-gen llvm: array writes inside a FOR loop don't advance the element address (every element reads back the last iteration's value)

[dalibor31]

fbc version: 1.10.1 (2023-12-24), linux-x86_64 (official SourceForge binary release, checksum-verified against sha1sums.txt)

Command: fbc -gen llvm -O 0 repro.bas -x repro (also reproduces with -O 2)

Minimal repro

dim shared a(1) as long
dim i as long
for i = 0 to 1
    a(i) = i + 1
next i
print a(0)
print a(1)

Expected output

 1
 2

(this is exactly what -gen gcc produces for the identical source, at both -O 0 and -O 2)

Actual output with -gen llvm

 2
 0

a(0) comes back holding the last iteration's value, and a(1) comes back as 0 (its true zero-initialized value) — meaning both loop iterations wrote to the same element (element 0), and element 1 was never actually written at all.

This reproduces identically at -O 0 and -O 2, so it is not an LLVM optimizer artifact — the IR that fbc itself emits for the array-element address inside the loop does not appear to depend on the loop variable's current value per iteration. It also reproduces the same way with a longer loop (every element ends up holding the last iteration's value) and with a plain, non-shared module-level array.

-gen gcc (the default backend) is unaffected and produces correct output for this exact same source.

Found while benchmarking fbc against another LLVM-emitting BASIC compiler for an unrelated project.