Saturating casts between integers and floats

src/librustc/session/config.rs

@@ -1107,6 +1107,9 @@ options! {DebuggingOptions, DebuggingSetter, basic_debugging_options,
         "control whether #[inline] functions are in all cgus"),
     tls_model: Option<String> = (None, parse_opt_string, [TRACKED],
          "choose the TLS model to use (rustc --print tls-models for details)"),
+    saturating_float_casts: bool = (false, parse_bool, [TRACKED],
+        "make casts between integers and floats safe: clip out-of-range inputs to the min/max \
+         integer or to infinity respectively, and turn `NAN` into 0 when casting to integers"),
 }
 
 pub fn default_lib_output() -> CrateType {

src/librustc_apfloat/lib.rs

@@ -96,7 +96,7 @@ impl Status {
 }
 
 impl<T> StatusAnd<T> {
-    fn map<F: FnOnce(T) -> U, U>(self, f: F) -> StatusAnd<U> {
+    pub fn map<F: FnOnce(T) -> U, U>(self, f: F) -> StatusAnd<U> {
         StatusAnd {
             status: self.status,
             value: f(self.value),
@@ -378,7 +378,7 @@ pub trait Float
     fn from_bits(input: u128) -> Self;
     fn from_i128_r(input: i128, round: Round) -> StatusAnd<Self> {
         if input < 0 {
-            Self::from_u128_r(-input as u128, -round).map(|r| -r)
+            Self::from_u128_r(input.wrapping_neg() as u128, -round).map(|r| -r)
         } else {
             Self::from_u128_r(input as u128, round)
         }

src/librustc_const_math/float.rs

@@ -203,3 +203,11 @@ impl ::std::ops::Neg for ConstFloat {
         ConstFloat { bits, ty: self.ty }
     }
 }
+
+/// This is `f32::MAX + (0.5 ULP)` as an integer. Numbers greater or equal to this
+/// are rounded to infinity when converted to `f32`.
+///
+/// NB: Computed as maximum significand with an extra 1 bit added (for the half ULP)
+/// shifted by the maximum exponent (accounting for normalization).
+pub const MAX_F32_PLUS_HALF_ULP: u128 = ((1 << (Single::PRECISION + 1)) - 1)
+                                        << (Single::MAX_EXP - Single::PRECISION as i16);

src/librustc_trans/Cargo.toml

@@ -19,6 +19,7 @@ owning_ref = "0.3.3"
 rustc-demangle = "0.1.4"
 rustc = { path = "../librustc" }
 rustc_allocator = { path = "../librustc_allocator" }
+rustc_apfloat = { path = "../librustc_apfloat" }
 rustc_back = { path = "../librustc_back" }
 rustc_const_math = { path = "../librustc_const_math" }
 rustc_data_structures = { path = "../librustc_data_structures" }

src/librustc_trans/lib.rs

@@ -24,6 +24,7 @@
 #![feature(custom_attribute)]
 #![allow(unused_attributes)]
 #![feature(i128_type)]
+#![feature(i128)]
 #![feature(libc)]
 #![feature(quote)]
 #![feature(rustc_diagnostic_macros)]
@@ -43,6 +44,7 @@ extern crate libc;
 extern crate owning_ref;
 #[macro_use] extern crate rustc;
 extern crate rustc_allocator;
+extern crate rustc_apfloat;
 extern crate rustc_back;
 extern crate rustc_data_structures;
 extern crate rustc_incremental;

