G53CMP: Lecture 20 - Nottinghampsznhn/G53CMP/LectureNotes-2016/lecture20.pdf · Result SEM G53CMP...

G53CMP: Lecture 20LLVM: A Real Compiler Backend

Henrik Nilsson

University of Nottingham, UK

G53CMP: Lecture 20 – p.1/23

Result SEM G53CMP 2016/17 (1)

Average scores.Scale: 5 is agree/positive; 1 is disagree/negative.

# Question G53CMP All modules

1 Clear objectives 4.43 4.12

2 Teaching methods 4.29 4.01

3 Assessment methods 4.00 4.00

4 Module level 4.07 4.05

5 Module workload 3.79 4.05



• On the whole, you were happy with the module:




- All but one would recommend it to otherstudents.





- Dissenting concern: more about Haskellthan compilers.






• Most disagreement over module workload:







- Most were happy, but 2 students stronglyfelt too much content and/or work.







- Most were happy, but 2 students stronglyfelt too much content and/or work.

- Key question: Overall, around 100 h of work?



• Specific comments on the coursework weight:

25 % too little, should be 50 %.





• I agree! And weight is effectively 50 % as partof the exam is directly related to coursework.





• I agree! And weight is effectively 50 % as partof the exam is directly related to coursework.

• Generally, coursework weight not necessarilyindicative of amount of work involved: themark is mainly to encourage engagement,and the question is if the amount of work for amodule overall is in line with 10 h/credit.


LLVM (1)

LLVM (formerly Low Level Virtual Machine) is acompiler infrastructure project:


LLVM (1)


• Highly modular and extensible; at its core:

- Set of reusable libraries

- Well-defined interfaces


LLVM (1)





• Designed for static and dynamic (JIT)compilation and optimzation: compile-time,link-time, load/installation-time, run-time.


LLVM (1)





• Designed for static and dynamic (JIT)compilation and optimzation: compile-time,link-time, load/installation-time, run-time.

• Language agnostic: LLVM-based compilersfor Ada, C, C++, Fortran, Haskell, Java bytecode,OpenGL Shading Language, Python, Scala.


LLVM (2)

Some background:

• The LLVM project started in 2000 at theUniversity of Illinois at Urbana-Champaign.

• Directed by Vikram Adve and Chris Lattner.

• Lattner later hired by Apple Inc.

• LLVM integral part of Apple’s developmenttools for OS X and iOS.

• LLVM is Open Source.

• Adve, Lattner, and Evan Cheng awarded theACM Software System Award for LLVM in 2012.


Motivations for LLVM (1)

When LLVM started:



When LLVM started:

• Open-source language implementationstended to be monolithic; e.g. GCC:

- Extremely hard to reuse individual parts

- Not even a self-contained intermediaterepresentation



When LLVM started:




• Implementations tended to either supportstatic or JIT compilation.



When LLVM started:




• Implementations tended to either supportstatic or JIT compilation.

• The text-book vision of multiple independentfront-ends and back-ends around a sharedcompiler core hardly ever realised in practice.


LLVM IR (1)

The LLVM Intermediate Representation (IR) isthe “glue” that holds LLVM together.


LLVM IR (1)


• Complete, self-contained representation: afirst-class language with well-definedsemantics.


LLVM IR (1)



• Designed to host mid-level analyses andtransformations.


LLVM IR (1)




• RISC-like code.


LLVM IR (1)




• RISC-like code.

• Sufficiently low-level to be a suitabletranslation target for any language.


LLVM IR (2)

• Sufficiently high-level to allow targetingarbitrary concrete architectures; e.g.:


LLVM IR (2)


- Unbounded number of registers


LLVM IR (2)



- Abstraction over calling conventions


LLVM IR (2)




• Typed:


LLVM IR (2)




• Typed:

- Base types: integers (of different sizes),floating point numbers


LLVM IR (2)




• Typed:


- Derived types: pointers, arrays, vectors,structures, functions


LLVM IR (2)




• Typed:


- Derived types: pointers, arrays, vectors,structures, functions

• Static Single Assignment (SSA): SSA formfor all scalar registers (everything except memory).


LLVM IR (3)

• Three isomorphic forms:

- Textual format (.ll)

- Compact, on-disk, “bitcode” format (.bc)

- In-memory data structure.


LLVM IR (3)

• Three isomorphic forms:

- Textual format (.ll)

- Compact, on-disk, “bitcode” format (.bc)

- In-memory data structure.

Some tools:

- llvm-as: .ll⇒ .bc

- llvm-dis: .bc ⇒ .ll


LLVM Modularity (1)

• Each LLVM pass, such as optimizations, is alibrary component transforming LLVM IR; e.g.

- constant folding

- loop unrolling

- motion of loop-invariant code

- inliner


LLVM Modularity (1)

• Each LLVM pass, such as optimizations, is alibrary component transforming LLVM IR; e.g.

- constant folding

- loop unrolling

- motion of loop-invariant code

- inliner

• Passes are written to be as independent aspossible; any dependences are declaredexplicitly.


LLVM Modularity (2)

• A pass manager can run the availablepasses in a suitable order, subject only todeclared constraints.


LLVM Modularity (2)

• A pass manager can run the availablepasses in a suitable order, subject only todeclared constraints.

• Any particular application only needs toinclude exactly the those passes that arerelevant making for small footprint.


