[clang] [Clang] Add elementwise conversions from encoded FP8 values (PR #212647)

[Chinmay Deshpande via cfe-commits]

================
@@ -0,0 +1,30 @@
+//===--- ArbitraryFPFormats.def - Arbitrary FP format database --*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Source encodings exposed by the __builtin_elementwise_convert_from_* family.
+// Src is the source half of the builtin name suffix; LLVMName is the
+// llvm.convert.from.arbitrary.fp interpretation string.
+//
+// Each entry needs a matching instantiation in Builtins.td: expansions name
+// builtin IDs directly, so a stale entry fails to compile. An entry must also
+// be accepted by APFloat::getArbitraryFPSemantics, or lowering fails.
+//
+// The sub-byte encodings (Float6E3M2FN, Float6E2M3FN, Float4E2M1FN) lower but
+// are not exposed; Clang cannot spell their vector element types coherently.
----------------
chinmaydd wrote:

As I understand it, #140253 allows power-of-two `_BitInt` vector elements, so `_BitInt(6)` vectors are still rejected.

FP4 vectors can be declared, but Clang and LLVM disagree on their sizes.
```
typedef unsigned _BitInt(4) v8u4 __attribute__((ext_vector_type(8)));
sizeof(v8u4)        // 8: Clang gives each element a byte
load <8 x i4>, ...  // LLVM packs elements: 32 bits = 4 bytes
```

Scalar `_BitInt(6)` and `_BitInt(4)` would work but I lean towards leaning them out for now such that every exposed format supports both scalars and vectors.

I've cleaned up the comment to specify the underlying problem.

---

Also, filed https://github.com/llvm/llvm-project/issues/230878

https://github.com/llvm/llvm-project/pull/212647