Google Git Google Git
Sign in
rust/rust/32ce99c38ee6a9d4a397e97a9017d69052d63d0f/./compiler/rustc_codegen_llvm/src/intrinsic.rs
blob: 0d11f1e4fb69b2ea2d8e5e65f7b78c31d43d3108 [file]
use std::cmp::Ordering;
use std::ffi::c_uint;
use std::{assert_matches, iter, ptr};
use rustc_abi::{
AddressSpace, Align, BackendRepr, CVariadicStatus, Float, HasDataLayout, NumScalableVectors,
Primitive, Size, WrappingRange,
};
use rustc_codegen_ssa::RetagInfo;
use rustc_codegen_ssa::base::{compare_simd_types, wants_msvc_seh, wants_wasm_eh};
use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
use rustc_codegen_ssa::errors::{ExpectedPointerMutability, InvalidMonomorphization};
use rustc_codegen_ssa::mir::IntrinsicResult;
use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
use rustc_codegen_ssa::mir::place::{PlaceRef, PlaceValue};
use rustc_codegen_ssa::traits::*;
use rustc_hir as hir;
use rustc_hir::def_id::LOCAL_CRATE;
use rustc_hir::find_attr;
use rustc_middle::mir::BinOp;
use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt, HasTypingEnv, LayoutOf};
use rustc_middle::ty::offload_meta::OffloadMetadata;
use rustc_middle::ty::{self, GenericArgsRef, Instance, SimdAlign, Ty, TyCtxt, TypingEnv};
use rustc_middle::{bug, span_bug};
use rustc_session::config::CrateType;
use rustc_session::diagnostics::feature_err;
use rustc_session::lint::builtin::DEPRECATED_LLVM_INTRINSIC;
use rustc_span::{ErrorGuaranteed, Span, Symbol, sym};
use rustc_symbol_mangling::{mangle_internal_symbol, symbol_name_for_instance_in_crate};
use rustc_target::callconv::PassMode;
use rustc_target::spec::Arch;
use tracing::debug;
use crate::abi::FnAbiLlvmExt;
use crate::builder::Builder;
use crate::builder::autodiff::{adjust_activity_to_abi, generate_enzyme_call};
use crate::builder::gpu_offload::{
OffloadKernelDims, gen_call_handling, gen_define_handling, register_offload,
};
use crate::context::CodegenCx;
use crate::declare::declare_raw_fn;
use crate::errors::{
AutoDiffWithoutEnable, AutoDiffWithoutLto, IntrinsicSignatureMismatch, IntrinsicWrongArch,
OffloadWithoutEnable, OffloadWithoutFatLTO, UnknownIntrinsic,
};
use crate::intrinsic::ty::typetree::fnc_typetrees;
use crate::llvm::{self, Attribute, AttributePlace, Type, Value};
use crate::type_of::LayoutLlvmExt;
use crate::va_arg::emit_va_arg;
fn call_simple_intrinsic<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
name: Symbol,
args: &[OperandRef<'tcx, &'ll Value>],
) -> Option<&'ll Value> {
let (base_name, type_params): (&'static str, &[&'ll Type]) = match name {
sym::sqrtf16 => ("llvm.sqrt", &[bx.type_f16()]),
sym::sqrtf32 => ("llvm.sqrt", &[bx.type_f32()]),
sym::sqrtf64 => ("llvm.sqrt", &[bx.type_f64()]),
sym::sqrtf128 => ("llvm.sqrt", &[bx.type_f128()]),
sym::powif16 => ("llvm.powi", &[bx.type_f16(), bx.type_i32()]),
sym::powif32 => ("llvm.powi", &[bx.type_f32(), bx.type_i32()]),
sym::powif64 => ("llvm.powi", &[bx.type_f64(), bx.type_i32()]),
sym::powif128 => ("llvm.powi", &[bx.type_f128(), bx.type_i32()]),
sym::sinf16 => ("llvm.sin", &[bx.type_f16()]),
sym::sinf32 => ("llvm.sin", &[bx.type_f32()]),
sym::sinf64 => ("llvm.sin", &[bx.type_f64()]),
sym::sinf128 => ("llvm.sin", &[bx.type_f128()]),
sym::cosf16 => ("llvm.cos", &[bx.type_f16()]),
sym::cosf32 => ("llvm.cos", &[bx.type_f32()]),
sym::cosf64 => ("llvm.cos", &[bx.type_f64()]),
sym::cosf128 => ("llvm.cos", &[bx.type_f128()]),
sym::powf16 => ("llvm.pow", &[bx.type_f16()]),
sym::powf32 => ("llvm.pow", &[bx.type_f32()]),
sym::powf64 => ("llvm.pow", &[bx.type_f64()]),
sym::powf128 => ("llvm.pow", &[bx.type_f128()]),
sym::expf16 => ("llvm.exp", &[bx.type_f16()]),
sym::expf32 => ("llvm.exp", &[bx.type_f32()]),
sym::expf64 => ("llvm.exp", &[bx.type_f64()]),
sym::expf128 => ("llvm.exp", &[bx.type_f128()]),
sym::exp2f16 => ("llvm.exp2", &[bx.type_f16()]),
sym::exp2f32 => ("llvm.exp2", &[bx.type_f32()]),
sym::exp2f64 => ("llvm.exp2", &[bx.type_f64()]),
sym::exp2f128 => ("llvm.exp2", &[bx.type_f128()]),
sym::logf16 => ("llvm.log", &[bx.type_f16()]),
sym::logf32 => ("llvm.log", &[bx.type_f32()]),
sym::logf64 => ("llvm.log", &[bx.type_f64()]),
sym::logf128 => ("llvm.log", &[bx.type_f128()]),
sym::log10f16 => ("llvm.log10", &[bx.type_f16()]),
sym::log10f32 => ("llvm.log10", &[bx.type_f32()]),
sym::log10f64 => ("llvm.log10", &[bx.type_f64()]),
sym::log10f128 => ("llvm.log10", &[bx.type_f128()]),
sym::log2f16 => ("llvm.log2", &[bx.type_f16()]),
sym::log2f32 => ("llvm.log2", &[bx.type_f32()]),
sym::log2f64 => ("llvm.log2", &[bx.type_f64()]),
sym::log2f128 => ("llvm.log2", &[bx.type_f128()]),
sym::fmaf16 => ("llvm.fma", &[bx.type_f16()]),
sym::fmaf32 => ("llvm.fma", &[bx.type_f32()]),
sym::fmaf64 => ("llvm.fma", &[bx.type_f64()]),
sym::fmaf128 => ("llvm.fma", &[bx.type_f128()]),
sym::fmuladdf16 => ("llvm.fmuladd", &[bx.type_f16()]),
sym::fmuladdf32 => ("llvm.fmuladd", &[bx.type_f32()]),
sym::fmuladdf64 => ("llvm.fmuladd", &[bx.type_f64()]),
sym::fmuladdf128 => ("llvm.fmuladd", &[bx.type_f128()]),
sym::minimumf16 => ("llvm.minimum", &[bx.type_f16()]),
sym::minimumf32 => ("llvm.minimum", &[bx.type_f32()]),
// FIXME: LLVM currently mis-compile those intrinsics, re-enable them
// when llvm/llvm-project#{139380,139381,140445} are fixed.
//sym::minimumf64 => ("llvm.minimum", &[bx.type_f64()]),
//sym::minimumf128 => ("llvm.minimum", &[cx.type_f128()]),
//
sym::maximumf16 => ("llvm.maximum", &[bx.type_f16()]),
sym::maximumf32 => ("llvm.maximum", &[bx.type_f32()]),
// FIXME: LLVM currently mis-compile those intrinsics, re-enable them
// when llvm/llvm-project#{139380,139381,140445} are fixed.
//sym::maximumf64 => ("llvm.maximum", &[bx.type_f64()]),
//sym::maximumf128 => ("llvm.maximum", &[cx.type_f128()]),
//
sym::copysignf16 => ("llvm.copysign", &[bx.type_f16()]),
sym::copysignf32 => ("llvm.copysign", &[bx.type_f32()]),
sym::copysignf64 => ("llvm.copysign", &[bx.type_f64()]),
sym::copysignf128 => ("llvm.copysign", &[bx.type_f128()]),
sym::floorf16 => ("llvm.floor", &[bx.type_f16()]),
sym::floorf32 => ("llvm.floor", &[bx.type_f32()]),
sym::floorf64 => ("llvm.floor", &[bx.type_f64()]),
sym::floorf128 => ("llvm.floor", &[bx.type_f128()]),
sym::ceilf16 => ("llvm.ceil", &[bx.type_f16()]),
sym::ceilf32 => ("llvm.ceil", &[bx.type_f32()]),
sym::ceilf64 => ("llvm.ceil", &[bx.type_f64()]),
sym::ceilf128 => ("llvm.ceil", &[bx.type_f128()]),
sym::truncf16 => ("llvm.trunc", &[bx.type_f16()]),
sym::truncf32 => ("llvm.trunc", &[bx.type_f32()]),
sym::truncf64 => ("llvm.trunc", &[bx.type_f64()]),
sym::truncf128 => ("llvm.trunc", &[bx.type_f128()]),
// We could use any of `rint`, `nearbyint`, or `roundeven`
// for this -- they are all identical in semantics when
// assuming the default FP environment.
// `rint` is what we used for $forever.
sym::round_ties_even_f16 => ("llvm.rint", &[bx.type_f16()]),
sym::round_ties_even_f32 => ("llvm.rint", &[bx.type_f32()]),
sym::round_ties_even_f64 => ("llvm.rint", &[bx.type_f64()]),
sym::round_ties_even_f128 => ("llvm.rint", &[bx.type_f128()]),
sym::roundf16 => ("llvm.round", &[bx.type_f16()]),
sym::roundf32 => ("llvm.round", &[bx.type_f32()]),
sym::roundf64 => ("llvm.round", &[bx.type_f64()]),
sym::roundf128 => ("llvm.round", &[bx.type_f128()]),
_ => return None,
};
Some(bx.call_intrinsic(
base_name,
type_params,
&args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
))
}
impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
fn codegen_intrinsic_call(
&mut self,
instance: ty::Instance<'tcx>,
args: &[OperandRef<'tcx, &'ll Value>],
result_layout: ty::layout::TyAndLayout<'tcx>,
result_place: Option<PlaceValue<&'ll Value>>,
span: Span,
) -> IntrinsicResult<'tcx, &'ll Value> {
let tcx = self.tcx;
let llvm_version = crate::llvm_util::get_version();
let name = tcx.item_name(instance.def_id());
let fn_args = instance.args;
let simple = call_simple_intrinsic(self, name, args);
let llval = match name {
_ if simple.is_some() => simple.unwrap(),
sym::minimum_number_nsz_f16
| sym::minimum_number_nsz_f32
| sym::minimum_number_nsz_f64
| sym::minimum_number_nsz_f128
| sym::maximum_number_nsz_f16
| sym::maximum_number_nsz_f32
| sym::maximum_number_nsz_f64
| sym::maximum_number_nsz_f128
// Need at least LLVM 22 for `min/maximumnum` to not crash LLVM.
if llvm_version >= (22, 0, 0) =>
{
let intrinsic_name = if name.as_str().starts_with("min") {
"llvm.minimumnum"
} else {
"llvm.maximumnum"
};
let call = self.call_intrinsic(
intrinsic_name,
&[args[0].layout.immediate_llvm_type(self.cx)],
&[args[0].immediate(), args[1].immediate()],
);
// `nsz` on minimumnum/maximumnum is special: its only effect is to make
// signed-zero ordering non-deterministic.
unsafe { llvm::LLVMRustSetNoSignedZeros(call) };
call
}
sym::ptr_mask => {
let ptr = args[0].immediate();
self.call_intrinsic(
"llvm.ptrmask",
&[self.val_ty(ptr), self.type_isize()],
&[ptr, args[1].immediate()],
)
}
sym::autodiff => {
return codegen_autodiff(self, instance, args, result_layout, result_place);
}
sym::offload => {
if tcx.sess.opts.unstable_opts.offload.is_empty() {
let _ = tcx.dcx().emit_almost_fatal(OffloadWithoutEnable);
}
if tcx.sess.lto() != rustc_session::config::Lto::Fat {
let _ = tcx.dcx().emit_almost_fatal(OffloadWithoutFatLTO);
}
codegen_offload(self, tcx, instance, args);
// offload *has* a return type, but somehow works without mentioning the place
return IntrinsicResult::WroteIntoPlace;
}
sym::is_val_statically_known => {
if let OperandValue::Immediate(imm) = args[0].val {
self.call_intrinsic(
"llvm.is.constant",
&[args[0].layout.immediate_llvm_type(self.cx)],
&[imm],
)
} else {
self.const_bool(false)
}
}
sym::select_unpredictable => {
let cond = args[0].immediate();
assert_eq!(args[1].layout, args[2].layout);
let select = |bx: &mut Self, true_val, false_val| {
let result = bx.select(cond, true_val, false_val);
bx.set_unpredictable(&result);
result
};
match (args[1].val, args[2].val) {
(OperandValue::Ref(true_val), OperandValue::Ref(false_val)) => {
assert!(true_val.llextra.is_none());
assert!(false_val.llextra.is_none());
assert_eq!(true_val.align, false_val.align);
let ptr = select(self, true_val.llval, false_val.llval);
let selected =
OperandValue::Ref(PlaceValue::new_sized(ptr, true_val.align));
let result = PlaceRef {
val: result_place.unwrap(),
layout: result_layout,
};
selected.store(self, result);
return IntrinsicResult::WroteIntoPlace;
}
(OperandValue::Immediate(_), OperandValue::Immediate(_))
| (OperandValue::Pair(_, _), OperandValue::Pair(_, _)) => {
let true_val = args[1].immediate_or_packed_pair(self);
let false_val = args[2].immediate_or_packed_pair(self);
select(self, true_val, false_val)
}
(OperandValue::ZeroSized, OperandValue::ZeroSized) => return IntrinsicResult::Operand(OperandValue::ZeroSized),
_ => span_bug!(span, "Incompatible OperandValue for select_unpredictable"),
}
}
sym::catch_unwind => {
catch_unwind_intrinsic(
self,
args[0].immediate(),
args[1].immediate(),
args[2].immediate(),
)
}
sym::breakpoint => self.call_intrinsic("llvm.debugtrap", &[], &[]),
sym::va_arg => {
let target = &self.cx.tcx.sess.target;
let stability = target.supports_c_variadic_definitions();
if let CVariadicStatus::Unstable { feature } = stability
&& !self.tcx.features().enabled(feature)
{
let msg =
format!("C-variadic function definitions on this target are unstable");
feature_err(&*self.sess(), feature, span, msg).emit();
}
let BackendRepr::Scalar(scalar) = result_layout.backend_repr else {
bug!("the va_arg intrinsic does not support non-scalar types")
};
// We reject types that would never be passed as varargs in C because
// they get promoted to a larger type, specifically integers smaller than
// c_int and float type smaller than c_double.
match scalar.primitive() {
Primitive::Pointer(_) => {
// Pointers are always OK.
