Instructions for Installing and Running Trimaran, a Compiler on Linux - Prof. Scott Mahlke, Assignments of Electrical and Electronics Engineering

Download Instructions for Installing and Running Trimaran, a Compiler on Linux - Prof. Scott Mahlke and more Assignments Electrical and Electronics Engineering in PDF only on Docsity!

EECS 583 – Lecture 3 Control Flow Optimization

University of Michigan January 13, 2003

• 1 -

Agenda for Today

Y

Trimaran overview^ »

What you need to know for HW 1

»^

See http://www.trimaran.org

, documentation page for more

Y

Finish off leftover from last time

Y

Simple control flow transformations^ »

Loop unrolling

»^

General code

Y

Start on regions if we have time

• 3 -

Needed Software Y

TCL/TK – v8.0p2 or higher (tclsh, wish)

Y

perl – v5.0p3 or higher

Y

gnumake (should be called make, gmake, gnumake, add links!)

Y

imake (generally part of X11)

Y

gzip/gunzip

Y

gcc^ »

3.0.4 works for sure »^

2.8.1 - 3.2 should work

Y

Gnu ranlib, ar (binutils 2.9 – 2.13)

Y

tcsh – Use this or you will have to translate some stuff

Y

emacs – Real compiler writers don’t use vi or pico

Y

Graph visualizing tools^ »

dot – http://www.research.att.com/sw/tools/graphviz »^

xvcg – source on course webpage (./configure, make)

• 4 -

Installation – Step 1

Y

Source on the course webpage, trimaran.w03.tgz

Y

In your home directory, mkdir trimaran

Y

tar zxvf trimaran.w03.tgz^ »

Will create 6 subdirectories

»^

impact, elcor, simu, scripts, benchmarks, gui

Y

Enviroment variables/path – this is critical^ »

trimaran/scripts/envrc – should have everything you need

»^

Or, trimaran/scripts/envrc.bash for you bash users

»^

Modify first line if you don’t put trimaran in $HOME/trimaran

»^

Add to your .cshrc, source trimaran/scripts/envrc^ y

So you don’t forget

• 6 -

Trimaran Overview

C source

Frontend – parsing, profiling function inlining, memory dep analysis. Also, does some backend stuff – region formation, optimization

Impact

Lcode

Backend – analysis, optimization, scheduling register allocation, code generation

Elcor

Rebel

Emulator – emulate VLIW semantics using C code – “compiled simulation”

Simu

C emulation code

• 7 -

Running Trimaran

Y

trimaran-gui – I don’t like/use this^ »

Compile strcpy with basic blocks to test install

»^

Lots of documentation at http://www.trimaran.org

Y

tcc – gcc equivalent, sorta!^ »

Command line way to invoke entire system – Gui calls this

»^

“tcc –help” to get options

»^

“tcc –bench strcpy” to run strcpy with basic blocks

Y

trimaran/benchmarks^ »

To run trimaran on a benchmark, need some setup^ y

source, compile flags, link flags y input for profiling the code y output to check it ran properly y To add a new benchmark, follow the organization that is there

»^

583hw1a, 583hw1b, 583hw1c, 583hw1d for HW

• 9 -

Elcor – Directory Structure

Y

bin, lib – the obvious stuff

Y

src – where you will spend most of your time^ »

type make in src to build elcor

Y

mdes^ »

Machine descriptions for the target processor

»^

hmdes2 = high level machine description – for human

»^

lmdes2 = low-level machine description – for compiler

»^

hc = script to convert hmdes2 to lmdes

»^

hpl_pd_elcor_std.hmdes2 – includes other files

Y

parms^ »

Compiler switches broken into groups^ y

DRIVER_DEFAULTS

, ANALYSIS_DEFAULTS, etc.

y^

Changing switch values^ X

Modify ***_DEFAULTS file X^

-Fswitch=value on the command-line

• 10 -

Running Elcor Manually

Y

Use trimaran-gui or tcc, compile a benchmark^ »

tcc –bench 583hw1a

Y

583hw1a_O directory created^ »

impact_intermediate

»^

elcor_intermediate^ y

583test1.O_tr.tgz – impact output / elcor input (bridge code)

X^

Impact assembly (called Lcode or Mcode) X^

This is what you will run Elcor manually on X^

Untar this file^ ±

file1.O_tr, file2.O_tr, … ±

1 file per original C source file

y^

583test1.O_el.tgz – elcor output^ X

Rebel

y^

host_layout_info.md – data structure organization stuff

»^

simu_intermediate

• 12 -

Running Simu Manually

Y

simu^ »

Converts Rebel to C

»^

Models VLIW architecture properly^ y

predicates, NUAL, etc.

