Understanding Compiler Passes

Contributing to the GCC compiler requires that I understand the compilation process and the many source code passes it makes in order to produce the final executable.

I only have a basic idea about the compilation process, so it was quite tricky for me to understand how compilation passes work. Our professor directed us to try out the -fdump-tree-all and -fdump-rtl-all compiler options, so I did just that.

From what I understand so far (and with the help of ChatGPT) the source code goes through a lot of compilation passes before it turns into the fully compiled executable.

We can use the -fdump-tree-all option to see the intermediate representation (IR) of our code during each compilation pass. We can also use -fdump-tree-<pass> to get just the singular dump of a pass

let's take this simple "Hello World" program in C for example.

#include <stdio.h>

int main() {
    printf("Hello World!\n");
}

and compile it with gcc using the -fdump-tree-all option

$ gcc -fdump-tree-all -o hello hello.c

we get all these files in the format of <filename>.<pass num>t.<pass>. Each file shows the IR of our source code during each compilation pass; mainly in the GIMPLE format. These passes occur in ascending order of t.

$ ls -la
total 140
drwxr-xr-x. 2 jmendoza23 jmendoza23   4096 Nov  3 07:34 .
drwxr-xr-x. 7 jmendoza23 jmendoza23   4096 Nov  3 07:21 ..
-rwxr-xr-x. 1 jmendoza23 jmendoza23 202376 Nov  3 07:34 hello
-rw-r--r--. 1 jmendoza23 jmendoza23     69 Nov  3 07:17 hello.c
-rw-r--r--. 1 jmendoza23 jmendoza23    124 Nov  3 07:34 hello.c.005t.original
-rw-r--r--. 1 jmendoza23 jmendoza23    103 Nov  3 07:34 hello.c.006t.gimple
-rw-r--r--. 1 jmendoza23 jmendoza23    188 Nov  3 07:34 hello.c.009t.omplower
-rw-r--r--. 1 jmendoza23 jmendoza23    217 Nov  3 07:34 hello.c.010t.lower
-rw-r--r--. 1 jmendoza23 jmendoza23    217 Nov  3 07:34 hello.c.013t.eh
-rw-r--r--. 1 jmendoza23 jmendoza23    345 Nov  3 07:34 hello.c.015t.cfg
-rw-r--r--. 1 jmendoza23 jmendoza23    218 Nov  3 07:34 hello.c.017t.ompexp
-rw-r--r--. 1 jmendoza23 jmendoza23    218 Nov  3 07:34 hello.c.021t.fixup_cfg1
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.022t.ssa
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.023t.walloca1
-rw-r--r--. 1 jmendoza23 jmendoza23    436 Nov  3 07:34 hello.c.026t.waccess1
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.030t.fixup_cfg2
-rw-r--r--. 1 jmendoza23 jmendoza23    624 Nov  3 07:34 hello.c.031t.local-fnsummary1
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.032t.einline
-rw-r--r--. 1 jmendoza23 jmendoza23      0 Nov  3 07:34 hello.c.052t.profile_estimate
-rw-r--r--. 1 jmendoza23 jmendoza23    261 Nov  3 07:34 hello.c.056t.release_ssa
-rw-r--r--. 1 jmendoza23 jmendoza23    624 Nov  3 07:34 hello.c.057t.local-fnsummary2
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.098t.fixup_cfg3
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.105t.adjust_alignment
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.251t.veclower
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.252t.cplxlower0
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.253t.bitintlower0
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.255t.switchlower_O0
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.263t.isel
-rw-r--r--. 1 jmendoza23 jmendoza23    436 Nov  3 07:34 hello.c.266t.waccess3
-rw-r--r--. 1 jmendoza23 jmendoza23    220 Nov  3 07:34 hello.c.267t.optimized
-rw-r--r--. 1 jmendoza23 jmendoza23      0 Nov  3 07:34 hello.c.364t.statistics
-rw-r--r--. 1 jmendoza23 jmendoza23      0 Nov  3 07:34 hello.c.365t.earlydebug
-rw-r--r--. 1 jmendoza23 jmendoza23      0 Nov  3 07:34 hello.c.366t.debug
-rw-r--r--. 1 jmendoza23 jmendoza23    485 Nov  3 07:17 Makefile

If we want to see the passes that affect the Register Transfer Language (RTL) representation of our code, we can pass the option -fdump-rtl-all or -fdump-rtl-<pass> to dump just a specific pass.

This gives us an output similar to -fdump-tree-all but for the RTL passes of our code. Passes also occur in ascending order of r.

