Add LLVM minnum/maxnum intrinsics

LLVM’s intrinsics `minnum` and `maxnum` are now used for `min` and
`max` on `f32` and `f64`. This resolves #18384.

Author: varkor

src/libcore/intrinsics.rs

@@ -1142,6 +1142,19 @@ extern "rust-intrinsic" {
     /// Returns the absolute value of an `f64`.
     pub fn fabsf64(x: f64) -> f64;
 
+    /// Returns the minimum of two `f32` values.
+    #[cfg(not(stage0))]
+    pub fn minnumf32(x: f32, y: f32) -> f32;
+    /// Returns the minimum of two `f64` values.
+    #[cfg(not(stage0))]
+    pub fn minnumf64(x: f64, y: f64) -> f64;
+    /// Returns the maximum of two `f32` values.
+    #[cfg(not(stage0))]
+    pub fn maxnumf32(x: f32, y: f32) -> f32;
+    /// Returns the maximum of two `f64` values.
+    #[cfg(not(stage0))]
+    pub fn maxnumf64(x: f64, y: f64) -> f64;
+
     /// Copies the sign from `y` to `x` for `f32` values.
     pub fn copysignf32(x: f32, y: f32) -> f32;
     /// Copies the sign from `y` to `x` for `f64` values.
@@ -1393,3 +1406,25 @@ extern "rust-intrinsic" {
     #[cfg(not(stage0))]
     pub fn nontemporal_store<T>(ptr: *mut T, val: T);
 }
+
+// Simple bootstrap implementations for stage0 compilation
+/// Returns the minimum of two `f32` values.
+#[cfg(stage0)]
+pub unsafe fn minnumf32(x: f32, y: f32) -> f32 {
+    (if x < y || y != y { x } else { y }) * 1.0
+}
+/// Returns the minimum of two `f64` values.
+#[cfg(stage0)]
+pub unsafe fn minnumf64(x: f64, y: f64) -> f64 {
+    (if x < y || y != y { x } else { y }) * 1.0
+}
+/// Returns the maximum of two `f32` values.
+#[cfg(stage0)]
+pub unsafe fn maxnumf32(x: f32, y: f32) -> f32 {
+    (if x < y || x != x { y } else { x }) * 1.0
+}
+/// Returns the maximum of two `f64` values.
+#[cfg(stage0)]
+pub unsafe fn maxnumf64(x: f64, y: f64) -> f64 {
+    (if x < y || x != x { y } else { x }) * 1.0
+}

src/libcore/num/f32.rs

@@ -250,28 +250,12 @@ impl Float for f32 {
     /// Returns the maximum of the two numbers.
     #[inline]
     fn max(self, other: f32) -> f32 {
-        // IEEE754 says: maxNum(x, y) is the canonicalized number y if x < y, x if y < x, the
-        // canonicalized number if one operand is a number and the other a quiet NaN. Otherwise it
-        // is either x or y, canonicalized (this means results might differ among implementations).
-        // When either x or y is a signalingNaN, then the result is according to 6.2.
-        //
-        // Since we do not support sNaN in Rust yet, we do not need to handle them.
-        // FIXME(nagisa): due to https://bugs.llvm.org/show_bug.cgi?id=33303 we canonicalize by
-        // multiplying by 1.0. Should switch to the `canonicalize` when it works.
-        (if self < other || self.is_nan() { other } else { self }) * 1.0
+        unsafe { intrinsics::maxnumf32(self, other) }
     }
 
     /// Returns the minimum of the two numbers.
     #[inline]
     fn min(self, other: f32) -> f32 {
-        // IEEE754 says: minNum(x, y) is the canonicalized number x if x < y, y if y < x, the
-        // canonicalized number if one operand is a number and the other a quiet NaN. Otherwise it
-        // is either x or y, canonicalized (this means results might differ among implementations).
-        // When either x or y is a signalingNaN, then the result is according to 6.2.
-        //
-        // Since we do not support sNaN in Rust yet, we do not need to handle them.
-        // FIXME(nagisa): due to https://bugs.llvm.org/show_bug.cgi?id=33303 we canonicalize by
-        // multiplying by 1.0. Should switch to the `canonicalize` when it works.
-        (if self < other || other.is_nan() { self } else { other }) * 1.0
+        unsafe { intrinsics::minnumf32(self, other) }
     }
 }

