Support for BFloat16s.

Project.toml

@@ -7,9 +7,19 @@ CEnum = "fa961155-64e5-5f13-b03f-caf6b980ea82"
 LLVMExtra_jll = "dad2f222-ce93-54a1-a47d-0025e8a3acab"
 Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
+Requires = "ae029012-a4dd-5104-9daa-d747884805df"
 Unicode = "4ec0a83e-493e-50e2-b9ac-8f72acf5a8f5"
 
+[weakdeps]
+BFloat16s = "ab4f0b2a-ad5b-11e8-123f-65d77653426b"
+
+[extensions]
+BFloat16sExt = "BFloat16s"
+
 [compat]
 CEnum = "0.2, 0.3, 0.4"
 LLVMExtra_jll = "=0.0.26"
 julia = "1.8"
+
+[extras]
+BFloat16s = "ab4f0b2a-ad5b-11e8-123f-65d77653426b"

ext/BFloat16sExt.jl

@@ -0,0 +1,18 @@
+module BFloat16sExt
+
+using LLVM
+using LLVM: API
+
+isdefined(Base, :get_extension) ? (using BFloat16s) : (using ..BFloat16s)
+
+## constant values
+
+LLVM.ConstantFP(val::BFloat16) = ConstantFP(BFloatType(), val)
+
+Base.convert(::Type{BFloat16}, val::ConstantFP) =
+    convert(BFloat16, API.LLVMConstRealGetDouble(val, Ref{API.LLVMBool}()))
+
+ConstantDataArray(data::AbstractVector{BFloat16}) =
+    ConstantDataArray(BFloatType(), data)
+
+end

src/LLVM.jl

@@ -4,6 +4,10 @@ using Unicode
 using Printf
 using Libdl
 
+if !isdefined(Base, :get_extension)
+    using Requires: @require
+end
+
 
 ## source code includes
 
@@ -106,6 +110,12 @@ function __init__()
                   Please re-compile Julia and LLVM.jl (but note that USE_SYSTEM_LLVM is not a supported configuration)."""
     end
 
+    @static if !isdefined(Base, :get_extension)
+        @require BFloat16s = "ab4f0b2a-ad5b-11e8-123f-65d77653426b" begin
+            include("../ext/BFloat16sExt.jl")
+        end
+    end
+
     _install_handlers()
     _install_handlers(GlobalContext())
 end

src/core/type.jl

@@ -85,7 +85,7 @@ width(inttyp::IntegerType) = API.LLVMGetIntTypeWidth(inttyp)
 #       we add it for convenience of typechecking generic values (see execution.jl)
  abstract type FloatingPointType <: LLVMType end
 
-for T in [:Half, :Float, :Double, :FP128, :X86_FP80, :PPC_FP128]
+for T in [:Half, :Float, :Double, :BFloat, :FP128, :X86_FP80, :PPC_FP128]
     CleanT = Symbol(replace(String(T), "_"=>""))    # only the type kind retains the underscore
     jl_fname = Symbol(CleanT, :Type)
     api_typename = Symbol(:LLVM, CleanT)

src/core/value/constant.jl

@@ -101,12 +101,9 @@ register(ConstantFP, API.LLVMConstantFPValueKind)
 ConstantFP(typ::FloatingPointType, val::Real) =
     ConstantFP(API.LLVMConstReal(typ, Cdouble(val)))
 
-ConstantFP(val::Float16) =
-    ConstantFP(HalfType(), val)
-ConstantFP(val::Float32) =
-    ConstantFP(FloatType(), val)
-ConstantFP(val::Float64) =
-    ConstantFP(DoubleType(), val)
+ConstantFP(val::Float64) = ConstantFP(DoubleType(), val)
+ConstantFP(val::Float32) = ConstantFP(FloatType(), val)
+ConstantFP(val::Float16) = ConstantFP(HalfType(), val)
 
 Base.convert(::Type{T}, val::ConstantFP) where {T<:AbstractFloat} =
     convert(T, API.LLVMConstRealGetDouble(val, Ref{API.LLVMBool}()))
@@ -166,12 +163,12 @@ ConstantDataArray(data::AbstractVector{T}) where {T<:Integer} =
     ConstantDataArray(IntType(sizeof(T)*8), data)
 ConstantDataArray(data::AbstractVector{Core.Bool}) =
     ConstantDataArray(Int1Type(), data)
-ConstantDataArray(data::AbstractVector{Float16}) =
-    ConstantDataArray(HalfType(), data)
-ConstantDataArray(data::AbstractVector{Float32}) =
-    ConstantDataArray(FloatType(), data)
 ConstantDataArray(data::AbstractVector{Float64}) =
     ConstantDataArray(DoubleType(), data)
+ConstantDataArray(data::AbstractVector{Float32}) =
+    ConstantDataArray(FloatType(), data)
+ConstantDataArray(data::AbstractVector{Float16}) =
+    ConstantDataArray(HalfType(), data)
 
 @checked struct ConstantDataVector <: ConstantDataSequential
     ref::API.LLVMValueRef

test/Project.toml

@@ -1,4 +1,5 @@
 [deps]
+BFloat16s = "ab4f0b2a-ad5b-11e8-123f-65d77653426b"
 InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 LLVMExtra_jll = "dad2f222-ce93-54a1-a47d-0025e8a3acab"
 ReTestItems = "817f1d60-ba6b-4fd5-9520-3cf149f6a823"

test/core_tests.jl

@@ -1,5 +1,7 @@
 @testitem "core" setup=[TestHelpers] begin
 
+using BFloat16s
+
 struct TestStruct
     x::Bool
     y::Int64
@@ -410,6 +412,11 @@ end
         c = ConstantFP(typ, 1.1)
         @test convert(Float64, c) == 1.1
     end
+    let
+        typ = LLVM.BFloatType()
+        c = ConstantFP(typ, BFloat16(1.1))
+        @test convert(BFloat16, c) == BFloat16(1.1)
+    end
     let
         typ = LLVM.X86FP80Type()
         # TODO: how to construct full-width constants?