Add lints to detect inaccurate and inefficient FP operations

Add lints to detect floating point computations that are either
inaccurate or inefficient and suggest better alternatives.

Author: krishna-veerareddy

CHANGELOG.md

@@ -1169,6 +1169,7 @@ Released 2018-09-13
 [`ifs_same_cond`]: https://rust-lang.github.io/rust-clippy/master/index.html#ifs_same_cond
 [`implicit_hasher`]: https://rust-lang.github.io/rust-clippy/master/index.html#implicit_hasher
 [`implicit_return`]: https://rust-lang.github.io/rust-clippy/master/index.html#implicit_return
+[`inaccurate_floating_point_computation`]: https://rust-lang.github.io/rust-clippy/master/index.html#inaccurate_floating_point_computation
 [`inconsistent_digit_grouping`]: https://rust-lang.github.io/rust-clippy/master/index.html#inconsistent_digit_grouping
 [`indexing_slicing`]: https://rust-lang.github.io/rust-clippy/master/index.html#indexing_slicing
 [`ineffective_bit_mask`]: https://rust-lang.github.io/rust-clippy/master/index.html#ineffective_bit_mask

clippy_lints/src/floating_point_arithmetic.rs

@@ -0,0 +1,228 @@
+use crate::consts::{
+    constant,
+    Constant::{F32, F64},
+};
+use crate::utils::*;
+use if_chain::if_chain;
+use rustc::declare_lint_pass;
+use rustc::hir::*;
+use rustc::lint::{LateContext, LateLintPass, LintArray, LintPass};
+use rustc_errors::Applicability;
+use rustc_session::declare_tool_lint;
+use std::f32::consts as f32_consts;
+use std::f64::consts as f64_consts;
+
+declare_clippy_lint! {
+    /// **What it does:** Looks for numerically unstable floating point
+    /// computations and suggests better alternatives.
+    ///
+    /// **Why is this bad?** Numerically unstable floating point computations
+    /// cause rounding errors to magnify and distorts the results strongly.
+    ///
+    /// **Known problems:** None
+    ///
+    /// **Example:**
+    ///
+    /// ```rust
+    /// use std::f32::consts::E;
+    ///
+    /// let a = 1f32.log(2.0);
+    /// let b = 1f32.log(10.0);
+    /// let c = 1f32.log(E);
+    /// ```
+    ///
+    /// is better expressed as
+    ///
+    /// ```rust
+    /// let a = 1f32.log2();
+    /// let b = 1f32.log10();
+    /// let c = 1f32.ln();
+    /// ```
+    pub INACCURATE_FLOATING_POINT_COMPUTATION,
+    nursery,
+    "checks for numerically unstable floating point computations"
+}
+
+declare_clippy_lint! {
+    /// **What it does:** Looks for inefficient floating point computations
+    /// and suggests faster alternatives.
+    ///
+    /// **Why is this bad?** Lower performance.
+    ///
+    /// **Known problems:** None
+    ///
+    /// **Example:**
+    ///
+    /// ```rust
+    /// use std::f32::consts::E;
+    ///
+    /// let a = (2f32).powf(3.0);
+    /// let c = E.powf(3.0);
+    /// ```
+    ///
+    /// is better expressed as
+    ///
+    /// ```rust
+    /// let a = (3f32).exp2();
+    /// let b = (3f32).exp();
+    /// ```
+    pub SLOW_FLOATING_POINT_COMPUTATION,
+    nursery,
+    "checks for inefficient floating point computations"
+}
+
+declare_lint_pass!(FloatingPointArithmetic => [
+    INACCURATE_FLOATING_POINT_COMPUTATION,
+    SLOW_FLOATING_POINT_COMPUTATION
+]);
+
+fn check_log_base(cx: &LateContext<'_, '_>, expr: &Expr, args: &HirVec<Expr>) {
+    let recv = &args[0];
+    let arg = sugg::Sugg::hir(cx, recv, "..").maybe_par();
+
+    if let Some((value, _)) = constant(cx, cx.tables, &args[1]) {
+        let method;
+
+        if F32(2.0) == value || F64(2.0) == value {
+            method = "log2";
+        } else if F32(10.0) == value || F64(10.0) == value {
+            method = "log10";
+        } else if F32(f32_consts::E) == value || F64(f64_consts::E) == value {
+            method = "ln";
+        } else {
+            return;
+        }
+
+        span_lint_and_sugg(
+            cx,
+            INACCURATE_FLOATING_POINT_COMPUTATION,
+            expr.span,
+            "logarithm for bases 2, 10 and e can be computed more accurately",
+            "consider using",
+            format!("{}.{}()", arg, method),
+            Applicability::MachineApplicable,
+        );
+    }
+}
+
+// TODO: Lint expressions of the form `(x + 1).ln()` and `(x + y).ln()`
+// where y > 1 and suggest usage of `(x + (y - 1)).ln_1p()` instead
