Add SIMD support postmortem

papers/P2728.md

@@ -1202,6 +1202,54 @@ There's precedent for this kind of approach in the `views::reverse` CPO, which s
 gives back the original underlying view if it detects that it's reversing another
 `reverse_view`.
 
+## Why no SIMD support?
+
+I created an experimental implementation with SIMD support for UTF-16 to UTF-8
+transcoding, using a 64-byte readahead buffer, and benchmarked it against both
+the scalar implementation and the highly optimized bulk `simdutf` library by
+Daniel Lemire et al., obtaining the following numbers in GiB/s of input
+consumed (GCC 16.1, `-O3 -march=native`, x86-64 AVX2, AMD Ryzen 9 5950X).
+
+| Corpus | Scalar view | Prototype SIMD view | Bulk `simdutf` |
+|---|---:|---:|---:|
+| Latin | 1.86 | 4.20 | 60.4 |
+| Arabic | 0.81 | 0.96 | 10.9 |
+| Chinese | 0.82 | 1.00 | 8.9 |
+| Japanese | 0.77 | 0.85 | 8.7 |
+| Korean | 0.89 | 0.86 | 8.0 |
+
+The inputs used were Daniel Lemire's
+[unicode_lipsum](https://github.com/lemire/unicode_lipsum) corpora.
+
+The reason the performance improvements are so limited is that a view delivers
+its output one code unit at a time: every element costs an iterator increment,
+a dereference, and a buffer-index check, no matter how cheaply the buffer was
+filled. SIMD accelerates only the buffer refill; once that cost is amortized
+away, throughput is bounded by the per-element delivery loop. That leads to the
+relatively disappointing results seen above as compared to `simdutf`, which has
+an eager, algorithm-style API instead of a view API.
+
+In addition to the comparatively unimpressive performance numbers, some SG9
+reviewers were also concerned that the SIMD implementation needed to eagerly
+buffer more readahead than the UTF transcoding algorithm strictly requires in
+order to fill a SIMD register, which some considered to violate the spirit of
+laziness that the view library is designed around.
+
+Furthermore, we have a potential mechanism for adding SIMD support later if we
+change our minds, by adding a new enumerator to `to_utf_view_kind`.
+
+Based on this information, SG9 voted against supporting SIMD:
+
+**Poll: We want to leave implementations the freedom to eagerly buffer more code units to allow a SIMD transcoding technique.**
+
+ SF| F | N | A | SA
+---|---|---|---|---
+  0| 0 | 3 | 3 | 1
+
+Attendance: 8
+Author: N
+Strong consensus against.
+
 ## `.base_code_units()`
 
 ### Proposal
@@ -1588,7 +1636,7 @@ An experimental implementation of `.base_code_units()` is available on the
 
 ## Changes since R14
 
-- Revert SIMD support
+- Revert SIMD support and add design discussion of why it was rejected
 
 ## Changes since R13