize(count);
let start = dest.project_index(&mut self, zero).llval;
let end = dest.project_index(&mut self, count).llval;
let mut header_bx = self.build_sibling_block("repeat_loop_header");
let mut body_bx = self.build_sibling_block("repeat_loop_body");
let next_bx = self.build_sibling_block("repeat_loop_next");
self.br(header_bx.llbb());
let current = header_bx.phi(self.val_ty(start), &[start], &[self.llbb()]);
let keep_going = header_bx.icmp(IntPredicate::IntNE, current, end);
header_bx.cond_br(keep_going, body_bx.llbb(), next_bx.llbb());
2018-09-14 15:48:57 +00:00
let align = dest.align.restrict_for_offset(dest.layout.field(self.cx(), 0).size);
2019-12-22 22:42:04 +00:00
cg_elem
.val
.store(&mut body_bx, PlaceRef::new_sized_aligned(current, cg_elem.layout, align));
let next = body_bx.inbounds_gep(current, &[self.const_usize(1)]);
body_bx.br(header_bx.llbb());
header_bx.add_incoming_to_phi(current, next, body_bx.llbb());
next_bx
}
fn range_metadata(&mut self, load: &'ll Value, range: Range<u128>) {
if self.sess().target.arch == "amdgpu" {
// amdgpu/LLVM does something weird and thinks a i64 value is
// split into a v2i32, halving the bitwidth LLVM expects,
// tripping an assertion. So, for now, just disable this
// optimization.
return;
}
unsafe {
let llty = self.cx.val_ty(load);
let v = [
self.cx.const_uint_big(llty, range.start),
2019-12-22 22:42:04 +00:00
self.cx.const_uint_big(llty, range.end),
];
2019-12-22 22:42:04 +00:00
llvm::LLVMSetMetadata(
load,
llvm::MD_range as c_uint,
llvm::LLVMMDNodeInContext(self.cx.llcx, v.as_ptr(), v.len() as c_uint),
);
}
}
fn nonnull_metadata(&mut self, load: &'ll Value) {
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMSetMetadata(
load,
llvm::MD_nonnull as c_uint,
llvm::LLVMMDNodeInContext(self.cx.llcx, ptr::null(), 0),
);
}
}
fn store(&mut self, val: &'ll Value, ptr: &'ll Value, align: Align) -> &'ll Value {
self.store_with_flags(val, ptr, align, MemFlags::empty())
}
fn store_with_flags(
&mut self,
2018-08-23 13:23:48 +00:00
val: &'ll Value,
ptr: &'ll Value,
align: Align,
flags: MemFlags,
2018-08-23 13:23:48 +00:00
) -> &'ll Value {
debug!("Store {:?} -> {:?} ({:?})", val, ptr, flags);
let ptr = self.check_store(val, ptr);
unsafe {
let store = llvm::LLVMBuildStore(self.llbuilder, val, ptr);
2019-12-22 22:42:04 +00:00
let align =
if flags.contains(MemFlags::UNALIGNED) { 1 } else { align.bytes() as c_uint };
2018-07-14 22:28:39 +00:00
llvm::LLVMSetAlignment(store, align);
if flags.contains(MemFlags::VOLATILE) {
llvm::LLVMSetVolatile(store, llvm::True);
}
if flags.contains(MemFlags::NONTEMPORAL) {
// According to LLVM [1] building a nontemporal store must
// *always* point to a metadata value of the integer 1.
//
// [1]: http://llvm.org/docs/LangRef.html#store-instruction
let one = self.cx.const_i32(1);
let node = llvm::LLVMMDNodeInContext(self.cx.llcx, &one, 1);
llvm::LLVMSetMetadata(store, llvm::MD_nontemporal as c_uint, node);
}
store
}
}
2019-12-22 22:42:04 +00:00
fn atomic_store(
&mut self,
val: &'ll Value,
ptr: &'ll Value,
order: rustc_codegen_ssa::common::AtomicOrdering,
size: Size,
) {
debug!("Store {:?} -> {:?}", val, ptr);
let ptr = self.check_store(val, ptr);
unsafe {
let store = llvm::LLVMRustBuildAtomicStore(
self.llbuilder,
val,
ptr,
AtomicOrdering::from_generic(order),
);
// LLVM requires the alignment of atomic stores to be at least the size of the type.
