Document:       N1396

Date:                2009/09/25

References:      WG14 reference documents: N1362, N1382, DR 290, N1353

Authors:           Jim Thomas and Fred Tydeman

Subject:           Wide function return values

 

Here’s the relevant text from N1362:

 

6.8.6.4 [3]

 

If a return statement with an expression is executed, the value of the expression is returned to the caller as the value of the function call expression. If the expression has a type different from the return type of the function in which it appears, the value is converted as if by assignment to an object having the return type of the function.142)

 

142) The return statement is not an assignment. The overlap restriction of subclause 6.5.16.1 does not apply to the case of function return. The representation of floating-point values may have wider range or precision [deleted by N1382: and is determined by FLT_EVAL_METHOD]. A cast may be used to remove this extra range and precision.

 

Problems:

 

I believe C99’s not requiring function returns to be narrowed to the type of the function was an oversight. A primary goal of Annex F and the standard evaluation methods was to require fully determined results from floating-point computation using IEC 60559 (IEEE 754-1985) basic formats and operations. The license for widened function returns undermines this goal. To my knowledge, the goal is otherwise met if __STDC_IEC_559__ is defined and FLT_EVAL_METHOD is 0 or 1.

 

Also, there are other problems with the current specification. Conversion as if by assignment to the type of the function is required if the return expression has a different type than the function, but not if the return expression has a wider value only because of wide evaluation. This allows seemingly inconsistent and confusing behavior. Consider

 

float f(float x) { return x * 0.1f; }

float g(float x) { return x * 0.1; }

 

Function f is allowed to return a value wider than float, but function g (which uses the wider constant) is not.

 

Although the current text does not require narrowing return expressions of the same type as the function, it does not clearly state what is allowed. Is it allowed to narrow the result? Is it allowed to narrow the result sometimes but not always? Is it allowed to partially narrow the result (e.g., if the ABI returns floats in double format but a float function has a float return expression evaluated to wider than double)? An aggressive implementation could argue “yes” to all these, though the resulting behavior would complicate debugging and error analysis.

 

Footnote 142 says a cast may be used to remove extra range and precision from the return expression. This means a predictable program must have casts on all floating-point function calls (except where the function directly feeds an operator like assignment that implies the conversion). With type-generic math (tgmath), the programmer has to reason through the tgmath resolution rules to determine which casts to apply. These are significant obstacles to writing predictable code.

 

 

Recommendation:

 

Require return expressions to be converted as if by assignment to the type of the function.

 

5.2.4.2.2 [9]: change “Except for assignment and cast” to “Except for assignment, cast,

and return”.

 

6.8.6.4 [3]: change the second sentence to “If the expression is evaluated to a format

different from the return type of the function in which it appears …”

 

Footnote 142: remove the last two sentences.

 

Alternate Recommendation:

 

Require return expressions to be converted as if by assignment to the type of the function, but only in Annex F. This is a compromise that addresses the problems for Annex F implementations while not impacting non-Annex F implementations that exercise the license for wide returns.

 

Insert the following new subclause after F.5 (and increment subsequent subclause numbers):

 

F.6 The return statement

 

If the return expression is evaluated in a floating-point format different from the return type, then the expression is converted to the return type of the function and the resulting value is returned to the caller.