}
Primitive::Int(..) => {
let int_width = self.cx().size_of(result_layout.ty).bits();
let target_c_int_width = self.cx().sess().target.options.c_int_width;
if int_width < u64::from(target_c_int_width) {
// Smaller integer types are automatically promototed and `va_arg`
// should not be called on them.
bug!(
"va_arg got i{} but needs at least c_int (an i{})",
int_width,
target_c_int_width
);
}
}
Primitive::Float(Float::F16) => {
bug!("the va_arg intrinsic does not support `f16`")
}
Primitive::Float(Float::F32) => {
// c_double is actually f32 on avr.
if self.cx().sess().target.arch != Arch::Avr {
bug!("the va_arg intrinsic does not support `f32` on this target")
}
}
Primitive::Float(Float::F64) => {
// 64-bit floats are always OK.
}
Primitive::Float(Float::F128) => {
// FIXME(f128) figure out whether we should support this.
bug!("the va_arg intrinsic does not support `f128`")
}
}
emit_va_arg(self, args[0], result_layout.ty)
}
sym::volatile_load | sym::unaligned_volatile_load => {
// Note that we cannot just load the `llvm_type` because we should never load non-scalars.
// Trying to do so blows up horribly in some cases -- for example loading a
// `MaybeUninint<&dyn Trait>` would load as `{ [i64x2] }` which gives assertions later
// (if we're lucky) from things not being pointers that ought to be.
let ptr = args[0].immediate();
let abi_align = result_layout.align.abi;
let ptr_align = if name == sym::volatile_load { abi_align } else { Align::ONE };
if result_layout.is_zst() {
return IntrinsicResult::Operand(OperandValue::ZeroSized);
} else if let BackendRepr::Scalar(scalar) = result_layout.backend_repr {
let load = self.volatile_load(self.type_from_scalar(scalar), ptr, ptr_align);
self.to_immediate_scalar(load, scalar)
} else {
// One day Rust will probably want to define how we split up a volatile load
// of something that's *not* just an ordinary scalar, but for now we can just
// use an LLVM integer type of the correct width and let it split it however.
let llty = self.type_ix(result_layout.size.bits());
let temp = if let Some(result_place) = result_place {
PlaceRef {
val: result_place,
layout: result_layout,
}
} else {
PlaceRef::alloca(self, result_layout)
};
let llval = self.volatile_load(llty, ptr, ptr_align);
self.store(llval, temp.val.llval, abi_align);
return if result_place.is_none() {
IntrinsicResult::Operand(self.load_operand(temp).val)
} else {
IntrinsicResult::WroteIntoPlace
};
}
}
sym::prefetch_read_data
| sym::prefetch_write_data
| sym::prefetch_read_instruction
| sym::prefetch_write_instruction => {
let (rw, cache_type) = match name {
sym::prefetch_read_data => (0, 1),
sym::prefetch_write_data => (1, 1),
sym::prefetch_read_instruction => (0, 0),
sym::prefetch_write_instruction => (1, 0),
_ => bug!(),
};
let ptr = args[0].immediate();
let locality = fn_args.const_at(1).to_leaf().to_i32();
self.call_intrinsic(
"llvm.prefetch.p0",
&[self.val_ty(ptr)],
&[
ptr,
self.const_i32(rw),
self.const_i32(locality),
self.const_i32(cache_type),
],
);
return IntrinsicResult::Operand(OperandValue::ZeroSized);
}
sym::carrying_mul_add => {
let (size, signed) = fn_args.type_at(0).int_size_and_signed(self.tcx);
let wide_llty = self.type_ix(size.bits() * 2);
let args = args.as_array().unwrap();
let [a, b, c, d] = args.map(|a| self.intcast(a.immediate(), wide_llty, signed));
let wide = if signed {
let prod = self.unchecked_smul(a, b);
let acc = self.unchecked_sadd(prod, c);
self.unchecked_sadd(acc, d)
} else {
let prod = self.unchecked_umul(a, b);
let acc = self.unchecked_uadd(prod, c);
self.unchecked_uadd(acc, d)
};
let narrow_llty = self.type_ix(size.bits());
let low = self.trunc(wide, narrow_llty);
let bits_const = self.const_uint(wide_llty, size.bits());
// No need for ashr when signed; LLVM changes it to lshr anyway.
let high = self.lshr(wide, bits_const);
// FIXME: could be `trunc nuw`, even for signed.
let high = self.trunc(high, narrow_llty);
let pair_llty = self.type_struct(&[narrow_llty, narrow_llty], false);
let pair = self.const_poison(pair_llty);
let pair = self.insert_value(pair, low, 0);
let pair = self.insert_value(pair, high, 1);
pair
}
// FIXME move into the branch below when LLVM 22 is the lowest version we support.
sym::carryless_mul if llvm_version >= (22, 0, 0) => {
let ty = args[0].layout.ty;
if !ty.is_integral() {
let err = tcx.dcx().emit_err(InvalidMonomorphization::BasicIntegerType {
span,
name,
ty,
});
return IntrinsicResult::Err(err);
}
let (size, _) = ty.int_size_and_signed(self.tcx);
let width = size.bits();
let llty = self.type_ix(width);
let lhs = args[0].immediate();
let rhs = args[1].immediate();
self.call_intrinsic("llvm.clmul", &[llty], &[lhs, rhs])
}
sym::ctlz
| sym::ctlz_nonzero
| sym::cttz
| sym::cttz_nonzero
| sym::ctpop
| sym::bswap
| sym::bitreverse
| sym::saturating_add
| sym::saturating_sub
| sym::unchecked_funnel_shl
| sym::unchecked_funnel_shr => {
let ty = args[0].layout.ty;
if !ty.is_integral() {
let err = tcx.dcx().emit_err(InvalidMonomorphization::BasicIntegerType {
span,
name,
ty,
});
return IntrinsicResult::Err(err);
}
let (size, signed) = ty.int_size_and_signed(self.tcx);
let width = size.bits();
let llty = self.type_ix(width);
match name {
sym::ctlz | sym::ctlz_nonzero | sym::cttz | sym::cttz_nonzero => {
let y =
self.const_bool(name == sym::ctlz_nonzero || name == sym::cttz_nonzero);
let llvm_name = if name == sym::ctlz || name == sym::ctlz_nonzero {
"llvm.ctlz"
} else {
"llvm.cttz"
};
let ret =
self.call_intrinsic(llvm_name, &[llty], &[args[0].immediate(), y]);
self.intcast(ret, result_layout.llvm_type(self), false)
}
sym::ctpop => {
let ret =
self.call_intrinsic("llvm.ctpop", &[llty], &[args[0].immediate()]);
self.intcast(ret, result_layout.llvm_type(self), false)
}
sym::bswap => {
if width == 8 {
args[0].immediate() // byte swap a u8/i8 is just a no-op
} else {
self.call_intrinsic("llvm.bswap", &[llty], &[args[0].immediate()])
}
}
sym::bitreverse => {
self.call_intrinsic("llvm.bitreverse", &[llty], &[args[0].immediate()])
}
sym::unchecked_funnel_shl | sym::unchecked_funnel_shr => {
let is_left = name == sym::unchecked_funnel_shl;
let lhs = args[0].immediate();
let rhs = args[1].immediate();
let raw_shift = args[2].immediate();
let llvm_name = format!("llvm.fsh{}", if is_left { 'l' } else { 'r' });
// llvm expects shift to be the same type as the values, but rust
// always uses `u32`.
let raw_shift = self.intcast(raw_shift, self.val_ty(lhs), false);
self.call_intrinsic(llvm_name, &[llty], &[lhs, rhs, raw_shift])
}
sym::saturating_add | sym::saturating_sub => {
let is_add = name == sym::saturating_add;
let lhs = args[0].immediate();
let rhs = args[1].immediate();
let llvm_name = format!(
"llvm.{}{}.sat",
if signed { 's' } else { 'u' },
if is_add { "add" } else { "sub" },
);
self.call_intrinsic(llvm_name, &[llty], &[lhs, rhs])
}
_ => bug!(),
}
}
sym::fabs => {
let ty = args[0].layout.ty;
let ty::Float(f) = ty.kind() else {
span_bug!(span, "the `fabs` intrinsic requires a floating-point argument, got {:?}", ty);
};
let llty = self.type_float_from_ty(*f);
let llvm_name = "llvm.fabs";
self.call_intrinsic(
llvm_name,
&[llty],
&args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
)
}
sym::raw_eq => {
use BackendRepr::*;
let tp_ty = fn_args.type_at(0);
let layout = self.layout_of(tp_ty).layout;
let use_integer_compare = match layout.backend_repr() {
Scalar(_) | ScalarPair { a: _, b: _, b_offset: _ } => true,
SimdVector { .. } => false,
SimdScalableVector { .. } => {
let err = tcx.dcx().emit_err(InvalidMonomorphization::NonScalableType {
span,
name: sym::raw_eq,
ty: tp_ty,
});
return IntrinsicResult::Err(err);
}
Memory { .. } => {
// For rusty ABIs, small aggregates are actually passed
// as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
// so we re-use that same threshold here.
layout.size() <= self.data_layout().pointer_size() * 2
}
};
let a = args[0].immediate();
let b = args[1].immediate();
if layout.size().bytes() == 0 {
self.const_bool(true)
} else if use_integer_compare {
let integer_ty = self.type_ix(layout.size().bits());
let a_val = self.load(integer_ty, a, layout.align().abi);
let b_val = self.load(integer_ty, b, layout.align().abi);
self.icmp(IntPredicate::IntEQ, a_val, b_val)
} else {
let n = self.const_usize(layout.size().bytes());
let cmp = self.call_intrinsic("memcmp", &[], &[a, b, n]);
self.icmp(IntPredicate::IntEQ, cmp, self.const_int(self.type_int(), 0))
}
}
sym::compare_bytes => {
// Here we assume that the `memcmp` provided by the target is a NOP for size 0.
let cmp = self.call_intrinsic(
"memcmp",
&[],
&[args[0].immediate(), args[1].immediate(), args[2].immediate()],
);
// Some targets have `memcmp` returning `i16`, but the intrinsic is always `i32`.
self.sext(cmp, self.type_ix(32))
}
sym::black_box => {
let result = PlaceRef {
val: result_place.unwrap(),
layout: result_layout,
};
args[0].val.store(self, result);
let result_val_span = [result.val.llval];
// We need to "use" the argument in some way LLVM can't introspect, and on
// targets that support it we can typically leverage inline assembly to do
// this. LLVM's interpretation of inline assembly is that it's, well, a black
// box. This isn't the greatest implementation since it probably deoptimizes
// more than we want, but it's so far good enough.
//
// For zero-sized types, the location pointed to by the result may be
// uninitialized. Do not "use" the result in this case; instead just clobber
// the memory.
let (constraint, inputs): (&str, &[_]) = if result.layout.is_zst() {
("~{memory}", &[])
} else {
("r,~{memory}", &result_val_span)
};
crate::asm::inline_asm_call(
self,
"",
constraint,
inputs,
self.type_void(),
&[],
true,
false,
llvm::AsmDialect::Att,
&[span],
false,
None,
None,
)
.unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
// We have copied the value to `result` already.
return IntrinsicResult::WroteIntoPlace;
}
sym::gpu_launch_sized_workgroup_mem => {
// Generate an anonymous global per call, with these properties:
// 1. The global is in the address space for workgroup memory
// 2. It is an `external` global
// 3. It is correctly aligned for the pointee `T`
// All instances of extern addrspace(gpu_workgroup) globals are merged in the LLVM backend.
// The name is irrelevant.
// See https://docs.nvidia.com/cuda/cuda-c-programming-guide/#shared
let name = if llvm_version < (23, 0, 0) && tcx.sess.target.arch == Arch::Nvptx64 {
// The auto-assigned name for extern shared globals in the nvptx backend does
// not compile in ptxas. Workaround this issue by assigning a name.
// Fixed in LLVM 23.
"gpu_launch_sized_workgroup_mem"
} else {
""
};
let global = self.declare_global_in_addrspace(
name,
self.type_array(self.type_i8(), 0),
AddressSpace::GPU_WORKGROUP,
);
let ty::RawPtr(inner_ty, _) = result_layout.ty.kind() else { unreachable!() };
// The alignment of the global is used to specify the *minimum* alignment that
// must be obeyed by the GPU runtime.
// When multiple of these global variables are used by a kernel, the maximum alignment is taken.
// See https://github.com/llvm/llvm-project/blob/a271d07488a85ce677674bbe8101b10efff58c95/llvm/lib/Target/AMDGPU/AMDGPULowerModuleLDSPass.cpp#L821
let alignment = self.align_of(*inner_ty).bytes() as u32;
unsafe {
// FIXME Workaround the above issue by taking maximum alignment if the global existed
if tcx.sess.target.arch == Arch::Nvptx64 {
if alignment > llvm::LLVMGetAlignment(global) {
llvm::LLVMSetAlignment(global, alignment);
}
} else {
llvm::LLVMSetAlignment(global, alignment);
}
}
self.cx().const_pointercast(global, self.type_ptr())
}
sym::amdgpu_dispatch_ptr => {
let val = self.call_intrinsic("llvm.amdgcn.dispatch.ptr", &[], &[]);
// Relying on `LLVMBuildPointerCast` to produce an addrspacecast
self.pointercast(val, self.type_ptr())
}
sym::sve_tuple_create2 => {
assert_matches!(
self.layout_of(fn_args.type_at(0)).backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(1),
..
}
);
let tuple_ty = self.layout_of(fn_args.type_at(1));
assert_matches!(
tuple_ty.backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(2),
..
