LLVM-Middle-End-Optimizations

Machine independent passes to optimise LLVM intermediate representation.

Local Optimizations

Installation

In order to run the optimization passes, it is necessary having LLVM 17.0.6 source code installed.
Official repository here.

The following steps assume that LLVM 17.0.6 has been installed.
Clone the current repository:

git clone https://github.com/RaffaeleTranfaglia/LLVM-Middle-End-Optimizations.git

src/LocalOpts/LocalOpts.cpp file must be moved to the following directory:

$SRC/llvm/lib/Transforms/Utils

Add LocalOpts.cpp in $SRC/llvm/lib/Transforms/Utils/CMakeLists.txt.
src/LocalOpts/LocalOpts.h file must be moved to the following directory:

$SRC/llvm/include/llvm/Transforms/Utils

Replace SRC/llvm/lib/Passes/PassRegistry.def with the provided src/PassRegistry.def.
Add the follwing line to SRC/llvm/lib/Passes/PassBuilder.cpp:

#include "llvm/Transforms/Utils/LocalOpts.h"

Where $SRC is the source folder of the project.
To compile the source code:

cd $ROOT/BUILD
make opt
make install

Otherwise, to install the source code already containig the optimized passes' files: here.

Introduced Optimizations

Constant Folding

Constant Folding execute at compile time operations involving contant values.
Examples:

• x = 4 + 9 → x = 13 + 0
• x = 6 * 2 → x = 12 + 0

Observation: The assignment of the computed constant is carryed out using an addition of the given constant with zero.
Afterwards the introduced add will furtherly be optimized by the Algebric Identity optimization.

Algebraic Identity optimization

Algebraic Identity aims to optimise operation containing neutral values.
Examples:

• y = x + 0 → every use of y is replaced with x
• a = b * 1 → every use of a is replaced with b

Strength Reduction

A strength reduction pass replace mul instructions with shift instruction to reduce computational complexity. Example:

• x * 15 → (x << 4) - x
• x * 17 → (x << 5) - 15x
• x / 16 → x >> 4

Observations:
In case the element "subtracted" can be optimized, the algorithm will optimize further more the operation. In case it is used on divisions, the divisor must be an exact multiple of 2.

Multi-instruction optimization

Multi instruction optimization operates in cases where, given a SSA register, the same fixed amount is addend and subtracted from it. Examples:

• y = x + 2; z = y - 2 → every use of z is replaced with x
• y = x + 2; z = y / 2 → every use of z is replaced with x

Global Optimizations

DataFlowAnalysis folder contains global optimizations algorithms.
Optimization tasks addressed:

• Very Busy Expressions
• Dominator Analysis
• Constant Propagation

Loop Invariant Code Motion (LICM)

Instructions that does not change from one iteration to another can be moved outside the loop in order to be executed only once.

Instructions that meet the following requirements are candidate for code motion:

• are invariant
• dominate their uses
• dominate the exits of the loop or are dead outside the loop

Candidated instruction may be moved outside the loop (just before the loop header, in the so called preheader block) in order to be executed only one time.

LoopOpts.cpp and LoopOpts.h files contain the Loop Invariant Code Motion pass.
In order to make the pass work, src/GlobalOpts/LoopOpts.cpp file must be moved to the following directory:

$SRC/llvm/lib/Transforms/Utils

Add LoopOpts.cpp in $SRC/llvm/lib/Transforms/Utils/CMakeLists.txt.
src/GlobalOpts/LoopOpts.h file must be moved to the following directory:

$SRC/llvm/include/llvm/Transforms/Utils

Replace SRC/llvm/lib/Passes/PassRegistry.def with the provided src/PassRegistry.def.
Add the follwing line to SRC/llvm/lib/Passes/PassBuilder.cpp:

#include "llvm/Transforms/Utils/LoopFusion.h"

Where $SRC is the source folder of the project.
To compile the source code:

cd $ROOT/BUILD
make opt
make install

Loop Fusion

Given two loops that satisfy the follwing requirements:

• are adjacent
  • there cannot be any statement that execute between them
• have the same number of iterations
• are control flow equivalent
  • given two loop Lj and Lk, Lj before Lk, Lk dominates Lj and Lj post-dominates Lk
• are distance independent or have a non negative dependence distance
  • a negative distance dependence occurs between Lj and Lk, Lj before Lk, when at iteration m from Lk uses a value that is computed by Lj at a future iteration m+n (where n > 0).

Then they can be fused, i.e. the body of the latter is connected after the body of the former.

LoopFusion.cpp and LoopFusion.h files contain the Loop Fusion pass.
In order to make the pass work, src/GlobalOpts/LoopFusion.cpp file must be moved to the following directory:

$SRC/llvm/lib/Transforms/Utils

Add LoopFusion.cpp in $SRC/llvm/lib/Transforms/Utils/CMakeLists.txt.
src/LoopOpts.h file must be moved to the following directory:

$SRC/llvm/include/llvm/Transforms/Utils

Replace SRC/llvm/lib/Passes/PassRegistry.def with the provided src/PassRegistry.def.
Add the follwing line to SRC/llvm/lib/Passes/PassBuilder.cpp:

#include "llvm/Transforms/Utils/LoopFusion.h"

Where $SRC is the source folder of the project.
To compile the source code:

cd $ROOT/BUILD
make opt
make install

Example

The example defined in Test/loop_fus_ex1_virtualregs.ll shows the loop fusion pass in action.

Testing

Tests folder contains different LLVM files to test the optimizations.

Compile the source code:

cd $ROOT/BUILD
make opt
make install

To create a file that uses virtual registries starting from the file written in C:

opt -p mem2reg <file_name>.c -o <file_name_virtualmem>.bc
llvm-dis <file_name_virtualmem>.bc -o <file_name_virtualmem>.ll

This step is needed only for global passes.
To directly start from the .ll file with virtual registers, both source files (.c and .ll) are provided in the Tests/ directory.

To run tests:

opt -p <optimization_pass_name> <file_name>.ll -o <file_name_optimized>.bc
llvm-dis <file_name_optimized>.bc -o <file_name_optimized>.ll

Note: Implemented passes names are localopts, loopopts and loopfusion.

Authors

• Raffaele Tranfaglia
• Matteo Venturi
• Mirco Piccinini