Writing a programming language in Bangla

alveelan is a small programming language where every keyword is Bangla. This is a tour of the parts that turned out to be interesting.

ফাংশন শুরু() {
    ধরি ঘর: সংখ্যা = ১
    যতক্ষণ (ঘর <= ১০) {
        দেখাও(৫ * ঘর)
        ঘর = ঘর + ১
    }
}

That’s the 5 times table: ধরি declares, যতক্ষণ loops and দেখাও prints.

Two digit systems

Kids write ১০ and adults sometimes write 10, so the lexer accepts both and normalises Bangla digits (U+09E6–U+09EF) to ASCII as it reads a number. Everything after the lexer only ever sees one digit system.

Answering in Bangla

The first version printed numbers with C’s printf, so a Bangla program answered 30. Now the compiler links a tiny C runtime with every program, and দেখাও calls it for numbers. It formats the number, then swaps each digit for its Bangla glyph. The program says ৩০.

Errors a child can read

Error messages are Bangla too, and they name things the way the language does:

ত্রুটি: 'বয়স' চলকটির ধরন হল সংখ্যা, কিন্তু আপনি একে লেখা করার চেষ্টা করছেন।
ত্রুটি [লাইন ২]: ')' প্রত্যাশিত ছিল, কিন্তু '{' পাওয়া গেছে।

The types print as সংখ্যা and লেখা rather than the compiler’s internal names, and the line number is in Bangla digits.

The bug in nested loops

An if without else, inside a while, produced invalid LLVM IR: “Basic Block does not have terminator”. The code generator checked whether the block it started in already ended with a branch:

if then_bb.get_terminator().is_none() { /* branch to merge */ }

But nested control flow moves the builder into new blocks, so the block that needs the branch is the current one:

fn current_block_open(&self) -> bool {
    self.builder.get_insert_block().is_some_and(|bb| bb.get_terminator().is_none())
}

Every example program is now compiled and run in CI, with its output checked, so this kind of bug can’t come back quietly.

Next

Reading input (নাও is already a reserved word), joining text with +, and a browser playground for classrooms.