Difference between revisions of "Melting Ice Technology"

Revision as of 13:56, 9 January 2007

Warning: Warning: Article under development

There are two main criteria for any good compiler. Both the compilation and the compiled program need to be fast. There is a trade off, the more time a compiler takes to make optimizations, the faster the compiled program is and the slower the compilation process becomes.

Fortunately the need for fast compilation and fast compiled programs occurs at different times during the development cycle. When the developer is incrementally writing and testing a program he needs short compilation time to be productive. At the end, when his work is thoroughly tested and ready to ship there can a slower compilation that generates a very fast delivery.

EiffelStudio exploits this by providing two basic compilation modes: workbench and finalized mode. The C code generated by the compiler looks different for workbench and finalized mode. The terms workbench code and finalized code are used to refer to the corresponding generated C code.

Workbench code has the following properties:

• It is easily testable (by debugging).
• It compiles and recompiles very fast (due to melting ice technology).
• It supports precompiles.

Finalized code has only two advantages, smallness and speed. Whereas the former contributes to the latter due too better cache efficiency. Finalized code can only be debugged at the C level. This article focuses on the workbench mode of EiffelStudio

Compiled versus interpreted

EiffelStudio was designed with the following principle in mind: When a programmer makes a small change he expects a short recompilation time, when he makes a big change he will accept some waiting time.

Traditionally, the fastest programming environments (in terms of recompilation time) were interpreted languages (like VisualBasic). Unfortunately interpreted languages are not competitive performance wise.

The designers of EiffelStudio circumnavigated the trade off by taking the best of both approaches. Thus, in workbench mode EiffelStudio is half compiler and half interpreter. The technology behind this is called melting ice.

An Eiffel System compiled in workbench mode consists of both frozen and melted code. Frozen code is code, that is translated to C code. Melted code is not yet translated to C code but to a form of byte code that has to be interpreted. This byte code is called EiffelStudio byte code (EBC).

This yields two requirements to the EiffelStudio runtime. It needs both to be able to execute the EiffelStudio byte code and to handle calls from frozen code into melted code and vice versa.

The interpreter and the byte code

The EiffelStudio byte code interpreter is a stack machine. The byte code itself is not very different from .NET byte code or Java Byte Code.

This article won't explain the byte codes. This is partially done in Byte Code. To give a general idea the Fibonacci feature and its EiffelStudio byte code are shown:

fibonacci (n: INTEGER): INTEGER
   require
      n >= 0
   do
      if n = 0 then
         Result := 0
      elseif n = 1 then
         Result := 1
      else
         Result := 
            fibonacci (n - 1) + 
            fibonacci (n - 2)
      end
   end

1: BC_START 
Routine Id   : 186
Body Id      : 185
Result Type  : [INTEGER_32]
Nr. args     : 1
Nr. locals   : 0
  18: BC_NO_CLONE_ARG 
Routine name : fibonacci
Written      : 9
 32: BC_PRECOND offset 23
 37: BC_HOOK 1
 42: BC_ASSERT <8, 16>
 45: BC_ARG 1 
 48: BC_INT32 0
 53: BC_GE 
 54: BC_END_PRE offset 0
 59: BC_RAISE_PREC 
 60: BC_HOOK 2
 65: BC_ARG 1 
 68: BC_INT32 0
 73: BC_EQ 
 74: BC_JMP_F 16
 79: BC_HOOK 3
 84: BC_INT32 0
 89: BC_RASSIGN 
 90: BC_JMP 80

90: BC_JMP 80
 95: BC_HOOK 4
100: BC_ARG 1 
103: BC_INT32 1
108: BC_EQ 
109: BC_JMP_F 16
114: BC_HOOK 5
119: BC_INT32 1
124: BC_RASSIGN 
125: BC_JMP 45
130: BC_HOOK 6
135: BC_ARG 1 
138: BC_INT32 1
143: BC_MINUS 
144: BC_CURRENT 
145: BC_FEATURE fid 36 [9: HELLO_WORLD]
154: BC_ARG 1 
157: BC_INT32 2
162: BC_MINUS 
163: BC_CURRENT 
164: BC_FEATURE fid 36 [9: HELLO_WORLD]
173: BC_PLUS 
174: BC_RASSIGN 
175: BC_HOOK 7
180: BC_NULL

The byte code is not that difficult to understand. Most instructions have an effect on the interpreter stack. Instruction 124 for example pops one value from the stack and saves it to the result of the current feature.