From eeb881dee88cfdfaa24f9fa3bdfcb05439aa6b0a Mon Sep 17 00:00:00 2001
From: Jed Brown <jed@jedbrown.org>
Date: Fri, 5 Jan 2024 21:04:41 -0700
Subject: [PATCH] intrinsics.fmuladdf{16,32,64,128}: expose llvm.fmuladd.*
 semantics

Add intrinsics `fmuladd{f16,f32,f64,f128}`. 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

MIRI support is included for f32 and f64.

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).
---
 core/src/intrinsics.rs | 53 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 53 insertions(+)

diff --git a/core/src/intrinsics.rs b/core/src/intrinsics.rs
index d7a2f1909cabe..061fba9a1f785 100644
--- a/core/src/intrinsics.rs
+++ b/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