diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs
index 5217fa2758f79..8a9450c20dda4 100644
--- a/compiler/rustc_codegen_llvm/src/builder.rs
+++ b/compiler/rustc_codegen_llvm/src/builder.rs
@@ -731,27 +731,11 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
}
fn fptoui_sat(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> Option<&'ll Value> {
- if !self.fptoint_sat_broken_in_llvm() {
- let src_ty = self.cx.val_ty(val);
- let float_width = self.cx.float_width(src_ty);
- let int_width = self.cx.int_width(dest_ty);
- let name = format!("llvm.fptoui.sat.i{}.f{}", int_width, float_width);
- return Some(self.call_intrinsic(&name, &[val]));
- }
-
- None
+ self.fptoint_sat(false, val, dest_ty)
}
fn fptosi_sat(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> Option<&'ll Value> {
- if !self.fptoint_sat_broken_in_llvm() {
- let src_ty = self.cx.val_ty(val);
- let float_width = self.cx.float_width(src_ty);
- let int_width = self.cx.int_width(dest_ty);
- let name = format!("llvm.fptosi.sat.i{}.f{}", int_width, float_width);
- return Some(self.call_intrinsic(&name, &[val]));
- }
-
- None
+ self.fptoint_sat(true, val, dest_ty)
}
fn fptoui(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value {
@@ -1455,4 +1439,43 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> {
_ => false,
}
}
+
+ fn fptoint_sat(
+ &mut self,
+ signed: bool,
+ val: &'ll Value,
+ dest_ty: &'ll Type,
+ ) -> Option<&'ll Value> {
+ if !self.fptoint_sat_broken_in_llvm() {
+ let src_ty = self.cx.val_ty(val);
+ let (float_ty, int_ty, vector_length) = if self.cx.type_kind(src_ty) == TypeKind::Vector
+ {
+ assert_eq!(self.cx.vector_length(src_ty), self.cx.vector_length(dest_ty));
+ (
+ self.cx.element_type(src_ty),
+ self.cx.element_type(dest_ty),
+ Some(self.cx.vector_length(src_ty)),
+ )
+ } else {
+ (src_ty, dest_ty, None)
+ };
+ let float_width = self.cx.float_width(float_ty);
+ let int_width = self.cx.int_width(int_ty);
+
+ let instr = if signed { "fptosi" } else { "fptoui" };
+ let name = if let Some(vector_length) = vector_length {
+ format!(
+ "llvm.{}.sat.v{}i{}.v{}f{}",
+ instr, vector_length, int_width, vector_length, float_width
+ )
+ } else {
+ format!("llvm.{}.sat.i{}.f{}", instr, int_width, float_width)
+ };
+ let f =
+ self.declare_cfn(&name, llvm::UnnamedAddr::No, self.type_func(&[src_ty], dest_ty));
+ Some(self.call(self.type_func(&[src_ty], dest_ty), f, &[val], None))
+ } else {
+ None
+ }
+ }
}
diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs
index 07d49b6e72996..8cfa0a6482805 100644
--- a/compiler/rustc_codegen_llvm/src/intrinsic.rs
+++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs
@@ -1689,7 +1689,7 @@ unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
bitwise_red!(simd_reduce_all: vector_reduce_and, true);
bitwise_red!(simd_reduce_any: vector_reduce_or, true);
- if name == sym::simd_cast {
+ if name == sym::simd_cast || name == sym::simd_as {
require_simd!(ret_ty, "return");
let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
require!(
@@ -1715,14 +1715,26 @@ unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
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(true), i.bit_width().unwrap()),
- ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
+ ty::Int(i) => (
+ Style::Int(true),
+ i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
+ ),
+ ty::Uint(u) => (
+ Style::Int(false),
+ 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(true), i.bit_width().unwrap()),
- ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
+ ty::Int(i) => (
+ Style::Int(true),
+ i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
+ ),
+ ty::Uint(u) => (
+ Style::Int(false),
+ u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
+ ),
ty::Float(f) => (Style::Float, f.bit_width()),
_ => (Style::Unsupported, 0),
};
@@ -1749,10 +1761,10 @@ unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
});
}
