IR is better than assembly (2013)

24 July 2013

Inthe LLVM the compilation takes three stages (image fromthe AOSA book):

Schema

The stages are:

• The frontend, parsing original language and spiting out LLVM Intermediate Representation (IR) code¹.
• The optimiser, mangling one IR into optimised equivalent IR. This stage does all the usual optimisations like constant propagation, dead code removal and so on.
• The backend, taking IR and producing machine code optimised for a specific CPU.

The crucial part is IR. It's a common language that sits between the high-level program and the low-level backend. IR is used to express high level concepts and is specific enough that any backend can produce a fast machine code.

IR is the heart of LLVM.

What is IR?

IR is a low-level programming language, pretty similar to assembly. From the AOSA book:

IR code is usually generated by the frontend, but nothing stops us from writing it by hand. Let's do it!

You'll need clang and LLVM (llc, opt):

$ sudo aptitude install llvm clang

C to IR

The learning curve for IR, like for any assembly, is a bit steep. When starting with IR is it's easiest to compile a normal program, for example in C, to IR. That can be done using an LLVM frontend. Once we have IR we can tweak it and use an LLVM backend to produce a real machine code.

Consider the following C code:

unsignedsquare_int(unsigneda){returna*a;}

Using clang we can generate IR. Flag -Os makes sure the produced IR is shortest and should be easiest to read:

$ clang -Os -S -emit-llvm sample.c -o sample.ll

After removing some unnecessary boilerplate we get:

definei32@square_unsigned(i32%a){%1=muli32%a,%areti32%1}

IR is strongly typed and you can see types being repeated everywhere. In line 1 we define a function that takes a single i32 parameter %a and returns type i32. In line 2 we assign the result of multiplication to register %1 and we return it in the next line.

For more details please seethe official documentation of the IR language.

Optimising

If you wish, you can run LLVM optimisations manually. The tool opt can transform unoptimised IR to an optimised one. Our code is already optimised, we used the -Os flag, but you can run opt anyway:

IR to machine code

Finally, given IR we can use a backend to generate a machine code for a real CPU. Let's compile our IR:

$ llc-3.0 -O3 sample.ll -march=x86-64 -o sample-x86-64.s

This generates:

square_unsigned:
        imull   %edi, %edi
        movl    %edi, %eax
        ret

Fancy x86 32 bit assembler? Nothing simpler:

$ llc-3.0 -O3 sample.ll -march=x86 -o sample-x86.s

square_unsigned:
        movl    4(%esp), %eax
        imull   %eax, %eax
        ret

How about ARM?

$ llc-3.0 -O3 sample.ll -march=arm -o sample-arm.s

square_unsigned:
        mov     r1, r0
        mul     r0, r1, r1
        mov     pc, lr

Given IR it's trivial to compile it to a decent machine code for any architecture. Your hand crafted assembly code might be faster if you're lucky, but writing IR can have advantages:

• Your assembly code will not use faster instructions available in future CPU generations.
• Your code won't work on older CPU's.
• Porting assembly to another CPU architecture is tedious.
• You even need to spend time porting your assembler to different operating systems. For example global symbols need underscores on mac. Dealing with #include's can be painful.
• You need to be a serious expert to write assembly better than an LLVM compiler.

If you really can write better assembly than LLVM, please:

It'll benefit everybody, not only you. Think about it - you won't need to write the same optimisations all over again on your next project!

Vectors in IR

The full power the IR language is visible when dealing with vectors. You can use any vector type you wish and not worry about the machine architecture underneath. For example, let's consider multiplying four 32 bit integers in parallel:

define<4xi32>@multiply_four(<4xi32>%a,<4xi32>%b){%1=mul<4xi32>%a,%bret<4xi32>%1}

In my opinion this code looks much nicer than its C equivalent (using vector extensions). The resulting assembly is long when LLVM compiles it without option -march=avx,sse41, but with this option turned on it becomes:

multiply_four:
        vpmulld %xmm1, %xmm0, %xmm0
        ret

Similarly the result for ARM with -march=neon is decent:

multiply_four:
        mov     r12, sp
        vmov    d19, r2, r3
        vldmia  r12,{d16, d17}
        vmov    d18, r0, r1
        vmul.i32        q8, q9, q8
        vmov    r0, r1, d16
        vmov    r2, r3, d17
        mov     pc, lr

Finally

I think it's amazing that once written pseudo-assembly code can be retargetted to any architecture². For me IR has all the advantages of assembly without any of its problems: fast, expressive, retargettable and maintainable.

If for some reason you're not happy with the machine code LLVM produces - writean LLVM optimiser (AOSA chapter 11.3.1).

Comments

From: Vitaly Vidmirov> "Similarly the result for ARM with -march=neon is decent:"

Are you joking? Right? ARM code in your examples is ridiculous. It is sad, that LLVM can't eliminate useless movs even in such trivial examples.

square_unsigned:
   mul r0,r0,r0
   mov pc,lr

Your examples use soft-float-abi then floating point data transfered in integer registers. For compatibility with FPU-less processors.  ARM processors before Cortex A15 has decoupled FPU, so fp<->int transfers cause huge stalls to serialize pipelines.

multiply_four:
   vmul.i32  q0, q0, q1
   mov pc,lr

From: giampaolo eusebi

To the attention of Vitaly:
From the ARM information center Assembler reference:
For the MUL and MLA instructions, Rn must be different from Rd in architectures before ARMv6.