src/libcore/num/f64.rs

@@ -248,28 +248,12 @@ impl Float for f64 {
     /// Returns the maximum of the two numbers.
     #[inline]
     fn max(self, other: f64) -> f64 {
-        // IEEE754 says: maxNum(x, y) is the canonicalized number y if x < y, x if y < x, the
-        // canonicalized number if one operand is a number and the other a quiet NaN. Otherwise it
-        // is either x or y, canonicalized (this means results might differ among implementations).
-        // When either x or y is a signalingNaN, then the result is according to 6.2.
-        //
-        // Since we do not support sNaN in Rust yet, we do not need to handle them.
-        // FIXME(nagisa): due to https://bugs.llvm.org/show_bug.cgi?id=33303 we canonicalize by
-        // multiplying by 1.0. Should switch to the `canonicalize` when it works.
-        (if self < other || self.is_nan() { other } else { self }) * 1.0
+        unsafe { intrinsics::maxnumf64(self, other) }
     }
 
     /// Returns the minimum of the two numbers.
     #[inline]
     fn min(self, other: f64) -> f64 {
-        // IEEE754 says: minNum(x, y) is the canonicalized number x if x < y, y if y < x, the
-        // canonicalized number if one operand is a number and the other a quiet NaN. Otherwise it
-        // is either x or y, canonicalized (this means results might differ among implementations).
-        // When either x or y is a signalingNaN, then the result is according to 6.2.
-        //
-        // Since we do not support sNaN in Rust yet, we do not need to handle them.
-        // FIXME(nagisa): due to https://bugs.llvm.org/show_bug.cgi?id=33303 we canonicalize by
-        // multiplying by 1.0. Should switch to the `canonicalize` when it works.
-        (if self < other || other.is_nan() { self } else { other }) * 1.0
+        unsafe { intrinsics::minnumf64(self, other) }
     }
 }

src/librustc_trans/context.rs

@@ -782,6 +782,11 @@ fn declare_intrinsic(ccx: &CrateContext, key: &str) -> Option<ValueRef> {
     ifn!("llvm.fabs.f32", fn(t_f32) -> t_f32);
     ifn!("llvm.fabs.f64", fn(t_f64) -> t_f64);
 
+    ifn!("llvm.minnum.f32", fn(t_f32, t_f32) -> t_f32);
+    ifn!("llvm.minnum.f64", fn(t_f64, t_f64) -> t_f64);
+    ifn!("llvm.maxnum.f32", fn(t_f32, t_f32) -> t_f32);
+    ifn!("llvm.maxnum.f64", fn(t_f64, t_f64) -> t_f64);
+
     ifn!("llvm.floor.f32", fn(t_f32) -> t_f32);
     ifn!("llvm.floor.f64", fn(t_f64) -> t_f64);
     ifn!("llvm.ceil.f32", fn(t_f32) -> t_f32);

src/librustc_trans/intrinsic.rs

@@ -61,6 +61,10 @@ fn get_simple_intrinsic(ccx: &CrateContext, name: &str) -> Option<ValueRef> {
         "fmaf64" => "llvm.fma.f64",
         "fabsf32" => "llvm.fabs.f32",
         "fabsf64" => "llvm.fabs.f64",
+        "minnumf32" => "llvm.minnum.f32",
+        "minnumf64" => "llvm.minnum.f64",
+        "maxnumf32" => "llvm.maxnum.f32",
+        "maxnumf64" => "llvm.maxnum.f64",
         "copysignf32" => "llvm.copysign.f32",
         "copysignf64" => "llvm.copysign.f64",
         "floorf32" => "llvm.floor.f32",

src/librustc_typeck/check/intrinsic.rs

@@ -255,6 +255,10 @@ pub fn check_intrinsic_type<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
             }
             "fabsf32"      => (0, vec![ tcx.types.f32 ], tcx.types.f32),
             "fabsf64"      => (0, vec![ tcx.types.f64 ], tcx.types.f64),
+            "minnumf32"    => (0, vec![ tcx.types.f32, tcx.types.f32 ], tcx.types.f32),
+            "minnumf64"    => (0, vec![ tcx.types.f64, tcx.types.f64 ], tcx.types.f64),
+            "maxnumf32"    => (0, vec![ tcx.types.f32, tcx.types.f32 ], tcx.types.f32),
+            "maxnumf64"    => (0, vec![ tcx.types.f64, tcx.types.f64 ], tcx.types.f64),
             "copysignf32"  => (0, vec![ tcx.types.f32, tcx.types.f32 ], tcx.types.f32),
             "copysignf64"  => (0, vec![ tcx.types.f64, tcx.types.f64 ], tcx.types.f64),
             "floorf32"     => (0, vec![ tcx.types.f32 ], tcx.types.f32),