src/librustc_trans/mir/constant.rs

@@ -11,7 +11,7 @@
 use llvm::{self, ValueRef};
 use rustc::middle::const_val::{ConstEvalErr, ConstVal, ErrKind};
 use rustc_const_math::ConstInt::*;
-use rustc_const_math::{ConstInt, ConstMathErr};
+use rustc_const_math::{ConstInt, ConstMathErr, MAX_F32_PLUS_HALF_ULP};
 use rustc::hir::def_id::DefId;
 use rustc::infer::TransNormalize;
 use rustc::traits;
@@ -21,6 +21,7 @@ use rustc::ty::{self, Ty, TyCtxt, TypeFoldable};
 use rustc::ty::layout::{self, LayoutTyper};
 use rustc::ty::cast::{CastTy, IntTy};
 use rustc::ty::subst::{Kind, Substs, Subst};
+use rustc_apfloat::{ieee, Float, Status};
 use rustc_data_structures::indexed_vec::{Idx, IndexVec};
 use {adt, base, machine};
 use abi::{self, Abi};
@@ -689,20 +690,18 @@ impl<'a, 'tcx> MirConstContext<'a, 'tcx> {
                                     llvm::LLVMConstIntCast(llval, ll_t_out.to_ref(), s)
                                 }
                                 (CastTy::Int(_), CastTy::Float) => {
-                                    if signed {
-                                        llvm::LLVMConstSIToFP(llval, ll_t_out.to_ref())
-                                    } else {
-                                        llvm::LLVMConstUIToFP(llval, ll_t_out.to_ref())
-                                    }
+                                    cast_const_int_to_float(self.ccx, llval, signed, ll_t_out)
                                 }
                                 (CastTy::Float, CastTy::Float) => {
                                     llvm::LLVMConstFPCast(llval, ll_t_out.to_ref())
                                 }
                                 (CastTy::Float, CastTy::Int(IntTy::I)) => {
-                                    llvm::LLVMConstFPToSI(llval, ll_t_out.to_ref())
+                                    cast_const_float_to_int(self.ccx, &operand,
+                                                            true, ll_t_out, span)
                                 }
                                 (CastTy::Float, CastTy::Int(_)) => {
-                                    llvm::LLVMConstFPToUI(llval, ll_t_out.to_ref())
+                                    cast_const_float_to_int(self.ccx, &operand,
+                                                            false, ll_t_out, span)
                                 }
                                 (CastTy::Ptr(_), CastTy::Ptr(_)) |
                                 (CastTy::FnPtr, CastTy::Ptr(_)) |
@@ -955,6 +954,64 @@ pub fn const_scalar_checked_binop<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
     }
 }
 
+unsafe fn cast_const_float_to_int(ccx: &CrateContext,
+                                  operand: &Const,
+                                  signed: bool,
+                                  int_ty: Type,
+                                  span: Span) -> ValueRef {
+    let llval = operand.llval;
+    let float_bits = match operand.ty.sty {
+        ty::TyFloat(fty) => fty.bit_width(),
+        _ => bug!("cast_const_float_to_int: operand not a float"),
+    };
+    // Note: this breaks if llval is a complex constant expression rather than a simple constant.
+    // One way that might happen would be if addresses could be turned into integers in constant
+    // expressions, but that doesn't appear to be possible?
+    // In any case, an ICE is better than producing undef.
+    let llval_bits = consts::bitcast(llval, Type::ix(ccx, float_bits as u64));
+    let bits = const_to_opt_u128(llval_bits, false).unwrap_or_else(|| {
+        panic!("could not get bits of constant float {:?}",
+               Value(llval));
+    });
+    let int_width = int_ty.int_width() as usize;
+    // Try to convert, but report an error for overflow and NaN. This matches HIR const eval.
+    let cast_result = match float_bits {
+        32 if signed => ieee::Single::from_bits(bits).to_i128(int_width).map(|v| v as u128),
+        64 if signed => ieee::Double::from_bits(bits).to_i128(int_width).map(|v| v as u128),
+        32 => ieee::Single::from_bits(bits).to_u128(int_width),
+        64 => ieee::Double::from_bits(bits).to_u128(int_width),
+        n => bug!("unsupported float width {}", n),
+    };
+    if cast_result.status.contains(Status::INVALID_OP) {
+        let err = ConstEvalErr { span: span, kind: ErrKind::CannotCast };
+        err.report(ccx.tcx(), span, "expression");
+    }
+    C_big_integral(int_ty, cast_result.value)
+}
+
+unsafe fn cast_const_int_to_float(ccx: &CrateContext,
+                                  llval: ValueRef,
+                                  signed: bool,
+                                  float_ty: Type) -> ValueRef {
+    // Note: this breaks if llval is a complex constant expression rather than a simple constant.
+    // One way that might happen would be if addresses could be turned into integers in constant
+    // expressions, but that doesn't appear to be possible?
+    // In any case, an ICE is better than producing undef.
+    let value = const_to_opt_u128(llval, signed).unwrap_or_else(|| {
+        panic!("could not get z128 value of constant integer {:?}",
+               Value(llval));
+    });
+    if signed {
+        llvm::LLVMConstSIToFP(llval, float_ty.to_ref())
+    } else if float_ty.float_width() == 32 && value >= MAX_F32_PLUS_HALF_ULP {
+        // We're casting to f32 and the value is > f32::MAX + 0.5 ULP -> round up to infinity.
+        let infinity_bits = C_u32(ccx, ieee::Single::INFINITY.to_bits() as u32);
+        consts::bitcast(infinity_bits, float_ty)
+    } else {
+        llvm::LLVMConstUIToFP(llval, float_ty.to_ref())
+    }
+}
+
 impl<'a, 'tcx> MirContext<'a, 'tcx> {
     pub fn trans_constant(&mut self,
                           bcx: &Builder<'a, 'tcx>,