llvm::LLVMSetAlignment(store, size.bytes() as c_uint);
}
}
fn gep(&mut self, ptr: &'ll Value, indices: &[&'ll Value]) -> &'ll Value {
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMBuildGEP(
self.llbuilder,
ptr,
indices.as_ptr(),
indices.len() as c_uint,
UNNAMED,
)
}
}
fn inbounds_gep(&mut self, ptr: &'ll Value, indices: &[&'ll Value]) -> &'ll Value {
unsafe {
llvm::LLVMBuildInBoundsGEP(
2019-12-22 22:42:04 +00:00
self.llbuilder,
ptr,
indices.as_ptr(),
indices.len() as c_uint,
UNNAMED,
)
}
}
2018-12-04 19:20:45 +00:00
fn struct_gep(&mut self, ptr: &'ll Value, idx: u64) -> &'ll Value {
assert_eq!(idx as c_uint as u64, idx);
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildStructGEP(self.llbuilder, ptr, idx as c_uint, UNNAMED) }
2018-12-04 19:20:45 +00:00
}
/* Casts */
fn trunc(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildTrunc(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn sext(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildSExt(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn fptoui_sat(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> Option<&'ll Value> {
2020-06-27 12:03:32 +00:00
// WebAssembly has saturating floating point to integer casts if the
// `nontrapping-fptoint` target feature is activated. We'll use those if
// they are available.
if self.sess().target.arch == "wasm32"
2020-07-08 01:54:27 +00:00
&& self.sess().target_features.contains(&sym::nontrapping_dash_fptoint)
{
let src_ty = self.cx.val_ty(val);
let float_width = self.cx.float_width(src_ty);
let int_width = self.cx.int_width(dest_ty);
let name = match (int_width, float_width) {
(32, 32) => Some("llvm.wasm.trunc.saturate.unsigned.i32.f32"),
(32, 64) => Some("llvm.wasm.trunc.saturate.unsigned.i32.f64"),
(64, 32) => Some("llvm.wasm.trunc.saturate.unsigned.i64.f32"),
(64, 64) => Some("llvm.wasm.trunc.saturate.unsigned.i64.f64"),
_ => None,
};
if let Some(name) = name {
let intrinsic = self.get_intrinsic(name);
return Some(self.call(intrinsic, &[val], None));
}
}
None
}
fn fptosi_sat(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> Option<&'ll Value> {
2020-06-27 12:03:32 +00:00
// WebAssembly has saturating floating point to integer casts if the
// `nontrapping-fptoint` target feature is activated. We'll use those if
// they are available.
if self.sess().target.arch == "wasm32"
2020-07-08 01:54:27 +00:00
&& self.sess().target_features.contains(&sym::nontrapping_dash_fptoint)
{
let src_ty = self.cx.val_ty(val);
let float_width = self.cx.float_width(src_ty);
let int_width = self.cx.int_width(dest_ty);
let name = match (int_width, float_width) {
(32, 32) => Some("llvm.wasm.trunc.saturate.signed.i32.f32"),
(32, 64) => Some("llvm.wasm.trunc.saturate.signed.i32.f64"),
(64, 32) => Some("llvm.wasm.trunc.saturate.signed.i64.f32"),
(64, 64) => Some("llvm.wasm.trunc.saturate.signed.i64.f64"),
_ => None,
};
if let Some(name) = name {
let intrinsic = self.get_intrinsic(name);
return Some(self.call(intrinsic, &[val], None));
}
}
None
}
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
fn fptosui_may_trap(&self, val: &'ll Value, dest_ty: &'ll Type) -> bool {
// Most of the time we'll be generating the `fptosi` or `fptoui`
// instruction for floating-point-to-integer conversions. These
// instructions by definition in LLVM do not trap. For the WebAssembly
// target, however, we'll lower in some cases to intrinsic calls instead
// which may trap. If we detect that this is a situation where we'll be
// using the intrinsics then we report that the call map trap, which
// callers might need to handle.
if !self.wasm_and_missing_nontrapping_fptoint() {
return false;
}
let src_ty = self.cx.val_ty(val);
let float_width = self.cx.float_width(src_ty);
let int_width = self.cx.int_width(dest_ty);
2020-10-27 01:02:48 +00:00
matches!((int_width, float_width), (32, 32) | (32, 64) | (64, 32) | (64, 64))
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
}
fn fptoui(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
// When we can, use the native wasm intrinsics which have tighter
// codegen. Note that this has a semantic difference in that the
// intrinsic can trap whereas `fptoui` never traps. That difference,
// however, is handled by `fptosui_may_trap` above.
//
// Note that we skip the wasm intrinsics for vector types where `fptoui`
// must be used instead.
