[SYCL] Matching sizeof long double of SYCL device with that of host

[schittir]

static_assert of sizeof(long double)>8 fails to compile with -fsycl
switch because sizeof(long double) for SYCL device is 8 bytes whereas
the host is 16 bytes. This commit sets sizeof(long double) width to 16
bytes to match that of SYCL host for Linux gnu SYCLdevice x86_64 and x86

[Fznamznon]

AFAIK community does such things in a bit other way. See 8557d1a .

[erichkeane]

@Fznamznon : That one is interesting! I wonder if we should just do that for SYCL as well (modify that 'if' statement). I think this one is 'the right' way, but if community has already made that selection perhaps we need to as well.

Additionally, I think that method actually correctly handles the double-size flags, which this doesn't seem to do.

[Fznamznon]

@Fznamznon : That one is interesting! I wonder if we should just do that for SYCL as well (modify that 'if' statement). I think this one is 'the right' way, but if community has already made that selection perhaps we need to as well.

I think we should be aligned with community code and generalize approaches between SYCL and other offloading models which already is presented in master (like OpenMP), because they solve similar problems quite often. One problem I know with this community method - it crashes if user didn't specify AUX triple, but It is easy to fix by adding check on AUX triple presence.
If the method which this PR exposes is better (I cannot say for sure, I'm not an expert in this place), I think we should generalize approach between OpenMP and SYCL and upstream it.

[erichkeane]

Please stop force pushing. I've sent you an email on how to fix this.