Error when instantiating ints

[thirtythreeforty]

After building LLVM and Clang from these repos, I wrote a test program:

int main(void)
{
        int i = 0;
        return i;
}

Then I attempted to compile it. Note that I only want the assembly (I have my own assembler):

$ clang -S -emit-llvm hello.c
$ llc -march=z80 hello.ll
LLVM ERROR: Cannot select: 0x21434a0: i16 = FrameIndex<1> [ID=4]
In function: main
$

If I replace all the ints with chars or unsigned chars, llc doesn't complain (although clang warns about main's return type).

Is this an issue, or has it just not been implemented?

[earl1k]

You need add some clang options. Try compile like this: clang -target z80 -S -emit-llvm Or like this (with optimization code option): clang -target z80 -S -emit-llvm -Os By default clang generate code to x86, where int size is 32 bits. z80 backend can return only 8 or 16 bits result from function, now.

[thirtythreeforty]

:+1: that fixed the issue. (I was unaware that LLVM bytecode was that specific; I thought it was more generic. :)

A more general question regarding -target z80. If I use that switch, is this usable as a C compiler, for experimental day-to-day use?

[earl1k]

Yes. Use this llvm z80 backend and clang only as experimental, because not all llvm bytecode may be lowered to the z80 native code.

[thirtythreeforty]

What won't work? Is it something I would be able to write?

[earl1k]

A list of feature, which need write is not small. If you want, you can join to this project and help me write it.

[tommythorn]

On Nov 8, 2013, at 08:13 , Dmitriy Limonov notifications@github.com wrote:

A list of feature, which need write is not small. If you want, you can join to this project and help me write it.

I can't remember what I ran into that didn't work, but general C support was pretty good, but unfortunately the generated code is less efficient than what I got from sdcc.

I think the Z80 is hard in general to compile for (as many pre-RISC architectures) and the best sequences are rarely obvious, see fx. http://www.chilliant.com/z80shift.html. I think a a super optimizer-like solution could help here, but I haven't had time to do much about it.

Tommy

[thirtythreeforty]

Less efficient than SDCC? That's worrisome; SDCC isn't very bright at all.

Alas, I'm not very familiar with compiler internals, although I am pretty good at assembly.

[tommythorn]

On Nov 8, 2013, at 12:17 , thirtythreeforty notifications@github.com wrote:

Less efficient than SDCC? That's worrisome; SDCC isn't very bright at all.

I apologize not having the time to produce examples of what I'm talking about, but let me try elaborate none the less:

While LLVM obviously have real and respectable optimizations, at the micro level, the instruction selection is in many instances worse than the, hardly-optimizing and very limited, sdcc.

It's quite possible that, in the large, the former more than compensates for the latter, but on my code it didn't.

For a simple RISC, generating near-optimal code is so much easier.

Note, this isn't a complaint! I think llvm-z80 is fantastic and holds great potential.

[Ayymoose]

@thirtythreeforty I know this is an old post but could you tell me how you built clang from this repository?

[thirtythreeforty]

@Ayymoose I have no idea anymore, sorry. Your best bet is to look through the LLVM documentation.

[earl1k]

@Ayymoose, clang can be built in a standard way for projects using cmake. How do you build it?