[PATCH] libstdc++: Improve simd::cat implementation strategy

[Matthias Kretz]

This goes on top of the main [simd] patch and is in preparation of an 
implementation for [simd.permute.dynamic].

How could/should I test this change? I have a test that compiles

auto f10(simd::vec<short, 2> a, simd::vec<short, 2> b,
	 simd::vec<short, 2> c, simd::vec<short, 2> d,
	 simd::vec<short, 8> e) {
  g(simd::cat(a, b, c, d, e));
}

to
	vmovdqa	xmm3, xmm0
	vmovd	xmm1, esi
	vmovd	xmm0, edi
	vinsertps	xmm0, xmm0, xmm1, 16
	vmovd	xmm2, ecx
	vmovd	xmm1, edx
	vinsertps	xmm1, xmm1, xmm2, 16
	vmovq	xmm0, xmm0
	vmovq	xmm1, xmm1
	vpunpcklqdq	xmm0, xmm0, xmm1
	vmovdqa	xmm1, xmm3
	vmovdqa	xmm0, xmm0
	vperm2i128	ymm0, ymm0, ymm1, 32

The 2x insertps -> unpck -> 128-bit concat sequence shows it's doing the 
expected sequence. But I really hope this will turn into
        vmovd   xmm2, edi
        vmovd   xmm1, edx
        vpinsrd xmm2, xmm2, esi, 1
        vpinsrd xmm1, xmm1, ecx, 1
        vpunpcklqdq     xmm1, xmm2, xmm1
        vinserti128     ymm0, ymm1, xmm0, 1

at some point. I don't think we want such detailed code-gen tests here. A test 
like this belongs into the gcc/testsuite.

---

cat(a, b, c, d) where each argument is e.g. 2 elements wide, would fold
from the left before this change: (2, 2, 2, 2) -> (4, 2, 2) -> (6, 2) ->
(8). It is better for ILP (and to avoid load and store instructions) to
go via (4, 2, 2) -> (4, 4) -> (8).

In theory, for even larger number of arguments, the current
implementation still isn't good enough. But larger number of arguments
is something users shouldn't be doing anyway.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>

libstdc++-v3/ChangeLog:

	* include/bits/vec_ops.h (__vec_concat_sized): Add an overload
	that concatenates the second and third operand, if they are
	smaller than the first.
---
 libstdc++-v3/include/bits/vec_ops.h | 23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

diff --git a/libstdc++-v3/include/bits/vec_ops.h b/libstdc++-v3/include/bits/
vec_ops.h
index 0e89c89b7af5..e5bf2f1497cd 100644
--- a/libstdc++-v3/include/bits/vec_ops.h
+++ b/libstdc++-v3/include/bits/vec_ops.h
@@ -187,7 +187,30 @@ __vec_concat(_TV __a, _TV __b)
    * with the elements from applying this function recursively to @p __rest.
    *
    * @pre _N0 <= __width_of<_TV0> && _N1 <= __width_of<_TV1> && _Ns <= 
__width_of<_TVs> && ...
+   *
+   * Strategy: Aim for a power-of-2 tree concat. E.g.
+   * - cat(2, 2, 2, 2) -> cat(4, 2, 2) -> cat(4, 4)
+   * - cat(2, 2, 2, 2, 8) -> cat(4, 2, 2, 8) -> cat(4, 4, 8) -> cat(8, 8)
    */
+  template <int _N0, int _N1, int... _Ns, __vec_builtin _TV0, __vec_builtin 
_TV1,
+	   __vec_builtin... _TVs>
+    [[__gnu__::__always_inline__]]
+    constexpr __vec_builtin_type<__vec_value_type<_TV0>,
+				 __bit_ceil(unsigned(_N0 + (_N1 + ... + 
_Ns)))>
+    __vec_concat_sized(const _TV0& __a, const _TV1& __b, const _TVs&... 
__rest);
+
+  template <int _N0, int _N1, int _N2, int... _Ns, __vec_builtin _TV0, 
__vec_builtin _TV1,
+	    __vec_builtin _TV2, __vec_builtin... _TVs>
+    requires (__has_single_bit(unsigned(_N0))) && (_N0 >= (_N1 + _N2))
+    [[__gnu__::__always_inline__]]
+    constexpr __vec_builtin_type<__vec_value_type<_TV0>,
+				 __bit_ceil(unsigned(_N0 + _N1 + (_N2 + ... + 
_Ns)))>
+    __vec_concat_sized(const _TV0& __a, const _TV1& __b, const _TV2& __c, 
const _TVs&... __rest)
+    {
+      return __vec_concat_sized<_N0, _N1 + _N2, _Ns...>(
+	       __a, __vec_concat_sized<_N1, _N2>(__b, __c), __rest...);
+    }
+
   template <int _N0, int _N1, int... _Ns, __vec_builtin _TV0, __vec_builtin 
_TV1,
 	   __vec_builtin... _TVs>
     [[__gnu__::__always_inline__]]
-- 
──────────────────────────────────────────────────────────────────────────
 Dr. Matthias Kretz                           https://mattkretz.github.io
 GSI Helmholtz Center for Heavy Ion Research               https://gsi.de
 std::simd
──────────────────────────────────────────────────────────────────────────

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: This is a digitally signed message part.
URL: <https://gcc.gnu.org/pipermail/libstdc++/attachments/20260220/459595c6/attachment-0001.sig>