(Style::Float, Style::Int(out_is_signed)) => {
- return Ok(if out_is_signed {
- bx.fptosi(args[0].immediate(), llret_ty)
- } else {
- bx.fptoui(args[0].immediate(), llret_ty)
+ return Ok(match (out_is_signed, name == sym::simd_as) {
+ (false, false) => bx.fptoui(args[0].immediate(), llret_ty),
+ (true, false) => bx.fptosi(args[0].immediate(), llret_ty),
+ (_, true) => bx.cast_float_to_int(out_is_signed, args[0].immediate(), llret_ty),
});
}
(Style::Float, Style::Float) => {
diff --git a/compiler/rustc_codegen_ssa/src/mir/rvalue.rs b/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
index 679c45767018d..68decce82ab52 100644
--- a/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
@@ -3,11 +3,10 @@ use super::place::PlaceRef;
use super::{FunctionCx, LocalRef};
use crate::base;
-use crate::common::{self, IntPredicate, RealPredicate};
+use crate::common::{self, IntPredicate};
use crate::traits::*;
use crate::MemFlags;
-use rustc_apfloat::{ieee, Float, Round, Status};
use rustc_middle::mir;
use rustc_middle::ty::cast::{CastTy, IntTy};
use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf};
@@ -368,10 +367,10 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
bx.inttoptr(usize_llval, ll_t_out)
}
(CastTy::Float, CastTy::Int(IntTy::I)) => {
- cast_float_to_int(&mut bx, true, llval, ll_t_in, ll_t_out)
+ bx.cast_float_to_int(true, llval, ll_t_out)
}
(CastTy::Float, CastTy::Int(_)) => {
- cast_float_to_int(&mut bx, false, llval, ll_t_in, ll_t_out)
+ bx.cast_float_to_int(false, llval, ll_t_out)
}
_ => bug!("unsupported cast: {:?} to {:?}", operand.layout.ty, cast.ty),
};
@@ -768,146 +767,3 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
// (*) this is only true if the type is suitable
}
}
-
-fn cast_float_to_int<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
- bx: &mut Bx,
- signed: bool,
- x: Bx::Value,
- float_ty: Bx::Type,
- int_ty: Bx::Type,
-) -> Bx::Value {
- if let Some(false) = bx.cx().sess().opts.debugging_opts.saturating_float_casts {
- return if signed { bx.fptosi(x, int_ty) } else { bx.fptoui(x, int_ty) };
- }
-
- let try_sat_result = if signed { bx.fptosi_sat(x, int_ty) } else { bx.fptoui_sat(x, int_ty) };
- if let Some(try_sat_result) = try_sat_result {
- return try_sat_result;
- }
-
- let int_width = bx.cx().int_width(int_ty);
- let float_width = bx.cx().float_width(float_ty);
- // LLVM's fpto[su]i returns undef when the input x is infinite, NaN, or does not fit into the
- // destination integer type after rounding towards zero. This `undef` value can cause UB in
- // safe code (see issue #10184), so we implement a saturating conversion on top of it:
- // Semantically, the mathematical value of the input is rounded towards zero to the next
- // mathematical integer, and then the result is clamped into the range of the destination
- // integer type. Positive and negative infinity are mapped to the maximum and minimum value of
- // the destination integer type. NaN is mapped to 0.
- //
- // Define f_min and f_max as the largest and smallest (finite) floats that are exactly equal to
- // a value representable in int_ty.
- // They are exactly equal to int_ty::{MIN,MAX} if float_ty has enough significand bits.
- // Otherwise, int_ty::MAX must be rounded towards zero, as it is one less than a power of two.
- // int_ty::MIN, however, is either zero or a negative power of two and is thus exactly
- // representable. Note that this only works if float_ty's exponent range is sufficiently large.
- // f16 or 256 bit integers would break this property. Right now the smallest float type is f32
- // with exponents ranging up to 127, which is barely enough for i128::MIN = -2^127.
- // On the other hand, f_max works even if int_ty::MAX is greater than float_ty::MAX. Because
- // we're rounding towards zero, we just get float_ty::MAX (which is always an integer).
- // This already happens today with u128::MAX = 2^128 - 1 > f32::MAX.