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
if self.wasm_and_missing_nontrapping_fptoint() {
let src_ty = self.cx.val_ty(val);
if self.cx.type_kind(src_ty) != TypeKind::Vector {
let float_width = self.cx.float_width(src_ty);
let int_width = self.cx.int_width(dest_ty);
let name = match (int_width, float_width) {
(32, 32) => Some("llvm.wasm.trunc.unsigned.i32.f32"),
(32, 64) => Some("llvm.wasm.trunc.unsigned.i32.f64"),
(64, 32) => Some("llvm.wasm.trunc.unsigned.i64.f32"),
(64, 64) => Some("llvm.wasm.trunc.unsigned.i64.f64"),
_ => None,
};
if let Some(name) = name {
let intrinsic = self.get_intrinsic(name);
return self.call(intrinsic, &[val], None);
}
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
}
}
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildFPToUI(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn fptosi(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
if self.wasm_and_missing_nontrapping_fptoint() {
let src_ty = self.cx.val_ty(val);
if self.cx.type_kind(src_ty) != TypeKind::Vector {
let float_width = self.cx.float_width(src_ty);
let int_width = self.cx.int_width(dest_ty);
let name = match (int_width, float_width) {
(32, 32) => Some("llvm.wasm.trunc.signed.i32.f32"),
(32, 64) => Some("llvm.wasm.trunc.signed.i32.f64"),
(64, 32) => Some("llvm.wasm.trunc.signed.i64.f32"),
(64, 64) => Some("llvm.wasm.trunc.signed.i64.f64"),
_ => None,
};
if let Some(name) = name {
let intrinsic = self.get_intrinsic(name);
return self.call(intrinsic, &[val], None);
}
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
}
}
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildFPToSI(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn uitofp(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildUIToFP(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn sitofp(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildSIToFP(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn fptrunc(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildFPTrunc(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn fpext(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildFPExt(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn ptrtoint(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildPtrToInt(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn inttoptr(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildIntToPtr(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn bitcast(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildBitCast(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn intcast(&mut self, val: &'ll Value, dest_ty: &'ll Type, is_signed: bool) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMRustBuildIntCast(self.llbuilder, val, dest_ty, is_signed) }
}
fn pointercast(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildPointerCast(self.llbuilder, val, dest_ty, UNNAMED) }
}
/* Comparisons */
fn icmp(&mut self, op: IntPredicate, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
let op = llvm::IntPredicate::from_generic(op);
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildICmp(self.llbuilder, op as c_uint, lhs, rhs, UNNAMED) }
}
fn fcmp(&mut self, op: RealPredicate, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildFCmp(self.llbuilder, op as c_uint, lhs, rhs, UNNAMED) }
}
/* Miscellaneous instructions */
2019-12-22 22:42:04 +00:00
fn memcpy(
&mut self,
dst: &'ll Value,
dst_align: Align,
src: &'ll Value,
src_align: Align,
size: &'ll Value,
flags: MemFlags,
) {
if flags.contains(MemFlags::NONTEMPORAL) {
// HACK(nox): This is inefficient but there is no nontemporal memcpy.
let val = self.load(src, src_align);
let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val)));
self.store_with_flags(val, ptr, dst_align, flags);
return;
}
let size = self.intcast(size, self.type_isize(), false);
let is_volatile = flags.contains(MemFlags::VOLATILE);
let dst = self.pointercast(dst, self.type_i8p());
let src = self.pointercast(src, self.type_i8p());
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMRustBuildMemCpy(
self.llbuilder,
dst,
dst_align.bytes() as c_uint,
src,
src_align.bytes() as c_uint,
size,
is_volatile,
);
}
}
2019-12-22 22:42:04 +00:00
fn memmove(
&mut self,
dst: &'ll Value,
dst_align: Align,
src: &'ll Value,
src_align: Align,
size: &'ll Value,
flags: MemFlags,
) {
if flags.contains(MemFlags::NONTEMPORAL) {
// HACK(nox): This is inefficient but there is no nontemporal memmove.
