[llvm-dev] Global or Local variable different builder behavior

[Tim Northover via llvm-dev]

Hi Duat,

On Mon, 30 Sep 2019 at 17:23, Duat via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> I posted this question on stackoverflow : https://stackoverflow.com/questions/58147549/global-or-local-variable-different-builder-behavior
>
> with no response, but I can't understand the different behavior based on whether the variable is global or local.
> Can you help me to understand?

LLVM has roughly[1] two kinds of Value: Constants and Instructions.

Constants are things like literal constants, (addresses of) global
variables, and various expressions based just on those things; they're
designed to be values that can be directly calculated by the compiler
and/or linker without any CPU instructions actually being executed[2].

Instructions on the other hand sit inside Functions as real entities,
they  produce %whatever Values and, unless optimized away, will be
turned into real CPU instructions in the end.

So, you were asking for "GEP something, 0, 49". If that "something" is
a Constant (e.g. a GlobalVariable) then that GEP only depends on
Constants so it can be a ConstantExpr too, written
"getelementptr([1024 x i64], [1024 x i64]* @var, i32 0, i32 49)". That
Constant is then not inserted into a block (it's not an instruction so
it can't be). Instead it's written directly in any instruction that
uses it, so if you actually use the GEP you might see something like:

    %val = load i64, i64* getelementptr([1024 x i64], [1024 x i64]*
@var, i32 0, i32 49)

Until you use it, it's not actually in the function anywhere though.
You just have a handle when needed.

On the other hand if the "something" is a local variable, then the GEP
needs to be an actual instruction inside a function and the API you're
using will insert it automatically.

In the Constant case, you can manually create an instruction anyway,
at least in C++. I'm afraid I haven't used the C API and couldn't see
an obvious way there, but you probably don't want to since
optimization would quickly undo it and turn it back into a Constant.

Cheers.

Tim.

[1] There's also Arguments, representing function parameters. They
behave like Instructions for these purposes.
[2] But you can build pathological Constants that no linker really
could calculate like 4 * @global. That tends to result in a compiler
error.