Optimizing C Compiler Sought

[mash at mips.UUCP]

In article <160 at bobkat.UUCP> m5 at bobkat.UUCP (Mr Mike McNally) writes:
>In article <4467 at brl-smoke.ARPA> NET-RELAY.ARPA>@brl-smoke.ARPA writes:
>>Does anyone know of a C compiler that will use its own profiling output
>>to generate faster runtime code?  I've long thought that a compiler that
>>didn't accept data about the runtime behavior of a program shouldn't
>>call itself a execution-time optimizing compiler - space maybe but not time.
>
>
>In the case of optimizing C, there are various well-discussed semantic
>characteristics of the language which prevent all but the most local
>optimizations (I personally have never been convinced that this is 
>always a ``feature'').  Even if it were possible to do local or global
>common subexpression elimination, loop unravelling, induction variable
>elimination, etc., I think that allowing the compiler to optimize 
>anything it can will result in good code.  Useful profile information
>might be somewhat difficult to generate, given a program with several
>possible flow characteristics and large possible input sets.  The only
>use of such information would be to make space/time tradeoff decisions,
>and these probably are more easily done through some sort of ``pragma''
>construct.

1) It certainly is possible to do the optimizations mentioned, although they
don't do you as much good in C as they do for FORTRAN or PASCAL.

2) Profiling feedback is usable.  The MIPSco compiler system has a
[not released, not counted in benchmarks, optional, etc] compiler
phase that can take profiling feedback files, then use the information for
selective inlining of procedures.  The feedback files get down to the
basic block, i.e., effective statment-by-statement counting.
"Selective inlining" means that you grab procedures inline only where they're
called frequently.  Besides eliminating function calls, this lets a serious
optimizer move newly-exposed-constant-subexpressions out of loops,
and eliminate newly-visible dead code. [Note that many arguments to functions
are constants, tests upon which can be eliminated at compile time.]
-- 
-john mashey	DISCLAIMER: <generic disclaimer, I speak for me only, etc>
UUCP: 	{decvax,ucbvax,ihnp4}!decwrl!mips!mash, DDD:  	408-720-1700, x253
USPS: 	MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086