let val = self.load(src, src_align);
let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val)));
self.store_with_flags(val, ptr, dst_align, flags);
return;
}
let size = self.intcast(size, self.type_isize(), false);
let is_volatile = flags.contains(MemFlags::VOLATILE);
let dst = self.pointercast(dst, self.type_i8p());
let src = self.pointercast(src, self.type_i8p());
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMRustBuildMemMove(
self.llbuilder,
dst,
dst_align.bytes() as c_uint,
src,
src_align.bytes() as c_uint,
size,
is_volatile,
);
}
}
fn memset(
&mut self,
ptr: &'ll Value,
fill_byte: &'ll Value,
size: &'ll Value,
align: Align,
flags: MemFlags,
) {
let is_volatile = flags.contains(MemFlags::VOLATILE);
let ptr = self.pointercast(ptr, self.type_i8p());
unsafe {
llvm::LLVMRustBuildMemSet(
self.llbuilder,
ptr,
align.bytes() as c_uint,
fill_byte,
size,
is_volatile,
);
}
}
fn select(
2019-12-22 22:42:04 +00:00
&mut self,
cond: &'ll Value,
2018-08-23 13:23:48 +00:00
then_val: &'ll Value,
else_val: &'ll Value,
) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildSelect(self.llbuilder, cond, then_val, else_val, UNNAMED) }
}
2018-12-04 19:20:45 +00:00
fn va_arg(&mut self, list: &'ll Value, ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildVAArg(self.llbuilder, list, ty, UNNAMED) }
2018-12-04 19:20:45 +00:00
}
fn extract_element(&mut self, vec: &'ll Value, idx: &'ll Value) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildExtractElement(self.llbuilder, vec, idx, UNNAMED) }
}
fn vector_splat(&mut self, num_elts: usize, elt: &'ll Value) -> &'ll Value {
unsafe {
let elt_ty = self.cx.val_ty(elt);
let undef = llvm::LLVMGetUndef(self.type_vector(elt_ty, num_elts as u64));
let vec = self.insert_element(undef, elt, self.cx.const_i32(0));
let vec_i32_ty = self.type_vector(self.type_i32(), num_elts as u64);
self.shuffle_vector(vec, undef, self.const_null(vec_i32_ty))
}
}
fn extract_value(&mut self, agg_val: &'ll Value, idx: u64) -> &'ll Value {
assert_eq!(idx as c_uint as u64, idx);
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildExtractValue(self.llbuilder, agg_val, idx as c_uint, UNNAMED) }
}
2019-12-22 22:42:04 +00:00
fn insert_value(&mut self, agg_val: &'ll Value, elt: &'ll Value, idx: u64) -> &'ll Value {
assert_eq!(idx as c_uint as u64, idx);
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildInsertValue(self.llbuilder, agg_val, elt, idx as c_uint, UNNAMED) }
}
2019-12-22 22:42:04 +00:00
fn landing_pad(
&mut self,
ty: &'ll Type,
pers_fn: &'ll Value,
num_clauses: usize,
) -> &'ll Value {
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMBuildLandingPad(self.llbuilder, ty, pers_fn, num_clauses as c_uint, UNNAMED)
}
}
fn set_cleanup(&mut self, landing_pad: &'ll Value) {
unsafe {
llvm::LLVMSetCleanup(landing_pad, llvm::True);
}
}
fn resume(&mut self, exn: &'ll Value) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildResume(self.llbuilder, exn) }
}
2019-12-22 22:42:04 +00:00
fn cleanup_pad(&mut self, parent: Option<&'ll Value>, args: &[&'ll Value]) -> Funclet<'ll> {
2021-02-13 11:17:15 +00:00
let name = cstr!("cleanuppad");
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
let ret = unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMRustBuildCleanupPad(
self.llbuilder,
parent,
args.len() as c_uint,
args.as_ptr(),
name.as_ptr(),
)
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
};
Funclet::new(ret.expect("LLVM does not have support for cleanuppad"))
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
}
fn cleanup_ret(
2019-12-22 22:42:04 +00:00
&mut self,
funclet: &Funclet<'ll>,
2018-08-23 13:23:48 +00:00
unwind: Option<&'ll BasicBlock>,
) -> &'ll Value {
2019-12-22 22:42:04 +00:00
let ret =
unsafe { llvm::LLVMRustBuildCleanupRet(self.llbuilder, funclet.cleanuppad(), unwind) };
ret.expect("LLVM does not have support for cleanupret")
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
}
2019-12-22 22:42:04 +00:00
fn catch_pad(&mut self, parent: &'ll Value, args: &[&'ll Value]) -> Funclet<'ll> {
2021-02-13 11:17:15 +00:00
let name = cstr!("catchpad");
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
let ret = unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMRustBuildCatchPad(
self.llbuilder,
parent,
args.len() as c_uint,