src/librustc_trans/mir/rvalue.rs

@@ -15,11 +15,15 @@ use rustc::ty::layout::{Layout, LayoutTyper};
 use rustc::mir::tcx::LvalueTy;
 use rustc::mir;
 use rustc::middle::lang_items::ExchangeMallocFnLangItem;
+use rustc_apfloat::{ieee, Float, Status, Round};
+use rustc_const_math::MAX_F32_PLUS_HALF_ULP;
+use std::{u128, i128};
 
 use base;
 use builder::Builder;
 use callee;
-use common::{self, val_ty, C_bool, C_i32, C_null, C_usize, C_uint};
+use common::{self, val_ty, C_bool, C_i32, C_u32, C_u64, C_null, C_usize, C_uint, C_big_integral};
+use consts;
 use adt;
 use machine;
 use monomorphize;
@@ -333,14 +337,12 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
                                 bcx.ptrtoint(llval, ll_t_out),
                             (CastTy::Int(_), CastTy::Ptr(_)) =>
                                 bcx.inttoptr(llval, ll_t_out),
-                            (CastTy::Int(_), CastTy::Float) if signed =>
-                                bcx.sitofp(llval, ll_t_out),
                             (CastTy::Int(_), CastTy::Float) =>
-                                bcx.uitofp(llval, ll_t_out),
+                                cast_int_to_float(&bcx, signed, llval, ll_t_in, ll_t_out),
                             (CastTy::Float, CastTy::Int(IntTy::I)) =>
-                                bcx.fptosi(llval, ll_t_out),
+                                cast_float_to_int(&bcx, true, llval, ll_t_in, ll_t_out),
                             (CastTy::Float, CastTy::Int(_)) =>
-                                bcx.fptoui(llval, ll_t_out),
+                                cast_float_to_int(&bcx, false, llval, ll_t_in, ll_t_out),
                             _ => bug!("unsupported cast: {:?} to {:?}", operand.ty, cast_ty)
                         };
                         OperandValue::Immediate(newval)
@@ -815,3 +817,158 @@ fn get_overflow_intrinsic(oop: OverflowOp, bcx: &Builder, ty: Ty) -> ValueRef {
 
