[SYCL][libdevice] Keep sign bit of result for std::sin(-0) (#17440)

For input '-0', __spirv_ocl_sin returns +0 for float and returns -0 for
double, MSVC complex math implementation does depend on this, discarding
the signbit for sin(-0) will lead to some corner case failures in
exp(std::complex<float>) e2e tests. This patch is a workaround for the
issue.

---------

Signed-off-by: jinge90 <[email protected]>

Author: jinge90

libdevice/fallback-cmath.cpp

@@ -168,8 +168,13 @@ float __devicelib_fmaf(float x, float y, float z) {
   return __spirv_ocl_fma(x, y, z);
 }
 
+#if defined(__SPIR__) || defined(__SPIRV__)
+DEVICE_EXTERN_C_INLINE
+float __devicelib_sinf(float x) { return (x == 0.0f) ? x : __spirv_ocl_sin(x); }
+#else
 DEVICE_EXTERN_C_INLINE
 float __devicelib_sinf(float x) { return __spirv_ocl_sin(x); }
+#endif
 
 DEVICE_EXTERN_C_INLINE
 float __devicelib_cosf(float x) { return __spirv_ocl_cos(x); }