[PATCH] D135930: [X86] Add AVX-NE-CONVERT instructions.

Mon Oct 31 01:29:36 PDT 2022

FreddyYe marked an inline comment as done.
FreddyYe added inline comments.

================
Comment at: clang/lib/Headers/avx512vlbf16intrin.h:164
+#define _mm_cvtneps_pbh(A) \
+  ((__m128bh)__builtin_ia32_vcvtneps2bf16128((__v4sf)(A)))

----------------
pengfei wrote:
> pengfei wrote:
> > RKSimon wrote:
> > > Is there no way for __attribute__ to allow different attribute permutations?
> > > 
> > > Also, can we keep the __builtin_ia32_cvtneps2bf16_128 naming convention?
> > > Is there no way for attribute to allow different attribute permutations?
> > 
> > We have discussed this problem with GCC folks. There are two problems here:
> > 1. Unlike builtins, function attributes are more generic. It may introduce a lot of checks between callers and callees. I had a research to limit it to `__always_inline__` functions only. However, Clang handles inlining in middle-end, we don't have such information in the front-end. Besides, we don't know how to merge different permutations if they are inlining to the same function.
> > 2. We don't know how to put the permutations into IR's function attributes. We need to preserve all permutations for inlining reference, but the backend needs a determine feature list rather than selective.
> It's better to use `__builtin_ia32_cvtneps2bf16_128`.
I think __builtin_ia32_vcvtneps2bf16128 is also a "right" name.

See __builtin_ia32_vfmaddsubph256, __builtin_ia32_minph256...

And I admit naming conventions of clang builtins as well as LLVM IR builtins are confusing right now.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135930/new/

https://reviews.llvm.org/D135930