From 0d8a978e8a55b08778ec6ee861c2c5ed6703eb6c Mon Sep 17 00:00:00 2001 From: Jed Brown Date: Fri, 5 Jan 2024 21:04:41 -0700 Subject: 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). --- compiler/rustc_codegen_llvm/src/intrinsic.rs | 5 +++++ 1 file changed, 5 insertions(+) (limited to 'compiler/rustc_codegen_llvm/src/intrinsic.rs') diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs index 30c6f08e894..bfe623e7fc3 100644 --- a/compiler/rustc_codegen_llvm/src/intrinsic.rs +++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs @@ -86,6 +86,11 @@ fn get_simple_intrinsic<'ll>( sym::fmaf64 => "llvm.fma.f64", sym::fmaf128 => "llvm.fma.f128", + sym::fmuladdf16 => "llvm.fmuladd.f16", + sym::fmuladdf32 => "llvm.fmuladd.f32", + sym::fmuladdf64 => "llvm.fmuladd.f64", + sym::fmuladdf128 => "llvm.fmuladd.f128", + sym::fabsf16 => "llvm.fabs.f16", sym::fabsf32 => "llvm.fabs.f32", sym::fabsf64 => "llvm.fabs.f64", -- cgit 1.4.1-3-g733a5