International Business Machines Corporation

Robert Ralph Roediger

William Jon Schmidt

Edward Curtis Prosser

Robert John Donovan

Roy N Bar-Haim

Nava Arela Aizikowitz

Vita Bortnikov

Bilha Mendelson

Scott Patrick Hanson

James Albert Kryka

Orit Edelstein

Inbal Shavit-Lottem

Peter Jerome Steinmetz

Craig Arthur Orcutt

Roger Wayne Southwick

David John Lambert

Liviu Asnash

Daniel Rodman Hicks

Mikko Herman Lipasti

Joseph Patrick Reynolds

Scott Thomas Robinson

Tim Clayton Muehe

Patrick Todd Haugen

Mark Novick

Paul Lu Verne Godtland

Curtis P Meier

Mark Jonathon Hessler

Sherri Lynn McMeeking

Keith Vernon Besaw

Mircea Namolaru

Nava E Aizikowitz

A method, apparatus and program product are provided for parallelizing analysis and optimization in a compiler. A plurality of basic blocks and a subset of data points of a computer program is prepared for processing by a main thread selected from a plurality of hardware threads. The plurality of prepared basic blocks and subset of data points are placed in a shared data structure by the main thread. A prepared basic block of the plurality of prepared basic blocks and/or a tuple associated with the subset of data points is concurrently retrieved from the shared data structure by a work thread selected from the plurality of hardware threads. A compiler analysis or optimization is performed on the prepared basic block or tuple by the work thread.

Leveraging multicore systems when compiling procedures

A method and apparatuses for allowing additional tail call optimizations. The compiler generates both optimized and non-optimized code for tail call candidates. At a later time when there is more information (e.g. regarding program bounds), a decision and a modification to the executable code is made, implementing one of the two earlier codes.

Refining tail call optimizations at link-time

A program feature set is compared to a processor feature set. The comparison may be performed in response to restoring the program onto a system or requesting its execution. The processor feature set represents zero, one or more optional hardware features supported by the processor, whereas the program feature set represents zero, one or more optional hardware features the program relies upon in its generated code. Comparison of the feature sets determines whether a particular program may run on a particular processor. Programs may be automatically or manually rebuilt to achieve full compatibility. If the comparison indicates that the program requires a feature not supported by the processor, the program is rebuilt based on the processor feature set. Alternatively, the program may be rebuilt in response to a rebuild request, rather than the comparison. The program is preferably rebuilt from an intermediate representation (IR) stored with or locatable from the program.

Adaptively generating code for a computer program

Methods and apparatus are disclosed for improved loop unrolling by a compiler. A large class of loops exists for which effective loop unrolling has not previously been performed because they are too large to be completely unrolled, but which do not have a single hot trace that covers an entire loop iteration. The present invention recognizes such loops that have partial hot traces identified using profile data. A set of instructions which constitute a proper superset of the hot trace and a proper subset of the entire loop, and which forms a complete loop iteration is identified. This set of instructions can then be unrolled without unrolling the entire loop.

Unrolling loops with partial hot traces

A superblock unroller creates a superblock in a first instruction stream, and unrolls the superblock using different methods depending on the suitability of the superblock to different types of unrolling. The superblock unroller uses complete unrolling if a number of iterations of the superblock is sufficiently small and if the number of iterations is known at compile-time. The superblock unroller uses static unrolling if the number of iterations of the superblock is too large for complete unrolling and if the number of iterations is known at compile-time. The superblock unroller uses dynamic unrolling if the number of iterations of the superblock is unknown at compile-time. Each of these unrolling methods may include the insertion of count rectification code into the first instruction stream to account for execution that exits the superblock. Code performance may be improved by providing the sophisticated unrolling of superblocks disclosed herein.

Compiler apparatus and method for unrolling a superblock in a computer program

A method, apparatus and article of manufacture for performing automatic intermodule call linkage optimization. In one embodiment, the run time is optimized for an object code generated from a source code. Initially, information is extracted for each procedure call in the source code. The extracted information is used to select a call linkage for each procedure call. The call linkages are selected to minimize the run time of the object code generated from the source code. Once the object code is generated form the source code, the object code is run using the selected call linkages for each procedure call.

Method, article of manufacture and apparatus for performing automatic intermodule call linkage optimization

Automatic template instantiation for cross compilation is provided. More particularly, a routine () on a development computer system () extracts template information from source code (). The routine () uses the template information to generate template source files (A,B), which may then be compiled with a cross compiler () to provide object code (A,B). Owing to one or more templates being dependent on one or more other templates, the routine is repeated until all templates, and their associated dependencies, are compiled into object code (B) with the cross compiler (). This object code is then portable from the development computer system () to a target computer system () for automatic template instantiation. Moreover, a template repository () is used to hold in part the template information. By centrally locating template information, duplicate templates may be avoided to conserve data resources.

