fix bswap for odd number of bytes

Fixes #38.

`bswap` was seeming to work well since Julia v1.6, but LLVM documents that
`bswap` is valid only for an even number of bytes
(https://llvm.org/docs/LangRef.html#llvm-bswap-intrinsics),
and a julia build with LLVM asserts enabled catches that.

NB: to enable LLVM asserts, put the following in "Make.user" :
```
FORCE_ASSERTIONS=1
LLVM_ASSERTIONS=1
```

Project.toml

@@ -1,7 +1,7 @@
 name = "BitIntegers"
 uuid = "c3b6d118-76ef-56ca-8cc7-ebb389d030a1"
 authors = ["Rafael Fourquet <fourquet.rafael@gmail.com>"]
-version = "0.3.0"
+version = "0.3.1"
 
 [deps]
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"

README.md

@@ -51,8 +51,13 @@ recompiled (fixed by https://github.com/JuliaLang/julia/pull/30830);
 to errors and segfaults (cf. e.g. https://github.com/rfourquet/BitIntegers.jl/issues/1, fixed by
 https://github.com/JuliaLang/julia/pull/33283).
 
+
 ## Release notes
 
+### v0.3.1
+
+* fix incorrect `bswap` for odd byte-sizes ([#41](https://github.com/rfourquet/BitIntegers.jl/pull/41))
+
 ### v0.3.0
 
 * change and document how `promote_rule` is implemented ([#36](https://github.com/rfourquet/BitIntegers.jl/pull/36))

src/BitIntegers.jl

@@ -363,14 +363,26 @@ leading_zeros( x::XBI) = Int(ctlz_int(x))
 trailing_zeros(x::XBI) = Int(cttz_int(x))
 
 function bswap(x::XBI)
-    if VERSION < v"1.6" && sizeof(x) % 2 != 0
+    if isodd(sizeof(x))
         # llvm instruction is invalid
-        error("unimplemented")
+        bswap_simple(x)
     else
         bswap_int(x)
     end
 end
 
+# llvm is clever enough to transform that into `bswap` of the input truncated to the correct
+# size (8 bits less), followed by a "funnel shift left" `fshl`
+function bswap_simple(x::XBI)
+    y = zero(x)
+    for _ = 1:sizeof(x)
+        y <<= 8
+        y |= x % UInt8
+        x >>>= 8
+    end
+    y
+end
+
 flipsign(x::T, y::T) where {T<:XBS} = flipsign_int(x, y)
 
 # this doesn't catch flipsign(x::BBS, y::BBS), which is more specific in Base

test/runtests.jl

@@ -180,15 +180,25 @@ end
     k, l = rand(Int(typemin(Int8)):-1, 2)
     for X in XInts
         for op in (~, bswap)
-            if VERSION < v"1.6" && sizeof(X) % 2 != 0 && op == bswap
-                @test_throws ErrorException op(X(i))
-                continue
-            end
             x = op(X(i))
             @test x isa X
             @test x != X(i) # we assume sizeof(X) > 8
             @test op(x) == X(i)
         end
+        # bswap specific
+        if VERSION >= v"1.6"
+            for y = rand(X, 20)
+                x = bswap(y)
+                if iseven(sizeof(x))
+                    # test that default implemented matches bswap_simple
+                    @test x == BitIntegers.bswap_simple(y)
+                else
+                    # test that default implemented (i.e. bswap_simple) matches bswap_odd
+                    @test x == bswap_odd(y)
+                end
+            end
+        end
+
         for op in (count_ones, leading_zeros, trailing_zeros, leading_ones, trailing_ones)
             r = op(X(i))
             @test r isa Int

test/setup.jl

@@ -48,4 +48,29 @@ const TypeCombos =
     [((X, Y) for X in Ints for Y in (X ∈ BInts ? XInts : Ints))...,
      (Int, Int), (Int, UInt), (UInt, Int), (UInt, UInt)]
 
+
+# initial attempt at implementing bswap for odd byte-sizes, but it's less efficient
+# than the `bswap_simple` version after being optimized by llvm!
+# let's keep this version around to test values
+if VERSION >= v"1.6"
+    @generated function bswap_odd(x::BitIntegers.XBI)
+        @assert isodd(sizeof(x))
+        nb = sizeof(x) * 8
+        ix = "i$nb"
+        iy = "i$(nb+8)"
+        quote
+            Base.llvmcall(($"""
+                           declare $iy @llvm.bswap.$iy($iy %Val)
+                           define $ix @entry($ix) {
+                               %y = zext $ix %0 to $iy
+                               %y2 = call $iy @llvm.bswap.$iy($iy %y)
+                               %y3 = lshr $iy %y2, 8
+                               %x2 = trunc $iy %y3 to $ix
+                               ret $ix %x2
+                           }
+                           """, "entry"), $x, Tuple{$x}, x)
+        end
+    end
+end
+
 nothing