}
);
let ret = self.const_poison(self.backend_type(tuple_ty));
let ret = self.insert_value(ret, args[0].immediate(), 0);
self.insert_value(ret, args[1].immediate(), 1)
}
sym::sve_tuple_create3 => {
assert_matches!(
self.layout_of(fn_args.type_at(0)).backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(1),
..
}
);
let tuple_ty = self.layout_of(fn_args.type_at(1));
assert_matches!(
tuple_ty.backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(3),
..
}
);
let ret = self.const_poison(self.backend_type(tuple_ty));
let ret = self.insert_value(ret, args[0].immediate(), 0);
let ret = self.insert_value(ret, args[1].immediate(), 1);
self.insert_value(ret, args[2].immediate(), 2)
}
sym::sve_tuple_create4 => {
assert_matches!(
self.layout_of(fn_args.type_at(0)).backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(1),
..
}
);
let tuple_ty = self.layout_of(fn_args.type_at(1));
assert_matches!(
tuple_ty.backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(4),
..
}
);
let ret = self.const_poison(self.backend_type(tuple_ty));
let ret = self.insert_value(ret, args[0].immediate(), 0);
let ret = self.insert_value(ret, args[1].immediate(), 1);
let ret = self.insert_value(ret, args[2].immediate(), 2);
self.insert_value(ret, args[3].immediate(), 3)
}
sym::sve_tuple_get => {
assert_matches!(
self.layout_of(fn_args.type_at(0)).backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(2 | 3 | 4 | 5 | 6 | 7 | 8),
..
}
);
assert_matches!(
self.layout_of(fn_args.type_at(1)).backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(1),
..
}
);
self.extract_value(
args[0].immediate(),
fn_args.const_at(2).to_leaf().to_i32() as u64,
)
}
sym::sve_tuple_set => {
assert_matches!(
self.layout_of(fn_args.type_at(0)).backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(2 | 3 | 4 | 5 | 6 | 7 | 8),
..
}
);
assert_matches!(
self.layout_of(fn_args.type_at(1)).backend_repr,
BackendRepr::SimdScalableVector {
number_of_vectors: NumScalableVectors(1),
..
}
);
self.insert_value(
args[0].immediate(),
args[1].immediate(),
fn_args.const_at(2).to_leaf().to_i32() as u64,
)
}
_ if name.as_str().starts_with("simd_") => {
// Unpack non-power-of-2 #[repr(packed, simd)] arguments.
// This gives them the expected layout of a regular #[repr(simd)] vector.
let mut loaded_args = Vec::new();
for arg in args {
loaded_args.push(
// #[repr(packed, simd)] vectors are passed like arrays (as references,
// with reduced alignment and no padding) rather than as immediates.
// We can use a vector load to fix the layout and turn the argument
// into an immediate.
if arg.layout.ty.is_simd()
&& let OperandValue::Ref(place) = arg.val
{
let (size, elem_ty) = arg.layout.ty.simd_size_and_type(self.tcx());
let elem_ll_ty = match elem_ty.kind() {
ty::Float(f) => self.type_float_from_ty(*f),
ty::Int(i) => self.type_int_from_ty(*i),
ty::Uint(u) => self.type_uint_from_ty(*u),
ty::RawPtr(_, _) => self.type_ptr(),
_ => unreachable!(),
};
let loaded =
self.load_from_place(self.type_vector(elem_ll_ty, size), place);
OperandRef::from_immediate_or_packed_pair(self, loaded, arg.layout)
} else {
*arg
},
);
}
let llret_ty = if result_layout.ty.is_simd()
&& let BackendRepr::Memory { .. } = result_layout.backend_repr
{
let (size, elem_ty) = result_layout.ty.simd_size_and_type(self.tcx());
let elem_ll_ty = match elem_ty.kind() {
ty::Float(f) => self.type_float_from_ty(*f),
ty::Int(i) => self.type_int_from_ty(*i),
ty::Uint(u) => self.type_uint_from_ty(*u),
ty::RawPtr(_, _) => self.type_ptr(),
_ => unreachable!(),
};
self.type_vector(elem_ll_ty, size)
} else {
result_layout.llvm_type(self)
};
match generic_simd_intrinsic(
self,
name,
fn_args,
&loaded_args,
result_layout.ty,
llret_ty,
span,
) {
Ok(llval) => llval,
// If there was an error, just skip this invocation... we'll abort compilation
// anyway, but we can keep codegen'ing to find more errors.
Err(err) => return IntrinsicResult::Err(err),
}
}
sym::return_address => {
match self.sess().target.arch {
// Expand this list as needed
| Arch::Wasm32
| Arch::Wasm64 => {
let ty = self.type_ptr();
self.const_null(ty)
}
_ => {
let ty = self.type_ix(32);
let val = self.const_int(ty, 0);
let type_params: &[&'ll Type] = if llvm_version < (23, 0, 0) {
&[]
} else {
&[self.type_ptr()]
};
self.call_intrinsic("llvm.returnaddress", type_params, &[val])
}
}
}
_ => {
debug!("unknown intrinsic '{}' -- falling back to default body", name);
// Call the fallback body instead of generating the intrinsic code
let fallback = ty::Instance::new_raw(instance.def_id(), instance.args);
return IntrinsicResult::Fallback(fallback);
}
};
if let BackendRepr::Memory { .. } = result_layout.backend_repr {
// We have an llvm immediate, but that's not what cg_ssa expects,
// so write it into the place (that always exists for memory)
if !result_layout.is_zst() {
self.store_to_place(llval, result_place.unwrap());
}
IntrinsicResult::WroteIntoPlace
} else {
IntrinsicResult::Operand(
OperandRef::from_immediate_or_packed_pair(self, llval, result_layout).val,
)
}
}
fn codegen_llvm_intrinsic_call(
&mut self,
instance: ty::Instance<'tcx>,
args: &[OperandRef<'tcx, Self::Value>],
_is_cleanup: bool,
) -> Self::Value {
let tcx = self.tcx();
let fn_ty = instance.ty(tcx, self.typing_env());
let fn_sig = match *fn_ty.kind() {
ty::FnDef(def_id, args) => tcx.instantiate_bound_regions_with_erased(
tcx.fn_sig(def_id).instantiate(tcx, args.no_bound_vars().unwrap()).skip_norm_wip(),
),
_ => unreachable!(),
};
assert!(!fn_sig.c_variadic());
let ret_layout = self.layout_of(fn_sig.output());
let llreturn_ty = if ret_layout.is_zst() {
self.type_void()
} else {
ret_layout.immediate_llvm_type(self)
};
let mut llargument_tys = Vec::with_capacity(fn_sig.inputs().len());
for &arg in fn_sig.inputs() {
let arg_layout = self.layout_of(arg);
if arg_layout.is_zst() {
continue;
}
llargument_tys.push(arg_layout.immediate_llvm_type(self));
}
let fn_ptr = if let Some(&llfn) = self.intrinsic_instances.borrow().get(&instance) {
llfn
} else {
let sym = tcx.symbol_name(instance).name;
let llfn = if let Some(llfn) = self.get_declared_value(sym) {
llfn
} else {
intrinsic_fn(self, sym, llreturn_ty, llargument_tys, instance)
};
self.intrinsic_instances.borrow_mut().insert(instance, llfn);
llfn
};
let fn_ty = self.get_type_of_global(fn_ptr);
let mut llargs = vec![];
for arg in args {
match arg.val {
OperandValue::ZeroSized => {}
OperandValue::Immediate(a) => llargs.push(a),
OperandValue::Pair(a, b) => {
llargs.push(a);
llargs.push(b);
}
OperandValue::Ref(op_place_val) => {
let mut llval = op_place_val.llval;
// We can't use `PlaceRef::load` here because the argument
// may have a type we don't treat as immediate, but the ABI
// used for this call is passing it by-value. In that case,
// the load would just produce `OperandValue::Ref` instead
// of the `OperandValue::Immediate` we need for the call.
llval = self.load(self.backend_type(arg.layout), llval, op_place_val.align);
if let BackendRepr::Scalar(scalar) = arg.layout.backend_repr {
if scalar.is_bool() {
self.range_metadata(llval, WrappingRange { start: 0, end: 1 });
}
// We store bools as `i8` so we need to truncate to `i1`.
llval = self.to_immediate_scalar(llval, scalar);
}
llargs.push(llval);
}
}
}
debug!("call intrinsic {:?} with args ({:?})", instance, llargs);
for (dest_ty, arg) in iter::zip(self.func_params_types(fn_ty), &mut llargs) {
let src_ty = self.val_ty(arg);
assert!(
can_autocast(self, src_ty, dest_ty),
"Cannot match `{dest_ty:?}` (expected) with {src_ty:?} (found) in `{fn_ptr:?}"
);
*arg = autocast(self, arg, src_ty, dest_ty);
}
let llret = unsafe {
llvm::LLVMBuildCallWithOperandBundles(
self.llbuilder,
fn_ty,
fn_ptr,
llargs.as_ptr(),
llargs.len() as c_uint,
ptr::dangling(),
0,
c"".as_ptr(),
)
};
let src_ty = self.val_ty(llret);
let dest_ty = llreturn_ty;
assert!(
can_autocast(self, dest_ty, src_ty),
"Cannot match `{src_ty:?}` (expected) with `{dest_ty:?}` (found) in `{fn_ptr:?}`"
);
autocast(self, llret, src_ty, dest_ty)
}
fn abort(&mut self) {
self.call_intrinsic("llvm.trap", &[], &[]);
}
fn assume(&mut self, val: Self::Value) {
if self.cx.sess().opts.optimize != rustc_session::config::OptLevel::No {
self.call_intrinsic("llvm.assume", &[], &[val]);
}
}
fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
if self.cx.sess().opts.optimize != rustc_session::config::OptLevel::No {
self.call_intrinsic(
"llvm.expect",
&[self.type_i1()],
&[cond, self.const_bool(expected)],
)
} else {
cond
}
}
fn type_checked_load(
&mut self,
llvtable: &'ll Value,
vtable_byte_offset: u64,
typeid: &[u8],
) -> Self::Value {
let typeid = self.create_metadata(typeid);
let typeid = self.get_metadata_value(typeid);
let vtable_byte_offset = self.const_i32(vtable_byte_offset as i32);
let type_checked_load = self.call_intrinsic(
"llvm.type.checked.load",
&[],
&[llvtable, vtable_byte_offset, typeid],
);
self.extract_value(type_checked_load, 0)
}
fn va_start(&mut self, va_list: &'ll Value) {
self.call_intrinsic("llvm.va_start", &[self.val_ty(va_list)], &[va_list]);
}
fn retag_reg(&mut self, ptr: Self::Value, info: &RetagInfo<Self::Value>) -> Self::Value {
codegen_retag_inner(self, "__rust_retag_reg", ptr, info)
}
fn retag_mem(&mut self, ptr: Self::Value, info: &RetagInfo<Self::Value>) {
codegen_retag_inner(self, "__rust_retag_mem", ptr, info);
}
}
fn llvm_arch_for(rust_arch: &Arch) -> Option<&'static str> {
Some(match rust_arch {
Arch::AArch64 | Arch::Arm64EC => "aarch64",
Arch::AmdGpu => "amdgcn",
Arch::Arm => "arm",
Arch::Bpf => "bpf",
Arch::Hexagon => "hexagon",
Arch::LoongArch32 | Arch::LoongArch64 => "loongarch",
Arch::Mips | Arch::Mips32r6 | Arch::Mips64 | Arch::Mips64r6 => "mips",
Arch::Nvptx64 => "nvvm",
Arch::PowerPC | Arch::PowerPC64 => "ppc",
Arch::RiscV32 | Arch::RiscV64 => "riscv",
Arch::S390x => "s390",
Arch::SpirV => "spv",
Arch::Wasm32 | Arch::Wasm64 => "wasm",
Arch::X86 | Arch::X86_64 => "x86",
_ => return None, // fallback for unknown archs
})
}
fn can_autocast<'ll>(cx: &CodegenCx<'ll, '_>, rust_ty: &'ll Type, llvm_ty: &'ll Type) -> bool {
if rust_ty == llvm_ty {
return true;
}
match cx.type_kind(llvm_ty) {
// Some LLVM intrinsics return **non-packed** structs, but they can't be mimicked from Rust
// due to auto field-alignment in non-packed structs (packed structs are represented in LLVM
// as, well, packed structs, so they won't match with those either)
TypeKind::Struct if cx.type_kind(rust_ty) == TypeKind::Struct => {
let rust_element_tys = cx.struct_element_types(rust_ty);
let llvm_element_tys = cx.struct_element_types(llvm_ty);
if rust_element_tys.len() != llvm_element_tys.len() {
return false;
}
iter::zip(rust_element_tys, llvm_element_tys).all(
|(rust_element_ty, llvm_element_ty)| {
can_autocast(cx, rust_element_ty, llvm_element_ty)
},
)
}
TypeKind::Vector => {
let llvm_element_ty = cx.element_type(llvm_ty);
let element_count = cx.vector_length(llvm_ty) as u64;
if llvm_element_ty == cx.type_bf16() {
rust_ty == cx.type_vector(cx.type_i16(), element_count)
} else if llvm_element_ty == cx.type_i1() {
let int_width = element_count.next_power_of_two().max(8);
rust_ty == cx.type_ix(int_width)
} else {
false
}
}
TypeKind::BFloat => rust_ty == cx.type_i16(),
TypeKind::X86_AMX if cx.type_kind(rust_ty) == TypeKind::Vector => {
let element_ty = cx.element_type(rust_ty);
let element_count = cx.vector_length(rust_ty) as u64;
let element_size_bits = match cx.type_kind(element_ty) {
TypeKind::Half => 16,
TypeKind::Float => 32,
TypeKind::Double => 64,
TypeKind::FP128 => 128,
TypeKind::Integer => cx.int_width(element_ty),
TypeKind::Pointer => cx.int_width(cx.isize_ty),
_ => bug!(
"Vector element type `{element_ty:?}` not one of integer, float or pointer"
),
};
element_size_bits * element_count == 8192
}
_ => false,
}
}
fn autocast<'ll>(
bx: &mut Builder<'_, 'll, '_>,
val: &'ll Value,
src_ty: &'ll Type,
dest_ty: &'ll Type,
) -> &'ll Value {
if src_ty == dest_ty {
return val;
}
match (bx.type_kind(src_ty), bx.type_kind(dest_ty)) {
// re-pack structs
(TypeKind::Struct, TypeKind::Struct) => {
let mut ret = bx.const_poison(dest_ty);
for (idx, (src_element_ty, dest_element_ty)) in
iter::zip(bx.struct_element_types(src_ty), bx.struct_element_types(dest_ty))
.enumerate()
{
let elt = bx.extract_value(val, idx as u64);
let casted_elt = autocast(bx, elt, src_element_ty, dest_element_ty);
ret = bx.insert_value(ret, casted_elt, idx as u64);
}
ret
}
// cast from the i1xN vector type to the primitive type
(TypeKind::Vector, TypeKind::Integer) if bx.element_type(src_ty) == bx.type_i1() => {
let vector_length = bx.vector_length(src_ty) as u64;
let int_width = vector_length.next_power_of_two().max(8);