- let int_max = |signed: bool, int_width: u64| -> u128 {
- let shift_amount = 128 - int_width;
- if signed { i128::MAX as u128 >> shift_amount } else { u128::MAX >> shift_amount }
- };
- let int_min = |signed: bool, int_width: u64| -> i128 {
- if signed { i128::MIN >> (128 - int_width) } else { 0 }
- };
-
- let compute_clamp_bounds_single = |signed: bool, int_width: u64| -> (u128, u128) {
- let rounded_min = ieee::Single::from_i128_r(int_min(signed, int_width), Round::TowardZero);
- assert_eq!(rounded_min.status, Status::OK);
- let rounded_max = ieee::Single::from_u128_r(int_max(signed, int_width), Round::TowardZero);
- assert!(rounded_max.value.is_finite());
- (rounded_min.value.to_bits(), rounded_max.value.to_bits())
- };
- let compute_clamp_bounds_double = |signed: bool, int_width: u64| -> (u128, u128) {
- let rounded_min = ieee::Double::from_i128_r(int_min(signed, int_width), Round::TowardZero);
- assert_eq!(rounded_min.status, Status::OK);
- let rounded_max = ieee::Double::from_u128_r(int_max(signed, int_width), Round::TowardZero);
- assert!(rounded_max.value.is_finite());
- (rounded_min.value.to_bits(), rounded_max.value.to_bits())
- };
-
- let mut float_bits_to_llval = |bits| {
- let bits_llval = match float_width {
- 32 => bx.cx().const_u32(bits as u32),
- 64 => bx.cx().const_u64(bits as u64),
- n => bug!("unsupported float width {}", n),
- };
- bx.bitcast(bits_llval, float_ty)
- };
- let (f_min, f_max) = match float_width {
- 32 => compute_clamp_bounds_single(signed, int_width),
- 64 => compute_clamp_bounds_double(signed, int_width),
- n => bug!("unsupported float width {}", n),
- };
- let f_min = float_bits_to_llval(f_min);
- let f_max = float_bits_to_llval(f_max);
- // To implement saturation, we perform the following steps:
- //
- // 1. Cast x to an integer with fpto[su]i. This may result in undef.
- // 2. Compare x to f_min and f_max, and use the comparison results to select:
- // a) int_ty::MIN if x < f_min or x is NaN
- // b) int_ty::MAX if x > f_max
- // c) the result of fpto[su]i otherwise
- // 3. If x is NaN, return 0.0, otherwise return the result of step 2.
- //
- // This avoids resulting undef because values in range [f_min, f_max] by definition fit into the
- // destination type. It creates an undef temporary, but *producing* undef is not UB. Our use of
- // undef does not introduce any non-determinism either.
- // More importantly, the above procedure correctly implements saturating conversion.
- // Proof (sketch):
- // If x is NaN, 0 is returned by definition.
- // Otherwise, x is finite or infinite and thus can be compared with f_min and f_max.
- // This yields three cases to consider:
- // (1) if x in [f_min, f_max], the result of fpto[su]i is returned, which agrees with
- // saturating conversion for inputs in that range.
- // (2) if x > f_max, then x is larger than int_ty::MAX. This holds even if f_max is rounded
- // (i.e., if f_max < int_ty::MAX) because in those cases, nextUp(f_max) is already larger
- // than int_ty::MAX. Because x is larger than int_ty::MAX, the return value of int_ty::MAX
- // is correct.
- // (3) if x < f_min, then x is smaller than int_ty::MIN. As shown earlier, f_min exactly equals
- // int_ty::MIN and therefore the return value of int_ty::MIN is correct.
- // QED.
-
- let int_max = bx.cx().const_uint_big(int_ty, int_max(signed, int_width));
- let int_min = bx.cx().const_uint_big(int_ty, int_min(signed, int_width) as u128);
- let zero = bx.cx().const_uint(int_ty, 0);
-
- // Step 1 ...
- let fptosui_result = if signed { bx.fptosi(x, int_ty) } else { bx.fptoui(x, int_ty) };
- let less_or_nan = bx.fcmp(RealPredicate::RealULT, x, f_min);
- let greater = bx.fcmp(RealPredicate::RealOGT, x, f_max);
-
- // Step 2: We use two comparisons and two selects, with %s1 being the
- // result:
- // %less_or_nan = fcmp ult %x, %f_min
- // %greater = fcmp olt %x, %f_max
- // %s0 = select %less_or_nan, int_ty::MIN, %fptosi_result
- // %s1 = select %greater, int_ty::MAX, %s0
- // Note that %less_or_nan uses an *unordered* comparison. This
- // comparison is true if the operands are not comparable (i.e., if x is
- // NaN). The unordered comparison ensures that s1 becomes int_ty::MIN if
- // x is NaN.
