[SYCL] Enable SPV_INTEL_fp_max_error by default

[MrSidims]

Since a pass that cleans up unnecessary llvm.fpbuiltin intrinsics was removed from
#16107 we need to enable the extension by default to
successfully translate them.

[maarquitos14]

LGTM. Just one question, why was the pass removed?

[MrSidims]

LGTM. Just one question, why was the pass removed?

The reason we originally added it in the first place is the following: clang now generated llvm.fpbuiltin.fdiv/sqrt by default (with 2.5 and 3.0 ULP precision). So the llvm-spirv translator would generate brand new decorations for them by default for every SYCL code that uses div or sqrt operations, making it impossible to use old drivers. So the idea behind the pass was: if LLVM IR module contains only llvm.fpbuiltin.fdiv/sqrt with 2.5 and 3.0 max error attributes, without any fdiv or llvm.sqrt instructions/intrinsics or without any other llvm.fpbuiltin.* intrinsics or without any llvm.fpbuiltin.fdiv/sqrt with other precision (aka default case) - then it means, that we would pass non-precise option to devices backend anyway and on LLVM IR level we can replace llvm.fpbuiltin.fdiv/sqrt with just fdiv and llvm.sqrt preventing generation of new decorations.

On the one hand it seem to be a nice idea, on the other hand oneAPI documentation says: please install the latest GPU drivers on your system, when install oneAPI version xx.yy.zz. So the extension should be implicitly supported by the user's system drivers (until we change oneAPI release criteria) and the pass is not needed from this perspective. And at the same time such pass, that iterates over all intrinsic declarations, analyses their usages and conditionally replaces calls to them is affecting compilation time.

IF we see a case, when OLD drivers must be used - the pass will be restored.

[MrSidims]

@intel/dpcpp-clang-driver-reviewers @intel/dpcpp-cfe-reviewers please take a look

[asudarsa]

LGTM. Thanks

[MrSidims]

@intel/dpcpp-clang-driver-reviewers @intel/dpcpp-cfe-reviewers friendly ping, there are several tests failing without this patch

[mdtoguchi]

OK for driver

[elizabethandrews]

I'm not sure why FE review was required for this since the changes are to the driver and tools (I think) has other code owners. In any case this looks ok to me.

[MrSidims]

@intel/llvm-gatekeepers please help with merge