SSA Form (1)

Static Single Assignment (SSA):• SSA form is a property of intermediate

representations where:


SSA Form (1)



- each variable is assigned exactly once


SSA Form (1)




- every variable defined before used.


SSA Form (1)





• Developed at IBM in the 1980s by researchersRon Cytron, Jeanne Ferrante, Barry K.Rosen, Mark N. Wegman, Kenneth Zadeck.


SSA Form (1)





• Developed at IBM in the 1980s by researchersRon Cytron, Jeanne Ferrante, Barry K.Rosen, Mark N. Wegman, Kenneth Zadeck.

• Compilers using SSA include: GCC, LLVM,Oracle’s HotSpot JVM, Android’s Dalvik andRuntime.


SSA Form (2)

• SSA form can for some purposes be seen asa purely functional representation.


SSA Form (2)


• Indeed, there is a formal correspondencebetween SSA form and purely functionalrepresentations, notably ContinuationPassing Style (CPS).


SSA Form (2)


• Indeed, there is a formal correspondencebetween SSA form and purely functionalrepresentations, notably ContinuationPassing Style (CPS).

• As a result, many compiler optimizations aresimplified and improved.


Static Single Assignment (SSA) Form (3)

Conversion to SSA form by splitting eachvariable into versions. For example:

y := 1; y := 2; x := y




y := 1; y := 2; x := y

In SSA form:

y1 := 1; y2 := 2; x1 := y2




y := 1; y := 2; x := y

In SSA form:

y1 := 1; y2 := 2; x1 := y2

Note that it now is manifest (no flow analysisneeded) where the value assigned to x comesfrom and that the first assignment to y is deadcode.


What about Control Flow Joins? (1)

The obvious question is how to handle joins inthe control flow.

Consider:

Before SSA conversion:

x := ...;

if x > 0 then

x := 1

else

x := 2;

y := x;


What about Control Flow Joins? (1)

The obvious question is how to handle joins inthe control flow.

Consider:


x := ...;

if x > 0 then

x := 1

else

x := 2;

y := x;

SSA form:

x1 := ...;

if x1 > 0 then

x2 := 1

else

x3 := 2;

y1 := x???;


What about Control Flow Joins (2)

Or consider:


x := ...;

while x < 100 do

x := x * 2;

y := x


What about Control Flow Joins (2)

Or consider:


x := ...;

while x < 100 do

x := x * 2;

y := x

SSA form:

x1 := ...;

while x??? < 100 do

x2 := x??? * 2;

y1 := x???


φ-Functions (1)

A φ-function (originally “phoney function”) selectsand returns exactly one of its arguments.Assume first it always picks the “right” argument.Then we can solve our dilemma as follows:

x1 := ...;

if x1 > 0 then

x2 := 1

else

x3 := 2;

x4 := φ(x2,x3);

y1 := x4


φ-Functions (1)

A φ-function (originally “phoney function”) selectsand returns exactly one of its arguments.Assume first it always picks the “right” argument.Then we can solve our dilemma as follows:

x1 := ...;

if x1 > 0 then

x2 := 1

else Clearly in SSA form!

x3 := 2;

x4 := φ(x2,x3);

y1 := x4


φ-Functions (2)

And:

x1 := ...;

while (x2 := φ1(x1,x3), x2 < 100) do

x3 := x2 * 2;

y1 := x2

Also clearly in SSA form!


φ-Functions (3)

• A φ-function selects an argument according tothe dynamically preceding basic block: fromwhere did the control reach the φ-function?


φ-Functions (3)


• “Translating out of” SSA is essentially amatter of joining up the different versions of avariable.


φ-Functions (3)



• A φ-function translates into no code if thearguments and results can be stored in thesame place (register).


φ-Functions (3)



• A φ-function translates into no code if thearguments and results can be stored in thesame place (register).

• Otherwise extra copy instructions (assignments)are needed to translate out of SSA.


Where Do φ-Functions Go? (1)

φ-functions are placed by constructing andanalysing the control flow graph:




• A node A strictly dominates a different nodeB iff all paths to B go through A.





• A node A dominates B iff A strictlydominates B or A = B.






• A node B is in the dominance frontier of a nodeA iff A does not strictly dominate B, but doesdominate an immediate predecessor of B.






• A node B is in the dominance frontier of a nodeA iff A does not strictly dominate B, but doesdominate an immediate predecessor of B.

φ-functions are placed on the dominance frontier.



Observation: we only need φ-functions for livevariables.




• Pruned SSA: Use live-variable information todecide whether a particular φ-function isneeded. Expensive computation.




• Pruned SSA: Use live-variable information todecide whether a particular φ-function isneeded. Expensive computation.

• Semi-pruned SSA: Identify variables that arenever live on entry to a block and omitφ-functions for such “block-local” variables.Cheaper to compute.


LLVM Demo

We will translate the following C-code into LLVMIR using the Clang compiler, study the result andrun some optimizations on it.

int i, m, n;

int main(int argc, char* argv[]) {

sscanf(argv[1], "%d", &m);

for (i = 0; i < m; i++) {

n += i;

}

printf("n = %d\n", n);

return 0;

}G53CMP: Lecture 20 – p.23/23

G53CMP: Lecture 20 - Nottinghampsznhn/G53CMP/LectureNotes-2016/lecture20.pdf · Result SEM G53CMP...

Documents

Transcript of G53CMP: Lecture 20 - Nottinghampsznhn/G53CMP/LectureNotes-2016/lecture20.pdf · Result SEM G53CMP...