[SYCL] Use union to do bitcast after bif removal in 8be6b10

Author: jsji

sycl/include/sycl/ext/oneapi/experimental/cuda/masked_shuffles.hpp

@@ -39,9 +39,15 @@ namespace detail {
           mask, static_cast<int>(tmp_b16), shfl_param, c);                     \
       asm("mov.b16 %0,%1;" : "=h"(res) : "h"(static_cast<short>(tmp_b32)));    \
     } else if constexpr (std::is_same_v<T, float>) {                           \
-      auto tmp_b32 = __nvvm_shfl_sync_##SHUFFLE_INSTR(                         \
-          mask, __nvvm_bitcast_f2i(val), shfl_param, c);                       \
-      res = __nvvm_bitcast_i2f(tmp_b32);                                       \
+      union {                                                                  \
+        int i;                                                                 \
+        float f;                                                               \
+      } u;                                                                     \
+      u.f = val;                                                               \
+      auto tmp_b32 =                                                           \
+          __nvvm_shfl_sync_##SHUFFLE_INSTR(mask, u.i, shfl_param, c);          \
+      u.i = tmp_b32;                                                           \
+      res = u.f;                                                               \
     } else {                                                                   \
       res = __nvvm_shfl_sync_##SHUFFLE_INSTR(mask, val, shfl_param, c);        \
     }                                                                          \

sycl/include/sycl/ext/oneapi/matrix/matrix-unified.hpp

@@ -580,7 +580,12 @@ inline __SYCL_ALWAYS_INLINE float round_to_tf32(const float &a) {
 #if defined(__SYCL_DEVICE_ONLY__)
 #if defined(__NVPTX__)
   int32_t tmp_int = __nvvm_f2tf32_rna(a);
-  return __nvvm_bitcast_i2f(tmp_int);
+  union {
+    int32_t i;
+    float f;
+  } u;
+  u.i = tmp_int;
+  return u.f;
 #else
   return __spirv_RoundFToTF32INTEL(a);
 #endif // defined(__NVPTX__)