let val = if vector_length == int_width {
val
} else {
// zero-extends vector
let shuffle_indices = match vector_length {
0 => unreachable!("zero length vectors are not allowed"),
1 => vec![0, 1, 1, 1, 1, 1, 1, 1],
2 => vec![0, 1, 2, 2, 2, 2, 2, 2],
3 => vec![0, 1, 2, 3, 3, 3, 3, 3],
4.. => (0..int_width as i32).collect(),
};
let shuffle_mask =
shuffle_indices.into_iter().map(|i| bx.const_i32(i)).collect::<Vec<_>>();
bx.shuffle_vector(val, bx.const_null(src_ty), bx.const_vector(&shuffle_mask))
};
bx.bitcast(val, dest_ty)
}
// cast from the primitive type to the i1xN vector type
(TypeKind::Integer, TypeKind::Vector) if bx.element_type(dest_ty) == bx.type_i1() => {
let vector_length = bx.vector_length(dest_ty) as u64;
let int_width = vector_length.next_power_of_two().max(8);
let intermediate_ty = bx.type_vector(bx.type_i1(), int_width);
let intermediate = bx.bitcast(val, intermediate_ty);
if vector_length == int_width {
intermediate
} else {
let shuffle_mask: Vec<_> =
(0..vector_length).map(|i| bx.const_i32(i as i32)).collect();
bx.shuffle_vector(
intermediate,
bx.const_poison(intermediate_ty),
bx.const_vector(&shuffle_mask),
)
}
}
(TypeKind::Vector, TypeKind::X86_AMX) => {
bx.call_intrinsic("llvm.x86.cast.vector.to.tile", &[src_ty], &[val])
}
(TypeKind::X86_AMX, TypeKind::Vector) => {
bx.call_intrinsic("llvm.x86.cast.tile.to.vector", &[dest_ty], &[val])
}
_ => bx.bitcast(val, dest_ty), // for `bf16(xN)` <-> `u16(xN)`
}
}
fn intrinsic_fn<'ll, 'tcx>(
bx: &Builder<'_, 'll, 'tcx>,
name: &str,
rust_return_ty: &'ll Type,
rust_argument_tys: Vec<&'ll Type>,
instance: ty::Instance<'tcx>,
) -> &'ll Value {
let tcx = bx.tcx;
let rust_fn_ty = bx.type_func(&rust_argument_tys, rust_return_ty);
let intrinsic = llvm::Intrinsic::lookup(name.as_bytes());
if let Some(intrinsic) = intrinsic
&& intrinsic.is_target_specific()
{
let (llvm_arch, _) = name[5..].split_once('.').unwrap();
let rust_arch = &tcx.sess.target.arch;
if let Some(correct_llvm_arch) = llvm_arch_for(rust_arch)
&& llvm_arch != correct_llvm_arch
{
tcx.dcx().emit_fatal(IntrinsicWrongArch {
name,
target_arch: rust_arch.desc(),
span: tcx.def_span(instance.def_id()),
});
}
}
if let Some(intrinsic) = intrinsic
&& !intrinsic.is_overloaded()
{
// FIXME: also do this for overloaded intrinsics
let llfn = intrinsic.get_declaration(bx.llmod, &[]);
let llvm_fn_ty = bx.get_type_of_global(llfn);
let llvm_return_ty = bx.get_return_type(llvm_fn_ty);
let llvm_argument_tys = bx.func_params_types(llvm_fn_ty);
let llvm_is_variadic = bx.func_is_variadic(llvm_fn_ty);
let is_correct_signature = !llvm_is_variadic
&& rust_argument_tys.len() == llvm_argument_tys.len()
&& iter::once((rust_return_ty, llvm_return_ty))
.chain(iter::zip(rust_argument_tys, llvm_argument_tys))
.all(|(rust_ty, llvm_ty)| can_autocast(bx, rust_ty, llvm_ty));
if !is_correct_signature {
tcx.dcx().emit_fatal(IntrinsicSignatureMismatch {
name,
llvm_fn_ty: &format!("{llvm_fn_ty:?}"),
rust_fn_ty: &format!("{rust_fn_ty:?}"),
span: tcx.def_span(instance.def_id()),
});
}
return llfn;
}
// Function addresses in Rust are never significant, allowing functions to be merged.
let llfn = declare_raw_fn(
bx,
name,
llvm::CCallConv,
llvm::UnnamedAddr::Global,
llvm::Visibility::Default,
rust_fn_ty,
);
if intrinsic.is_none() {
let mut new_llfn = None;
let can_upgrade = unsafe { llvm::LLVMRustUpgradeIntrinsicFunction(llfn, &mut new_llfn) };
if !can_upgrade {
// This is either plain wrong, or this can be caused by incompatible LLVM versions
tcx.dcx().emit_fatal(UnknownIntrinsic { name, span: tcx.def_span(instance.def_id()) });
} else if let Some(def_id) = instance.def_id().as_local() {
// we can emit diagnostics only for local crates
let hir_id = tcx.local_def_id_to_hir_id(def_id);
// not all intrinsics are upgraded to some other intrinsics, most are upgraded to instruction sequences
let msg = if let Some(new_llfn) = new_llfn {
format!(
"using deprecated intrinsic `{name}`, `{}` can be used instead",
str::from_utf8(&llvm::get_value_name(new_llfn)).unwrap()
)
} else {
format!("using deprecated intrinsic `{name}`")
};
tcx.emit_node_lint(
DEPRECATED_LLVM_INTRINSIC,
hir_id,
rustc_errors::DiagDecorator(|d| {
d.primary_message(msg).span(tcx.hir_span(hir_id));
}),
);
}
}
llfn
}
fn catch_unwind_intrinsic<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
try_func: &'ll Value,
data: &'ll Value,
catch_func: &'ll Value,
) -> &'ll Value {
if !bx.sess().panic_strategy().unwinds() {
let try_func_ty = bx.type_func(&[bx.type_ptr()], bx.type_void());
bx.call(try_func_ty, None, None, try_func, &[data], None, None);
// Return 0 unconditionally from the intrinsic call;
// we can never unwind.
bx.const_bool(false)
} else if wants_msvc_seh(bx.sess()) {
codegen_msvc_try(bx, try_func, data, catch_func)
} else if wants_wasm_eh(bx.sess()) {
codegen_wasm_try(bx, try_func, data, catch_func)
} else {
codegen_gnu_try(bx, try_func, data, catch_func)
}
}
// MSVC's definition of the `rust_try` function.
//
// This implementation uses the new exception handling instructions in LLVM
// which have support in LLVM for SEH on MSVC targets. Although these
// instructions are meant to work for all targets, as of the time of this
// writing, however, LLVM does not recommend the usage of these new instructions
// as the old ones are still more optimized.
fn codegen_msvc_try<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
try_func: &'ll Value,
data: &'ll Value,
catch_func: &'ll Value,
) -> &'ll Value {
let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
bx.set_personality_fn(bx.eh_personality());
let normal = bx.append_sibling_block("normal");
let catchswitch = bx.append_sibling_block("catchswitch");
let catchpad_rust = bx.append_sibling_block("catchpad_rust");
let catchpad_foreign = bx.append_sibling_block("catchpad_foreign");
let caught = bx.append_sibling_block("caught");
let try_func = llvm::get_param(bx.llfn(), 0);
let data = llvm::get_param(bx.llfn(), 1);
let catch_func = llvm::get_param(bx.llfn(), 2);
// We're generating an IR snippet that looks like:
//
// declare bool @rust_try(%try_func, %data, %catch_func) {
// %slot = alloca i8*
// invoke %try_func(%data) to label %normal unwind label %catchswitch
//
// normal:
// ret i1 false
//
// catchswitch:
// %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
//
// catchpad_rust:
// %tok = catchpad within %cs [%type_descriptor, 8, %slot]
// %ptr = load %slot
// call %catch_func(%data, %ptr)
// catchret from %tok to label %caught
//
// catchpad_foreign:
// %tok = catchpad within %cs [null, 64, null]
// call %catch_func(%data, null)
// catchret from %tok to label %caught
//
// caught:
// ret i1 true
// }
//
// This structure follows the basic usage of throw/try/catch in LLVM.
// For example, compile this C++ snippet to see what LLVM generates:
//
// struct rust_panic {
// rust_panic(const rust_panic&);
// ~rust_panic();
//
// void* x[2];
// };
//
// int __rust_try(
// void (*try_func)(void*),
// void *data,
// void (*catch_func)(void*, void*) noexcept
// ) {
// try {
// try_func(data);
// return 0;
// } catch(rust_panic& a) {
// catch_func(data, &a);
// return 1;
// } catch(...) {
// catch_func(data, NULL);
// return 1;
// }
// }
//
// More information can be found in libstd's seh.rs implementation.
let ptr_size = bx.tcx().data_layout.pointer_size();
let ptr_align = bx.tcx().data_layout.pointer_align().abi;
let slot = bx.alloca(ptr_size, ptr_align);
let try_func_ty = bx.type_func(&[bx.type_ptr()], bx.type_void());
bx.invoke(try_func_ty, None, None, try_func, &[data], normal, catchswitch, None, None);
bx.switch_to_block(normal);
bx.ret(bx.const_bool(false));
bx.switch_to_block(catchswitch);
let cs = bx.catch_switch(None, None, &[catchpad_rust, catchpad_foreign]);
// We can't use the TypeDescriptor defined in libpanic_unwind because it
// might be in another DLL and the SEH encoding only supports specifying
// a TypeDescriptor from the current module.
//
// However this isn't an issue since the MSVC runtime uses string
// comparison on the type name to match TypeDescriptors rather than
// pointer equality.
//
// So instead we generate a new TypeDescriptor in each module that uses
// `try` and let the linker merge duplicate definitions in the same
// module.
//
// When modifying, make sure that the type_name string exactly matches
// the one used in library/panic_unwind/src/seh.rs.
let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_ptr());
let type_name = bx.const_bytes(b"rust_panic\0");
let type_info =
bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_ptr()), type_name], false);
let tydesc = bx.declare_global(
&mangle_internal_symbol(bx.tcx, "__rust_panic_type_info"),
bx.val_ty(type_info),
);
llvm::set_linkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
if bx.cx.tcx.sess.target.supports_comdat() {
llvm::SetUniqueComdat(bx.llmod, tydesc);
}
llvm::set_initializer(tydesc, type_info);
// The flag value of 8 indicates that we are catching the exception by
// reference instead of by value. We can't use catch by value because
// that requires copying the exception object, which we don't support
// since our exception object effectively contains a Box.
//
// Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
bx.switch_to_block(catchpad_rust);
let flags = bx.const_i32(8);
let funclet = bx.catch_pad(cs, &[tydesc, flags, slot]);
let ptr = bx.load(bx.type_ptr(), slot, ptr_align);
let catch_ty = bx.type_func(&[bx.type_ptr(), bx.type_ptr()], bx.type_void());
bx.call(catch_ty, None, None, catch_func, &[data, ptr], Some(&funclet), None);
bx.catch_ret(&funclet, caught);
// The flag value of 64 indicates a "catch-all".
bx.switch_to_block(catchpad_foreign);
let flags = bx.const_i32(64);
let null = bx.const_null(bx.type_ptr());
let funclet = bx.catch_pad(cs, &[null, flags, null]);
bx.call(catch_ty, None, None, catch_func, &[data, null], Some(&funclet), None);
bx.catch_ret(&funclet, caught);
bx.switch_to_block(caught);
bx.ret(bx.const_bool(true));
});
// Note that no invoke is used here because by definition this function
// can't panic (that's what it's catching).
let ret = bx.call(llty, None, None, llfn, &[try_func, data, catch_func], None, None);
ret
}
// WASM's definition of the `rust_try` function.
fn codegen_wasm_try<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
try_func: &'ll Value,
data: &'ll Value,
catch_func: &'ll Value,
) -> &'ll Value {
let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
bx.set_personality_fn(bx.eh_personality());
let normal = bx.append_sibling_block("normal");
let catchswitch = bx.append_sibling_block("catchswitch");
let catchpad = bx.append_sibling_block("catchpad");
let caught = bx.append_sibling_block("caught");
let try_func = llvm::get_param(bx.llfn(), 0);
let data = llvm::get_param(bx.llfn(), 1);
let catch_func = llvm::get_param(bx.llfn(), 2);
// We're generating an IR snippet that looks like:
//
// declare i1 @rust_try(%try_func, %data, %catch_func) {
// %slot = alloca i8*
// invoke %try_func(%data) to label %normal unwind label %catchswitch
//
// normal:
// ret i1 false
//
// catchswitch:
// %cs = catchswitch within none [%catchpad] unwind to caller
//
// catchpad:
// %tok = catchpad within %cs [null]
// %ptr = call @llvm.wasm.get.exception(token %tok)
// %sel = call @llvm.wasm.get.ehselector(token %tok)
// call %catch_func(%data, %ptr)
// catchret from %tok to label %caught
//
// caught:
// ret i1 true
// }
//
let try_func_ty = bx.type_func(&[bx.type_ptr()], bx.type_void());
bx.invoke(try_func_ty, None, None, try_func, &[data], normal, catchswitch, None, None);
bx.switch_to_block(normal);
bx.ret(bx.const_bool(false));
bx.switch_to_block(catchswitch);
let cs = bx.catch_switch(None, None, &[catchpad]);
bx.switch_to_block(catchpad);
let null = bx.const_null(bx.type_ptr());
let funclet = bx.catch_pad(cs, &[null]);
let ptr = bx.call_intrinsic("llvm.wasm.get.exception", &[], &[funclet.cleanuppad()]);
let _sel = bx.call_intrinsic("llvm.wasm.get.ehselector", &[], &[funclet.cleanuppad()]);
let catch_ty = bx.type_func(&[bx.type_ptr(), bx.type_ptr()], bx.type_void());
bx.call(catch_ty, None, None, catch_func, &[data, ptr], Some(&funclet), None);
bx.catch_ret(&funclet, caught);
bx.switch_to_block(caught);
bx.ret(bx.const_bool(true));
});
// Note that no invoke is used here because by definition this function
// can't panic (that's what it's catching).
let ret = bx.call(llty, None, None, llfn, &[try_func, data, catch_func], None, None);
ret
}
// Definition of the standard `try` function for Rust using the GNU-like model
// of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
// instructions).