- //
- // Performance note: Unordered comparison can be lowered to a "flipped"
- // comparison and a negation, and the negation can be merged into the
- // select. Therefore, it not necessarily any more expensive than an
- // ordered ("normal") comparison. Whether these optimizations will be
- // performed is ultimately up to the backend, but at least x86 does
- // perform them.
- let s0 = bx.select(less_or_nan, int_min, fptosui_result);
- let s1 = bx.select(greater, int_max, s0);
-
- // Step 3: NaN replacement.
- // For unsigned types, the above step already yielded int_ty::MIN == 0 if x is NaN.
- // Therefore we only need to execute this step for signed integer types.
- if signed {
- // LLVM has no isNaN predicate, so we use (x == x) instead
- let cmp = bx.fcmp(RealPredicate::RealOEQ, x, x);
- bx.select(cmp, s1, zero)
- } else {
- s1
- }
-}
diff --git a/compiler/rustc_codegen_ssa/src/traits/builder.rs b/compiler/rustc_codegen_ssa/src/traits/builder.rs
index 48d88095855d1..5a06fb4610587 100644
--- a/compiler/rustc_codegen_ssa/src/traits/builder.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/builder.rs
@@ -1,18 +1,21 @@
use super::abi::AbiBuilderMethods;
use super::asm::AsmBuilderMethods;
+use super::consts::ConstMethods;
use super::coverageinfo::CoverageInfoBuilderMethods;
use super::debuginfo::DebugInfoBuilderMethods;
use super::intrinsic::IntrinsicCallMethods;
-use super::type_::ArgAbiMethods;
+use super::misc::MiscMethods;
+use super::type_::{ArgAbiMethods, BaseTypeMethods};
use super::{HasCodegen, StaticBuilderMethods};
use crate::common::{
- AtomicOrdering, AtomicRmwBinOp, IntPredicate, RealPredicate, SynchronizationScope,
+ AtomicOrdering, AtomicRmwBinOp, IntPredicate, RealPredicate, SynchronizationScope, TypeKind,
};
use crate::mir::operand::OperandRef;
use crate::mir::place::PlaceRef;
use crate::MemFlags;
+use rustc_apfloat::{ieee, Float, Round, Status};
use rustc_middle::ty::layout::{HasParamEnv, TyAndLayout};
use rustc_middle::ty::Ty;
use rustc_span::Span;
@@ -202,6 +205,179 @@ pub trait BuilderMethods<'a, 'tcx>:
fn intcast(&mut self, val: Self::Value, dest_ty: Self::Type, is_signed: bool) -> Self::Value;
fn pointercast(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value;
+ fn cast_float_to_int(
+ &mut self,
+ signed: bool,
+ x: Self::Value,
+ dest_ty: Self::Type,
+ ) -> Self::Value {
+ let in_ty = self.cx().val_ty(x);
+ let (float_ty, int_ty) = if self.cx().type_kind(dest_ty) == TypeKind::Vector
+ && self.cx().type_kind(in_ty) == TypeKind::Vector
+ {
+ (self.cx().element_type(in_ty), self.cx().element_type(dest_ty))
+ } else {
+ (in_ty, dest_ty)
+ };
+ assert!(matches!(self.cx().type_kind(float_ty), TypeKind::Float | TypeKind::Double));
+ assert_eq!(self.cx().type_kind(int_ty), TypeKind::Integer);
+
+ if let Some(false) = self.cx().sess().opts.debugging_opts.saturating_float_casts {
+ return if signed { self.fptosi(x, dest_ty) } else { self.fptoui(x, dest_ty) };
+ }
+
+ let try_sat_result =
+ if signed { self.fptosi_sat(x, dest_ty) } else { self.fptoui_sat(x, dest_ty) };
+ if let Some(try_sat_result) = try_sat_result {
+ return try_sat_result;
+ }
+
+ let int_width = self.cx().int_width(int_ty);
+ let float_width = self.cx().float_width(float_ty);
+ // LLVM's fpto[su]i returns undef when the input x is infinite, NaN, or does not fit into the
+ // destination integer type after rounding towards zero. This `undef` value can cause UB in
+ // safe code (see issue #10184), so we implement a saturating conversion on top of it:
+ // Semantically, the mathematical value of the input is rounded towards zero to the next
+ // mathematical integer, and then the result is clamped into the range of the destination
+ // integer type. Positive and negative infinity are mapped to the maximum and minimum value of
+ // the destination integer type. NaN is mapped to 0.
