[Clang][AArch64][SME] Add ZA zeroing intrinsics

[bryanpkc]

Summary

This patch adds support for the following SME ACLE intrinsics (as defined
in https://arm-software.github.io/acle/main/acle.html):

• svzero_mask_za
• svzero_za

Co-authored-by: Sagar Kulkarni <[e-mail]>

[bryanpkc]

Update patch with more context.

[sdesmalen]

clang/lib/CodeGen/CGBuiltin.cpp

9968–9969

Given that the type flags are a little precious (we've already used 42 out of the 64 bits) and there only being a single intrinsic for svzero, can you do something similar to what's done in EmitAArch64SVEBuiltinExpr and create a switch on BuiltinID instead?

clang/test/CodeGen/aarch64-sme-intrinsics/acle_sme_zero.c

14

These tests are missing some kind of ARM_SHARED_ZA_ATTR macro (similar to what you've done for ARM_STREAMING_ATTR in D127910 and D128648), because svzero requires ZA state to be available.

clang/test/Sema/aarch64-sme-intrinsics/acle_sme_imm.cpp

195

svzero is streaming-compatible, but does require shared_za. So you should probably change this to ARM_SHARED_ZA_ATTR (and add the #define for it)

[bryanpkc]

Addressed @sdesmalen's comments.

[bryanpkc]

Rebased. NFC.

[sdesmalen]

Other than my comment on the test, the patch looks good to me.

clang/test/CodeGen/aarch64-sme-intrinsics/acle_sme_zero.c

21

The new keyword attributes are a bit stricter on the placement.

Because this is a type attribute (as opposed to a declaration attribute), it should be placed like this:

void test_svzero_mask_za() ARM_SHARED_ZA_ATTR {

As I mentioned on D128648, I think it's better to remove all these macros for now and only add the attributes when we add diagnostics in Sema for missing attributes.

[bryanpkc]

Removed the attribute macro as suggested.

[sdesmalen]

LGTM