global optimizing compilers vs. unix/make/rcs development methodology

metzger@bach.convex.com (Robert Metzger)
Mon, 2 Aug 1993 14:22:15 GMT

          From comp.compilers

Related articles
global optimizing compilers vs. unix/make/rcs development methodology merlin@neuro.usc.edu (1993-08-01)
Re: global optimizing compilers vs. unix/make/rcs development methodol cliffc@rice.edu (1993-08-02)
Re: global optimizing compilers vs. unix/make/rcs development methodol preston@dawn.cs.rice.edu (1993-08-02)
global optimizing compilers vs. unix/make/rcs development methodology metzger@bach.convex.com (1993-08-02)
| List of all articles for this month |
Newsgroups: comp.compilers
From: metzger@bach.convex.com (Robert Metzger)
Keywords: tools, optimize
Organization: Compilers Central
References: 93-08-007
Date: Mon, 2 Aug 1993 14:22:15 GMT

>[what do people do about inter-procedural optimization
>and minimal recompilation tools like make?]
>is such optimization accomplished by existing and/or planned compilers?


Interprocedural optimization technology has been an active research topic
since the early 1980's. The researchers at Rice Univ. have probably
written the most on the topic, and from your perspective, might make the
most interesting reading, since they have been concerned about the cost
and effectiveness of such optimizations, not merely their theoretical
formulation.


The Convex Application Compiler is a commercially available
interprocedural optimizer that works on very large applications. The
first production release occurred in the spring of 1991, the second in the
fall of 1992. Source files can contain both Fortran and C. It currently
generates code for Convex C-series vector-parallel computers. When we
ship our scalable parallel processor, based on HP PA-RISC chips, it will
also generate code for them.


Currently, it performs automatic inlining, interprocedural constant
propagation, procedure cloning, storage optimization, pointer tracking.
It has successfully compiled and optimized applications in the range of
500,000 lines of Fortran. The improvement seen in real applications has
ranged from 0 to 60x speedup. Excluding parallel speedup, the median is
in the 10-30% range. It all depends on how the application was coded. In
general, the more modular the implementation, the more speedup observed.


/Bob
-------------------------------------------------------------------------------
Robert Metzger CONVEX Computer Corp.
Voice: 214-497-4437 Email: metzger@convex.com
--


Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.