args.as_ptr(),
name.as_ptr(),
)
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
};
Funclet::new(ret.expect("LLVM does not have support for catchpad"))
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
}
fn catch_switch(
&mut self,
2018-08-23 13:23:48 +00:00
parent: Option<&'ll Value>,
unwind: Option<&'ll BasicBlock>,
num_handlers: usize,
2018-08-23 13:23:48 +00:00
) -> &'ll Value {
2021-02-13 11:17:15 +00:00
let name = cstr!("catchswitch");
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
let ret = unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMRustBuildCatchSwitch(
self.llbuilder,
parent,
unwind,
num_handlers as c_uint,
name.as_ptr(),
)
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
};
ret.expect("LLVM does not have support for catchswitch")
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
}
fn add_handler(&mut self, catch_switch: &'ll Value, handler: &'ll BasicBlock) {
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
unsafe {
llvm::LLVMRustAddHandler(catch_switch, handler);
}
}
fn set_personality_fn(&mut self, personality: &'ll Value) {
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
unsafe {
llvm::LLVMSetPersonalityFn(self.llfn(), personality);
trans: Reimplement unwinding on MSVC This commit transitions the compiler to using the new exception handling instructions in LLVM for implementing unwinding for MSVC. This affects both 32 and 64-bit MSVC as they're both now using SEH-based strategies. In terms of standard library support, lots more details about how SEH unwinding is implemented can be found in the commits. In terms of trans, this change necessitated a few modifications: * Branches were added to detect when the old landingpad instruction is used or the new cleanuppad instruction is used to `trans::cleanup`. * The return value from `cleanuppad` is not stored in an `alloca` (because it cannot be). * Each block in trans now has an `Option<LandingPad>` instead of `is_lpad: bool` for indicating whether it's in a landing pad or not. The new exception handling intrinsics require that on MSVC each `call` inside of a landing pad is annotated with which landing pad that it's in. This change to the basic block means that whenever a `call` or `invoke` instruction is generated we know whether to annotate it as part of a cleanuppad or not. * Lots of modifications were made to the instruction builders to construct the new instructions as well as pass the tagging information for the call/invoke instructions. * The translation of the `try` intrinsics for MSVC has been overhauled to use the new `catchpad` instruction. The filter function is now also a rustc-generated function instead of a purely libstd-defined function. The libstd definition still exists, it just has a stable ABI across architectures and leaves some of the really weird implementation details to the compiler (e.g. the `localescape` and `localrecover` intrinsics).
2015-10-24 01:18:44 +00:00
}
}
// Atomic Operations
fn atomic_cmpxchg(
&mut self,
2018-08-23 13:23:48 +00:00
dst: &'ll Value,
cmp: &'ll Value,
src: &'ll Value,
order: rustc_codegen_ssa::common::AtomicOrdering,
failure_order: rustc_codegen_ssa::common::AtomicOrdering,
weak: bool,
2018-08-23 13:23:48 +00:00
) -> &'ll Value {
let weak = if weak { llvm::True } else { llvm::False };
unsafe {
llvm::LLVMRustBuildAtomicCmpXchg(
self.llbuilder,
dst,
cmp,
src,
AtomicOrdering::from_generic(order),
AtomicOrdering::from_generic(failure_order),
2019-12-22 22:42:04 +00:00
weak,
)
}
}
fn atomic_rmw(
&mut self,
op: rustc_codegen_ssa::common::AtomicRmwBinOp,
2018-08-23 13:23:48 +00:00
dst: &'ll Value,
src: &'ll Value,
order: rustc_codegen_ssa::common::AtomicOrdering,
2018-08-23 13:23:48 +00:00
) -> &'ll Value {
unsafe {
llvm::LLVMBuildAtomicRMW(
self.llbuilder,
AtomicRmwBinOp::from_generic(op),
dst,
src,
AtomicOrdering::from_generic(order),
2019-12-22 22:42:04 +00:00
False,
)
}
}
2013-07-28 07:48:16 +00:00
2018-09-28 10:18:03 +00:00
fn atomic_fence(
&mut self,
order: rustc_codegen_ssa::common::AtomicOrdering,
2019-12-22 22:42:04 +00:00
scope: rustc_codegen_ssa::common::SynchronizationScope,
2018-09-28 10:18:03 +00:00
) {
2013-07-28 07:48:16 +00:00
unsafe {
llvm::LLVMRustBuildAtomicFence(
self.llbuilder,
AtomicOrdering::from_generic(order),
