[PATCH] D69250: [ARM][AArch64] Implement __cls and __clsl intrinsics from ACLE

[Saleem Abdulrasool via Phabricator via cfe-commits]

compnerd added inline comments.


================
Comment at: clang/lib/Headers/arm_acle.h:150
+__clsl(unsigned long __t) {
+#if __SIZEOF_LONG__ == 4
+  return __builtin_arm_cls(__t);
----------------
vhscampos wrote:
> compnerd wrote:
> > vhscampos wrote:
> > > compnerd wrote:
> > > > I don't see a pattern match for the `cls64` on ARM32, would that not fail to lower?
> > > Yes. However, for now, I am not enabling support for `cls64` on ARM32 as it is not done yet.
> > Is the difference not just the parameter type?  I think that implementing it should be a trivial change to the existing implementation.  Is there a reason that you are not implementing that?
> At clang's side, yes, but not in the backend: Arm32 does not have a `cls` instruction, thus the CLS operations need to be custom lowered. In the `llvm.arm.cls(i32)` case, lowering is quite simple, and it's been included in this patch. For `llvm.arm.cls64(i64)`, on the other hand, it is not as trivial since it's necessary to break its logic into 32-bit instructions.
> 
> So the reason not to implement that (yet) is just to split work in two different efforts.
Would it not be sufficient to do the top half (after a shift right of 32-bits), and if it is exactly 32, then do the bottom 32-bits, otherwise, you're done?


================
Comment at: clang/lib/Headers/arm_acle.h:155
+#endif
+}
+
----------------
vhscampos wrote:
> compnerd wrote:
> > vhscampos wrote:
> > > compnerd wrote:
> > > > Should we have a `__clsll` extension, otherwise these two are the same in LLP64?  I'm thinking about the LLP64 environments, where `long` and `long long` are different (32-bit vs 64-bit).
> > > ACLE does provide a `long long` version of `cls` called `__clsll`. But since the support for `cls64` on Arm32 is not done yet, I decided not to write support for `__clsll`. If I did, it would work for 64-bit but not for 32-bit.
> > > 
> > > Please let me know what you think.
> > clang supports Windows where `long` is 4-bytes even on 64-bit targets, and this means that this doesn't work for that target.  I think that we need to add `__clsll` so that 64-bit ARM at least is covered.
> I'm not sure if I am following you. On AArch64-Windows, `__clsl` will be lowered to `llvm.aarch64.cls(i32)` which will then be custom lowered correctly. Let me know if I am thinking this wrong.
Windows is LLP64, `__cls` and `__clsl` are the same thing.  It needs a `__clsll` for the 64-bit value.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69250/new/

https://reviews.llvm.org/D69250