From e8b62e8d72d24171157a230a67d124d150443429 Mon Sep 17 00:00:00 2001
From: Alex Guteniev <gutenev@gmail.com>
Date: Thu, 2 May 2024 09:57:38 +0300
Subject: [PATCH 1/4] Help the compiler vectorize `ranges::iota`

Ranges version of #4627
---
 benchmarks/src/iota.cpp |  2 +-
 stl/inc/numeric         | 16 ++++++++++++++++
 2 files changed, 17 insertions(+), 1 deletion(-)

diff --git a/benchmarks/src/iota.cpp b/benchmarks/src/iota.cpp
index 9d96d12d773..d6fcab9a369 100644
--- a/benchmarks/src/iota.cpp
+++ b/benchmarks/src/iota.cpp
@@ -14,7 +14,7 @@ void bm(benchmark::State& state) {
     std::vector<T> a(size);
 
     for (auto _ : state) {
-        std::iota(a.begin(), a.end(), T{22});
+        std::ranges::iota(a, T{22});
         benchmark::DoNotOptimize(a);
     }
 }
diff --git a/stl/inc/numeric b/stl/inc/numeric
index a0186094e52..31085f86ee5 100644
--- a/stl/inc/numeric
+++ b/stl/inc/numeric
@@ -567,6 +567,22 @@ namespace ranges {
             _STL_INTERNAL_STATIC_ASSERT(weakly_incrementable<_Ty>);
             _STL_INTERNAL_STATIC_ASSERT(indirectly_writable<_It, const _Ty&>);
 
+            if constexpr (_Iterator_is_contiguous<_It> && sized_sentinel_for<_Se, _It> && is_integral_v<_Ty>
+                          && sizeof(_Ty) >= 4) {
+                // TRANSITION, DevCom-10593477: help the compiler vectorize
+                const auto _Ptr  = _To_address(_First);
+                const auto _Size = static_cast<size_t>(_Last - _First);
+
+                if (_STD _In_range<_Ty>(_Size)) {
+                    const auto _Size_typed = static_cast<_Ty>(_Size);
+                    for (_Ty _Ix = 0; _Ix != _Size_typed; ++_Ix) {
+                        _Ptr[_Ix] = _Val + _Ix;
+                    }
+
+                    return _First + _Size;
+                }
+            }
+
             const _Ty& _Const_val = _Val;
             for (; _First != _Last; ++_First, (void) ++_Val) {
                 *_First = _Const_val;

From cedc297372dddb8907244ff396016981f657e3b3 Mon Sep 17 00:00:00 2001
From: Alex Guteniev <gutenev@gmail.com>
Date: Thu, 2 May 2024 11:50:06 +0300
Subject: [PATCH 2/4] missing `const`

---
 stl/inc/numeric | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/stl/inc/numeric b/stl/inc/numeric
index 31085f86ee5..dfa240e82eb 100644
--- a/stl/inc/numeric
+++ b/stl/inc/numeric
@@ -576,7 +576,8 @@ namespace ranges {
                 if (_STD _In_range<_Ty>(_Size)) {
                     const auto _Size_typed = static_cast<_Ty>(_Size);
                     for (_Ty _Ix = 0; _Ix != _Size_typed; ++_Ix) {
-                        _Ptr[_Ix] = _Val + _Ix;
+                        const _Ty _Const_val = _Val + _Ix;
+                        _Ptr[_Ix]            = _Const_val;
                     }
 
                     return _First + _Size;

From c87c40358a2f870ed06fdc3650a1662dd2c322c6 Mon Sep 17 00:00:00 2001
From: Alex Guteniev <gutenev@gmail.com>
Date: Thu, 2 May 2024 19:16:09 +0300
Subject: [PATCH 3/4] Aha!

---
 stl/inc/numeric | 1 +
 1 file changed, 1 insertion(+)

diff --git a/stl/inc/numeric b/stl/inc/numeric
index dfa240e82eb..9f4af6eb911 100644
--- a/stl/inc/numeric
+++ b/stl/inc/numeric
@@ -580,6 +580,7 @@ namespace ranges {
                         _Ptr[_Ix]            = _Const_val;
                     }
 
+                    _Val += _Size_typed;
                     return _First + _Size;
                 }
             }

From 8d4e563285b2de571379eeb3b79acc087d94944e Mon Sep 17 00:00:00 2001
From: "Stephan T. Lavavej" <stl@nuwen.net>
Date: Wed, 15 May 2024 18:59:22 -0700
Subject: [PATCH 4/4] Benchmark both algorithms.

---
 benchmarks/src/iota.cpp | 21 +++++++++++++++++----
 1 file changed, 17 insertions(+), 4 deletions(-)

diff --git a/benchmarks/src/iota.cpp b/benchmarks/src/iota.cpp
index d6fcab9a369..5860b1b605e 100644
--- a/benchmarks/src/iota.cpp
+++ b/benchmarks/src/iota.cpp
@@ -7,14 +7,24 @@
 #include <numeric>
 #include <vector>
 
-template <class T>
+enum class Alg {
+    Std,
+    Rng,
+};
+
+template <class T, Alg Algorithm>
 void bm(benchmark::State& state) {
     const auto size = static_cast<std::size_t>(state.range(0));
 
     std::vector<T> a(size);
 
     for (auto _ : state) {
-        std::ranges::iota(a, T{22});
+        if constexpr (Algorithm == Alg::Std) {
+            std::iota(a.begin(), a.end(), T{22});
+        } else if constexpr (Algorithm == Alg::Rng) {
+            std::ranges::iota(a, T{22});
+        }
+
         benchmark::DoNotOptimize(a);
     }
 }
@@ -23,7 +33,10 @@ void common_args(auto bm) {
     bm->Arg(7)->Arg(18)->Arg(43)->Arg(131)->Arg(315)->Arg(1212);
 }
 
-BENCHMARK(bm<std::uint32_t>)->Apply(common_args);
-BENCHMARK(bm<std::uint64_t>)->Apply(common_args);
+BENCHMARK(bm<std::uint32_t, Alg::Std>)->Apply(common_args);
+BENCHMARK(bm<std::uint64_t, Alg::Std>)->Apply(common_args);
+
+BENCHMARK(bm<std::uint32_t, Alg::Rng>)->Apply(common_args);
+BENCHMARK(bm<std::uint64_t, Alg::Rng>)->Apply(common_args);
 
 BENCHMARK_MAIN();