2019-12-22 22:42:04 +00:00
SynchronizationScope::from_generic(scope),
);
2013-07-28 07:48:16 +00:00
}
}
fn set_invariant_load(&mut self, load: &'ll Value) {
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMSetMetadata(
load,
llvm::MD_invariant_load as c_uint,
llvm::LLVMMDNodeInContext(self.cx.llcx, ptr::null(), 0),
);
}
}
fn lifetime_start(&mut self, ptr: &'ll Value, size: Size) {
self.call_lifetime_intrinsic("llvm.lifetime.start.p0i8", ptr, size);
}
fn lifetime_end(&mut self, ptr: &'ll Value, size: Size) {
self.call_lifetime_intrinsic("llvm.lifetime.end.p0i8", ptr, size);
}
fn instrprof_increment(
&mut self,
fn_name: &'ll Value,
hash: &'ll Value,
num_counters: &'ll Value,
index: &'ll Value,
) {
debug!(
"instrprof_increment() with args ({:?}, {:?}, {:?}, {:?})",
fn_name, hash, num_counters, index
);
let llfn = unsafe { llvm::LLVMRustGetInstrProfIncrementIntrinsic(self.cx().llmod) };
let args = &[fn_name, hash, num_counters, index];
let args = self.check_call("call", llfn, args);
unsafe {
let _ = llvm::LLVMRustBuildCall(
self.llbuilder,
llfn,
args.as_ptr() as *const &llvm::Value,
args.len() as c_uint,
None,
);
}
}
fn call(
&mut self,
llfn: &'ll Value,
args: &[&'ll Value],
funclet: Option<&Funclet<'ll>>,
) -> &'ll Value {
2019-12-22 22:42:04 +00:00
debug!("call {:?} with args ({:?})", llfn, args);
let args = self.check_call("call", llfn, args);
let bundle = funclet.map(|funclet| funclet.bundle());
let bundle = bundle.as_ref().map(|b| &*b.raw);
unsafe {
llvm::LLVMRustBuildCall(
self.llbuilder,
llfn,
args.as_ptr() as *const &llvm::Value,
args.len() as c_uint,
2019-12-22 22:42:04 +00:00
bundle,
)
}
}
fn zext(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildZExt(self.llbuilder, val, dest_ty, UNNAMED) }
}
fn cx(&self) -> &CodegenCx<'ll, 'tcx> {
2018-11-07 10:08:41 +00:00
self.cx
}
unsafe fn delete_basic_block(&mut self, bb: &'ll BasicBlock) {
llvm::LLVMDeleteBasicBlock(bb);
}
fn do_not_inline(&mut self, llret: &'ll Value) {
llvm::Attribute::NoInline.apply_callsite(llvm::AttributePlace::Function, llret);
}
}
impl StaticBuilderMethods for Builder<'a, 'll, 'tcx> {
fn get_static(&mut self, def_id: DefId) -> &'ll Value {
2019-03-01 14:05:18 +00:00
// Forward to the `get_static` method of `CodegenCx`
self.cx().get_static(def_id)
}
}
impl Builder<'a, 'll, 'tcx> {
2018-12-04 19:20:45 +00:00
pub fn llfn(&self) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMGetBasicBlockParent(self.llbb()) }
2018-12-04 19:20:45 +00:00
}
fn position_at_start(&mut self, llbb: &'ll BasicBlock) {
unsafe {
llvm::LLVMRustPositionBuilderAtStart(self.llbuilder, llbb);
}
}
pub fn minnum(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildMinNum(self.llbuilder, lhs, rhs) }
}
pub fn maxnum(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildMaxNum(self.llbuilder, lhs, rhs) }
}
pub fn insert_element(
2019-12-22 22:42:04 +00:00
&mut self,
vec: &'ll Value,
elt: &'ll Value,
idx: &'ll Value,
) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildInsertElement(self.llbuilder, vec, elt, idx, UNNAMED) }
}
pub fn shuffle_vector(
&mut self,
v1: &'ll Value,
v2: &'ll Value,
mask: &'ll Value,
) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildShuffleVector(self.llbuilder, v1, v2, mask, UNNAMED) }
}
pub fn vector_reduce_fadd(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceFAdd(self.llbuilder, acc, src) }
}
pub fn vector_reduce_fmul(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceFMul(self.llbuilder, acc, src) }
}
pub fn vector_reduce_fadd_fast(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMRustBuildVectorReduceFAdd(self.llbuilder, acc, src);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
pub fn vector_reduce_fmul_fast(&mut self, acc: &'ll Value, src: &'ll Value) -> &'ll Value {
unsafe {
let instr = llvm::LLVMRustBuildVectorReduceFMul(self.llbuilder, acc, src);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
pub fn vector_reduce_add(&mut self, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceAdd(self.llbuilder, src) }
}
pub fn vector_reduce_mul(&mut self, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceMul(self.llbuilder, src) }
}
pub fn vector_reduce_and(&mut self, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceAnd(self.llbuilder, src) }
}
