Open64/ORC compilers Sbastian Pop Universit Louis Pasteur - - PowerPoint PPT Presentation

Open64/ORC compilers

Sébastian Pop Université Louis Pasteur Strasbourg, Project A3 INRIA FRANCE

Open64/ORC compilers – p.1

Short history

1994: Ragnarok compiler for MIPS R8000

Open64/ORC compilers – p.2

Short history

1994: Ragnarok compiler for MIPS R8000 designed for scientific applications

Open64/ORC compilers – p.2

Short history

1994: Ragnarok compiler for MIPS R8000 designed for scientific applications August 1994: start Mongoose compiler

Open64/ORC compilers – p.2

Short history

1994: Ragnarok compiler for MIPS R8000 designed for scientific applications August 1994: start Mongoose compiler scientific and non-scientific applications

Open64/ORC compilers – p.2

Short history

1994: Ragnarok compiler for MIPS R8000 designed for scientific applications August 1994: start Mongoose compiler scientific and non-scientific applications fast and stable for day-to-day development

Open64/ORC compilers – p.2

Short history

1994: Ragnarok compiler for MIPS R8000 designed for scientific applications August 1994: start Mongoose compiler scientific and non-scientific applications fast and stable for day-to-day development 1999: focus on IA-64: SGIpro 1.0 (alias osprey1.0)

Open64/ORC compilers – p.2

Short history

1994: Ragnarok compiler for MIPS R8000 designed for scientific applications August 1994: start Mongoose compiler scientific and non-scientific applications fast and stable for day-to-day development 1999: focus on IA-64: SGIpro 1.0 (alias osprey1.0) 2001: Intel and ICT Chinese Academy of Sc. ORC (Open Research Compiler)

Open64/ORC compilers – p.2

Compiler’s Structure

• 1. FE (Front-ends)
• 2. WHIRL (Intermediate Representation)
• 3. IPA (Inter Procedural Analysis)
• 4. LNO (Loop Nest Optimizer)
• 5. WOPT (Global Optimizer)
• 6. CG (Code Generator)
• 7. ORC (Open Research Compiler)

Compiler’s Structure – p.3

Front Ends

Use GCC’s C/C++ and Cray F90 front ends

Compiler’s Structure – p.4

Front Ends

Use GCC’s C/C++ and Cray F90 front ends Each front end has its own specific trees

Compiler’s Structure – p.4

Front Ends

Use GCC’s C/C++ and Cray F90 front ends Each front end has its own specific trees Translation to WHIRL

Compiler’s Structure – p.4

Front Ends

Use GCC’s C/C++ and Cray F90 front ends Each front end has its own specific trees Translation to WHIRL Question: Is this translation valid?

Compiler’s Structure – p.4

Front Ends

Use GCC’s C/C++ and Cray F90 front ends Each front end has its own specific trees Translation to WHIRL Question: Is this translation valid? Test suites were not GPL-ed, could use GCC test suites (inappropriate)

Compiler’s Structure – p.4

Front Ends

Use GCC’s C/C++ and Cray F90 front ends Each front end has its own specific trees Translation to WHIRL Question: Is this translation valid? Test suites were not GPL-ed, could use GCC test suites (inappropriate) Bug data base wasn’t GPL-ed.

Compiler’s Structure – p.4

WHIRL

Winning Hierarchical Intermediate Representation Language

Compiler’s Structure – p.5

WHIRL

Winning Hierarchical Intermediate Representation Language 5 levels: VH, H, M, L, VL Lowering happens when needed Each optimization performed at the right level

Compiler’s Structure – p.5

WHIRL

Winning Hierarchical Intermediate Representation Language

whirl2c and whirl2f dump WHIRL in compilable

files.

whirl2a dump WHIRL in ASCII.

Compiler’s Structure – p.5

Inter Procedural Analysis

file2.cxx file3.f

Suppose that we want to build a project containing 3 files and use the IPA for

• ptimizing it.

file1.c

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end C front−end file1.c

the right front−end. The first step invokes

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c

This representation is then dumped into a .o file. These .o files behave like normal relocatable code (I.e. can be put in archives, etc.)

file3.o file2.o file1.o

The compiler transforms front−end specific trees into WHIRL trees.

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c file1.o file2.o file3.o lib1.a

The linker is called as usual

• n the last step of compilation.

Linker lib2.so

.a files can contain WHIRL normal .o files. trees, as well as

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c file1.o file2.o file3.o lib1.a

Some files contain WHIRL trees: the compilation is not complete, and the IPA is called.