//
// This codegen is a little surprising because we always call a shim
// function instead of inlining the call to `invoke` manually here. This is done
// because in LLVM we're only allowed to have one personality per function
// definition. The call to the `try` intrinsic is being inlined into the
// function calling it, and that function may already have other personality
// functions in play. By calling a shim we're guaranteed that our shim will have
// the right personality function.
fn codegen_gnu_try<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
try_func: &'ll Value,
data: &'ll Value,
catch_func: &'ll Value,
) -> &'ll Value {
let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
// Codegens the shims described above:
//
// bx:
// invoke %try_func(%data) normal %normal unwind %catch
//
// normal:
// ret 0
//
// catch:
// (%ptr, _) = landingpad
// call %catch_func(%data, %ptr)
// ret 1
let then = bx.append_sibling_block("then");
let catch = bx.append_sibling_block("catch");
let try_func = llvm::get_param(bx.llfn(), 0);
let data = llvm::get_param(bx.llfn(), 1);
let catch_func = llvm::get_param(bx.llfn(), 2);
let try_func_ty = bx.type_func(&[bx.type_ptr()], bx.type_void());
bx.invoke(try_func_ty, None, None, try_func, &[data], then, catch, None, None);
bx.switch_to_block(then);
bx.ret(bx.const_bool(false));
// Type indicator for the exception being thrown.
//
// The first value in this tuple is a pointer to the exception object
// being thrown. The second value is a "selector" indicating which of
// the landing pad clauses the exception's type had been matched to.
// rust_try ignores the selector.
bx.switch_to_block(catch);
let lpad_ty = bx.type_struct(&[bx.type_ptr(), bx.type_i32()], false);
let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 1);
let tydesc = bx.const_null(bx.type_ptr());
bx.add_clause(vals, tydesc);
let ptr = bx.extract_value(vals, 0);
let catch_ty = bx.type_func(&[bx.type_ptr(), bx.type_ptr()], bx.type_void());
bx.call(catch_ty, None, None, catch_func, &[data, ptr], None, None);
bx.ret(bx.const_bool(true));
});
// Note that no invoke is used here because by definition this function
// can't panic (that's what it's catching).
let ret = bx.call(llty, None, None, llfn, &[try_func, data, catch_func], None, None);
ret
}
// Helper function to give a Block to a closure to codegen a shim function.
// This is currently primarily used for the `try` intrinsic functions above.
fn gen_fn<'a, 'll, 'tcx>(
cx: &'a CodegenCx<'ll, 'tcx>,
name: &str,
rust_fn_sig: ty::PolyFnSig<'tcx>,
codegen: &mut dyn FnMut(Builder<'a, 'll, 'tcx>),
) -> (&'ll Type, &'ll Value) {
let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
let llty = fn_abi.llvm_type(cx);
let llfn = cx.declare_fn(name, fn_abi, None);
cx.set_frame_pointer_type(llfn);
cx.apply_target_cpu_attr(llfn);
// FIXME(eddyb) find a nicer way to do this.
llvm::set_linkage(llfn, llvm::Linkage::InternalLinkage);
let llbb = Builder::append_block(cx, llfn, "entry-block");
let bx = Builder::build(cx, llbb);
codegen(bx);
(llty, llfn)
}
// Helper function used to get a handle to the `__rust_try` function used to
// catch exceptions.
//
// This function is only generated once and is then cached.
fn get_rust_try_fn<'a, 'll, 'tcx>(
cx: &'a CodegenCx<'ll, 'tcx>,
codegen: &mut dyn FnMut(Builder<'a, 'll, 'tcx>),
) -> (&'ll Type, &'ll Value) {
if let Some(llfn) = cx.rust_try_fn.get() {
return llfn;
}
// Define the type up front for the signature of the rust_try function.
let tcx = cx.tcx;
let i8p = Ty::new_mut_ptr(tcx, tcx.types.i8);
// `unsafe fn(*mut Data) -> ()`
let try_fn_ty = Ty::new_fn_ptr(
tcx,
ty::Binder::dummy(tcx.mk_fn_sig_rust_abi([i8p], tcx.types.unit, hir::Safety::Unsafe)),
);
// `unsafe fn(*mut Data, *mut i8) -> ()`
let catch_fn_ty = Ty::new_fn_ptr(
tcx,
ty::Binder::dummy(tcx.mk_fn_sig_rust_abi([i8p, i8p], tcx.types.unit, hir::Safety::Unsafe)),
);
// `unsafe fn(unsafe fn(*mut Data) -> (), *mut Data, unsafe fn(*mut Data, *mut i8) -> ()) -> bool`
let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig_rust_abi(
[try_fn_ty, i8p, catch_fn_ty],
tcx.types.bool,
hir::Safety::Unsafe,
));
let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
if cx.sess().pointer_authentication() {
let cfg = cx.sess().pointer_auth_config.as_ref().unwrap();
let attrs: Vec<&Attribute> =
cfg.fn_attrs().into_iter().map(|name| llvm::CreateAttrString(cx.llcx, name)).collect();
let (_ty, rust_try_fn) = rust_try;
crate::attributes::apply_to_llfn(rust_try_fn, AttributePlace::Function, &attrs);
}
cx.rust_try_fn.set(Some(rust_try));
rust_try
}
fn codegen_retag_inner<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
name: &'static str,
ptr: &'ll Value,
info: &RetagInfo<&'ll Value>,
) -> &'ll Value {
let size = bx.const_usize(info.size.bytes());
let perms = bx.const_u8(info.flags.bits());
bx.call_intrinsic(
name,
// Retag intrinsics have special handling within `CodegenCx::declare_intrinsic`
// to ensure that each form has the correct return type.
&[bx.type_ptr(), bx.val_ty(size), bx.type_i8(), bx.type_ptr(), bx.type_ptr()],
&[ptr, size, perms, info.im_layout, info.pin_layout],
)
}
fn codegen_autodiff<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
instance: ty::Instance<'tcx>,
args: &[OperandRef<'tcx, &'ll Value>],
result_layout: ty::layout::TyAndLayout<'tcx>,
result_place: Option<PlaceValue<&'ll Value>>,
) -> IntrinsicResult<'tcx, &'ll Value> {
let tcx = bx.tcx;
if !tcx.sess.opts.unstable_opts.autodiff.contains(&rustc_session::config::AutoDiff::Enable) {
let _ = tcx.dcx().emit_almost_fatal(AutoDiffWithoutEnable);
}
let ct = tcx.crate_types();
let lto = tcx.sess.lto();
if ct.len() == 1 && ct.contains(&CrateType::Executable) {
if lto != rustc_session::config::Lto::Fat {
let _ = tcx.dcx().emit_almost_fatal(AutoDiffWithoutLto);
}
} else {
if lto != rustc_session::config::Lto::Fat && !tcx.sess.opts.cg.linker_plugin_lto.enabled() {
let _ = tcx.dcx().emit_almost_fatal(AutoDiffWithoutLto);
}
}
let fn_args = instance.args;
let callee_ty = instance.ty(tcx, bx.typing_env());
let sig = callee_ty.fn_sig(tcx).skip_binder();
let ret_ty = sig.output();
let llret_ty = bx.layout_of(ret_ty).llvm_type(bx);
let source_fn_ptr_ty = fn_args.into_type_list(tcx)[0];
let fn_to_diff = args[0].immediate();
let (diff_id, diff_args) = match fn_args.into_type_list(tcx)[1].kind() {
ty::FnDef(def_id, diff_args) => (def_id, diff_args.no_bound_vars().unwrap()),
_ => bug!("invalid args"),
};
let fn_diff = match Instance::try_resolve(tcx, bx.cx.typing_env(), *diff_id, diff_args) {
Ok(Some(instance)) => instance,
Ok(None) => bug!(
"could not resolve ({:?}, {:?}) to a specific autodiff instance",
diff_id,
diff_args
),
Err(err) => {
// An error has already been emitted
return IntrinsicResult::Err(err);
}
};
let val_arr = get_args_from_tuple(bx, args[2], fn_diff);
let diff_symbol = symbol_name_for_instance_in_crate(tcx, fn_diff.clone(), LOCAL_CRATE);
let Some(Some(mut diff_attrs)) =
find_attr!(tcx, fn_diff.def_id(), RustcAutodiff(attr) => attr.clone())
else {
bug!("could not find autodiff attrs")
};
adjust_activity_to_abi(
tcx,
source_fn_ptr_ty,
TypingEnv::fully_monomorphized(),
&mut diff_attrs.input_activity,
);
let fnc_tree = fnc_typetrees(tcx, source_fn_ptr_ty);
// Build body
generate_enzyme_call(
bx,
fn_to_diff,
&diff_symbol,
llret_ty,
&val_arr,
&diff_attrs,
result_layout,
result_place,
fnc_tree,
)
}
// Generates the LLVM code to offload a Rust function to a target device (e.g., GPU).
// For each kernel call, it generates the necessary globals (including metadata such as
// size and pass mode), manages memory mapping to and from the device, handles all
// data transfers, and launches the kernel on the target device.
fn codegen_offload<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
tcx: TyCtxt<'tcx>,
instance: ty::Instance<'tcx>,
args: &[OperandRef<'tcx, &'ll Value>],
) {
let cx = bx.cx;
let fn_args = instance.args;
let (target_id, target_args) = match fn_args.into_type_list(tcx)[0].kind() {
ty::FnDef(def_id, params) => (def_id, params.no_bound_vars().unwrap()),
_ => bug!("invalid offload intrinsic arg"),
};
let fn_target = match Instance::try_resolve(tcx, cx.typing_env(), *target_id, target_args) {
Ok(Some(instance)) => instance,
Ok(None) => bug!(
"could not resolve ({:?}, {:?}) to a specific offload instance",
target_id,
target_args
),
Err(_) => {
// An error has already been emitted
return;
}
};
let offload_dims = OffloadKernelDims::from_operands(bx, &args[1], &args[2]);
let dyn_cache = match args[3].val {
OperandValue::Immediate(val) => val,
_ => panic!("unparsable"),
};
let args = get_args_from_tuple(bx, args[4], fn_target);
let target_symbol = symbol_name_for_instance_in_crate(tcx, fn_target, LOCAL_CRATE);
let sig = tcx.fn_sig(fn_target.def_id()).skip_binder();
let sig = tcx.instantiate_bound_regions_with_erased(sig);
let inputs = sig.inputs();
let fn_abi = cx.fn_abi_of_instance(fn_target, ty::List::empty());
let mut metadata = Vec::new();
let mut types = Vec::new();
for (i, arg_abi) in fn_abi.args.iter().enumerate() {
let ty = inputs[i];
let decomposed = OffloadMetadata::handle_abi(cx, tcx, ty, arg_abi);
for (meta, entry_ty) in decomposed {
metadata.push(meta);
types.push(bx.cx.layout_of(entry_ty).llvm_type(bx.cx));
}
}
let offload_globals_ref = cx.offload_globals.borrow();
let offload_globals = match offload_globals_ref.as_ref() {
Some(globals) => globals,
None => {
// Offload is not initialized, cannot continue
return;
}
};
register_offload(cx);
let offload_data = gen_define_handling(&cx, &metadata, target_symbol, offload_globals);
gen_call_handling(
bx,
&offload_data,
&args,
&types,
&metadata,
offload_globals,
&offload_dims,
&dyn_cache,
);
}
fn get_args_from_tuple<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
tuple_op: OperandRef<'tcx, &'ll Value>,
fn_instance: Instance<'tcx>,
) -> Vec<&'ll Value> {
let cx = bx.cx;
let fn_abi = cx.fn_abi_of_instance(fn_instance, ty::List::empty());
match tuple_op.val {
OperandValue::Immediate(val) => vec![val],
OperandValue::Pair(v1, v2) => vec![v1, v2],
OperandValue::Ref(ptr) => {
let tuple_place = PlaceRef { val: ptr, layout: tuple_op.layout };
let mut result = Vec::with_capacity(fn_abi.args.len());
let mut tuple_index = 0;
for arg in &fn_abi.args {
match arg.mode {
PassMode::Ignore => {}
PassMode::Direct(_) | PassMode::Cast { .. } => {
let field = tuple_place.project_field(bx, tuple_index);
let llvm_ty = field.layout.llvm_type(bx.cx);
let val = bx.load(llvm_ty, field.val.llval, field.val.align);
result.push(val);
tuple_index += 1;
}
PassMode::Pair(_, _) => {
let field = tuple_place.project_field(bx, tuple_index);
let llvm_ty = field.layout.llvm_type(bx.cx);
let pair_val = bx.load(llvm_ty, field.val.llval, field.val.align);
result.push(bx.extract_value(pair_val, 0));
result.push(bx.extract_value(pair_val, 1));
tuple_index += 1;
}
PassMode::Indirect { .. } => {
let field = tuple_place.project_field(bx, tuple_index);
result.push(field.val.llval);
tuple_index += 1;
}
}
}
result
}
OperandValue::ZeroSized => vec![],
}
}
fn generic_simd_intrinsic<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
name: Symbol,
fn_args: GenericArgsRef<'tcx>,
args: &[OperandRef<'tcx, &'ll Value>],
ret_ty: Ty<'tcx>,
llret_ty: &'ll Type,
span: Span,
) -> Result<&'ll Value, ErrorGuaranteed> {
macro_rules! return_error {
($diag: expr) => {{
let err = bx.sess().dcx().emit_err($diag);
return Err(err);
}};
}
macro_rules! require {
($cond: expr, $diag: expr) => {
if !$cond {
return_error!($diag);
}
};
}
macro_rules! require_simd {
($ty: expr, $variant:ident) => {{
require!($ty.is_simd(), InvalidMonomorphization::$variant { span, name, ty: $ty });
$ty.simd_size_and_type(bx.tcx())
}};
}
macro_rules! require_simd_or_scalable {
($ty: expr, $variant:ident) => {{
require!(
$ty.is_simd() || $ty.is_scalable_vector(),
InvalidMonomorphization::$variant { span, name, ty: $ty }
);
if $ty.is_simd() {
let (len, ty) = $ty.simd_size_and_type(bx.tcx());
(len, ty, None)
} else {
let (count, ty, num_vecs) =
$ty.scalable_vector_parts(bx.tcx()).expect("`is_scalable_vector` was wrong");
(count as u64, ty, Some(num_vecs))
}
}};
}
/// Returns the bitwidth of the `$ty` argument if it is an `Int` or `Uint` type.
macro_rules! require_int_or_uint_ty {
($ty: expr, $diag: expr) => {
match $ty {
ty::Int(i) => {
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size().bits())
}
ty::Uint(i) => {
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size().bits())
}
_ => {
return_error!($diag);
}
}
};
}
let llvm_version = crate::llvm_util::get_version();
/// Converts a vector mask, where each element has a bit width equal to the data elements it is used with,
/// down to an i1 based mask that can be used by llvm intrinsics.