pub fn vector_reduce_or(&mut self, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceOr(self.llbuilder, src) }
}
pub fn vector_reduce_xor(&mut self, src: &'ll Value) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceXor(self.llbuilder, src) }
}
pub fn vector_reduce_fmin(&mut self, src: &'ll Value) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe {
llvm::LLVMRustBuildVectorReduceFMin(self.llbuilder, src, /*NoNaNs:*/ false)
}
}
pub fn vector_reduce_fmax(&mut self, src: &'ll Value) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe {
llvm::LLVMRustBuildVectorReduceFMax(self.llbuilder, src, /*NoNaNs:*/ false)
}
}
pub fn vector_reduce_fmin_fast(&mut self, src: &'ll Value) -> &'ll Value {
unsafe {
2019-12-22 22:42:04 +00:00
let instr =
llvm::LLVMRustBuildVectorReduceFMin(self.llbuilder, src, /*NoNaNs:*/ true);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
pub fn vector_reduce_fmax_fast(&mut self, src: &'ll Value) -> &'ll Value {
unsafe {
2019-12-22 22:42:04 +00:00
let instr =
llvm::LLVMRustBuildVectorReduceFMax(self.llbuilder, src, /*NoNaNs:*/ true);
llvm::LLVMRustSetHasUnsafeAlgebra(instr);
instr
}
}
pub fn vector_reduce_min(&mut self, src: &'ll Value, is_signed: bool) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceMin(self.llbuilder, src, is_signed) }
}
pub fn vector_reduce_max(&mut self, src: &'ll Value, is_signed: bool) -> &'ll Value {
unsafe { llvm::LLVMRustBuildVectorReduceMax(self.llbuilder, src, is_signed) }
}
pub fn add_clause(&mut self, landing_pad: &'ll Value, clause: &'ll Value) {
unsafe {
llvm::LLVMAddClause(landing_pad, clause);
}
}
pub fn catch_ret(&mut self, funclet: &Funclet<'ll>, unwind: &'ll BasicBlock) -> &'ll Value {
2019-12-22 22:42:04 +00:00
let ret =
unsafe { llvm::LLVMRustBuildCatchRet(self.llbuilder, funclet.cleanuppad(), unwind) };
ret.expect("LLVM does not have support for catchret")
}
fn check_store(&mut self, val: &'ll Value, ptr: &'ll Value) -> &'ll Value {
let dest_ptr_ty = self.cx.val_ty(ptr);
let stored_ty = self.cx.val_ty(val);
let stored_ptr_ty = self.cx.type_ptr_to(stored_ty);
assert_eq!(self.cx.type_kind(dest_ptr_ty), TypeKind::Pointer);
if dest_ptr_ty == stored_ptr_ty {
ptr
} else {
2019-12-22 22:42:04 +00:00
debug!(
"type mismatch in store. \
Expected {:?}, got {:?}; inserting bitcast",
2019-12-22 22:42:04 +00:00
dest_ptr_ty, stored_ptr_ty
);
self.bitcast(ptr, stored_ptr_ty)
}
}
2019-12-22 22:42:04 +00:00
fn check_call<'b>(
&mut self,
typ: &str,
llfn: &'ll Value,
args: &'b [&'ll Value],
) -> Cow<'b, [&'ll Value]> {
let mut fn_ty = self.cx.val_ty(llfn);
// Strip off pointers
while self.cx.type_kind(fn_ty) == TypeKind::Pointer {
fn_ty = self.cx.element_type(fn_ty);
}
2019-12-22 22:42:04 +00:00
assert!(
self.cx.type_kind(fn_ty) == TypeKind::Function,
"builder::{} not passed a function, but {:?}",
typ,
fn_ty
);
let param_tys = self.cx.func_params_types(fn_ty);
2019-12-22 22:42:04 +00:00
let all_args_match = param_tys
.iter()
.zip(args.iter().map(|&v| self.val_ty(v)))
.all(|(expected_ty, actual_ty)| *expected_ty == actual_ty);
if all_args_match {
return Cow::Borrowed(args);
}
2019-12-22 22:42:04 +00:00
let casted_args: Vec<_> = param_tys
.into_iter()
.zip(args.iter())
.enumerate()
.map(|(i, (expected_ty, &actual_val))| {
let actual_ty = self.val_ty(actual_val);
if expected_ty != actual_ty {
2019-12-22 22:42:04 +00:00
debug!(
"type mismatch in function call of {:?}. \
Expected {:?} for param {}, got {:?}; injecting bitcast",
2019-12-22 22:42:04 +00:00
llfn, expected_ty, i, actual_ty
);
self.bitcast(actual_val, expected_ty)
} else {
actual_val
}
})
.collect();
Cow::Owned(casted_args)
}
pub fn va_arg(&mut self, list: &'ll Value, ty: &'ll Type) -> &'ll Value {
2019-12-22 22:42:04 +00:00
unsafe { llvm::LLVMBuildVAArg(self.llbuilder, list, ty, UNNAMED) }
}
fn call_lifetime_intrinsic(&mut self, intrinsic: &str, ptr: &'ll Value, size: Size) {
let size = size.bytes();
if size == 0 {
return;
}
if !self.cx().sess().emit_lifetime_markers() {
return;
}
let lifetime_intrinsic = self.cx.get_intrinsic(intrinsic);
let ptr = self.pointercast(ptr, self.cx.type_i8p());
self.call(lifetime_intrinsic, &[self.cx.const_u64(size), ptr], None);
}
pub(crate) fn phi(
&mut self,
ty: &'ll Type,
vals: &[&'ll Value],
bbs: &[&'ll BasicBlock],
) -> &'ll Value {