Apparatus and method for template instantiation with a cross compiler

A profile-based loop optimizer generates an execution frequency table for each loop that gives more detailed profile data that allows making a more intelligent decision regarding if and how to optimize each loop in the computer program. The execution frequency table contains entries that correlate a number of times a loop is executed each time the loop is entered with a count of the occurrences of each number during the execution of an instrumented instruction stream. The execution frequency table is used to determine whether there is one dominant mode that appears in the profile data, and if so, optimizes the loop according to the dominant mode. The optimizer may perform optimizations by peeling a loop, by unrolling a loop, and by performing both peeling and unrolling on a loop according to the profile data in the execution frequency table for the loop. In this manner the execution time of the resulting code is minimized according to the detailed profile data in the execution frequency tables, resulting in a computer program with loops that are more fully optimized.

Compiler apparatus and method for optimizing loops in a computer program

A mechanism and method for hoisting invariant computations from loops analyzes the lifetimes of fixed processor resources defined by an instruction, and determines whether a group of computations present in multiple instructions within the lifetime are, taken together, loop-invariant and legal to hoist from the loop. If the group of computations within the lifetime of the fixed processor resource are loop-invariant and hoistable, all of the computations are hoisted out of the loop as a group. By determining the lifetimes of fixed processor resources defined in an instruction, the hoisting mechanism succeeds in hoisting out groups of computations that cannot be individually hoisted out of a loop, thereby achieving better performance when the computer program executes.

Lifetime-sensitive mechanism and method for hoisting invariant computations out of loops in a computer program

Disclosed is a system and method for a profiling system wherein profile data is stored in a separable hierarchical fashion such that profile data for each compiled procedure in a computer program can be readily identified and utilized. In particular, each source module has a corresponding profile data file and each procedure has a corresponding procedure profile area. The system and method also includes a mechanism for verifying the existence and validity of profile information, and a mechanism for handling invalid profile information.

Generating and utilizing organized profile information

A profiler that operates in a multi-stage environment is disclosed. As program code undergoes a series of transformations, branches of interest are selected and tracked. Regardless of how many transformations are involved only a single instrumentation/data gathering phase is required. The gathered profile data is then used to perform various optimizations at the differing transformation stages.

Multi-stage profiler

An instruction scheduler in an optimizing compiler schedules instructions in a computer program by determining the lifetimes of fixed registers in the computer program. By determining the lifetimes of fixed registers, the instruction scheduler can achieve a schedule that has a higher degree of parallelism by relaxing dependences between instructions in independent lifetimes of a fixed register so that instructions can be scheduled earlier than would otherwise be possible if those dependences were precisely honored.

Lifetime-sensitive instruction scheduling mechanism and method

A compiler system and method is provided that can 1) generate a second instruction stream from a first instruction stream, 2) read in and process predetermined external information regarding the basic blocks that makes up the second instruction stream and 3) place certain of the basic blocks on cache line boundaries based on predicted execution frequencies. In particular, the compiler system and method utilize profile information containing predicted block execution or edge-weight execution frequencies to determine which of the basic blocks to align on cache line boundaries. One method for obtaining profile information includes precompiling the source code, creating an executable program, executing the program with test inputs, and outputting a profile containing execution frequency information. Once the profile information is obtained, the source code can then be recompiled using the profile information. The compiler can then selectively cache align those blocks identified as important.

Instruction cache alignment mechanism for branch targets based on predicted execution frequencies

Apparatus and a method are provided for defining visibility boundaries in compiled code of an optimizing compiler used with a digital computer for converting a source program to an object code. The source program is received and a first intermediate representation of the source program is generated. A pseudo operation representing a predetermined event is inserted in the first intermediate representation. Sequential operations of the first intermediate representation including the inserted pseudo operations are processed for defining an internal data structure of the source program. A visibility boundary for a selected variable is identified responsive to identifying a pseudo operation in the first intermediate representation.

Optimizing apparatus and method for defining visibility boundaries in

A compiler and method of compiling provide enhanced performance by inlining one or more frequently executed paths through a child procedure into a parent procedure without inlining the entire child procedure. Accordingly, a substantial improvement in speed of execution of the program can be achieved by reducing procedure call overhead, with reduced expense in terms of program size as compared to traditional inlining. Various criteria for determining whether to inline particular child procedures are also described.

Growing community of inventors

Rochester, MN, United States of America

Robert Ralph Roediger