»^

Allows testing the compiler is correct!

Y

running simu (aka codegen)^ »

codegen –el test1.O_el –o test1.c –Finput_format=rebel \ –Femulate_unscheduled=yes –Femulate_virtual_regs=yes^ y

test1.c y benchmark_data_init.simu.c

»^

Note need host_layout_info.md in the directory

»^

gcc –I$HOME/trimaran/simu/src/Emulib \ –L$HOME/trimaran/simu/lib *.c –lequals –lm

»^

./a.out > output, diff output verses expected output

• 13 -

Elcor Organization Y

src/Tools

y^

“homegrown” version of STL - list, dlist, slist, hash_set, hash_map, …

Y

src/Graph

y^

All the core compiler data structs in here y^

operand.h, port.h, op.h, region.h, edge.h, *_utilities.h, opcode_properties.h

Y

Main

y^

process_function.cpp is the main driver

Y

Mdes

y^

target processor info – built from .lmdes2 file

Y

Analysis

y^

control flow, dataflow analysis, predicate analysis, edge drawing

Y

Control

y^

Control flow transformations

Y

Templates

y^

Whenever you create a List that has never been done before y^

Must instantiate template – look in the files for example y^

Hw1 – can add to template_control.cpp

• 15 -

Back to Loops – Assembly Generation Schema

for (i=x; i<y; i+=z) {

body;

while-do schema

do-while schema

loop:

if (i >= y) goto done body; i += z; goto loop

if (i >= y) goto done body; i += z; if (i < y) goto loop

loop:

done:

done:

Question: which schema is better and why?

• 16 -

Loop Induction Variables

Y

Induction variables are variables such that every time theychanges value, they are incremented/decremented bysome constant

Y

Basic induction variable

• induction variable whose only

assignments within a loop are of the form j = j+- C, whereC is a constant

Y

Primary induction variable

• basic induction variable that

controls the loop execution (for I=0; I<100; I++), I(virtual register holding I) is the primary inductionvariable

Y

Derived induction variable

• variable that is a linear

function of a basic induction variable

• 18 -

Loop Unrolling Y

Most renowned control flow opti

Y

Replicate the body of a loop N-1times (giving N total copies)»

Loop unrolled N times or Nxunrolled »^

Enable overlap of operations fromdifferent iterations »^

Increase potential for ILP(instruction level parallelism)

Y

3 variants»

Unroll multiple of known tripcount »^

Unroll with remainder loop »^

While loop unroll

Loop:r1 = MEM[r2 + 0]r4 = r1 * r5r6 = r4 << 2MEM[r3 + 0] = r6r2 = r2 + 1blt r2 100 Loop

• 19 -

Loop Unroll – Type 1

r1 = MEM[r2 + 0] r4 = r1 * r5 r6 = r4 << 2 MEM[r3 + 0] = r6 r2 = r2 + 1 blt r2 100 Loop

Loop: r1 = MEM[r2 + 0] r4 = r1 * r5 r6 = r4 << 2 MEM[r3 + 0] = r6 r2 = r2 + 1 blt r2 100 Loop r2 is the loop variable, Increment is 1 Initial value is 0 Final value is 100 Trip count is 100

Loop: r1 = MEM[r2 + 0] r4 = r1 * r5 r6 = r4 << 2 MEM[r3 + 0] = r6 r2 = r2 + 1

Loop: r1 = MEM[r2 + 0] r4 = r1 * r5 r6 = r4 << 2 MEM[r3 + 0] = r6 r1 = MEM[r2 + 1] r4 = r1 * r5 r6 = r4 << 2 MEM[r3 + 0] = r6 r2 = r2 + 2 blt r2 100 Loop

Counted loop All parms known

Remove r2 increments from first N-1 iterations and update last increment

Remove branch from first N-1 iterations