assert_eq!(vals.len(), bbs.len());
2019-12-22 22:42:04 +00:00
let phi = unsafe { llvm::LLVMBuildPhi(self.llbuilder, ty, UNNAMED) };
unsafe {
2019-12-22 22:42:04 +00:00
llvm::LLVMAddIncoming(phi, vals.as_ptr(), bbs.as_ptr(), vals.len() as c_uint);
phi
}
}
fn add_incoming_to_phi(&mut self, phi: &'ll Value, val: &'ll Value, bb: &'ll BasicBlock) {
unsafe {
llvm::LLVMAddIncoming(phi, &val, &bb, 1 as c_uint);
}
}
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
fn wasm_and_missing_nontrapping_fptoint(&self) -> bool {
self.sess().target.arch == "wasm32"
rustc: Improving safe wasm float->int casts This commit improves code generation for WebAssembly targets when translating floating to integer casts. This improvement is only relevant when the `nontrapping-fptoint` feature is not enabled, but the feature is not enabled by default right now. Additionally this improvement only affects safe casts since unchecked casts were improved in #74659. Some more background for this issue is present on #73591, but the general gist of the issue is that in LLVM the `fptosi` and `fptoui` instructions are defined to return an `undef` value if they execute on out-of-bounds values; they notably do not trap. To implement these instructions for WebAssembly the LLVM backend must therefore generate quite a few instructions before executing `i32.trunc_f32_s` (for example) because this WebAssembly instruction traps on out-of-bounds values. This codegen into wasm instructions happens very late in the code generator, so what ends up happening is that rustc inserts its own codegen to implement Rust's saturating semantics, and then LLVM also inserts its own codegen to make sure that the `fptosi` instruction doesn't trap. Overall this means that a function like this: #[no_mangle] pub unsafe extern "C" fn cast(x: f64) -> u32 { x as u32 } will generate this WebAssembly today: (func $cast (type 0) (param f64) (result i32) (local i32 i32) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.gt local.set 1 block ;; label = @1 block ;; label = @2 local.get 0 f64.const 0x0p+0 (;=0;) local.get 0 f64.const 0x0p+0 (;=0;) f64.gt select local.tee 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@2;) local.get 0 i32.trunc_f64_u local.set 2 br 1 (;@1;) end i32.const 0 local.set 2 end i32.const -1 local.get 2 local.get 1 select) This PR improves the situation by updating the code generation for float-to-int conversions in rustc, specifically only for WebAssembly targets and only for some situations (float-to-u8 still has not great codegen). The fix here is to use basic blocks and control flow to avoid speculatively executing `fptosi`, and instead LLVM's raw intrinsic for the WebAssembly instruction is used instead. This effectively extends the support added in #74659 to checked casts. After this commit the codegen for the above Rust function looks like: (func $cast (type 0) (param f64) (result i32) (local i32) block ;; label = @1 local.get 0 f64.const 0x0p+0 (;=0;) f64.ge local.tee 1 i32.const 1 i32.xor br_if 0 (;@1;) local.get 0 f64.const 0x1.fffffffep+31 (;=4.29497e+09;) f64.le i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const -1 i32.const 0 local.get 1 select) For reference, in Rust 1.44, which did not have saturating float-to-integer casts, the codegen LLVM would emit is: (func $cast (type 0) (param f64) (result i32) block ;; label = @1 local.get 0 f64.const 0x1p+32 (;=4.29497e+09;) f64.lt local.get 0 f64.const 0x0p+0 (;=0;) f64.ge i32.and i32.eqz br_if 0 (;@1;) local.get 0 i32.trunc_f64_u return end i32.const 0) So we're relatively close to the original codegen, although it's slightly different because the semantics of the function changed where we're emulating the `i32.trunc_sat_f32_s` instruction rather than always replacing out-of-bounds values with zero. There is still work that could be done to improve casts such as `f32` to `u8`. That form of cast still uses the `fptosi` instruction which generates lots of branch-y code. This seems less important to tackle now though. In the meantime this should take care of most use cases of floating-point conversion and as a result I'm going to speculate that this... Closes #73591
2020-07-23 19:20:42 +00:00
&& !self.sess().target_features.contains(&sym::nontrapping_dash_fptoint)
}
}