The Standard ML of New Jersey compiler has been under development for five years now. We have developed a robust and complete environment for Standard ML that supports the implementation of large software systems and generates efficient code. The compiler has also served as a laboratory for developing novel implementation techniques for a sophisticated type and module system, continuation based code generation, efficient pattern matching, and concurrent programming features.

Garbage collection can be done in vector mode on supercomputers like the Cray-2 and the Cyber 205. Both copying collection and mark-and-sweep can be expressed as breadth-first searches in which the "queue" can be processed in parallel. We have designed a copying garbage collector whose inner loop works entirely in vector mode. We give performance measurements of the algorithm as implemented for Lisp CONS cells on the Cyber 205. Vector-mode garbage collection performs up to 9 times faster than scalar-mode collection --- a worthwhile improvement. - 1. Automatic garbage collection on vector supercomputers Languages like Lisp with dynamic storage allocation and automatic garbage collection are increasingly being used on vector supercomputers. Implementations of Lisp have been done for Cray supercomputers [1], and fully supported supercomputer Lisp environments will soon be available (e.g. Common Lisp provided by Cray Research and Franz, Inc.)[2]. This is a natural development. Languages ...

Heap allocation with copying garbage collection is a general storage management technique for modern programming languages. It is believed to have poor memory subsystem performance. To investigate this, we conducted an in-depth study of the memory subsystem performance of heap allocation for memory subsystems found on many machines. We studied the performance of mostly-functional Standard ML programs which made heavy use of heap allocation. We found that most machines support heap allocation poorly. However, with the appropriate memory subsystem organization, heap allocation can have good performance. The memory subsystem property crucial for achieving good performance was the ability to allocate and initialize a new object into the cache without a penalty. This can be achieved by having subblock placement with a subblock size of one word with a write allocate policy, along with fast page-mode writes or a write buffer. For caches with subblock placement, the data cache overhead was under 9% for a 64K of larger data cache; without subblock placement the overhead was often higher than 50%.

Heap allocation with copying garbage collection 1s a general storage management technique for programming languages. It is believed to have poor memory system performance, To investigate this, we conducted an in-depth study of the memory system performance of heap allocation for memory systems found on many machines. We studied the performance of mostly functional Standard ML programs which made heavy use of heap allocation. We found that most machmes support heap allocation poorly. However, with the appropriate memory system organization, heap allocation can have good performance. The memory system property crucial for achieving good performance was the ability to allocate and initialize a new object into the cache without a penalty. This can be achieved by having subblock placement with a subblock size of one word with a write-allocate pohcy, along with fast page-mode writes or a write buffer. For caches with subblock placement, the data cache overhead was under 9T0 for a 64K or larger data cache; without subblock placement the overhead was often higher than 50’?o.

Many language theoreticians have taken great efforts in designing higher-level programming languages that are more elegant and more expressive than conventional languages. However, few of these new languages have been implemented very efficiently. The result is that most software engineers still prefer to use conventional languages, even though the new higherlevel languages offer a better and simpler programming model. This dissertation concentrates on improving the performance of programs written in Standard ML (SML)---a statically typed functional language---on today's RISC machines. SML poses tough challenges to efficient implementations: very frequent function calls, polymorphic types, recursive data structures, higher-order functions, and first-class continuations. This dissertation presents the design and evaluation of several new compilation techniques that meet these challenges by taking advantage of some of the higher-level language features in SML. Type-directed compilation ...

We describe a method of measuring lifetime characteristics of heap objects, and discuss ways in which such quantitative object behaviour measurements can help improve language implementations, especially garbage collection performance. For Standard ML of New Jersey, we find that certain primary aspects of object behaviour are qualitatively the same across benchmark programs, in particular the rapid object decay. We show that the heap-only allocation implementation model is the cause of this similarity. We confirm the weak generational hypothesis for SML/NJ and discuss garbage collector configuration tuning. Our approach is to obtain object statistics directly from program execution, rather than simulation, for reasons of simplicity and speed. Towards this end, we exploit the flexibility of the garbage collector toolkit as a measurement tool. Careful numerical analysis of the acquired data is necessary to arrive at relevant object lifetime measures. This study fills a gap in quantitativ...

We study the cache performance of a set of ML programs, compiled by the Standard ML of New Jersey compiler. We find that more than half of the reads are for objects that have just been allocated. We also consider the effects of varying software (garbage collection frequency) and hardware (cache) parameters. Confirming results of related experiments, we found that ML programs can have good cache performance when there is no penalty for allocation. Even on caches that have an allocation penalty, we found that ML programs can have lower miss ratios than the C and Fortran SPEC92 benchmarks. Topics: 4 benchmarks, performance analysis; 21 hardware design, measurements; 17 garbage collection, storage allocation; 46 runtime systems. 1 Introduction With the gap between CPU and memory speed widening, good cache performance is increasingly important for programs to take full advantage of the speed of current microprocessors. Most recent microprocessors come with a small on-chip cache, and many m...

Separate compilation is an important tool for coping with design complexity in large software projects. When done right it can also be used to create software libraries, thus promoting code reuse. But separate compilation comes in various flavors and has many facets: namespace management, linking, optimization, dependencies. Many programming languages identify modular units with units of compilation, while only a few extend this to permit hierarchies of language-level modules within individual compilation units. When the number of compilation units is large, then it becomes increasingly important that the mechanism of separate compilation itself can be used to control namespaces. The group model implemented in SML/NJ's compilation manager CM provides the necessary facilities to avoid unwanted interferences between unrelated parts of large programs. Compilation units are arranged into groups, and explicit export interfaces can be used to control namespaces. When there are many groups, t...

We study the cost of storage management for garbage-collected programs compiled with the Standard ML of New Jersey compiler. We show that the cost of storage management is not the same as the time spent garbage collecting. For many of the programs, the time spent garbage collecting is less than the time spent doing other storage-management tasks. 1 Authors' addresses: David Tarditi, Computer Science Department, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA, 152133891. e-mail: dtarditi@cs.cmu.edu. Amer Diwan, Department of Computer Science, University of Massachusetts, Amherst, MA 01003-4610. e-mail: diwan@cs.umass.edu. This research is sponsored by the Defense Advanced Research Projects Agency, DoD, through ARPA Order 8313, and monitored by ESD/AVS under contract F19628-91-C-0168. David Tarditi is also supported by an AT&T PhD Scholarship. Views and conclusions contained in this document are those of the authors and should not be interpreted as representing the offic...

ective recompilation for any SML program consisting of more than one compilation unit. But when the number of compilation units is large, it becomes increasingly important that they themselves can be grouped together and that explicit export interfaces can be used to control namespaces. When there are many groups, then one may want to organize them into supergroups, and so on. Figure 1 shows the dependency graph for CML's sources. CML [Rep91] is implemented as a regular SML/NJ program. Without the help from a compilation manager one would have to invoke a number of use commands---one for each source file---in order to load CML into SML/NJ. Of course, the order in which the commands are issued is important here, because identifiers can only be referred to after they are defined. Therefore, since run-cml.sml refers to objects defined in cml-compat.sml the former must be loaded after the latter. More precisely, the sequence of use<F