     bcx.ccx.get_intrinsic(&name)
 }
+
+fn cast_int_to_float(bcx: &Builder,
+                     signed: bool,
+                     x: ValueRef,
+                     int_ty: Type,
+                     float_ty: Type) -> ValueRef {
+    // Most integer types, even i128, fit into [-f32::MAX, f32::MAX] after rounding.
+    // It's only u128 -> f32 that can cause overflows (i.e., should yield infinity).
+    // LLVM's uitofp produces undef in those cases, so we manually check for that case.
+    let is_u128_to_f32 = !signed && int_ty.int_width() == 128 && float_ty.float_width() == 32;
+    if is_u128_to_f32 && bcx.sess().opts.debugging_opts.saturating_float_casts {
+        // All inputs greater or equal to (f32::MAX + 0.5 ULP) are rounded to infinity,
+        // and for everything else LLVM's uitofp works just fine.
+        let max = C_big_integral(int_ty, MAX_F32_PLUS_HALF_ULP);
+        let overflow = bcx.icmp(llvm::IntUGE, x, max);
+        let infinity_bits = C_u32(bcx.ccx, ieee::Single::INFINITY.to_bits() as u32);
+        let infinity = consts::bitcast(infinity_bits, float_ty);
+        bcx.select(overflow, infinity, bcx.uitofp(x, float_ty))
+    } else {
+        if signed {
+            bcx.sitofp(x, float_ty)
+        } else {
+            bcx.uitofp(x, float_ty)
+        }
+    }
+}
+
+fn cast_float_to_int(bcx: &Builder,
+                     signed: bool,
+                     x: ValueRef,
+                     float_ty: Type,
+                     int_ty: Type) -> ValueRef {
+    let fptosui_result = if signed {
+        bcx.fptosi(x, int_ty)
+    } else {
+        bcx.fptoui(x, int_ty)
+    };
+
+    if !bcx.sess().opts.debugging_opts.saturating_float_casts {
+        return fptosui_result;
+    }
+    // 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 sufficently 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.
+    fn compute_clamp_bounds<F: Float>(signed: bool, int_ty: Type) -> (u128, u128) {
+        let rounded_min = F::from_i128_r(int_min(signed, int_ty), Round::TowardZero);
+        assert_eq!(rounded_min.status, Status::OK);
+        let rounded_max = F::from_u128_r(int_max(signed, int_ty), Round::TowardZero);
+        assert!(rounded_max.value.is_finite());
+        (rounded_min.value.to_bits(), rounded_max.value.to_bits())
+    }
+    fn int_max(signed: bool, int_ty: Type) -> u128 {
+        let shift_amount = 128 - int_ty.int_width();
+        if signed {
+            i128::MAX as u128 >> shift_amount
+        } else {
+            u128::MAX >> shift_amount
+        }
+    }
+    fn int_min(signed: bool, int_ty: Type) -> i128 {
+        if signed {
+            i128::MIN >> (128 - int_ty.int_width())
+        } else {
+            0
+        }
+    }
+    let float_bits_to_llval = |bits| {
+        let bits_llval = match float_ty.float_width() {
+            32 => C_u32(bcx.ccx, bits as u32),
+            64 => C_u64(bcx.ccx, bits as u64),
+            n => bug!("unsupported float width {}", n),
+        };
+        consts::bitcast(bits_llval, float_ty)
+    };
+    let (f_min, f_max) = match float_ty.float_width() {
+        32 => compute_clamp_bounds::<ieee::Single>(signed, int_ty),
+        64 => compute_clamp_bounds::<ieee::Double>(signed, int_ty),
+        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.
+
+    // Step 1 was already performed above.
+
+    // 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 a ordered ("normal") comparison. Whether these optimizations will be
+    // performed is ultimately up to the backend, but at least x86 does perform them.
+    let less_or_nan = bcx.fcmp(llvm::RealULT, x, f_min);
+    let greater = bcx.fcmp(llvm::RealOGT, x, f_max);
+    let int_max = C_big_integral(int_ty, int_max(signed, int_ty));
+    let int_min = C_big_integral(int_ty, int_min(signed, int_ty) as u128);
+    let s0 = bcx.select(less_or_nan, int_min, fptosui_result);
+    let s1 = bcx.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
+        bcx.select(bcx.fcmp(llvm::RealOEQ, x, x), s1, C_uint(int_ty, 0))
+    } else {
+        s1
+    }
+}