From f6f706d7081b8006cc6e8e815b95018234720a67 Mon Sep 17 00:00:00 2001
From: Oleksandr Kulkov <adamant.pwn@gmail.com>
Date: Fri, 15 Nov 2024 00:01:35 +0100
Subject: [PATCH] Use bmi2 target for kth_set_bit, improve compress_coords and
 radix sort

---
 cp-algo/util/bit.hpp             |  2 +-
 cp-algo/util/compress_coords.hpp | 11 +++++++++--
 cp-algo/util/sort.hpp            |  8 ++++----
 3 files changed, 14 insertions(+), 7 deletions(-)

diff --git a/cp-algo/util/bit.hpp b/cp-algo/util/bit.hpp
index 99ed132..0ad96e6 100644
--- a/cp-algo/util/bit.hpp
+++ b/cp-algo/util/bit.hpp
@@ -11,7 +11,7 @@ namespace cp_algo {
     size_t order_of_bit(auto x, size_t k) {
         return k ? std::popcount(x << (bit_width<decltype(x)> - k)) : 0;
     }
-    // Requires GCC target("popcnt,bmi2")
+    [[gnu::target("bmi2")]]
     size_t kth_set_bit(uint64_t x, size_t k) {
         return std::countr_zero(_pdep_u64(1ULL << k, x));
     }
diff --git a/cp-algo/util/compress_coords.hpp b/cp-algo/util/compress_coords.hpp
index a9e36c2..39558d2 100644
--- a/cp-algo/util/compress_coords.hpp
+++ b/cp-algo/util/compress_coords.hpp
@@ -5,10 +5,17 @@
 namespace cp_algo {
     // coords is a range of reference_wrapper<T>
     auto compress_coords(auto &coords) {
-        std::vector<int> original;
+        using T = std::decay_t<std::unwrap_reference_t<
+            std::ranges::range_value_t<decltype(coords)>
+        >>;
+        std::vector<T> original;
+        if(empty(coords)) {
+            return original;
+        }
         original.reserve(size(coords));
         radix_sort(coords);
-        int idx = -1, prev = -1;
+        size_t idx = -1;
+        T prev = ~coords.front();
         for(auto &x: coords) {
             if(x != prev) {
                 idx++;
diff --git a/cp-algo/util/sort.hpp b/cp-algo/util/sort.hpp
index 9466da9..af5466b 100644
--- a/cp-algo/util/sort.hpp
+++ b/cp-algo/util/sort.hpp
@@ -24,12 +24,12 @@ namespace cp_algo {
         if(empty(a)) {
             return;
         }
-        auto mx = std::ranges::max(a);
-        with_bit_floor(size(a), [&]<size_t floor>() {
+        auto [mn, mx] = std::ranges::minmax(a);
+        with_bit_floor<1>(size(a), [&]<size_t floor>() {
             constexpr int base = std::min<size_t>(floor, 1 << 16);
-            for(int64_t i = 1; i <= mx; i *= base) {
+            for(int64_t i = 1; i <= mx - mn; i *= base) {
                 count_sort<base>(a, [&](auto x) {
-                    return x / i % base;
+                    return (x - mn) / i % base;
                 });
             }
         });