[SYCL] Ignore 'used' attribute when emitting sycl device code

[premanandrao]

Signed-off-by: Premanand M Rao [email protected]

[bader]

@premanandrao, what is the motivation for this change?
At some point I was thinking that we might use this attribute to implement SYCL_EXTERNAL macro to avoid adding another custom attribute.
Is it used in some legacy code, which we compile "as is" in SYCL mode?

[Fznamznon]

@premanandrao, what is the motivation for this change?
At some point I was thinking that we might use this attribute to implement SYCL_EXTERNAL macro to avoid adding another custom attribute.
Is it used in some legacy code, which we compile "as is" in SYCL mode?

The original problem was with crash due to address space mismatch. But there was emission of some global variable with initializers forced by used attribute, I don't think that used attribute should force emission of device code unless SYCL spec requires it (and it doesn't).
I also don't think that it is a good idea to use used attrubute to implement SYCL_EXTERNAL because they may have different semantics.

[erichkeane]

@premanandrao, what is the motivation for this change?
At some point I was thinking that we might use this attribute to implement SYCL_EXTERNAL macro to avoid adding another custom attribute.
Is it used in some legacy code, which we compile "as is" in SYCL mode?

The original problem was with crash due to address space mismatch. But there was emission of some global variable with initializers forced by used attribute, I don't think that used attribute should force emission of device code unless SYCL spec requires it (and it doesn't).
I also don't think that it is a good idea to use used attrubute to implement SYCL_EXTERNAL because they may have different semantics.

I agree that we shouldn't use 'used' as SYCL_EXTERNAL. It has the problem of being used in existing code, so we'll end up having some odd variety of non-SYCL_EXTERNAL code be valid/linkable/callable, simply because the user did attribute-used in the past.

I believe a new/novel attribute is the correct way to go here.

[premanandrao]

I believe a new/novel attribute is the correct way to go here.

+1

[premanandrao]

Thanks @bader