From bf6857c949f877ee0636626f8e8078aae1e7d754 Mon Sep 17 00:00:00 2001
From: Chris Elrod <elrodc@gmail.com>
Date: Mon, 15 Apr 2024 23:38:23 -0400
Subject: [PATCH] Fix ambiguity in vconvert (#110)

* Fix ambiguity in vconvert

* don't limit inference on vconvert recursion
---
 src/VectorizationBase.jl      |  9 +++++++++
 src/llvm_intrin/conversion.jl | 11 ++++++-----
 2 files changed, 15 insertions(+), 5 deletions(-)

diff --git a/src/VectorizationBase.jl b/src/VectorizationBase.jl
index 306e586..15f8dcc 100644
--- a/src/VectorizationBase.jl
+++ b/src/VectorizationBase.jl
@@ -550,6 +550,15 @@ end
 @inline reduce_to_onevec(f::F, vu::VecUnroll) where {F} =
   ArrayInterface.reduce_tup(f, data(vu))
 
+if VERSION >= v"1.7.0" && hasfield(Method, :recursion_relation)
+  dont_limit = Returns(true)
+  for f in (vconvert, _vconvert)
+    for m in methods(f)
+      m.recursion_relation = dont_limit
+    end
+  end
+end
+
 include("precompile.jl")
 _precompile_()
 
diff --git a/src/llvm_intrin/conversion.jl b/src/llvm_intrin/conversion.jl
index cf0e1ee..f2d2a7d 100644
--- a/src/llvm_intrin/conversion.jl
+++ b/src/llvm_intrin/conversion.jl
@@ -70,7 +70,7 @@ if (Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)
     ::Type{Vec{W,F}},
     v::Vec{W,T}
   )::Vec{W,F} where {W,F<:FloatingTypes,T<:IntegerTypesHW}
-    _vconvert(Vec{W,F}, v, True())
+    _vconvert(Vec{W,F}, v, True())::Vec{W,F}
   end
   @inline function vconvert(
     ::Type{Vec{W,F}},
@@ -80,7 +80,7 @@ if (Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)
       Vec{W,F},
       v,
       has_feature(Val(:x86_64_avx512dq)) | (!has_feature(Val(:x86_64_avx2)))
-    )
+    )::Vec{W,F}
   end
   @inline function vconvert(
     ::Type{F},
@@ -90,7 +90,7 @@ if (Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)
       Vec{W,F},
       v,
       has_feature(Val(:x86_64_avx512dq)) | (!has_feature(Val(:x86_64_avx2)))
-    )
+    )::VecUnroll{N,W,F,Vec{W,F}}
   end
   @inline function vconvert(
     ::Type{Vec{W,F}},
@@ -100,7 +100,7 @@ if (Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)
       Vec{W,F},
       v,
       has_feature(Val(:x86_64_avx512dq)) | (!has_feature(Val(:x86_64_avx2)))
-    )
+    )::VecUnroll{N,W,F,Vec{W,F}}
   end
   @inline function vconvert(
     ::Type{VecUnroll{N,W,F,Vec{W,F}}},
@@ -110,7 +110,7 @@ if (Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)
       Vec{W,F},
       v,
       has_feature(Val(:x86_64_avx512dq)) | (!has_feature(Val(:x86_64_avx2)))
-    )
+    )::VecUnroll{N,W,F,Vec{W,F}}
   end
 else
   @generated function vconvert(
@@ -232,6 +232,7 @@ end
 @inline vconvert(::Type{T}, s::Union{Float16,Float32,Float64}) where {T<:IntegerTypesHW} = Base.fptosi(T, Base.trunc_llvm(s))
 @inline vconvert(::Type{T}, s::IntegerTypesHW) where {T<:Union{Float16,Float32,Float64}} = convert(T, s)::T
 @inline vconvert(::Type{T}, s::Union{Float16,Float32,Float64}) where {T<:Union{Float16,Float32,Float64}} = convert(T, s)::T
+@inline vconvert(::Type{T}, s::T) where {T<:Union{Float16,Float32,Float64}} = s
 @inline vconvert(::Type{T}, s::IntegerTypesHW) where {T<:IntegerTypesHW} = s % T
 @inline vconvert(::Type{T}, v::AbstractSIMD{W,T}) where {T<:NativeTypes,W} = v
 @inline vconvert(::Type{T}, v::AbstractSIMD{W,S}) where {T<:NativeTypes,S,W} =