miri: make float min/max non-deterministic

Author: RalfJung

compiler/rustc_const_eval/src/interpret/intrinsics.rs

@@ -747,7 +747,13 @@ impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
     {
         let a: F = self.read_scalar(&args[0])?.to_float()?;
         let b: F = self.read_scalar(&args[1])?.to_float()?;
-        let res = self.adjust_nan(a.min(b), &[a, b]);
+        let res = if a == b {
+            // They are definitely not NaN (those are never equal), but they could be `+0` and `-0`.
+            // Let the machine decide which one to return.
+            M::equal_float_min_max(self, a, b)
+        } else {
+            self.adjust_nan(a.min(b), &[a, b])
+        };
         self.write_scalar(res, dest)?;
         interp_ok(())
     }
@@ -762,7 +768,13 @@ impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
     {
         let a: F = self.read_scalar(&args[0])?.to_float()?;
         let b: F = self.read_scalar(&args[1])?.to_float()?;
-        let res = self.adjust_nan(a.max(b), &[a, b]);
+        let res = if a == b {
+            // They are definitely not NaN (those are never equal), but they could be `+0` and `-0`.
+            // Let the machine decide which one to return.
+            M::equal_float_min_max(self, a, b)
+        } else {
+            self.adjust_nan(a.max(b), &[a, b])
+        };
         self.write_scalar(res, dest)?;
         interp_ok(())
     }

compiler/rustc_const_eval/src/interpret/machine.rs

@@ -278,6 +278,12 @@ pub trait Machine<'tcx>: Sized {
         F2::NAN
     }
 
+    /// Determines the result of `min`/`max` on floats when the arguments are equal.
+    fn equal_float_min_max<F: Float>(_ecx: &InterpCx<'tcx, Self>, a: F, _b: F) -> F {
+        // By default, we pick the left argument.
+        a
+    }
+
     /// Called before a basic block terminator is executed.
     #[inline]
     fn before_terminator(_ecx: &mut InterpCx<'tcx, Self>) -> InterpResult<'tcx> {

src/tools/miri/src/machine.rs

@@ -11,6 +11,7 @@ use std::{fmt, process};
 use rand::rngs::StdRng;
 use rand::{Rng, SeedableRng};
 use rustc_abi::{Align, ExternAbi, Size};
+use rustc_apfloat::{Float, FloatConvert};
 use rustc_attr_parsing::InlineAttr;
 use rustc_data_structures::fx::{FxHashMap, FxHashSet};
 #[allow(unused)]
@@ -1129,20 +1130,24 @@ impl<'tcx> Machine<'tcx> for MiriMachine<'tcx> {
     }
 
     #[inline(always)]
-    fn generate_nan<
-        F1: rustc_apfloat::Float + rustc_apfloat::FloatConvert<F2>,
-        F2: rustc_apfloat::Float,
-    >(
+    fn generate_nan<F1: Float + FloatConvert<F2>, F2: Float>(
         ecx: &InterpCx<'tcx, Self>,
         inputs: &[F1],
     ) -> F2 {
         ecx.generate_nan(inputs)
     }
 
+    #[inline(always)]
+    fn equal_float_min_max<F: Float>(ecx: &MiriInterpCx<'tcx>, a: F, b: F) -> F {
+        ecx.equal_float_min_max(a, b)
+    }
+
+    #[inline(always)]
     fn ub_checks(ecx: &InterpCx<'tcx, Self>) -> InterpResult<'tcx, bool> {
         interp_ok(ecx.tcx.sess.ub_checks())
     }
 
+    #[inline(always)]
     fn thread_local_static_pointer(
         ecx: &mut MiriInterpCx<'tcx>,
         def_id: DefId,

src/tools/miri/src/operator.rs

@@ -115,4 +115,11 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
             nan
         }
     }
+
+    fn equal_float_min_max<F: Float>(&self, a: F, b: F) -> F {
+        let this = self.eval_context_ref();
+        // Return one side non-deterministically.
+        let mut rand = this.machine.rng.borrow_mut();
+        if rand.gen() { a } else { b }
+    }
 }

src/tools/miri/tests/pass/float.rs

@@ -31,6 +31,7 @@ fn main() {
     test_fast();
     test_algebraic();
     test_fmuladd();
+    test_min_max_nondet();
 }
 
 trait Float: Copy + PartialEq + Debug {
@@ -1211,3 +1212,30 @@ fn test_fmuladd() {
     test_operations_f32(0.1, 0.2, 0.3);
     test_operations_f64(1.1, 1.2, 1.3);
 }
+
+/// `min` and `max` on equal arguments are non-deterministic.
+fn test_min_max_nondet() {
+    /// Ensure that if we call the closure often enough, we see both `true` and `false.`
+    #[track_caller]
+    fn ensure_both(f: impl Fn() -> bool) {
+        let rounds = 16;
+        let first = f();
+        for _ in 1..rounds {
+            if f() != first {
+                // We saw two different values!
+                return;
+            }
+        }
+        // We saw the same thing N times.
+        panic!("expected non-determinism, got {rounds} times the same result: {first:?}");
+    }
+
+    ensure_both(|| f16::min(0.0, -0.0).is_sign_positive());
+    ensure_both(|| f16::max(0.0, -0.0).is_sign_positive());
+    ensure_both(|| f32::min(0.0, -0.0).is_sign_positive());
+    ensure_both(|| f32::max(0.0, -0.0).is_sign_positive());
+    ensure_both(|| f64::min(0.0, -0.0).is_sign_positive());
+    ensure_both(|| f64::max(0.0, -0.0).is_sign_positive());
+    ensure_both(|| f128::min(0.0, -0.0).is_sign_positive());
+    ensure_both(|| f128::max(0.0, -0.0).is_sign_positive());
+}