+ //
+ // Define f_min and f_max as the largest and smallest (finite) floats that are exactly equal to
+ // a value representable in int_ty.
+ // They are exactly equal to int_ty::{MIN,MAX} if float_ty has enough significand bits.
+ // Otherwise, int_ty::MAX must be rounded towards zero, as it is one less than a power of two.
+ // int_ty::MIN, however, is either zero or a negative power of two and is thus exactly
+ // representable. Note that this only works if float_ty's exponent range is sufficiently large.
+ // f16 or 256 bit integers would break this property. Right now the smallest float type is f32
+ // with exponents ranging up to 127, which is barely enough for i128::MIN = -2^127.
+ // On the other hand, f_max works even if int_ty::MAX is greater than float_ty::MAX. Because
+ // we're rounding towards zero, we just get float_ty::MAX (which is always an integer).
+ // This already happens today with u128::MAX = 2^128 - 1 > f32::MAX.
+ let int_max = |signed: bool, int_width: u64| -> u128 {
+ let shift_amount = 128 - int_width;
+ if signed { i128::MAX as u128 >> shift_amount } else { u128::MAX >> shift_amount }
+ };
+ let int_min = |signed: bool, int_width: u64| -> i128 {
+ if signed { i128::MIN >> (128 - int_width) } else { 0 }
+ };
+
+ let compute_clamp_bounds_single = |signed: bool, int_width: u64| -> (u128, u128) {
+ let rounded_min =
+ ieee::Single::from_i128_r(int_min(signed, int_width), Round::TowardZero);
+ assert_eq!(rounded_min.status, Status::OK);
+ let rounded_max =
+ ieee::Single::from_u128_r(int_max(signed, int_width), Round::TowardZero);
+ assert!(rounded_max.value.is_finite());
+ (rounded_min.value.to_bits(), rounded_max.value.to_bits())
+ };
+ let compute_clamp_bounds_double = |signed: bool, int_width: u64| -> (u128, u128) {
+ let rounded_min =
+ ieee::Double::from_i128_r(int_min(signed, int_width), Round::TowardZero);
+ assert_eq!(rounded_min.status, Status::OK);
+ let rounded_max =
+ ieee::Double::from_u128_r(int_max(signed, int_width), Round::TowardZero);
+ assert!(rounded_max.value.is_finite());
+ (rounded_min.value.to_bits(), rounded_max.value.to_bits())
+ };
+ // To implement saturation, we perform the following steps:
+ //
+ // 1. Cast x to an integer with fpto[su]i. This may result in undef.
+ // 2. Compare x to f_min and f_max, and use the comparison results to select:
+ // a) int_ty::MIN if x < f_min or x is NaN
+ // b) int_ty::MAX if x > f_max
+ // c) the result of fpto[su]i otherwise
+ // 3. If x is NaN, return 0.0, otherwise return the result of step 2.
+ //
+ // This avoids resulting undef because values in range [f_min, f_max] by definition fit into the
+ // destination type. It creates an undef temporary, but *producing* undef is not UB. Our use of
+ // undef does not introduce any non-determinism either.
+ // More importantly, the above procedure correctly implements saturating conversion.
+ // Proof (sketch):
+ // If x is NaN, 0 is returned by definition.
+ // Otherwise, x is finite or infinite and thus can be compared with f_min and f_max.
+ // This yields three cases to consider:
+ // (1) if x in [f_min, f_max], the result of fpto[su]i is returned, which agrees with
+ // saturating conversion for inputs in that range.
+ // (2) if x > f_max, then x is larger than int_ty::MAX. This holds even if f_max is rounded
+ // (i.e., if f_max < int_ty::MAX) because in those cases, nextUp(f_max) is already larger
+ // than int_ty::MAX. Because x is larger than int_ty::MAX, the return value of int_ty::MAX
+ // is correct.
+ // (3) if x < f_min, then x is smaller than int_ty::MIN. As shown earlier, f_min exactly equals
+ // int_ty::MIN and therefore the return value of int_ty::MIN is correct.
+ // QED.