$ gcc -fdump-rtl-all -o hello hello.c
$ ls -la
total 184
drwxr-xr-x. 2 jmendoza23 jmendoza23   4096 Nov  3 08:38 .
drwxr-xr-x. 7 jmendoza23 jmendoza23   4096 Nov  3 07:21 ..
-rwxr-xr-x. 1 jmendoza23 jmendoza23 202376 Nov  3 08:38 hello
-rw-r--r--. 1 jmendoza23 jmendoza23     69 Nov  3 07:21 hello.c
-rw-r--r--. 1 jmendoza23 jmendoza23   2078 Nov  3 08:38 hello.c.268r.expand
-rw-r--r--. 1 jmendoza23 jmendoza23   1705 Nov  3 08:38 hello.c.269r.vregs
-rw-r--r--. 1 jmendoza23 jmendoza23   1706 Nov  3 08:38 hello.c.270r.into_cfglayout
-rw-r--r--. 1 jmendoza23 jmendoza23   3325 Nov  3 08:38 hello.c.271r.jump
-rw-r--r--. 1 jmendoza23 jmendoza23   1674 Nov  3 08:38 hello.c.283r.reginfo
-rw-r--r--. 1 jmendoza23 jmendoza23   1733 Nov  3 08:38 hello.c.306r.outof_cfglayout
-rw-r--r--. 1 jmendoza23 jmendoza23   1705 Nov  3 08:38 hello.c.307r.split1
-rw-r--r--. 1 jmendoza23 jmendoza23   3591 Nov  3 08:38 hello.c.309r.dfinit
-rw-r--r--. 1 jmendoza23 jmendoza23   3627 Nov  3 08:38 hello.c.310r.mode_sw
-rw-r--r--. 1 jmendoza23 jmendoza23   3545 Nov  3 08:38 hello.c.311r.asmcons
-rw-r--r--. 1 jmendoza23 jmendoza23   5769 Nov  3 08:38 hello.c.318r.ira
-rw-r--r--. 1 jmendoza23 jmendoza23   8449 Nov  3 08:38 hello.c.319r.reload
-rw-r--r--. 1 jmendoza23 jmendoza23   5411 Nov  3 08:38 hello.c.326r.pro_and_epilogue
-rw-r--r--. 1 jmendoza23 jmendoza23   4954 Nov  3 08:38 hello.c.329r.jump2
-rw-r--r--. 1 jmendoza23 jmendoza23   4922 Nov  3 08:38 hello.c.349r.zero_call_used_regs
-rw-r--r--. 1 jmendoza23 jmendoza23   4922 Nov  3 08:38 hello.c.350r.alignments
-rw-r--r--. 1 jmendoza23 jmendoza23   4892 Nov  3 08:38 hello.c.354r.barriers
-rw-r--r--. 1 jmendoza23 jmendoza23   4892 Nov  3 08:38 hello.c.356r.split5
-rw-r--r--. 1 jmendoza23 jmendoza23   4981 Nov  3 08:38 hello.c.359r.shorten
-rw-r--r--. 1 jmendoza23 jmendoza23   4981 Nov  3 08:38 hello.c.360r.nothrow
-rw-r--r--. 1 jmendoza23 jmendoza23   5661 Nov  3 08:38 hello.c.361r.dwarf2
-rw-r--r--. 1 jmendoza23 jmendoza23   4981 Nov  3 08:38 hello.c.362r.final
-rw-r--r--. 1 jmendoza23 jmendoza23   3269 Nov  3 08:38 hello.c.363r.dfinish
-rw-r--r--. 1 jmendoza23 jmendoza23    510 Nov  3 08:12 Makefile

These options will help us understand how passes work and how we can develop custom passes in the compiler which will be crucial in our development of the new Automatic Function Multi-Versioning (AFMV) feature of GCC.

Until then, thanks for reading!

Comments

Post a Comment

Popular posts from this blog

Lab 03 - Breakout Game

Image
This lab is quite special compared to previous ones, as it is more subjective and taps into our creativity. We were tasked with creating any program we want, as long as it meets these criteria: The program must work in the 6502 Emulator It must output to the character screen as well as the graphics (bitmapped) screen. It must accept user input from the keyboard in some form. It must use some arithmetic/math instructions (to add, subtract, do bitwise operations, or rotate/shift) Some suggestions given by our professor were to make either a simple game or a simple calculator/converter. With that in mind, I decided to remake the 1976 arcade game Breakout. What is Breakout? Based on its Wikipedia description , Breakout is a 1967 arcade game developed and published by Atari, Inc. It involves some bricks, a paddle, and a bouncing ball. The goal is to destroy all bricks with a bouncing ball, using the paddle as a sort of tool to keep the ball in play. You lose when the ...
Read more

Building GCC

In my last blog, I talked about contributing to the GCC compiler by helping out with the Automatic Function Multi-versioning (AFMV) for aarch64. In order to do that I must first learn how to build GCC myself. Unfortunately, since I do not have an aarch64 machine, I have to rely on the aarch64 systems my professor has setup in order to build/work with GCC. The system I decided to start my GGC build is an ARMv8 machine with 4 cores and 16 GiB ram. Now to start, the first step we must do is to clone the GCC git repository. Following the steps from my class wiki page I cloned the repo to my spo600 directory using the following command. $ git clone git://gcc.gnu.org/git/gcc.git ~/spo600/ After cloning, the next step is to create a build directory and configure our build to use that directory. Here are the commands I used to do so. $ mkdir ~/spo600/gcc-build-001 $ cd ~/spo600/gcc-build-001 $ ~/spo600/gcc/configure --prefix=$HOME/spo600/gcc-build-001 Now we can start the build ...
Read more

Lab 04 - Diving into 64-bit Assembly

Image
After having worked with 6502 Assembly, we now move on to the more complex 64-bit Assembly that is being used by the majority of computers today. In this lab we explore the 64-bit assembly language used in both aarch64 (ARM) and x86_64 (Intel/AMD) architectures. We will explore their differences, similarities, as well as the tradeoffs for using either. The task for this lab was to develop a simple program that prints the following output: Loop: 0 Loop: 1 Loop: 2 Loop: 3 Loop: 4 Loop: 5 Loop: 6 Loop: 7 Loop: 8 Loop: 9 Loop: 10 Loop: 11 Loop: 12 Loop: 13 Loop: 14 Loop: 15 Loop: 16 Loop: 17 Loop: 18 Loop: 19 Loop: 20 Loop: 21 Loop: 22 Loop: 23 Loop: 24 Loop: 25 Loop: 26 Loop: 27 Loop: 28 Loop: 29 Here is program I developed to do this in Aarch64 assembly. .text .globl _start min = 0 max = 30 _start: mov x19, min // Store to x19 the value of min (0) loop: /* ... body of the loop ... do something useful here ... */ // write() syscall takes 3 args; file descr...
Read more