+fn check_ln1p(cx: &LateContext<'_, '_>, expr: &Expr, args: &HirVec<Expr>) {
+    if_chain! {
+        if let ExprKind::Binary(op, ref lhs, ref rhs) = &args[0].kind;
+        if op.node == BinOpKind::Add;
+        if let Some((value, _)) = constant(cx, cx.tables, lhs);
+        if F32(1.0) == value || F64(1.0) == value;
+        then {
+            let arg = sugg::Sugg::hir(cx, rhs, "..").maybe_par();
+
+            span_lint_and_sugg(
+                cx,
+                INACCURATE_FLOATING_POINT_COMPUTATION,
+                expr.span,
+                "ln(1 + x) can be computed more accurately",
+                "consider using",
+                format!("{}.ln_1p()", arg),
+                Applicability::MachineApplicable,
+            );
+        }
+    }
+}
+
+fn check_powf(cx: &LateContext<'_, '_>, expr: &Expr, args: &HirVec<Expr>) {
+    // Check receiver
+    if let Some((value, _)) = constant(cx, cx.tables, &args[0]) {
+        let method;
+
+        if F32(f32_consts::E) == value || F64(f64_consts::E) == value {
+            method = "exp";
+        } else if F32(2.0) == value || F64(2.0) == value {
+            method = "exp2";
+        } else {
+            return;
+        }
+
+        span_lint_and_sugg(
+            cx,
+            SLOW_FLOATING_POINT_COMPUTATION,
+            expr.span,
+            "exponent for bases 2 and e can be computed more efficiently",
+            "consider using",
+            format!("{}.{}()", sugg::Sugg::hir(cx, &args[1], "..").maybe_par(), method),
+            Applicability::MachineApplicable,
+        );
+    }
+
+    // Check argument
+    if let Some((value, _)) = constant(cx, cx.tables, &args[1]) {
+        let help;
+        let method;
+
+        if F32(1.0 / 2.0) == value || F64(1.0 / 2.0) == value {
+            help = "square-root of a number can be computer more efficiently";
+            method = "sqrt";
+        } else if F32(1.0 / 3.0) == value || F64(1.0 / 3.0) == value {
+            help = "cube-root of a number can be computer more efficiently";
+            method = "cbrt";
+        } else {
+            return;
+        }
+
+        span_lint_and_sugg(
+            cx,
+            SLOW_FLOATING_POINT_COMPUTATION,
+            expr.span,
+            help,
+            "consider using",
+            format!("{}.{}()", sugg::Sugg::hir(cx, &args[0], ".."), method),
+            Applicability::MachineApplicable,
+        );
+    }
+}
+
+// TODO: Lint expressions of the form `x.exp() - y` where y > 1
+// and suggest usage of `x.exp_m1() - (y - 1)` instead
+fn check_expm1(cx: &LateContext<'_, '_>, expr: &Expr) {
+    if_chain! {
+        if let ExprKind::Binary(op, ref lhs, ref rhs) = expr.kind;
+        if op.node == BinOpKind::Sub;
+        if cx.tables.expr_ty(lhs).is_floating_point();
+        if let Some((value, _)) = constant(cx, cx.tables, rhs);
+        if F32(1.0) == value || F64(1.0) == value;
+        if let ExprKind::MethodCall(ref path, _, ref method_args) = lhs.kind;
+        if path.ident.name.as_str() == "exp";
+        then {
+            span_lint_and_sugg(
+                cx,
+                INACCURATE_FLOATING_POINT_COMPUTATION,
+                expr.span,
+                "(e.pow(x) - 1) can be computed more accurately",
+                "consider using",
+                format!(
+                    "{}.exp_m1()",
+                    sugg::Sugg::hir(cx, &method_args[0], "..")
+                ),
+                Applicability::MachineApplicable,
+            );
+        }
+    }
+}
+
+impl<'a, 'tcx> LateLintPass<'a, 'tcx> for FloatingPointArithmetic {
+    fn check_expr(&mut self, cx: &LateContext<'a, 'tcx>, expr: &'tcx Expr) {
+        if let ExprKind::MethodCall(ref path, _, args) = &expr.kind {
+            let recv_ty = cx.tables.expr_ty(&args[0]);
+
+            if recv_ty.is_floating_point() {
+                match &*path.ident.name.as_str() {
+                    "ln" => check_ln1p(cx, expr, args),
+                    "log" => check_log_base(cx, expr, args),
+                    "powf" => check_powf(cx, expr, args),
+                    _ => {},
+                }
+            }
+        } else {
+            check_expm1(cx, expr);
+        }
+    }
+}