+
+ let float_bits_to_llval = |bx: &mut Self, bits| {
+ let bits_llval = match float_width {
+ 32 => bx.cx().const_u32(bits as u32),
+ 64 => bx.cx().const_u64(bits as u64),
+ n => bug!("unsupported float width {}", n),
+ };
+ bx.bitcast(bits_llval, float_ty)
+ };
+ let (f_min, f_max) = match float_width {
+ 32 => compute_clamp_bounds_single(signed, int_width),
+ 64 => compute_clamp_bounds_double(signed, int_width),
+ n => bug!("unsupported float width {}", n),
+ };
+ let f_min = float_bits_to_llval(self, f_min);
+ let f_max = float_bits_to_llval(self, f_max);
+ let int_max = self.cx().const_uint_big(int_ty, int_max(signed, int_width));
+ let int_min = self.cx().const_uint_big(int_ty, int_min(signed, int_width) as u128);
+ let zero = self.cx().const_uint(int_ty, 0);
+
+ // If we're working with vectors, constants must be "splatted": the constant is duplicated
+ // into each lane of the vector. The algorithm stays the same, we are just using the
+ // same constant across all lanes.
+ let maybe_splat = |bx: &mut Self, val| {
+ if bx.cx().type_kind(dest_ty) == TypeKind::Vector {
+ bx.vector_splat(bx.vector_length(dest_ty), val)
+ } else {
+ val
+ }
+ };
+ let f_min = maybe_splat(self, f_min);
+ let f_max = maybe_splat(self, f_max);
+ let int_max = maybe_splat(self, int_max);
+ let int_min = maybe_splat(self, int_min);
+ let zero = maybe_splat(self, zero);
+
+ // Step 1 ...
+ let fptosui_result = if signed { self.fptosi(x, dest_ty) } else { self.fptoui(x, dest_ty) };
+ let less_or_nan = self.fcmp(RealPredicate::RealULT, x, f_min);
+ let greater = self.fcmp(RealPredicate::RealOGT, x, f_max);
+
+ // Step 2: We use two comparisons and two selects, with %s1 being the
+ // result:
+ // %less_or_nan = fcmp ult %x, %f_min
+ // %greater = fcmp olt %x, %f_max
+ // %s0 = select %less_or_nan, int_ty::MIN, %fptosi_result
+ // %s1 = select %greater, int_ty::MAX, %s0
+ // Note that %less_or_nan uses an *unordered* comparison. This
+ // comparison is true if the operands are not comparable (i.e., if x is
+ // NaN). The unordered comparison ensures that s1 becomes int_ty::MIN if
+ // x is NaN.
+ //
+ // Performance note: Unordered comparison can be lowered to a "flipped"
+ // comparison and a negation, and the negation can be merged into the
+ // select. Therefore, it not necessarily any more expensive than an
+ // ordered ("normal") comparison. Whether these optimizations will be
+ // performed is ultimately up to the backend, but at least x86 does
+ // perform them.
+ let s0 = self.select(less_or_nan, int_min, fptosui_result);
+ let s1 = self.select(greater, int_max, s0);
+
+ // Step 3: NaN replacement.
+ // For unsigned types, the above step already yielded int_ty::MIN == 0 if x is NaN.
+ // Therefore we only need to execute this step for signed integer types.
