llvm-spirv: Set Opaque Pointers support also when compiling LLVM to SPIR-V

[karolherbst]

LLVM implicitly upgrades modules to opaque pointers, so we have to do it also when translating LLVM to SPIR-V.

This should also get backported to branches containing 0fe726f (and backports).

[karolherbst]

Seems like a lot of tests are failing, because they expect -emit-opaque-pointers to be set. Maybe it's better to only do this for the llvm-16 based branch for now and hope everything is going to be fixed with llvm-17.

But maybe fixing the translator is the better approach here, though for llvm-16 I think it's better to fallback to keep opaque pointers disabled.

[karolherbst]

maybe we even want to disable it by default as this might break users? Not quite sure. Any thoughts on this?

[MrSidims]

AFAIK opaque pointers are not default for LLVM 16, so I'd vote not to enable them here. @jcranmer-intel WDYT?

[karolherbst]

AFAIK opaque pointers are not default for LLVM 16, so I'd vote not to enable them here. @jcranmer-intel WDYT?

they are on by default though, so that's why I need this change so the LLVMContext knows the translator wants them disabled.

LLVM tooling implicitly upgrades to opaque pointers in its tools as well (e.g. link or opt) unless you state you want them disabled, but that's a different problem.

It's already set for the "SPIRV to LLVM" path, just not in the "LLVM to SPIRV" one

[jcranmer-intel]

I've created #2074 for flipping the default of -emit-opaque-pointers on trunk.

The proper way to set the flag for LLVM-to-SPIR-V translation is to use the -opaque-pointers=0 flag for versions that don't autodetect LLVM IR on output (which is 16 and newer). I don't see a need to add a redundant -emit-opaque-pointers=0 for the forward translation, especially given that the name is inherently wrong.

[karolherbst]

I've created #2074 for flipping the default of -emit-opaque-pointers on trunk.

The proper way to set the flag for LLVM-to-SPIR-V translation is to use the -opaque-pointers=0 flag for versions that don't autodetect LLVM IR on output (which is 16 and newer). I don't see a need to add a redundant -emit-opaque-pointers=0 for the forward translation, especially given that the name is inherently wrong.

but that doesn't work? llvm-spirv consumes the LLVM IR (even with opaque pointers disabled) and upgrades it to opaque pointers breaking various things translating to SPIR-V without this change.