clippy_lints/src/lib.rs

@@ -1000,6 +1000,7 @@ pub fn register_plugins(store: &mut rustc_lint::LintStore, sess: &Session, conf:
     store.register_late_pass(|| box to_digit_is_some::ToDigitIsSome);
     let array_size_threshold = conf.array_size_threshold;
     store.register_late_pass(move || box large_stack_arrays::LargeStackArrays::new(array_size_threshold));
+    store.register_late_pass(move || box floating_point_arithmetic::FloatingPointArithmetic);
     store.register_early_pass(|| box as_conversions::AsConversions);
     store.register_early_pass(|| box utils::internal_lints::ProduceIce);
     store.register_late_pass(|| box let_underscore::LetUnderscore);
@@ -1648,6 +1649,8 @@ pub fn register_plugins(store: &mut rustc_lint::LintStore, sess: &Session, conf:
     store.register_group(true, "clippy::nursery", Some("clippy_nursery"), vec![
         LintId::of(&attrs::EMPTY_LINE_AFTER_OUTER_ATTR),
         LintId::of(&fallible_impl_from::FALLIBLE_IMPL_FROM),
+        LintId::of(&floating_point_arithmetic::INACCURATE_FLOATING_POINT_COMPUTATION),
+        LintId::of(&floating_point_arithmetic::SLOW_FLOATING_POINT_COMPUTATION),
         LintId::of(&missing_const_for_fn::MISSING_CONST_FOR_FN),
         LintId::of(&mul_add::MANUAL_MUL_ADD),
         LintId::of(&mutable_debug_assertion::DEBUG_ASSERT_WITH_MUT_CALL),

src/lintlist/mod.rs

@@ -749,6 +749,13 @@ pub const ALL_LINTS: [Lint; 357] = [
         deprecation: None,
         module: "implicit_return",
     },
+    Lint {
+        name: "inaccurate_floating_point_computation",
+        group: "nursery",
+        desc: "checks for numerically unstable floating point computations",
+        deprecation: None,
+        module: "floating_point_arithmetic",
+    },
     Lint {
         name: "inconsistent_digit_grouping",
         group: "style",

tests/ui/floating_point_arithmetic.rs

@@ -0,0 +1,76 @@
+#![allow(dead_code)]
+#![warn(
+    clippy::inaccurate_floating_point_computation,
+    clippy::slow_floating_point_computation
+)]
+
+const TWO: f32 = 2.0;
+const E: f32 = std::f32::consts::E;
+
+fn check_log_base() {
+    let x = 1f32;
+    let _ = x.log(2f32);
+    let _ = x.log(10f32);
+    let _ = x.log(std::f32::consts::E);
+    let _ = x.log(TWO);
+    let _ = x.log(E);
+
+    let x = 1f64;
+    let _ = x.log(2f64);
+    let _ = x.log(10f64);
+    let _ = x.log(std::f64::consts::E);
+}
+
+fn check_ln1p() {
+    let x = 1f32;
+    let _ = (1.0 + x).ln();
+    let _ = (1.0 + x * 2.0).ln();
+    let _ = (1.0 + x.powi(2)).ln();
+    let _ = (1.0 + x.powi(2) * 2.0).ln();
+    let _ = (1.0 + (std::f32::consts::E - 1.0)).ln();
+    // Cases where the lint shouldn't be applied
+    let _ = (x + 1.0).ln();
+    let _ = (1.0 + x + 2.0).ln();
+    let _ = (1.0 + x - 2.0).ln();
+
+    let x = 1f64;
+    let _ = (1.0 + x).ln();
+    let _ = (1.0 + x * 2.0).ln();
+    let _ = (1.0 + x.powi(2)).ln();
+    // Cases where the lint shouldn't be applied
+    let _ = (x + 1.0).ln();
+    let _ = (1.0 + x + 2.0).ln();
+    let _ = (1.0 + x - 2.0).ln();
+}
+
+fn check_powf() {
+    let x = 3f32;
+    let _ = 2f32.powf(x);
+    let _ = std::f32::consts::E.powf(x);
+    let _ = x.powf(1.0 / 2.0);
+    let _ = x.powf(1.0 / 3.0);
+
+    let x = 3f64;
+    let _ = 2f64.powf(x);
+    let _ = std::f64::consts::E.powf(x);
+    let _ = x.powf(1.0 / 2.0);
+    let _ = x.powf(1.0 / 3.0);
+}
+
+fn check_expm1() {
+    let x = 2f32;
+    let _ = x.exp() - 1.0;
+    let _ = x.exp() - 1.0 + 2.0;
+    // Cases where the lint shouldn't be applied
+    let _ = x.exp() - 2.0;
+    let _ = x.exp() - 1.0 * 2.0;
+
+    let x = 2f64;
+    let _ = x.exp() - 1.0;
+    let _ = x.exp() - 1.0 + 2.0;
+    // Cases where the lint shouldn't be applied
+    let _ = x.exp() - 2.0;
+    let _ = x.exp() - 1.0 * 2.0;
+}
+
+fn main() {}