AArch64 specific builtins/intrinsics

[bcardosolopes]

We currently don't emit ARM64 specific intrinsics/builtins, nor none for other arches as well. See clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp for the paths full of asserts. The suggested way to tackle these are by incrementally adding support to the missing pieces - there's a huge number of them.

[khoipkv]

Hey, is this issue still open?

[bcardosolopes]

Yep, and there's enough work for multiple folks!

[bcardosolopes]

See commented tests in clang/test/CIR/CodeGen/aarch64-neon-intrinsics.c. This is just a subset, there is even more if you search in clang/test/CodeGen.

[khoipkv]

If so, could you please assign this one to me? If anyone wants to join in, we add more.

[bcardosolopes]

@khoipkv Here's a list of ones I'd give priority:

Start with:
NEON::BI__builtin_neon_vld1_lane_v
NEON::BI__builtin_neon_vld1q_lane_v
NEON::BI__builtin_neon_vst1_lane_v
NEON::BI__builtin_neon_vst1q_lane_v
NEON::BI__builtin_neon_vld1_dup_v
NEON::BI__builtin_neon_vld1q_dup_v

Next ones:
NEON::BI__builtin_neon_vsetq_lane_f64
NEON::BI__builtin_neon_vdups_lane_i32
NEON::BI__builtin_neon_vrshrn_n_v
NEON::BI__builtin_neon_vget_lane_i8
NEON::BI__builtin_neon_vdupb_lane_i8
clang::AArch64::BI__builtin_arm_ldrex
clang::AArch64::BI__builtin_arm_ldaex
NEON::BI__builtin_neon_vqrshrun_n_v
NEON::BI__builtin_neon_vsra_n_v
NEON::BI__builtin_neon_vsraq_n_v
NEON::BI__builtin_neon_vget_lane_i64
NEON::BI__builtin_neon_vdupd_lane_i64
NEON::BI__builtin_neon_vgetq_lane_i64
NEON::BI__builtin_neon_vdupd_laneq_i64

I suggest you try to get one PR for each individually (or few relevant ones together), incremental is best to make progress and ramp up.

After you complete this list, you might gain a bit more familiarity with how builtins work. Then, you could go for the other bulk ones (table driven), like in clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp:2242:

  // Not all intrinsics handled by the common case work for AArch64 yet, so only
  // defer to common code if it's been added to our special map.
  Builtin = findARMVectorIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID,
                                        AArch64SIMDIntrinsicsProvenSorted);
  if (Builtin) {
    llvm_unreachable("NYI");
  }

(cc: @dkolsen-pgi)

[NoumanAmir657]

Is there work still left to be done on this?

[AmrDeveloper]

@NoumanAmir657 There is a lot of work still left, you can check the file clang/lib/CIR/CodeGen/CIRGenBuiltinAArch64.cpp and search for llvm_unreachable or case NEON::BI__builtin

[bcardosolopes]

@AmrDeveloper thanks a bunch for working on this, really appreciate your contributions :)

[AmrDeveloper]

@AmrDeveloper thanks a bunch for working on this, really appreciate your contributions :)

Thank you @bcardosolopes 🙏🏻

[bcardosolopes]

FWIW, just hit this NYI in our codebase: BI__builtin_neon_vzipq_v