Rollup merge of rust-lang#124874 - jedbrown:float-mul-add-fast, r=sae…

…thlin

intrinsics fmuladdf{32,64}: expose llvm.fmuladd.* semantics

Add intrinsics `fmuladd{f32,f64}`. This computes `(a * b) + c`, to be fused if the code generator determines that (i) the target instruction set has support for a fused operation, and (ii) that the fused operation is more efficient than the equivalent, separate pair of `mul` and `add` instructions.

https://llvm.org/docs/LangRef.html#llvm-fmuladd-intrinsic

The codegen_cranelift uses the `fma` function from libc, which is a correct implementation, but without the desired performance semantic. I think this requires an update to cranelift to expose a suitable instruction in its IR.

I have not tested with codegen_gcc, but it should behave the same way (using `fma` from libc).

---
This topic has been discussed a few times on Zulip and was suggested, for example, by `@workingjubilee` in [Effect of fma disabled](https://rust-lang.zulipchat.com/#narrow/stream/122651-general/topic/Effect.20of.20fma.20disabled/near/274179331).

core/src/intrinsics.rs

@@ -1795,6 +1795,59 @@ extern "rust-intrinsic" {
     #[rustc_nounwind]
     pub fn fmaf128(a: f128, b: f128, c: f128) -> f128;
 
+    /// Returns `a * b + c` for `f16` values, non-deterministically executing
+    /// either a fused multiply-add or two operations with rounding of the
+    /// intermediate result.
+    ///
+    /// The operation is fused if the code generator determines that target
+    /// instruction set has support for a fused operation, and that the fused
+    /// operation is more efficient than the equivalent, separate pair of mul
+    /// and add instructions. It is unspecified whether or not a fused operation
+    /// is selected, and that may depend on optimization level and context, for
+    /// example.
+    #[rustc_nounwind]
+    #[cfg(not(bootstrap))]
+    pub fn fmuladdf16(a: f16, b: f16, c: f16) -> f16;
+    /// Returns `a * b + c` for `f32` values, non-deterministically executing
+    /// either a fused multiply-add or two operations with rounding of the
+    /// intermediate result.
+    ///
+    /// The operation is fused if the code generator determines that target
+    /// instruction set has support for a fused operation, and that the fused
+    /// operation is more efficient than the equivalent, separate pair of mul
+    /// and add instructions. It is unspecified whether or not a fused operation
+    /// is selected, and that may depend on optimization level and context, for
+    /// example.
+    #[rustc_nounwind]
+    #[cfg(not(bootstrap))]
+    pub fn fmuladdf32(a: f32, b: f32, c: f32) -> f32;
+    /// Returns `a * b + c` for `f64` values, non-deterministically executing
+    /// either a fused multiply-add or two operations with rounding of the
+    /// intermediate result.
+    ///
+    /// The operation is fused if the code generator determines that target
+    /// instruction set has support for a fused operation, and that the fused
+    /// operation is more efficient than the equivalent, separate pair of mul
+    /// and add instructions. It is unspecified whether or not a fused operation
+    /// is selected, and that may depend on optimization level and context, for
+    /// example.
+    #[rustc_nounwind]
+    #[cfg(not(bootstrap))]
+    pub fn fmuladdf64(a: f64, b: f64, c: f64) -> f64;
+    /// Returns `a * b + c` for `f128` values, non-deterministically executing
+    /// either a fused multiply-add or two operations with rounding of the
+    /// intermediate result.
+    ///
+    /// The operation is fused if the code generator determines that target
+    /// instruction set has support for a fused operation, and that the fused
+    /// operation is more efficient than the equivalent, separate pair of mul
+    /// and add instructions. It is unspecified whether or not a fused operation
+    /// is selected, and that may depend on optimization level and context, for
+    /// example.
+    #[rustc_nounwind]
+    #[cfg(not(bootstrap))]
+    pub fn fmuladdf128(a: f128, b: f128, c: f128) -> f128;
+
     /// Returns the absolute value of an `f16`.
     ///
     /// The stabilized version of this intrinsic is