Each of the main TableGen files for the supported targets includes
include "llvm/Target/Target.td"
In turn, Target.td includes
include "llvm/IR/Intrinsics.td"
The final lines of Instrinsics.td are
include "llvm/IR/IntrinsicsPowerPC.td"
include "llvm/IR/IntrinsicsX86.td"
include "llvm/IR/IntrinsicsARM.td"
include "llvm/IR/IntrinsicsAArch64.td"
include "llvm/IR/IntrinsicsXCore.td"
include "llvm/IR/IntrinsicsHexagon.td"
include "llvm/IR/IntrinsicsNVVM.td"
include "llvm/IR/IntrinsicsMips.td"
include "llvm/IR/IntrinsicsAMDGPU.td"
include "llvm/IR/IntrinsicsBPF.td"
include "llvm/IR/IntrinsicsSystemZ.td"
include "llvm/IR/IntrinsicsWebAssembly.td"
include "llvm/IR/IntrinsicsRISCV.td"
Why does every target include the all the instrinsics for all the targets?
For example, when I process the ARC TableGen file, the records include 1,187 intrinsics for the X86.
the intrinsics of all backends together with the target-agnostic
intrinsics are all part of a single large enum, and there are some
subtle assumptions made e.g. about how the values in this enum are
sorted (in order to emit tables that are suitable for binary search).
So this is an area in which to tread carefully.
That said, I don't think there's any inherent reason why we
_shouldn't_ try to entangle this a bit.
I've been thinking for a while now that in a sense, the
target-specific intrinsics are like a dialect of LLVM IR. It would
make sense to treat them as such more explicitly.
If I decide to tackle this issue, I will certainly post a proposal first. I think it will be down the road a bit, as I learn more about how the system works.
Is there a place online where the results of a full build are posted?
But yes, they use the same enumerator space, and tablgen processes them all at once.
Processing intrinsics separately by target requires dividing the 32-bit intrinsic opcode space up front. For example, we could use 8 bits for target, 24 for intrinsics. This would make intrinsic ids no longer dense, and no longer suitable as array indices, but perhaps having distinct arrays per target is better anyway.
Another question: Is there a reason why all the TableGen intrinsic files have to be included in every run of TableGen, one per backend per target? Couldn't those files be included only when running backends pertinent to the intrinsics?