What algo is optimized for

The previous tutorials showed the mechanics of using the language. This page explains the design choices behind it.

algo is not trying to be a general-purpose language. It is designed for competitive programming, where the common problems are:

• repetitive input/output boilerplate
• repeated use of the same graph and data-structure patterns
• a need to move quickly without hiding what the final C++ will do

The features that matter most come from those three problems.

1. Problem-shaped I/O

Contest statements usually describe input in a declarative shape: n, then an array, then m edges, then q queries. algo mirrors that shape:

input:
    n, m: int
    edges: [(int, int)][m]

That is less code than hand-written cin loops, but it still compiles to ordinary reads in C++.

2. Builtins for common contest tools

Common algorithms and data structures are callable directly.

input:
    n, m: int
    edges: [(int, int, int)][m]

g = wgraph(n, undirected = true)
for (u, v, w) in edges:
    g.add(u, v, w)

d = g.dijkstra(0).dist

You write this kind of code in every contest. Instead of pasting a template each time, you call a function.

The same idea applies to other graph helpers:

g.bfs(src).dist          # BFS distances (-1 if unreachable)
wg.bellman_ford(src).dist # Bellman-Ford
g.topo()                 # topological order
g.scc()                  # strongly connected components
wg.mst()                 # minimum spanning tree

And to common data structures:

d = dsu(n)               # disjoint set union
d.unite(u, v)
d.same(u, v)

st = segtree(a, op = max)     # segment tree
st.update(i, val)
st.query(l, r)

b = bit(n)               # Fenwick tree
b.add(i, val)
b.sum(l, r)

3. Structured syntax for common patterns

DP is repetitive enough to get its own syntax:

input:
    n, W: int
    w: [int][n]
    v: [int][n]

dp[i: 1..n, j: 0..W] -> int, maximize:
    base: dp[0][j] = 0
    <- dp[i-1][j]
    <- dp[i-1][j-w[i-1]] + v[i-1] where j >= w[i-1]

println dp[n][W]

Likewise, reductions and binary search helpers let you express the operation directly:

println max(dp[i] for i in [0, n))
println sum(a[i] * b[i] for i in [0, n))
ans = bsearch(0, 1_000_000_000, |mid| feasible(mid))

These forms save time because they skip setup code, not because they look clever.

4. Generated C++ is still the source of truth

The language only works if you can still trust the lowered code. That is why algo compiles to readable C++17 instead of hiding everything in a runtime.

When the language is not enough, there is an explicit escape hatch:

cpp:
    #pragma GCC optimize("O3,unroll-loops")

Unknown method calls also pass through to C++:

a.reserve(n)

That boundary keeps the language useful without pretending it can replace every line of contest C++.

5. A few global switches for contest defaults

One example is #!int64:

#!int64

This makes int in your source emit as 64-bit in the generated C++. A contest convenience, nothing more.

Summary

If you remember one thing, it should be this:

• algo is meant to make contest code shorter
• not by hiding the solution shape
• but by baking common competitive-programming patterns into the language

Continue with the guides for task-focused docs, or jump to the reference when you need exact forms.