///
/// The rust simd semantics are that each element should either consist of all ones or all zeroes,
/// but this information is not available to llvm. Truncating the vector effectively uses the lowest bit,
/// but codegen for several targets is better if we consider the highest bit by shifting.
///
/// For x86 SSE/AVX targets this is beneficial since most instructions with mask parameters only consider the highest bit.
/// So even though on llvm level we have an additional shift, in the final assembly there is no shift or truncate and
/// instead the mask can be used as is.
///
/// For aarch64 and other targets there is a benefit because a mask from the sign bit can be more
/// efficiently converted to an all ones / all zeroes mask by comparing whether each element is negative.
fn vector_mask_to_bitmask<'a, 'll, 'tcx>(
bx: &mut Builder<'a, 'll, 'tcx>,
i_xn: &'ll Value,
in_elem_bitwidth: u64,
in_len: u64,
) -> &'ll Value {
// Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
let shift_idx = bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
let shift_indices = vec![shift_idx; in_len as _];
let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
// Truncate vector to an <i1 x N>
bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len))
}
// Sanity-check: all vector arguments must be immediates.
if cfg!(debug_assertions) {
for arg in args {
if arg.layout.ty.is_simd() {
assert_matches!(arg.val, OperandValue::Immediate(_));
}
}
}
if name == sym::simd_select_bitmask {
let (len, _) = require_simd!(args[1].layout.ty, SimdArgument);
let expected_int_bits = len.max(8).next_power_of_two();
let expected_bytes = len.div_ceil(8);
let mask_ty = args[0].layout.ty;
let mask = match mask_ty.kind() {
ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
ty::Array(elem, len)
if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
&& len
.try_to_target_usize(bx.tcx)
.expect("expected monomorphic const in codegen")
== expected_bytes =>
{
let place = PlaceRef::alloca(bx, args[0].layout);
args[0].val.store(bx, place);
let int_ty = bx.type_ix(expected_bytes * 8);
bx.load(int_ty, place.val.llval, Align::ONE)
}
_ => return_error!(InvalidMonomorphization::InvalidBitmask {
span,
name,
mask_ty,
expected_int_bits,
expected_bytes
}),
};
let i1 = bx.type_i1();
let im = bx.type_ix(len);
let i1xn = bx.type_vector(i1, len);
let m_im = bx.trunc(mask, im);
let m_i1s = bx.bitcast(m_im, i1xn);
return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
}
if name == sym::simd_splat {
let (out_len, out_ty) = require_simd!(ret_ty, SimdReturn);
require!(
args[0].layout.ty == out_ty,
InvalidMonomorphization::ExpectedVectorElementType {
span,
name,
expected_element: out_ty,
vector_type: ret_ty,
}
);
// `insertelement <N x elem> poison, elem %x, i32 0`
let poison_vec = bx.const_poison(llret_ty);
let idx0 = bx.const_i32(0);
let v0 = bx.insert_element(poison_vec, args[0].immediate(), idx0);
// `shufflevector <N x elem> v0, <N x elem> poison, <N x i32> zeroinitializer`
// The masks is all zeros, so this splats lane 0 (which has our element in it).
let mask_ty = bx.type_vector(bx.type_i32(), out_len);
let splat = bx.shuffle_vector(v0, poison_vec, bx.const_null(mask_ty));
return Ok(splat);
}
let supports_scalable = match name {
sym::simd_cast | sym::simd_select => true,
_ => false,
};
// Every intrinsic below takes a SIMD vector as its first argument. Some intrinsics also accept
// scalable vectors. `require_simd_or_scalable` is used regardless as it'll do the right thing
// for non-scalable vectors, and an additional check to prohibit scalable vectors for those
// intrinsics that do not support them is added.
if !supports_scalable {
let _ = require_simd!(args[0].layout.ty, SimdInput);
}
let (in_len, in_elem, in_num_vecs) = require_simd_or_scalable!(args[0].layout.ty, SimdInput);
let in_ty = args[0].layout.ty;
let comparison = match name {
sym::simd_eq => Some(BinOp::Eq),
sym::simd_ne => Some(BinOp::Ne),
sym::simd_lt => Some(BinOp::Lt),
sym::simd_le => Some(BinOp::Le),
sym::simd_gt => Some(BinOp::Gt),
sym::simd_ge => Some(BinOp::Ge),
_ => None,
};
if let Some(cmp_op) = comparison {
let (out_len, out_ty) = require_simd!(ret_ty, SimdReturn);
require!(
in_len == out_len,
InvalidMonomorphization::ReturnLengthInputType {
span,
name,
in_len,
in_ty,
ret_ty,
out_len
}
);
require!(
bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
InvalidMonomorphization::ReturnIntegerType { span, name, ret_ty, out_ty }
);
return Ok(compare_simd_types(
bx,
args[0].immediate(),
args[1].immediate(),
in_elem,
llret_ty,
cmp_op,
));
}
if name == sym::simd_shuffle_const_generic {
let idx = fn_args[2].expect_const().to_branch();
let n = idx.len() as u64;
let (out_len, out_ty) = require_simd!(ret_ty, SimdReturn);
require!(
out_len == n,
InvalidMonomorphization::ReturnLength { span, name, in_len: n, ret_ty, out_len }
);
require!(
in_elem == out_ty,
InvalidMonomorphization::ReturnElement { span, name, in_elem, in_ty, ret_ty, out_ty }
);
let total_len = in_len * 2;
let indices: Option<Vec<_>> = idx
.iter()
.enumerate()
.map(|(arg_idx, val)| {
let idx = val.to_leaf().to_i32();
if idx >= i32::try_from(total_len).unwrap() {
bx.sess().dcx().emit_err(InvalidMonomorphization::SimdIndexOutOfBounds {
span,
name,
arg_idx: arg_idx as u64,
total_len: total_len.into(),
});
None
} else {
Some(bx.const_i32(idx))
}
})
.collect();
let Some(indices) = indices else {
return Ok(bx.const_null(llret_ty));
};
return Ok(bx.shuffle_vector(
args[0].immediate(),
args[1].immediate(),
bx.const_vector(&indices),
));
}
if name == sym::simd_shuffle {
// Make sure this is actually a SIMD vector.
let idx_ty = args[2].layout.ty;
let n: u64 = if idx_ty.is_simd()
&& matches!(idx_ty.simd_size_and_type(bx.cx.tcx).1.kind(), ty::Uint(ty::UintTy::U32))
{
idx_ty.simd_size_and_type(bx.cx.tcx).0
} else {
return_error!(InvalidMonomorphization::SimdShuffle { span, name, ty: idx_ty })
};
let (out_len, out_ty) = require_simd!(ret_ty, SimdReturn);
require!(
out_len == n,
InvalidMonomorphization::ReturnLength { span, name, in_len: n, ret_ty, out_len }
);
require!(
in_elem == out_ty,
InvalidMonomorphization::ReturnElement { span, name, in_elem, in_ty, ret_ty, out_ty }
);
let total_len = u128::from(in_len) * 2;
// Check that the indices are in-bounds.
let indices = args[2].immediate();
for i in 0..n {
let val = bx.const_get_elt(indices, i as u64);
let idx = bx
.const_to_opt_u128(val, true)
.unwrap_or_else(|| bug!("typeck should have already ensured that these are const"));
if idx >= total_len {
return_error!(InvalidMonomorphization::SimdIndexOutOfBounds {
span,
name,
arg_idx: i,
total_len,
});
}
}
return Ok(bx.shuffle_vector(args[0].immediate(), args[1].immediate(), indices));
}
if name == sym::simd_insert || name == sym::simd_insert_dyn {
require!(
in_elem == args[2].layout.ty,
InvalidMonomorphization::InsertedType {
span,
name,
in_elem,
in_ty,
out_ty: args[2].layout.ty
}
);
let index_imm = if name == sym::simd_insert {
let idx = bx
.const_to_opt_u128(args[1].immediate(), false)
.expect("typeck should have ensure that this is a const");
if idx >= in_len.into() {
return_error!(InvalidMonomorphization::SimdIndexOutOfBounds {
span,
name,
arg_idx: 1,
total_len: in_len.into(),
});
}
bx.const_i32(idx as i32)
} else {
args[1].immediate()
};
return Ok(bx.insert_element(args[0].immediate(), args[2].immediate(), index_imm));
}
if name == sym::simd_extract || name == sym::simd_extract_dyn {
require!(
ret_ty == in_elem,
InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty }
);
let index_imm = if name == sym::simd_extract {
let idx = bx
.const_to_opt_u128(args[1].immediate(), false)
.expect("typeck should have ensure that this is a const");
if idx >= in_len.into() {
return_error!(InvalidMonomorphization::SimdIndexOutOfBounds {
span,
name,
arg_idx: 1,
total_len: in_len.into(),
});
}
bx.const_i32(idx as i32)
} else {
args[1].immediate()
};
return Ok(bx.extract_element(args[0].immediate(), index_imm));
}
if name == sym::simd_select {
let m_elem_ty = in_elem;
let m_len = in_len;
let (v_len, _, _) = require_simd_or_scalable!(args[1].layout.ty, SimdArgument);
require!(
m_len == v_len,
InvalidMonomorphization::MismatchedLengths { span, name, m_len, v_len }
);
let m_i1s = if args[1].layout.ty.is_scalable_vector() {
match m_elem_ty.kind() {
ty::Bool => {}
_ => return_error!(InvalidMonomorphization::MaskWrongElementType {
span,
name,
ty: m_elem_ty
}),
};
let i1 = bx.type_i1();
let i1xn = bx.type_scalable_vector(i1, m_len as u64);
bx.trunc(args[0].immediate(), i1xn)
} else {
let in_elem_bitwidth = require_int_or_uint_ty!(
m_elem_ty.kind(),
InvalidMonomorphization::MaskWrongElementType { span, name, ty: m_elem_ty }
);
vector_mask_to_bitmask(bx, args[0].immediate(), in_elem_bitwidth, m_len)
};
return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
}
if name == sym::simd_bitmask {
// The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a vector mask and
// returns one bit for each lane (which must all be `0` or `!0`) in the form of either:
// * an unsigned integer
// * an array of `u8`
// If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
//
// The bit order of the result depends on the byte endianness, LSB-first for little
// endian and MSB-first for big endian.
