Optimize intrinsics to avoid exceptions.

Author: maleadt

src/device/intrinsics/math.jl

@@ -150,25 +150,25 @@ end
 
 
 @device_function clz(x::Union{Int32, UInt32}) =
-    assume(within(UInt32(0), UInt32(32)),
-           ccall("extern __nv_clz", llvmcall, Int32, (UInt32,), x))
+    assume(within(0u32, 32u32),
+           ccall("extern __nv_clz", llvmcall, UInt32, (UInt32,), x))
 @device_function clz(x::Union{Int64, UInt64}) =
-    assume(within(UInt64(0), UInt64(64)),
-           ccall("extern __nv_clzll", llvmcall, Int32, (UInt64,), x))
+    assume(within(0u32, 64u32),
+           ccall("extern __nv_clzll", llvmcall, UInt32, (UInt64,), x))
 
 @device_function ffs(x::Union{Int32, UInt32}) =
-    assume(within(UInt32(0), UInt32(32)),
-           ccall("extern __nv_ffs", llvmcall, Int32, (UInt32,), x))
+    assume(within(0u32, 32u32),
+           ccall("extern __nv_ffs", llvmcall, UInt32, (UInt32,), x))
 @device_function ffs(x::Union{Int64, UInt64}) =
-    assume(within(UInt64(0), UInt64(64)),
-           ccall("extern __nv_ffsll", llvmcall, Int32, (UInt64,), x))
+    assume(within(0u32, 64u32),
+           ccall("extern __nv_ffsll", llvmcall, UInt32, (UInt64,), x))
 
 @device_function popc(x::Union{Int32, UInt32}) =
-    assume(within(UInt32(0), UInt32(32)),
-           ccall("extern __nv_popc", llvmcall, Int32, (UInt32,), x))
+    assume(within(0u32, 32u32),
+           ccall("extern __nv_popc", llvmcall, UInt32, (UInt32,), x))
 @device_function popc(x::Union{Int64, UInt64}) =
-    assume(within(UInt64(0), UInt64(64)),
-           ccall("extern __nv_popcll", llvmcall, Int32, (UInt64,), x))
+    assume(within(0u32, 64u32),
+           ccall("extern __nv_popcll", llvmcall, UInt32, (UInt64,), x))
 
 @device_function byte_perm(x::Union{Int32, UInt32}, y::Union{Int32, UInt32}, z::Union{Int32, UInt32}) =
     ccall("extern __nv_byte_perm", llvmcall, Int32, (UInt32, UInt32, UInt32), x, y, z)

src/device/intrinsics/warp_shuffle.jl

@@ -5,7 +5,7 @@
 # TODO: does not work on sub-word (ie. Int16) or non-word divisible sized types
 
 # TODO: these functions should dispatch based on the actual warp size
-const ws = Int32(32)
+const ws = 32u32
 
 
 # core intrinsics
@@ -18,7 +18,7 @@ const ws = Int32(32)
 for (name, mode, mask, offset) in (("_up",   :up,   UInt32(0x00), src->src),
                                    ("_down", :down, UInt32(0x1f), src->src),
                                    ("_xor",  :bfly, UInt32(0x1f), src->src),
-                                   ("",      :idx,  UInt32(0x1f), src->:($src-1)))
+                                   ("",      :idx,  UInt32(0x1f), src->:($src-(1u32))))
     fname = Symbol("shfl$(name)_sync")
     @eval export $fname
 
@@ -28,8 +28,8 @@ for (name, mode, mask, offset) in (("_up",   :up,   UInt32(0x00), src->src),
         @eval begin
             @inline $fname(mask, val::$T, src, width=$ws) =
                 ccall($intrinsic, llvmcall, $T,
-                    (UInt32, $T, UInt32, UInt32),
-                    mask, val, $(offset(:src)), pack(width, $mask))
+                      (UInt32, $T, UInt32, UInt32),
+                      mask, val, $(offset(:src)), pack(width, $mask))
         end
     end
 end

src/device/utils.jl

@@ -3,8 +3,9 @@
 # helper type for writing Int32 literals
 # TODO: upstream this
 struct Literal{T} end
-Base.:(*)(x, ::Type{Literal{T}}) where {T} = T(x)
+Base.:(*)(x, ::Type{Literal{T}}) where {T} = x%T
 const i32 = Literal{Int32}
+const u32 = Literal{UInt32}
 
 # local method table for device functions
 @static if isdefined(Base.Experimental, Symbol("@overlay"))