diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6dd60d49..1832f8f3 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -495,6 +495,7 @@ if (TESTS)
         crypto/tea
         crypto/xtea
         crypto/xxtea
+        endian
         exe
         extent
         fixed
diff --git a/endian.hpp b/endian.hpp
index aef5087e..e63d7ff9 100644
--- a/endian.hpp
+++ b/endian.hpp
@@ -33,7 +33,7 @@ namespace util {
     //-------------------------------------------------------------------------
     template <typename T>
     constexpr T
-    bswap (T);
+    bswap (T) noexcept;
 
     // CXX: Update using "if constexpr" when available. It doesn't seem to be
     // worth the dicking about to use enable_if_t to differentiate the
@@ -42,7 +42,7 @@ namespace util {
     template <>                             \
     constexpr                               \
     int##S##_t                              \
-    bswap (int##S##_t v) {                  \
+    bswap (int##S##_t v) noexcept {         \
         const union {                       \
              int##S##_t s;                  \
             uint##S##_t u;                  \
@@ -55,12 +55,31 @@ namespace util {
     SIGNED_BSWAP(32)
     SIGNED_BSWAP(64)
 
-    template <> constexpr  int8_t bswap ( int8_t v) { return v; }
-    template <> constexpr uint8_t bswap (uint8_t v) { return v; }
+    template <> constexpr  int8_t bswap ( int8_t v) noexcept { return v; }
+    template <> constexpr uint8_t bswap (uint8_t v) noexcept { return v; }
 
-    template <> constexpr uint16_t bswap (uint16_t v) { return __builtin_bswap16 (v); }
-    template <> constexpr uint32_t bswap (uint32_t v) { return __builtin_bswap32 (v); }
-    template <> constexpr uint64_t bswap (uint64_t v) { return __builtin_bswap64 (v); }
+    template <> constexpr uint16_t bswap (uint16_t v) noexcept { return __builtin_bswap16 (v); }
+    template <> constexpr uint32_t bswap (uint32_t v) noexcept { return __builtin_bswap32 (v); }
+    template <> constexpr uint64_t bswap (uint64_t v) noexcept { return __builtin_bswap64 (v); }
+
+
+    // use a type-punning union to punt the byte swapping operation to the
+    // unsigned integer code.
+    //
+    // this is debatably safe under production compilers like gcc + clang
+    // despite what the standard may say about unions.
+    template <>
+    constexpr float
+    bswap (float v) noexcept
+    {
+        union {
+            float f;
+            uint32_t u;
+        } temp { .f = v };
+
+        temp.u = bswap (temp.u);
+        return temp.f;
+    }
 
 
     //-------------------------------------------------------------------------
diff --git a/test/endian.cpp b/test/endian.cpp
new file mode 100644
index 00000000..d91f4ad2
--- /dev/null
+++ b/test/endian.cpp
@@ -0,0 +1,31 @@
+#include "tap.hpp"
+
+#include "endian.hpp"
+
+
+///////////////////////////////////////////////////////////////////////////////
+int
+main (void)
+{
+    util::TAP::logger tap;
+
+    {
+        uint32_t a = 0x12345678, b = 0x78563412;
+        tap.expect_eq (a, util::bswap (b), "u32 byteswap");
+    }
+
+    {
+        // try to byte swap the pattern for 1.0f
+        //
+        // it may not be the most robust test, but it'll catch the most
+        // egregious errors for the time being.
+        union {
+            uint32_t u;
+            float f;
+        } data { .u = 0x0000803f /* 0x3f800000 == 1.f */ };
+
+        tap.expect_eq (util::bswap (data.f), 1.f, "f32 byteswap");
+    };
+
+    return tap.status ();
+}
\ No newline at end of file