let expected_int_bits = in_len.max(8).next_power_of_two();
let expected_bytes = in_len.div_ceil(8);
// Integer vector <i{in_bitwidth} x in_len>:
let in_elem_bitwidth = require_int_or_uint_ty!(
in_elem.kind(),
InvalidMonomorphization::MaskWrongElementType { span, name, ty: in_elem }
);
let i1xn = vector_mask_to_bitmask(bx, args[0].immediate(), in_elem_bitwidth, in_len);
// Bitcast <i1 x N> to iN:
let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
match ret_ty.kind() {
ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
// Zero-extend iN to the bitmask type:
return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
}
ty::Array(elem, len)
if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
&& len
.try_to_target_usize(bx.tcx)
.expect("expected monomorphic const in codegen")
== expected_bytes =>
{
// Zero-extend iN to the array length:
let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
// Convert the integer to a byte array
let ptr = bx.alloca(Size::from_bytes(expected_bytes), Align::ONE);
bx.store(ze, ptr, Align::ONE);
let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
return Ok(bx.load(array_ty, ptr, Align::ONE));
}
_ => return_error!(InvalidMonomorphization::CannotReturn {
span,
name,
ret_ty,
expected_int_bits,
expected_bytes
}),
}
}
fn simd_simple_float_intrinsic<'ll, 'tcx>(
name: Symbol,
in_elem: Ty<'_>,
in_ty: Ty<'_>,
in_len: u64,
bx: &mut Builder<'_, 'll, 'tcx>,
span: Span,
args: &[OperandRef<'tcx, &'ll Value>],
) -> Result<&'ll Value, ErrorGuaranteed> {
macro_rules! return_error {
($diag: expr) => {{
let err = bx.sess().dcx().emit_err($diag);
return Err(err);
}};
}
let ty::Float(f) = in_elem.kind() else {
return_error!(InvalidMonomorphization::BasicFloatType { span, name, ty: in_ty });
};
let elem_ty = bx.cx.type_float_from_ty(*f);
let vec_ty = bx.type_vector(elem_ty, in_len);
let intr_name = match name {
sym::simd_ceil => "llvm.ceil",
sym::simd_fabs => "llvm.fabs",
sym::simd_fcos => "llvm.cos",
sym::simd_fexp2 => "llvm.exp2",
sym::simd_fexp => "llvm.exp",
sym::simd_flog10 => "llvm.log10",
sym::simd_flog2 => "llvm.log2",
sym::simd_flog => "llvm.log",
sym::simd_floor => "llvm.floor",
sym::simd_fma => "llvm.fma",
sym::simd_relaxed_fma => "llvm.fmuladd",
sym::simd_fsin => "llvm.sin",
sym::simd_fsqrt => "llvm.sqrt",
sym::simd_round => "llvm.round",
sym::simd_round_ties_even => "llvm.rint",
sym::simd_trunc => "llvm.trunc",
_ => return_error!(InvalidMonomorphization::UnrecognizedIntrinsic { span, name }),
};
Ok(bx.call_intrinsic(
intr_name,
&[vec_ty],
&args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
))
}
if std::matches!(
name,
sym::simd_ceil
| sym::simd_fabs
| sym::simd_fcos
| sym::simd_fexp2
| sym::simd_fexp
| sym::simd_flog10
| sym::simd_flog2
| sym::simd_flog
| sym::simd_floor
| sym::simd_fma
| sym::simd_fsin
| sym::simd_fsqrt
| sym::simd_relaxed_fma
| sym::simd_round
| sym::simd_round_ties_even
| sym::simd_trunc
) {
return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
}
fn llvm_vector_ty<'ll>(cx: &CodegenCx<'ll, '_>, elem_ty: Ty<'_>, vec_len: u64) -> &'ll Type {
let elem_ty = match *elem_ty.kind() {
ty::Int(v) => cx.type_int_from_ty(v),
ty::Uint(v) => cx.type_uint_from_ty(v),
ty::Float(v) => cx.type_float_from_ty(v),
ty::RawPtr(_, _) => cx.type_ptr(),
_ => unreachable!(),
};
cx.type_vector(elem_ty, vec_len)
}
if name == sym::simd_gather {
// simd_gather(values: <N x T>, pointers: <N x *_ T>,
// mask: <N x i{M}>) -> <N x T>
// * N: number of elements in the input vectors
// * T: type of the element to load
// * M: any integer width is supported, will be truncated to i1
// All types must be simd vector types
// The second argument must be a simd vector with an element type that's a pointer
// to the element type of the first argument
let (_, element_ty0) = require_simd!(in_ty, SimdFirst);
let (out_len, element_ty1) = require_simd!(args[1].layout.ty, SimdSecond);
// The element type of the third argument must be a signed integer type of any width:
let (out_len2, element_ty2) = require_simd!(args[2].layout.ty, SimdThird);
require_simd!(ret_ty, SimdReturn);
// Of the same length:
require!(
in_len == out_len,
InvalidMonomorphization::SecondArgumentLength {
span,
name,
in_len,
in_ty,
arg_ty: args[1].layout.ty,
out_len
}
);
require!(
in_len == out_len2,
InvalidMonomorphization::ThirdArgumentLength {
span,
name,
in_len,
in_ty,
arg_ty: args[2].layout.ty,
out_len: out_len2
}
);
// The return type must match the first argument type
require!(
ret_ty == in_ty,
InvalidMonomorphization::ExpectedReturnType { span, name, in_ty, ret_ty }
);
require!(
matches!(
*element_ty1.kind(),
ty::RawPtr(p_ty, _) if p_ty == in_elem && p_ty.kind() == element_ty0.kind()
),
InvalidMonomorphization::ExpectedElementType {
span,
name,
expected_element: element_ty1,
second_arg: args[1].layout.ty,
in_elem,
in_ty,
mutability: ExpectedPointerMutability::Not,
}
);
let mask_elem_bitwidth = require_int_or_uint_ty!(
element_ty2.kind(),
InvalidMonomorphization::MaskWrongElementType { span, name, ty: element_ty2 }
);
// Alignment of T, must be a constant integer value:
let alignment = bx.align_of(in_elem).bytes();
// Truncate the mask vector to a vector of i1s:
let mask = vector_mask_to_bitmask(bx, args[2].immediate(), mask_elem_bitwidth, in_len);
// Type of the vector of pointers:
let llvm_pointer_vec_ty = llvm_vector_ty(bx, element_ty1, in_len);
// Type of the vector of elements:
let llvm_elem_vec_ty = llvm_vector_ty(bx, element_ty0, in_len);
let args: &[&'ll Value] = if llvm_version < (22, 0, 0) {
let alignment = bx.const_i32(alignment as i32);
&[args[1].immediate(), alignment, mask, args[0].immediate()]
} else {
&[args[1].immediate(), mask, args[0].immediate()]
};
let call =
bx.call_intrinsic("llvm.masked.gather", &[llvm_elem_vec_ty, llvm_pointer_vec_ty], args);
if llvm_version >= (22, 0, 0) {
crate::attributes::apply_to_callsite(
call,
crate::llvm::AttributePlace::Argument(0),
&[crate::llvm::CreateAlignmentAttr(bx.llcx, alignment)],
)
}
return Ok(call);
}
fn llvm_alignment<'ll, 'tcx>(
bx: &mut Builder<'_, 'll, 'tcx>,
alignment: SimdAlign,
vector_ty: Ty<'tcx>,
element_ty: Ty<'tcx>,
) -> u64 {
match alignment {
SimdAlign::Unaligned => 1,
SimdAlign::Element => bx.align_of(element_ty).bytes(),
SimdAlign::Vector => bx.align_of(vector_ty).bytes(),
}
}
if name == sym::simd_masked_load {
// simd_masked_load<_, _, _, const ALIGN: SimdAlign>(mask: <N x i{M}>, pointer: *_ T, values: <N x T>) -> <N x T>
// * N: number of elements in the input vectors
// * T: type of the element to load
// * M: any integer width is supported, will be truncated to i1
// Loads contiguous elements from memory behind `pointer`, but only for
// those lanes whose `mask` bit is enabled.
// The memory addresses corresponding to the “off” lanes are not accessed.
let alignment = fn_args[3].expect_const().to_branch()[0].to_leaf().to_simd_alignment();
// The element type of the "mask" argument must be a signed integer type of any width
let mask_ty = in_ty;
let (mask_len, mask_elem) = (in_len, in_elem);
// The second argument must be a pointer matching the element type
let pointer_ty = args[1].layout.ty;
// The last argument is a passthrough vector providing values for disabled lanes
let values_ty = args[2].layout.ty;
let (values_len, values_elem) = require_simd!(values_ty, SimdThird);
require_simd!(ret_ty, SimdReturn);
// Of the same length:
require!(
values_len == mask_len,
InvalidMonomorphization::ThirdArgumentLength {
span,
name,
in_len: mask_len,
in_ty: mask_ty,
arg_ty: values_ty,
out_len: values_len
}
);
// The return type must match the last argument type
require!(
ret_ty == values_ty,
InvalidMonomorphization::ExpectedReturnType { span, name, in_ty: values_ty, ret_ty }
);
require!(
matches!(
*pointer_ty.kind(),
ty::RawPtr(p_ty, _) if p_ty == values_elem && p_ty.kind() == values_elem.kind()
),
InvalidMonomorphization::ExpectedElementType {
span,
name,
expected_element: values_elem,
second_arg: pointer_ty,
in_elem: values_elem,
in_ty: values_ty,
mutability: ExpectedPointerMutability::Not,
}
);
let m_elem_bitwidth = require_int_or_uint_ty!(
mask_elem.kind(),
InvalidMonomorphization::MaskWrongElementType { span, name, ty: mask_elem }
);
let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len);
// Alignment of T, must be a constant integer value:
let alignment = llvm_alignment(bx, alignment, values_ty, values_elem);
let llvm_pointer = bx.type_ptr();
// Type of the vector of elements:
let llvm_elem_vec_ty = llvm_vector_ty(bx, values_elem, values_len);
let args: &[&'ll Value] = if llvm_version < (22, 0, 0) {
let alignment = bx.const_i32(alignment as i32);
&[args[1].immediate(), alignment, mask, args[2].immediate()]
} else {
&[args[1].immediate(), mask, args[2].immediate()]
};
let call = bx.call_intrinsic("llvm.masked.load", &[llvm_elem_vec_ty, llvm_pointer], args);
if llvm_version >= (22, 0, 0) {
crate::attributes::apply_to_callsite(
call,
crate::llvm::AttributePlace::Argument(0),
&[crate::llvm::CreateAlignmentAttr(bx.llcx, alignment)],
)
}
return Ok(call);
}
if name == sym::simd_masked_store {
// simd_masked_store<_, _, _, const ALIGN: SimdAlign>(mask: <N x i{M}>, pointer: *mut T, values: <N x T>) -> ()
// * N: number of elements in the input vectors
// * T: type of the element to load
// * M: any integer width is supported, will be truncated to i1
// Stores contiguous elements to memory behind `pointer`, but only for
// those lanes whose `mask` bit is enabled.
// The memory addresses corresponding to the “off” lanes are not accessed.
let alignment = fn_args[3].expect_const().to_branch()[0].to_leaf().to_simd_alignment();
// The element type of the "mask" argument must be a signed integer type of any width
let mask_ty = in_ty;
let (mask_len, mask_elem) = (in_len, in_elem);
// The second argument must be a pointer matching the element type
let pointer_ty = args[1].layout.ty;
// The last argument specifies the values to store to memory
let values_ty = args[2].layout.ty;
let (values_len, values_elem) = require_simd!(values_ty, SimdThird);
// Of the same length:
require!(
values_len == mask_len,
InvalidMonomorphization::ThirdArgumentLength {
span,
name,
in_len: mask_len,
in_ty: mask_ty,
arg_ty: values_ty,
out_len: values_len
}
);
// The second argument must be a mutable pointer type matching the element type
require!(
matches!(
*pointer_ty.kind(),
ty::RawPtr(p_ty, p_mutbl)
if p_ty == values_elem && p_ty.kind() == values_elem.kind() && p_mutbl.is_mut()
),
InvalidMonomorphization::ExpectedElementType {
span,
name,
expected_element: values_elem,
second_arg: pointer_ty,
in_elem: values_elem,
in_ty: values_ty,
mutability: ExpectedPointerMutability::Mut,
}
);
let m_elem_bitwidth = require_int_or_uint_ty!(
mask_elem.kind(),
InvalidMonomorphization::MaskWrongElementType { span, name, ty: mask_elem }
);
let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len);
// Alignment of T, must be a constant integer value:
let alignment = llvm_alignment(bx, alignment, values_ty, values_elem);
let llvm_pointer = bx.type_ptr();
// Type of the vector of elements:
let llvm_elem_vec_ty = llvm_vector_ty(bx, values_elem, values_len);
let args: &[&'ll Value] = if llvm_version < (22, 0, 0) {
let alignment = bx.const_i32(alignment as i32);
&[args[2].immediate(), args[1].immediate(), alignment, mask]
} else {
&[args[2].immediate(), args[1].immediate(), mask]
};
let call = bx.call_intrinsic("llvm.masked.store", &[llvm_elem_vec_ty, llvm_pointer], args);
if llvm_version >= (22, 0, 0) {
crate::attributes::apply_to_callsite(
call,
crate::llvm::AttributePlace::Argument(1),
&[crate::llvm::CreateAlignmentAttr(bx.llcx, alignment)],
)
}
return Ok(call);
}
if name == sym::simd_scatter {
// simd_scatter(values: <N x T>, pointers: <N x *mut T>,
// mask: <N x i{M}>) -> ()
// * N: number of elements in the input vectors
// * T: type of the element to load
// * M: any integer width is supported, will be truncated to i1
// All types must be simd vector types
// The second argument must be a simd vector with an element type that's a pointer
// to the element type of the first argument
let (_, element_ty0) = require_simd!(in_ty, SimdFirst);
let (element_len1, element_ty1) = require_simd!(args[1].layout.ty, SimdSecond);
let (element_len2, element_ty2) = require_simd!(args[2].layout.ty, SimdThird);
// Of the same length:
require!(
in_len == element_len1,
InvalidMonomorphization::SecondArgumentLength {
span,
name,
in_len,
in_ty,
arg_ty: args[1].layout.ty,
out_len: element_len1
}
);
require!(
in_len == element_len2,
InvalidMonomorphization::ThirdArgumentLength {
span,
name,
in_len,
in_ty,
arg_ty: args[2].layout.ty,
out_len: element_len2
}
);
require!(
matches!(
*element_ty1.kind(),
ty::RawPtr(p_ty, p_mutbl)
if p_ty == in_elem && p_mutbl.is_mut() && p_ty.kind() == element_ty0.kind()
),
InvalidMonomorphization::ExpectedElementType {
span,
name,
expected_element: element_ty1,
second_arg: args[1].layout.ty,
in_elem,
in_ty,
mutability: ExpectedPointerMutability::Mut,
}
);
// The element type of the third argument must be an integer type of any width:
let mask_elem_bitwidth = require_int_or_uint_ty!(
element_ty2.kind(),
InvalidMonomorphization::MaskWrongElementType { span, name, ty: element_ty2 }
);
// Alignment of T, must be a constant integer value:
let alignment = bx.align_of(in_elem).bytes();
// Truncate the mask vector to a vector of i1s:
let mask = vector_mask_to_bitmask(bx, args[2].immediate(), mask_elem_bitwidth, in_len);
// Type of the vector of pointers:
let llvm_pointer_vec_ty = llvm_vector_ty(bx, element_ty1, in_len);
// Type of the vector of elements:
let llvm_elem_vec_ty = llvm_vector_ty(bx, element_ty0, in_len);
let args: &[&'ll Value] = if llvm_version < (22, 0, 0) {
let alignment = bx.const_i32(alignment as i32);
&[args[0].immediate(), args[1].immediate(), alignment, mask]
} else {
&[args[0].immediate(), args[1].immediate(), mask]
};
let call = bx.call_intrinsic(
"llvm.masked.scatter",
&[llvm_elem_vec_ty, llvm_pointer_vec_ty],
args,
);
if llvm_version >= (22, 0, 0) {
crate::attributes::apply_to_callsite(
call,
crate::llvm::AttributePlace::Argument(1),
&[crate::llvm::CreateAlignmentAttr(bx.llcx, alignment)],
)
}
return Ok(call);
}
macro_rules! arith_red {
($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
$identity:expr) => {
if name == sym::$name {
require!(
ret_ty == in_elem,
InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty }
);
return match in_elem.kind() {
ty::Int(_) | ty::Uint(_) => {
let r = bx.$integer_reduce(args[0].immediate());
if $ordered {
// if overflow occurs, the result is the
// mathematical result modulo 2^n:
Ok(bx.$op(args[1].immediate(), r))
} else {
Ok(bx.$integer_reduce(args[0].immediate()))
}
}
ty::Float(f) => {
let acc = if $ordered {
// ordered arithmetic reductions take an accumulator
args[1].immediate()
} else {
// unordered arithmetic reductions use the identity accumulator
match f.bit_width() {
32 => bx.const_real(bx.type_f32(), $identity),
64 => bx.const_real(bx.type_f64(), $identity),
v => return_error!(
InvalidMonomorphization::UnsupportedSymbolOfSize {
span,
name,
symbol: sym::$name,
in_ty,
in_elem,
size: v,
ret_ty
}
),
}
};
Ok(bx.$float_reduce(acc, args[0].immediate()))
}
_ => return_error!(InvalidMonomorphization::UnsupportedSymbol {
span,
name,
symbol: sym::$name,
in_ty,
in_elem,
ret_ty
}),
};
}
};
}
arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, -0.0);
arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
arith_red!(
simd_reduce_add_unordered: vector_reduce_add,
vector_reduce_fadd_reassoc,
false,
add,
-0.0
);
arith_red!(
simd_reduce_mul_unordered: vector_reduce_mul,
vector_reduce_fmul_reassoc,
false,
mul,
1.0
);
macro_rules! minmax_red {
($name:ident: $int_red:ident) => {
if name == sym::$name {
require!(
ret_ty == in_elem,
InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty }
);
return match in_elem.kind() {
ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
_ => return_error!(InvalidMonomorphization::UnsupportedSymbol {
span,
name,
symbol: sym::$name,
in_ty,
in_elem,
ret_ty
}),
};
}
};
}
// Currently no support for float due to <https://github.com/llvm/llvm-project/issues/185827>.
