Awesome

Squint: A peephole optimizer for stack VM compilers

Introduction

Short summary: This repo contains a highly functional, but not quite complete, C compiler (mc.c) that supports JIT execution, ELF executable generation, and peephole optimization. mc.c is a follow on to the AMaCC compiler. See the AMaCC documentation referenced below for more information.

This compiler was developed on a Raspberry Pi computer running 32 bit Buster Linux. See the "Discussion" tab above for instructions on using this compiler with an ARM Chromebook.

The MC C compiler found in this repository is a subset of the full MC HPC compiler which is being developed offline. That said, results from the full MC HPC compiler are shown below.

This compiler supports the following features beyond AMaCC:

• Float data types (AMaCC is an integer based compiler).

• Array declarations and initializers. (e.g. float foo[4][4] = { { 0.0, ... }, { 0.0 ... }, ... };

• The Squint peephole optimizer that roughly halves the number of executable instructions in compiled code. The tests/sieve.c benchmark provides an optimization example, and runs roughly 4x faster after peephole optimization.

• Greatly improved type checking, error checking, and IR code generation (try -si option).

Source code size (public code version, this repository):

• mc C compiler -- 4035 SLOC
• squint optimizer -- 3836 SLOC

The original AMaCC compiler is based on the phenomenal work of the team at https://github.com/jserv/amacc , and I strongly suggest you visit that site to see the history of AMaCC compiler development as stored in that repository. It shows how small changes can be added to a base complier, step-by-step, to create a truly marvelous educational compiler. The README there expands upon this README and is where you will learn the most about the AMaCC compiler that was used as a starting point for this work.

Example

The following time command uses the mc compiler to JIT compile an optimized verison of the mc compiler, using an external optimizer linked as a shared object library, and then that optimized JIT complier is used to JIT compile an optimized version of Sieve of Eratosthenes (again using a shared object optimizer), and then the sieve benchmark is JIT executed. The time shown is the time it takes to do everything in this paragraph, sieving 8388608 values using three different algorithms applied to bit arrays.

$ make mc-so      # use gcc to build an enhanced version of the mc compiler
  CC+LD		mc-so
$ time ./mc-so -DSQUINT_SO -Op mc.c -Op tests/sieve.c

real	0m1.031s
user	0m0.999s
sys	0m0.031s

Performance

Note: shock run with 8192 elements, 4096 timesteps, no output. Best of 20 runs.

Assembly language quality

Below is a comparison of assembly language quality of three compilers on the Raspberry Pi 4B when compiling tests/shock.c.

• GCC, specifically: "gcc -mfloat-abi=hard -mtune=cortex-a72 -O3 tests/shock.c -lm"

• The Squint compiler uses the compiler in this repository with the -Op option, followed by the Squint optimizer.

• The MC compiler is a non-public HPC version of Squint that I am working on offline.

For floating point, my HPC compiler is currently always faster than gcc with the above compiler options, by a minimum of 3%.

That said, make no mistake, my current optimizations are all crap** and yet I am still beating gcc. I am only one person with no resources, so I pick the path I see as most interesting and plod along at a snail's pace when I am not watching TV, playing video games, or reading/commenting on news articles.

If I had the resources to hire a small team and treat this as a real full time project, I could do much better! If you might be interested, let's talk.

** No common subexpression elimination, register renaming to reduce stalls, code motion to reduce stalls, or register coloring to reduce register pressure, etc. None of these things are hard to do, but all consume time to implement.

Below is a C function followed by the assembly language listing for the loop body, for all three compilers. The MC compiler creates only 3.05 assembly language instructions to represent each (complex) line of high level C code, on average:

void ComputeFaceInfo(int numFace, float *mass, float *momentum, float *energy,
                                  float *f0, float *f1, float *f2)
{
   int i;
   int contributor;
   float ev;
   float cLocal;

   for (i = 0; i < numFace; ++i)
   {
      /* each face has an upwind and downwind element. */
      int upWind   = i;     /* upwind element */
      int downWind = i + 1; /* downwind element */

      /* calculate face centered quantities */
      float massf =     0.5 * (mass[upWind]     + mass[downWind]);
      float momentumf = 0.5 * (momentum[upWind] + momentum[downWind]);
      float energyf =   0.5 * (energy[upWind]   + energy[downWind]);
      float pressuref = (gammaa - 1.0) *
                         (energyf - 0.5*momentumf*momentumf/massf);
      float c = sqrtf(gammaa*pressuref/massf);
      float v = momentumf/massf;

      /* Now that we have the wave speeds, we might want to */
      /* look for the max wave speed here, and update dt */
      /* appropriately right before leaving this function. */
      /* ... */

      /* OK, calculate face quantities */

      contributor = ((v >= 0.0) ? upWind : downWind);
      massf = mass[contributor];
      momentumf = momentum[contributor];
      energyf = energy[contributor];
      pressuref = energyf - 0.5*momentumf*momentumf/massf;
      ev = v*(gammaa - 1.0);

      f0[i] = ev*massf;
      f1[i] = ev*momentumf;
      f2[i] = ev*(energyf - pressuref);

      contributor = ((v + c >= 0.0) ? upWind : downWind);
      massf = mass[contributor];
      momentumf = momentum[contributor];
      energyf = energy[contributor];
      pressuref = (gammaa - 1.0)*(energyf - 0.5*momentumf*momentumf/massf);
      ev = 0.5*(v + c);
      cLocal = sqrtf(gammaa*pressuref/massf);

      f0[i] += ev*massf;
      f1[i] += ev*(momentumf + massf*cLocal);
      f2[i] += ev*(energyf + pressuref + momentumf*cLocal);

      contributor = ((v - c >= 0.0) ? upWind : downWind);
      massf = mass[contributor];
      momentumf = momentum[contributor];
      energyf = energy[contributor];
      pressuref = (gammaa - 1.0)*(energyf - 0.5*momentumf*momentumf/massf);
      ev = 0.5*(v - c);
      cLocal = sqrtf(gammaa*pressuref/massf);

      f0[i] += ev*massf;
      f1[i] += ev*(momentumf - massf*cLocal);
      f2[i] += ev*(energyf + pressuref - momentumf*cLocal);
   }
}

By the end of 2030, I expect to have automatic vectorization and/or parallelization working in my HPC compiler. The HPC extensions/restrictions make it "natural" to manage parallel partitions, unlike the mess created by C standard semantics.

Prerequisites

• This compiler project depends on several GNU/Linux behaviors, and it is necessary to have Arm/Linux installed in your build environment.
• Install GNU Toolchain for the A-profile Architecture
  • Select arm-linux-none-gnueabihf (AArch32 target with hard float)

Test environment:

• Raspberry Pi 4B (SoC: bcm2711, ARMv8-A architecture)
• Raspbian GNU/Linux, kernel 5.10.17-v7l+, gcc 8.3.0 (armv7l userland)
• See the "Discussion" tab in this repo to run Squint on your ARM Chromebook.

The architecture support targets armv7hf with Linux ABI, verified on Raspberry Pi 2/3/4 with GNU/Linux.

Acknowledgements

AMaCC is based on the infrastructure of c4. Please see the AMaCC repository AMaCC to learn more about the AMaCC compiler.