X3J3/97-125

Date:     February 11, 1997
To: 	    X3J3
From:	    /HPC
Subject:	 Assertions

Goals:
1. Allow programmers to say things about their program that a compiler is unlikely or unable to 
determine on its own.
2. Make explicit the difference between likelihood vs. truth.
3. Satisfy (at least partially) the request for "performance directives"

Non-goal: allow users to tell their processor things that are processor specific.

Illustrative syntax:

	ASSERT  N > 64
	DO I =1,N
	        STUFF
	END  DO

This would allow a processor to generate code that can 100% rely on N > 64. A processor may  provide a 
mode of operation where there is a runtime check. Typically this will be used for more exciting expressions 
(M>N to assure no-overlap,  mod(n,3)==0 to eliminate the need for "cleanup logic" when unrolling by a 
factor of 3 (and perhaps a hint to the optimizer that factors of 3 are a better choice than other factors, etc.)) 
This form of the assertion to apply to the next construct (in this case the DO LOOP) only.

In addition, a BLOCK ... END BLOCK may be specified to allow more flexible control of the range of an 
assertion. An assertion applies to code in the current scope only.

For the possibly true syntax might look something like:

	ASSERT (X)  logical_expression ! where x is a default kind real 0<= x <=1.0

0. means that the processor should assume that the probability of the expression is arbitrarily close to 0 
(that is, can happen but almost never will). 1. Means that the processor should assume that the 
probability of the logical expression is arbitrarily close to certainty. A specific example:

ASSERT (.95) N > 64
DO I = 1, N

Would suggest that a processor with vector registers may assume that 95% of the time doing this 
computation in the vector units would be profitable. (.95) N < 5 would strongly suggest scalar processing 
would be best.