minmax_red!(simd_reduce_min: vector_reduce_min);
minmax_red!(simd_reduce_max: vector_reduce_max);
macro_rules! bitwise_red {
($name:ident : $red:ident, $boolean:expr) => {
if name == sym::$name {
let input = if !$boolean {
require!(
ret_ty == in_elem,
InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty }
);
args[0].immediate()
} else {
let bitwidth = match in_elem.kind() {
ty::Int(i) => {
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size().bits())
}
ty::Uint(i) => {
i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size().bits())
}
_ => return_error!(InvalidMonomorphization::UnsupportedSymbol {
span,
name,
symbol: sym::$name,
in_ty,
in_elem,
ret_ty
}),
};
vector_mask_to_bitmask(bx, args[0].immediate(), bitwidth, in_len as _)
};
return match in_elem.kind() {
ty::Int(_) | ty::Uint(_) => {
let r = bx.$red(input);
Ok(r)
}
_ => return_error!(InvalidMonomorphization::UnsupportedSymbol {
span,
name,
symbol: sym::$name,
in_ty,
in_elem,
ret_ty
}),
};
}
};
}
bitwise_red!(simd_reduce_and: vector_reduce_and, false);
bitwise_red!(simd_reduce_or: vector_reduce_or, false);
bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
bitwise_red!(simd_reduce_all: vector_reduce_and, true);
bitwise_red!(simd_reduce_any: vector_reduce_or, true);
if name == sym::simd_cast_ptr {
let (out_len, out_elem) = require_simd!(ret_ty, SimdReturn);
require!(
in_len == out_len,
InvalidMonomorphization::ReturnLengthInputType {
span,
name,
in_len,
in_ty,
ret_ty,
out_len
}
);
match in_elem.kind() {
ty::RawPtr(p_ty, _) => {
let metadata = p_ty.ptr_metadata_ty(bx.tcx, |ty| {
bx.tcx.normalize_erasing_regions(bx.typing_env(), ty)
});
require!(
metadata.is_unit(),
InvalidMonomorphization::CastWidePointer { span, name, ty: in_elem }
);
}
_ => {
return_error!(InvalidMonomorphization::ExpectedPointer { span, name, ty: in_elem })
}
}
match out_elem.kind() {
ty::RawPtr(p_ty, _) => {
let metadata = p_ty.ptr_metadata_ty(bx.tcx, |ty| {
bx.tcx.normalize_erasing_regions(bx.typing_env(), ty)
});
require!(
metadata.is_unit(),
InvalidMonomorphization::CastWidePointer { span, name, ty: out_elem }
);
}
_ => {
return_error!(InvalidMonomorphization::ExpectedPointer { span, name, ty: out_elem })
}
}
return Ok(args[0].immediate());
}
if name == sym::simd_expose_provenance {
let (out_len, out_elem) = require_simd!(ret_ty, SimdReturn);
require!(
in_len == out_len,
InvalidMonomorphization::ReturnLengthInputType {
span,
name,
in_len,
in_ty,
ret_ty,
out_len
}
);
match in_elem.kind() {
ty::RawPtr(_, _) => {}
_ => {
return_error!(InvalidMonomorphization::ExpectedPointer { span, name, ty: in_elem })
}
}
match out_elem.kind() {
ty::Uint(ty::UintTy::Usize) => {}
_ => return_error!(InvalidMonomorphization::ExpectedUsize { span, name, ty: out_elem }),
}
return Ok(bx.ptrtoint(args[0].immediate(), llret_ty));
}
if name == sym::simd_with_exposed_provenance {
let (out_len, out_elem) = require_simd!(ret_ty, SimdReturn);
require!(
in_len == out_len,
InvalidMonomorphization::ReturnLengthInputType {
span,
name,
in_len,
in_ty,
ret_ty,
out_len
}
);
match in_elem.kind() {
ty::Uint(ty::UintTy::Usize) => {}
_ => return_error!(InvalidMonomorphization::ExpectedUsize { span, name, ty: in_elem }),
}
match out_elem.kind() {
ty::RawPtr(_, _) => {}
_ => {
return_error!(InvalidMonomorphization::ExpectedPointer { span, name, ty: out_elem })
}
}
return Ok(bx.inttoptr(args[0].immediate(), llret_ty));
}
if name == sym::simd_cast || name == sym::simd_as {
let (out_len, out_elem, out_num_vecs) = require_simd_or_scalable!(ret_ty, SimdReturn);
require!(
in_len == out_len,
InvalidMonomorphization::ReturnLengthInputType {
span,
name,
in_len,
in_ty,
ret_ty,
out_len
}
);
require!(
in_num_vecs == out_num_vecs,
InvalidMonomorphization::ReturnNumVecsInputType {
span,
name,
in_num_vecs: in_num_vecs.unwrap_or(NumScalableVectors(1)),
in_ty,
ret_ty,
out_num_vecs: out_num_vecs.unwrap_or(NumScalableVectors(1))
}
);
// Casting cares about nominal type, not just structural type
if in_elem == out_elem {
return Ok(args[0].immediate());
}
#[derive(Copy, Clone)]
enum Sign {
Unsigned,
Signed,
}
use Sign::*;
enum Style {
Float,
Int(Sign),
Unsupported,
}
let (in_style, in_width) = match in_elem.kind() {
// vectors of pointer-sized integers should've been
// disallowed before here, so this unwrap is safe.
ty::Int(i) => (
Style::Int(Signed),
i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
),
ty::Uint(u) => (
Style::Int(Unsigned),
u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
),
ty::Float(f) => (Style::Float, f.bit_width()),
_ => (Style::Unsupported, 0),
};
let (out_style, out_width) = match out_elem.kind() {
ty::Int(i) => (
Style::Int(Signed),
i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
),
ty::Uint(u) => (
Style::Int(Unsigned),
u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
),
ty::Float(f) => (Style::Float, f.bit_width()),
_ => (Style::Unsupported, 0),
};
match (in_style, out_style) {
(Style::Int(sign), Style::Int(_)) => {
return Ok(match in_width.cmp(&out_width) {
Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
Ordering::Equal => args[0].immediate(),
Ordering::Less => match sign {
Sign::Signed => bx.sext(args[0].immediate(), llret_ty),
Sign::Unsigned => bx.zext(args[0].immediate(), llret_ty),
},
});
}
(Style::Int(Sign::Signed), Style::Float) => {
return Ok(bx.sitofp(args[0].immediate(), llret_ty));
}
(Style::Int(Sign::Unsigned), Style::Float) => {
return Ok(bx.uitofp(args[0].immediate(), llret_ty));
}
(Style::Float, Style::Int(sign)) => {
return Ok(match (sign, name == sym::simd_as) {
(Sign::Unsigned, false) => bx.fptoui(args[0].immediate(), llret_ty),
(Sign::Signed, false) => bx.fptosi(args[0].immediate(), llret_ty),
(_, true) => bx.cast_float_to_int(
matches!(sign, Sign::Signed),
args[0].immediate(),
llret_ty,
),
});
}
(Style::Float, Style::Float) => {
return Ok(match in_width.cmp(&out_width) {
Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
Ordering::Equal => args[0].immediate(),
Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
});
}
_ => return_error!(InvalidMonomorphization::UnsupportedCast {
span,
name,
in_ty,
in_elem,
ret_ty,
out_elem
}),
}
}
macro_rules! arith_binary {
($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
$(if name == sym::$name {
match in_elem.kind() {
$($(ty::$p(_))|* => {
return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
})*
_ => {},
}
return_error!(
InvalidMonomorphization::UnsupportedOperation { span, name, in_ty, in_elem }
);
})*
}
}
arith_binary! {
simd_add: Uint, Int => add, Float => fadd;
simd_sub: Uint, Int => sub, Float => fsub;
simd_mul: Uint, Int => mul, Float => fmul;
simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
simd_rem: Uint => urem, Int => srem, Float => frem;
simd_shl: Uint, Int => shl;
simd_shr: Uint => lshr, Int => ashr;
simd_and: Uint, Int => and;
simd_or: Uint, Int => or;
simd_xor: Uint, Int => xor;
simd_maximum_number_nsz: Float => maximum_number_nsz;
simd_minimum_number_nsz: Float => minimum_number_nsz;
}
macro_rules! arith_unary {
($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
$(if name == sym::$name {
match in_elem.kind() {
$($(ty::$p(_))|* => {
return Ok(bx.$call(args[0].immediate()))
})*
_ => {},
}
return_error!(
InvalidMonomorphization::UnsupportedOperation { span, name, in_ty, in_elem }
);
})*
}
}
arith_unary! {
simd_neg: Int => neg, Float => fneg;
}
// Unary integer intrinsics
if matches!(
name,
sym::simd_bswap
| sym::simd_bitreverse
| sym::simd_ctlz
| sym::simd_ctpop
| sym::simd_cttz
| sym::simd_carryless_mul
| sym::simd_funnel_shl
| sym::simd_funnel_shr
) {
let vec_ty = bx.cx.type_vector(
match *in_elem.kind() {
ty::Int(i) => bx.cx.type_int_from_ty(i),
ty::Uint(i) => bx.cx.type_uint_from_ty(i),
_ => return_error!(InvalidMonomorphization::UnsupportedOperation {
span,
name,
in_ty,
in_elem
}),
},
in_len as u64,
);
let llvm_intrinsic = match name {
sym::simd_bswap => "llvm.bswap",
sym::simd_bitreverse => "llvm.bitreverse",
sym::simd_ctlz => "llvm.ctlz",
sym::simd_ctpop => "llvm.ctpop",
sym::simd_cttz => "llvm.cttz",
sym::simd_funnel_shl => "llvm.fshl",
sym::simd_funnel_shr => "llvm.fshr",
sym::simd_carryless_mul => "llvm.clmul",
_ => unreachable!(),
};
let int_size = in_elem.int_size_and_signed(bx.tcx()).0.bits();
return match name {
// byte swap is no-op for i8/u8
sym::simd_bswap if int_size == 8 => Ok(args[0].immediate()),
sym::simd_ctlz | sym::simd_cttz => {
// for the (int, i1 immediate) pair, the second arg adds `(0, true) => poison`
let dont_poison_on_zero = bx.const_int(bx.type_i1(), 0);
Ok(bx.call_intrinsic(
llvm_intrinsic,
&[vec_ty],
&[args[0].immediate(), dont_poison_on_zero],
))
}
sym::simd_bswap | sym::simd_bitreverse | sym::simd_ctpop => {
// simple unary argument cases
Ok(bx.call_intrinsic(llvm_intrinsic, &[vec_ty], &[args[0].immediate()]))
}
sym::simd_funnel_shl | sym::simd_funnel_shr => Ok(bx.call_intrinsic(
llvm_intrinsic,
&[vec_ty],
&[args[0].immediate(), args[1].immediate(), args[2].immediate()],
)),
sym::simd_carryless_mul => {
if crate::llvm_util::get_version() >= (22, 0, 0) {
Ok(bx.call_intrinsic(
llvm_intrinsic,
&[vec_ty],
&[args[0].immediate(), args[1].immediate()],
))
} else {
span_bug!(span, "`simd_carryless_mul` needs LLVM 22 or higher");
}
}
_ => unreachable!(),
};
}
if name == sym::simd_arith_offset {
// This also checks that the first operand is a ptr type.
let pointee = in_elem.builtin_deref(true).unwrap_or_else(|| {
span_bug!(span, "must be called with a vector of pointer types as first argument")
});
let layout = bx.layout_of(pointee);
let ptrs = args[0].immediate();
// The second argument must be a ptr-sized integer.
// (We don't care about the signedness, this is wrapping anyway.)
let (_offsets_len, offsets_elem) = args[1].layout.ty.simd_size_and_type(bx.tcx());
if !matches!(offsets_elem.kind(), ty::Int(ty::IntTy::Isize) | ty::Uint(ty::UintTy::Usize)) {
span_bug!(
span,
"must be called with a vector of pointer-sized integers as second argument"
);
}
let offsets = args[1].immediate();
return Ok(bx.gep(bx.backend_type(layout), ptrs, &[offsets]));
}
if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
let lhs = args[0].immediate();
let rhs = args[1].immediate();
let is_add = name == sym::simd_saturating_add;
let (signed, elem_ty) = match *in_elem.kind() {
ty::Int(i) => (true, bx.cx.type_int_from_ty(i)),
ty::Uint(i) => (false, bx.cx.type_uint_from_ty(i)),
_ => {
return_error!(InvalidMonomorphization::ExpectedVectorElementType {
span,
name,
expected_element: args[0].layout.ty.simd_size_and_type(bx.tcx()).1,
vector_type: args[0].layout.ty
});
}
};
let llvm_intrinsic = format!(
"llvm.{}{}.sat",
if signed { 's' } else { 'u' },
if is_add { "add" } else { "sub" },
);
let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
return Ok(bx.call_intrinsic(llvm_intrinsic, &[vec_ty], &[lhs, rhs]));
}
span_bug!(span, "unknown SIMD intrinsic");
}