Linker Inter Procedural Analysis (IPA) lib2.so

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c file1.o file2.o file3.o lib1.a Linker Inter Procedural Optimizations (IPO) Inter Procedural Analysis (IPA) lib2.so

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c file1.o file2.o file3.o lib1.a Linker Loop Nest Optimizer (LNO) Inter Procedural Optimizations (IPO) Inter Procedural Analysis (IPA) lib2.so

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c file1.o file2.o file3.o lib1.a Linker Loop Nest Optimizer (LNO) Inter Procedural Optimizations (IPO) Inter Procedural Analysis (IPA) Main Optimizer (WOPT) lib2.so

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c file1.o file2.o file3.o lib1.a Linker Code Generator (CG) Loop Nest Optimizer (LNO) Inter Procedural Optimizations (IPO) Inter Procedural Analysis (IPA) Main Optimizer (WOPT) lib2.so

Compiler’s Structure – p.6

Inter Procedural Analysis

file2.cxx C++ front−end file3.f F90 front−end WHIRL dumper C front−end file1.c file1.o file2.o file3.o lib1.a Executable file Linker Code Generator (CG) Loop Nest Optimizer (LNO) Inter Procedural Optimizations (IPO) Inter Procedural Analysis (IPA) Main Optimizer (WOPT) lib2.so

Compiler’s Structure – p.6

Inter Procedural Analysis

Idea: gather information over a whole project

Compiler’s Structure – p.7

Inter Procedural Analysis

Idea: gather information over a whole project Solution: save WHIRL trees in .o files build a global tree at link time perform all optimizations generate code

Compiler’s Structure – p.7

Loop Nest Optimizer

LNO works on High level WHIRL. Lowering removed unstructured control flow (gotos, switch, ...)

Compiler’s Structure – p.8

Loop Nest Optimizer

Analyzes extract information from WHIRL and construct specific Intermediate Representations (IRs): Array Dependence Graph LEGO: for data distributions Array and vectors accesses Vector space Systems of equations Polytope

Compiler’s Structure – p.8

Loop Nest Optimizer

Main optimizers in LNO: Loop unrolling Hoist conditionals Hoist varying lower bounds Dead store eliminate arrays Loop reversal / fission / fusion / tiling Array scalarization Prefetch Inter iteration Common Subexpression Elimination

Compiler’s Structure – p.8

Global Optimizer

WOPT works on Medium-level WHIRL (arrays lowered into load/store + offset, ...)

Compiler’s Structure – p.9

Global Optimizer

Main intermediate representations: CFG (Control Flow Graph) SSA (Static Single Assignement) Main optimizations: SSA-PRE (Partial Redundancy Elimination) DCE (Dead Code Elimination) IVR (Induction Variable Recognition) VNFRE (Value Numbering based Full Redundancy Elimination) Copy propagation

Compiler’s Structure – p.9

Code Generator

Code Generator works on CGIR. explicit CFG each BB contains a list of instructions each instruction is under the form OP_result OP_code OP_opnd This representation is close to assembler code.

Compiler’s Structure – p.10

Code Generator

Main optimizers in CG are: EBO: Extended Block Optimizer GRA: Global Register Allocation LRA: Local Register Allocation GCM: Global Code Motion SWP: Software Pipelining CIO: Cross Iteration loop Optimizations FREQ: execution frequencies of BBs and edges

Compiler’s Structure – p.10

Open Research Compiler

ORC is an extension of the Code Generator. ORC added the following infrastructure: IPFEC Regions: structures the CFG into a tree If-conversion PRDB: Predicate Relation DataBase Microscheduler Local/Global instruction scheduling

Compiler’s Structure – p.11

Partial Redundancy Elimination

Compiler’s Structure – p.12

Predicated code

IF_COND BB1 BB2

Compiler’s Structure – p.13

Predicated code

IF_COND <BB1, p1> <BB2, p2>

If−conversion converts control−dependences to data−dependences replaces "if" constructs with guarded statements IF−conversion =

Compiler’s Structure – p.13

Predicated code

IF_COND <BB1, p1> <BB2, p2> eval (BB1) (p1) eval (BB2) (p2) (p1, p2) = eval (IF_COND)

If−conversion DAG of the CFG linearized into a single block with guarded stmts Hyperblock = converts control−dependences to data−dependences replaces "if" constructs with guarded statements IF−conversion =

Compiler’s Structure – p.13

Predicated code

IF ELSE THEN

4 instructions can be executed in parallel

• n IA−64 architectures

ILP = Instruction Level Parallelism

• ✁

Compiler’s Structure – p.13

Predicated code

IF ELSE THEN

p0 p0 p0 p0 p0 p1 p2 p1 p1 p1 p1 p2 p2 p2 p0 p0 p0 p0 p0 p0 p0 Each instruction is predicated.

Compiler’s Structure – p.13

Predicated code

p0 p0 p0 p0 p0 p0 p0 p0 p1 p1 p1 p1 p1 p0 p0 p0 p0 p2 p2 p2 p2 Create a hyperblock

Compiler’s Structure – p.13

Compiler’s Structure – p.14

Compiler’s Structure – p.15