+ if signed {
+ // LLVM has no isNaN predicate, so we use (x == x) instead
+ let cmp = self.fcmp(RealPredicate::RealOEQ, x, x);
+ self.select(cmp, s1, zero)
+ } else {
+ s1
+ }
+ }
+
fn icmp(&mut self, op: IntPredicate, lhs: Self::Value, rhs: Self::Value) -> Self::Value;
fn fcmp(&mut self, op: RealPredicate, lhs: Self::Value, rhs: Self::Value) -> Self::Value;
diff --git a/compiler/rustc_span/src/symbol.rs b/compiler/rustc_span/src/symbol.rs
index 84cf8878af809..730438576176e 100644
--- a/compiler/rustc_span/src/symbol.rs
+++ b/compiler/rustc_span/src/symbol.rs
@@ -1186,6 +1186,7 @@ symbols! {
simd,
simd_add,
simd_and,
+ simd_as,
simd_bitmask,
simd_cast,
simd_ceil,
diff --git a/compiler/rustc_typeck/src/check/intrinsic.rs b/compiler/rustc_typeck/src/check/intrinsic.rs
index 6314f2aba4efe..4c612ed5be51a 100644
--- a/compiler/rustc_typeck/src/check/intrinsic.rs
+++ b/compiler/rustc_typeck/src/check/intrinsic.rs
@@ -453,7 +453,7 @@ pub fn check_platform_intrinsic_type(tcx: TyCtxt<'_>, it: &hir::ForeignItem<'_>)
sym::simd_scatter => (3, vec![param(0), param(1), param(2)], tcx.mk_unit()),
sym::simd_insert => (2, vec![param(0), tcx.types.u32, param(1)], param(0)),
sym::simd_extract => (2, vec![param(0), tcx.types.u32], param(1)),
- sym::simd_cast => (2, vec![param(0)], param(1)),
+ sym::simd_cast | sym::simd_as => (2, vec![param(0)], param(1)),
sym::simd_bitmask => (2, vec![param(0)], param(1)),
sym::simd_select | sym::simd_select_bitmask => {
(2, vec![param(0), param(1), param(1)], param(1))
diff --git a/src/test/ui/simd/intrinsic/generic-as.rs b/src/test/ui/simd/intrinsic/generic-as.rs
new file mode 100644
index 0000000000000..a975190a2fafd
--- /dev/null
+++ b/src/test/ui/simd/intrinsic/generic-as.rs
@@ -0,0 +1,48 @@
+// run-pass
+
+#![feature(repr_simd, platform_intrinsics)]
+
+extern "platform-intrinsic" {
+ fn simd_as<T, U>(x: T) -> U;
+}
+
+#[derive(Copy, Clone)]
+#[repr(simd)]
+struct V<T>([T; 2]);
+
+fn main() {
+ unsafe {
+ let u = V::<u32>([u32::MIN, u32::MAX]);
+ let i: V<i16> = simd_as(u);
+ assert_eq!(i.0[0], u.0[0] as i16);
+ assert_eq!(i.0[1], u.0[1] as i16);
+ }
+
+ unsafe {
+ let f = V::<f32>([f32::MIN, f32::MAX]);
+ let i: V<i16> = simd_as(f);
+ assert_eq!(i.0[0], f.0[0] as i16);
+ assert_eq!(i.0[1], f.0[1] as i16);
+ }
+
+ unsafe {
+ let f = V::<f32>([f32::MIN, f32::MAX]);
+ let u: V<u8> = simd_as(f);
+ assert_eq!(u.0[0], f.0[0] as u8);
+ assert_eq!(u.0[1], f.0[1] as u8);
+ }
+
+ unsafe {
+ let f = V::<f64>([f64::MIN, f64::MAX]);
+ let i: V<isize> = simd_as(f);
+ assert_eq!(i.0[0], f.0[0] as isize);
+ assert_eq!(i.0[1], f.0[1] as isize);
+ }
+
+ unsafe {
+ let f = V::<f64>([f64::MIN, f64::MAX]);
+ let u: V<usize> = simd_as(f);
+ assert_eq!(u.0[0], f.0[0] as usize);
+ assert_eq!(u.0[1], f.0[1] as usize);
+ }
+}
diff --git a/src/test/ui/simd/intrinsic/generic-cast-pointer-width.rs b/src/test/ui/simd/intrinsic/generic-cast-pointer-width.rs
new file mode 100644
index 0000000000000..b9382310deb2c
--- /dev/null
+++ b/src/test/ui/simd/intrinsic/generic-cast-pointer-width.rs
@@ -0,0 +1,21 @@
+// run-pass
+#![feature(repr_simd, platform_intrinsics)]
+
+extern "platform-intrinsic" {
+ fn simd_cast<T, U>(x: T) -> U;
+}
+
+#[derive(Copy, Clone)]
+#[repr(simd)]
+struct V<T>([T; 4]);
+
+fn main() {
+ let u = V::<usize>([0, 1, 2, 3]);
+ let uu32: V<u32> = unsafe { simd_cast(u) };
+ let ui64: V<i64> = unsafe { simd_cast(u) };
+
+ for (u, (uu32, ui64)) in u.0.iter().zip(uu32.0.iter().zip(ui64.0.iter())) {
+ assert_eq!(*u as u32, *uu32);
+ assert_eq!(*u as i64, *ui64);
+ }
+}