[SYCL] Add an attr tablegen feature for checking langopts

[AaronBallman]

For SYCL, we want some attributes to be ignored in host mode, but
without diagnosing the attribute as being ignored. This adds a bit to
the tablegen emitter to specify that a given language option opts into
this behavior. This allows us to more naturally specify that an
attribute is a device-only attribute from Attr.td without adding logic
in SemaDeclAttr.cpp to bail out in host mode (which looks like a bug
due to the lack of an ignored attribute warning).

As a drive-by, this also corrects a think-o with the
[[intel::reqd_sub_group_size]] attribute in OpenCL mode, which was
always being silently dropped.

[AaronBallman]

I should note: there are more attributes in Attr.td that should be investigated. I hit all the ones that should be an NFC change (aside from the OpenCL-related fix).

We have some attributes that have no spellings (those don't need any LangOpts) and we have a bunch that have no checking for them in SemaDeclAttr.cpp but specify both device and host in Attr.td.

[smanna12]

I should note: there are more attributes in Attr.td that should be investigated. I hit all the ones that should be an NFC change (aside from the OpenCL-related fix).

We have some attributes that have no spellings (those don't need any LangOpts) and we have a bunch that have no checking for them in SemaDeclAttr.cpp but specify both device and host in Attr.td.

inside SemaStmtAttr.cpp, i see checking for all loop attributes.
if (S.LangOpts.SYCLIsHost)
return nullptr;

I think they all will be removed as well.

[AaronBallman]

I should note: there are more attributes in Attr.td that should be investigated. I hit all the ones that should be an NFC change (aside from the OpenCL-related fix).
We have some attributes that have no spellings (those don't need any LangOpts) and we have a bunch that have no checking for them in SemaDeclAttr.cpp but specify both device and host in Attr.td.

inside SemaStmtAttr.cpp, i see checking for all loop attributes.
if (S.LangOpts.SYCLIsHost)
return nullptr;

I think they all will be removed as well.

Good call, I'll work on those for this patch as well.

[AaronBallman]

I should note: there are more attributes in Attr.td that should be investigated. I hit all the ones that should be an NFC change (aside from the OpenCL-related fix).
We have some attributes that have no spellings (those don't need any LangOpts) and we have a bunch that have no checking for them in SemaDeclAttr.cpp but specify both device and host in Attr.td.

inside SemaStmtAttr.cpp, i see checking for all loop attributes.
if (S.LangOpts.SYCLIsHost)
return nullptr;
I think they all will be removed as well.

Good call, I'll work on those for this patch as well.

I've added the right definitions to Attr.td, but that's as far as I think I'll go with that in this patch. Statement attributes do not have the same common attribute handling code as declaration attributes, so we don't take advantage of much of the tablegen goodies for those. I need to work on a patch to improve that, but that'll be an upstream patch first most likely.

[smanna12]

I should note: there are more attributes in Attr.td that should be investigated. I hit all the ones that should be an NFC change (aside from the OpenCL-related fix).
We have some attributes that have no spellings (those don't need any LangOpts) and we have a bunch that have no checking for them in SemaDeclAttr.cpp but specify both device and host in Attr.td.

inside SemaStmtAttr.cpp, i see checking for all loop attributes.
if (S.LangOpts.SYCLIsHost)
return nullptr;
I think they all will be removed as well.

Good call, I'll work on those for this patch as well.

I've added the right definitions to Attr.td, but that's as far as I think I'll go with that in this patch. Statement attributes do not have the same common attribute handling code as declaration attributes, so we don't take advantage of much of the tablegen goodies for those. I need to work on a patch to improve that, but that'll be an upstream patch first most likely.

I should note: there are more attributes in Attr.td that should be investigated. I hit all the ones that should be an NFC change (aside from the OpenCL-related fix).
We have some attributes that have no spellings (those don't need any LangOpts) and we have a bunch that have no checking for them in SemaDeclAttr.cpp but specify both device and host in Attr.td.

inside SemaStmtAttr.cpp, i see checking for all loop attributes.
if (S.LangOpts.SYCLIsHost)
return nullptr;
I think they all will be removed as well.

Good call, I'll work on those for this patch as well.

I've added the right definitions to Attr.td, but that's as far as I think I'll go with that in this patch. Statement attributes do not have the same common attribute handling code as declaration attributes, so we don't take advantage of much of the tablegen goodies for those. I need to work on a patch to improve that, but that'll be an upstream patch first most likely.

Thanks for the update and explanation.

[elizabethandrews]

Change looks ok to me but I am not familiar with TableGen code. So @erichkeane @premanandrao could you also please take a look?

[smanna12]

LGTM